This Week in All Things AI

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.

Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our

Sam Altman

@sama

I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks.

Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.

Dario Amodei

Elon Musk

Dario is right

Dario Amodei

📖 Choosing external research tools for AI agents

Dario Amodei - We Must Pace the Frontier

I have worked on AI for the last twelve years because I believe it could dramatically raise the quality of human life. I've written often about these incredible benefits: I believe that AI could cure most major diseases in the next 5-10 years, greatly accelerate economic growth rates, create a world of abundance and empowerment, and usher in a renaissance of democracy and freedom.

https://darioamodei.com

Harness Engineering Transforms AI Agent Development

A shift toward domain-specific harnesses is gaining speed, with Garry Tan noting that systems either become these customized frameworks or fade away. OpenAI launched its Agents API in public beta, offering managed services like context compaction and subagents while supporting any model and open-source flexibility. Builders like Harrison Chase from LangChain share practical guides on custom harnesses,

Harrison Chase

@hwchase17

in case you want to build a domain specific harness:

langchain.com/blog/how-to-bu…

great blog by @sydneyrunkle

Garry Tan

@garrytan

Either you die a system of record or you live long enough to become a domain-specific harness

elvis

@omarsar0

Learn to build a harness, folks.

It's not surprising to me that so many YC builders want to build domain-specific harnesses.

If you work long enough on a domain-specific problem, you quickly realize the opportunity. But you also realize how important that harness will be to

Garry Tan

@garrytan

Either you die a system of record or you live long enough to become a domain-specific harness

Min Choi

@minchoi

OpenAI just made it stupidly easy to build real AI agents.

The new Agents API basically gives developers the Codex harness as an API:

• One call → cloud agent
• Long-running sessions
• Automatic context compaction
• Parallel subagents
• MCP + built-in tools
•

OpenAI Developers

@OpenAIDevs

Go from idea to a working agent faster with the Agents API.

Build and run cloud agents with the Codex harness, fully managed by OpenAI.

We handle orchestration, long-running sessions, and context management. You focus on what makes your agent unique.

Available in public beta.

spect

@spectnfa

OpenAI engineers just dropped a free 1-hour course on harness engineering and coding agents

How one engineer with agents goes from writing code line by line to running a team of 5000, and how their own coding agent stays fast and safe while doing it:

0% → 0:00 - why one

spect

@spectnfa

x.com/i/article/2097…

🔄 Updated skills and guides

Inspired by OpenAI’s article on rethinking skills and prompts for GPT-6 Astra I had Astra audit and refine two skills I’ve previously shared here:

• Portable LLM Council — for reviewing business and technical decisions from multiple perspectives.

• Multi-Provider Research — for choosing and using external research tools with AI agents.

The article makes a useful point: as models improve, it’s worth revisiting older instructions and removing unnecessary complexity.

The updated skills are available in the linked repositories—feel free to grab the latest versions.

I also used Astra to audit and tidy up the Notion guide linked from the Multi-Provider Research repository. There are a few small updates, so it’s worth revisiting if you’ve read it before.

For the background and guides, here are my original newsletter posts:

📖 Introducing Portable LLM Council

🎬 A product brief, a few screenshots, your brand kit—what would it take for an AI agent to turn those into a finished explainer?

That’s the idea behind OpenMontage, an open-source project that connects research, scripting, visuals, narration and rendering into a video production workflow.

What caught my attention is how it brings together your own material, stock footage and AI-generated assets—with human approval along the way and cost estimates before spending.

The engine is public; the Studio app is still in private alpha. Paid generation can add costs.

🎬 See what OpenMontage can do

🛠 Explore the open-source project on GitHub

💬 For those building with AI: is coordinating the tools now the bigger challenge—or getting consistently good output from each one?

Monday 14th September 2026

Tuesday 15th September 2026

Wednesday 16th September 2026

TypeSafe AI Launches Jev for Fast Structured AI Decisions

Diogo Almeida, co-author of the InstructGPT paper that introduced RLHF at OpenAI, announces Jev from TypeSafe AI as a new "System One" model trained with RLCD for calibrated, structured decisions rather than text generation.

Jev outputs typed results like choices, scores, or true/false answers with confidence levels, enabling developers to handle multiple queries in one 150ms call. Priced at just $0.042 per million input tokens with free outputs, it promises 20-200x speed and 40-400x cost savings over typical models for decision tasks, eliminating hallucinations by avoiding string generation.

Diogo Almeida

@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x

Introduction - TypeSafe AI

Jev is TypeSafe's flagship model and the first System One model. Send state and typed questions; get structured answers your code can use directly.

https://docs.typesafe.ai

Odyssey AI lab unveils Odyssey-3, a foundation world model that controls robots and humanoids, drives cars on Indian roads, pilots drones, plays video games, and trains other AIs.

Built as an autoregressive diffusion transformer, it simulates diverse physical and virtual scenarios from visual data, enabling efficient adaptation with just tens of hours of demonstrations per task.

The model generates training environments for recursive learning, creating physical agents that natively interface with real-world systems and support iterative intelligence improvements.

Odyssey

@odysseyml

Today we’re unveiling Odyssey-3, a big step forward for foundation world models.

It can control robots, power humanoids, drive cars (on the roads of India!), train AIs, pilot drones, and even play video games.

We can’t wait to see what intelligent systems it enables.

Introducing Odyssey-3: A General-Purpose Physical Intelligence

Odyssey-3 is a foundation world model that can power robots, drive cars, train AIs, pilot drones, and even play video games.

https://odyssey.systems

Meta now lets AI agents handle the boring parts of WhatsApp Business setup

A new WhatsApp Business MCP server lets developers use AI coding agents like Claude, Cursor, Codex, and ChatGPT to handle setup, messaging templates, testing, and troubleshooting.

Meta now lets AI agents handle the boring parts of WhatsApp Business setup | TechCrunch

A new WhatsApp Business MCP server lets developers use AI coding agents like Claude, Cursor, Codex, and ChatGPT to handle setup, messaging templates, testing, and troubleshooting.

Thursday 17th September 2026

Claude is merging its dedicated Cowork workspace with the main chat into a single unified experience, allowing seamless handoff of quick questions or complex tasks like reports that Claude continues autonomously even when the app is closed.

New integrated tools enable creating and editing Claude Docs, Slides, and Design artifacts directly in conversations, with options for manual fixes, comments, presentations, exports, and single-link sharing while retaining user final approval.

The update preserves all existing Cowork chats, projects, and artifacts, rolling out first to Pro and Max users over the coming weeks starting September 16, 2026, with Docs, Slides, and Design in beta for paid plans.

Claude

You can also now make decks, docs, and designs in your conversation.

Draft the one-pager in Claude Docs, turn it into a deck with Claude Slides, and mock up a matching visual in Claude Design, all from one place.

Claude Cowork and chat are now one Claude | Claude by Anthropic

Starting today, Claude Cowork and chat are one Claude, rolling out on Pro and Max plans. Hand over the task and Claude does the work. You keep the final say.

https://claude.com

Xiaomi Livestreams Massive AI Model Training with Full Transparency

Led by Fuli Luo, Xiaomi's MiMo team is livestreaming reinforcement learning for MiMo-V2.6 models on a public dashboard, tracking compute at 2 billion tokens per step, multi-task environments, and advanced grading systems. The larger Pro version has 1 trillion total parameters and nears $700,000 in costs, while the Flash variant advances faster at lower expense, with daily runs hitting $493,000 and $247,000 respectively. Researchers praise the openness, calling it a rare glimpse into large-scale RL usually hidden by frontier labs, as benchmark scores climb steadily on tasks like DeepSWE.

Fuli Luo

Nearly half a year of silence. We spent it studying one problem: how far RL can scale.

MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task

Nathan Lambert

@natolambert

One of the coolest at-scale RL resources made public yet!
You love to see it.

Fuli Luo

Jev is now available on the Vercel AI Gateway

Jev is TypeSafe AI’s System One evaluation model for fast, structured decisions in software. It evaluates shared state against typed questions and returns choices, scores, and boolean probabilities, supporting classification, routing, rubric-based assessment, and automated verification. Multiple questions can be evaluated in parallel within a single request

Vercel Developers

@vercel_dev

Jev from @typesafeai is on AI Gateway.

Build agents that decide, route, score, and stop in milliseconds:
𝚊𝚠𝚊𝚒𝚝 𝚎𝚟𝚊𝚕𝚞𝚊𝚝𝚎({
𝚖𝚘𝚍𝚎𝚕: '𝚝𝚢𝚙𝚎𝚜𝚊𝚏𝚎-𝚊𝚒/𝚓𝚎𝚟',
𝚜𝚝𝚊𝚝𝚎,
𝚚𝚞𝚎𝚜𝚝𝚒𝚘𝚗𝚜,
});

vercel.com

TypeSafe AI's Jev now available on AI Gateway - Vercel

Use Jev from TypeSafe AI on Vercel AI Gateway to classify, route, score, and automate workflows with typed outputs, probabilities, and confidence.

TypeSafe AI's Jev now available on AI Gateway - Vercel

Use Jev from TypeSafe AI on Vercel AI Gateway to classify, route, score, and automate workflows with typed outputs, probabilities, and confidence.

https://vercel.com

Evaluation

Evaluate shared state against typed questions and get back structured choices, scores, and boolean probabilities through Vercel AI Gateway.

https://vercel.com

Friday 18th September 2026

Wispr Flow announces Canto, its new real-time speech model for dictation, claiming a 3.4% word error rate on 9.8 hours of real-world English audio from over 2,300 speakers—the lowest among models tested on their internal dataset.

Canto uses supervised fine-tuning plus reinforcement learning (GRPO) trained on user corrections and contextual vocabulary, targeting practical low-latency dictation rather than clean read speech.

It's now rolled out for everyone, with nothing to switch on.

Wispr says that they already training its successor at more than ten times the scale, with better recognition in difficult conditions and broader language coverage ahead.

Canto: a speech model built for the real world | Wispr Flow

Today, the Wispr Advanced Interfaces Lab is introducing Canto, our latest speech model for real-time dictation. On an evaluation of real-world dictations, Canto achieved the lowest word error rate among all the models we tested.

https://wisprflow.ai

Claude AI's redesigned Projects shifts from folders to a single conversation where users describe goals and Claude autonomously directs parallel cloud threads that persist and operate independently even offline.

Threads leverage shared project memory and a dedicated library for files/artifacts, enabling coordinated workflows like task delegation, conflict resolution, and context retention across sessions similar to briefing a chief of staff.

Now in beta for select Pro and Max cloud users in Claude Code, with rollout expanding soon and local device support planned; existing projects will upgrade gradually.

Claude

Projects now run from one conversation, starting in Claude Code. You describe what needs doing, and Claude directs parallel threads that keep working after you close your laptop.

In beta today for select Pro and Max users in cloud sessions; coming to all Claude users soon.

Addy Osmani

@addyosmani

Projects are rolling out in Claude Code today.

A Project is an ongoing conversation that coordinates a stream of related work for you. You describe a goal, Claude splits it into threads, runs them in parallel in the cloud and keeps going after you close your laptop.

Starting

ClaudeDevs

Today we're rolling out Projects in Claude Code on desktop and web.

A project is one conversation with Claude. It splits the work into threads itself, runs them as parallel cloud sessions, passes context between them, and keeps going when you leave.

In beta for select users.

cat

@_catwu

In the new Projects experience, Claude coordinates all of your sessions.

I've been using it every day since it lets me operate at a higher level of abstraction. I fire off a batch of tasks and move on. Claude has context on everything I'm working on and can give me an

ClaudeDevs

Boris Cherny

@bcherny

Projects have changed not only how I interact with Claude but how I code.

I stopped managing sessions. I just send thoughts as they come, Claude splits them into threads, and the project remembers how I work. It's where I do a ton of my coding now.

[screenshot: my actual

ClaudeDevs

https://docs.z.ai/guides/vlm/glm-5.3-flash

Just saw mention of GLM-5.3-FlashX on ZAI website, not seeing this in zcode or via OpenRouter as yet. Read some buzz on X

GLM-5.3-FlashX is now live, delivering inference speeds of 200 tokens/s for faster responses and a smoother experience.

Pricing of GLM-5.3-FlashX from ZAI [0.37/1.25/0.075] is more than double of GLM-5.3-Flash [0.15/0.5/0.03]

https://docs.z.ai/guides/overview/pricing

Saturday 19th September 2026

via Lisa Gu

Pulled multiple all nighters on this one 🫡 and so excited the world can finally see it

Rohan Nayak, CEO of Pocket FM, announces Sherpa, their advanced AI fiction writing tool that drove the company's ARR from $250M to $500M in one year by scaling content production 1200% with 550K+ creators generating 2.6M annualized hours.

Sherpa uses a Narrative World Model for long-term continuity, hierarchical planning across seasons-to-scenes, and a Prose Engine trained on 5.5B hours of playtime and retention data to maintain engagement in serialized stories.

Pocket FM's owned distribution enables one-tap audio production, performance-based revenue share, and localization, allowing 10% of writers to earn over $200K and creating blockbusters exceeding $100M while fostering a creator-AI feedback loop.

Rohan Nayak

@RohanNayak2

Introducing Sherpa: the most advanced fiction writing AI

We accelerated from $250M in ARR to $500M because Sherpa helped increase content production by 1200% in 1 year

Sherpa was trained on 5.5B hours of playtime with minute by minute dynamic retention data.

550K+ creators

Sherpa by Pocket FM | Free AI Fiction Writing Tool

Sherpa by Pocket FM is a free AI fiction writing tool that helps anyone turn ideas into stories. Write it, produce it into audio and publish it.

https://pocketfm.com

via Matt Weigand

Hey everyone - I’m Matt Weigand at Perplexity.

A lot of my interest comes back to search: if a model doesn’t find the right information, everything that follows starts with an incomplete picture.

If you are interested in what Perplexity is doing with web research, check out this study we just published - Q2D-Web. The study digs into evaluating how well search systems retrieve relevant documents at web scale. It’s work I’m excited to share more about here, along with what better retrieval means for people building and using AI.

Q2D-Web: Evaluating First-Stage Retrievers at Scale- https://www.perplexity.ai/hub/blog/q2d-web.

Look forward to connecting!

Finally, Anthropic taking small steps to use conventions used by other harnesses

Claude Code now [as of version 2.1.277] reads AGENTS.md if there is no CLAUDE.md

Hopefully they will in the future also read skills from .agents/skills so one doesn't have to do symlink jugglery

Thariq

@trq212

We're adding support for AGENTS.md to Claude Code.

Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md.

You can toggle this behavior in /config.

You choose your AI assistant. But did you know you can also choose how it searches?

The models that help you think, write, and reason don’t have to determine which tools find information for them.

In a compatible AI application, you can connect search and retrieval providers separately. That gives you more freedom to combine the models, tools, and costs that suit you.

I built the multi-provider-research skill to help your assistant work with those tools—choosing suitable providers, checking source evidence, and explaining gaps. You can start with just one provider.

I’ve put together a practical guide covering:

• What harnesses, skills, and providers actually mean

• How to get started with options such as Exa, TinyFish, Octen, and Perplexity

• Free allowances and lower-cost combinations worth exploring

📖 Choose Your AI. Choose How It Searches.
https://goolamabbas.github.io/separate-intelligence-and-search/

The guide links directly to the skill and further reading. I’d love your feedback, especially if this is your first time exploring skills or connected search tools.

Below is my personal website which aggregates links to many of my socials as well as the various content and community that I curate. Feel free to share this link to others who you think may find this content/community useful to them

The cover image of this newsletter via generated via the GPT Image 2.5 model within the Krea tool via the following prompt

Create an artistic representation of the Garden of Eden from the Book of Genesis, capturing its lushness and tranquility, Mysterious --no buildings, people

This Week in All Things AI - Week 37-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 13 Sep 2026 01:24:36 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 6th to 12th September 2026 brought advances in both AI performance and the tools that turn it into everyday work. DeepSeek introduced V4.1 Flash with native visual understanding and strong self-reported coding and agent benchmarks, while Inception’s Mercury 2.5 pushed diffusion-based language generation beyond 1,100 tokens per second. NVIDIA’s Sol-H3 demonstrated video generation faster than playback, and ChatGPT Images 2.5 and Eddie v4 expanded creative editing capabilities. Cursor Projects added another dimension: a persistent coordinator agent that manages subagents and ongoing work in the cloud. Alongside Hugging Face’s ML Intern and ChatGPT Work’s Data agent, it pointed toward workflows in which users delegate an extended task rather than prompt each individual step.

Making that delegation effective—and deciding where its limits belong—was a recurring theme. My Portable LLM Council offered a structured way to examine competing arguments before making a decision, while Spotify’s use of cheaper worker models and Tencent’s FlexKV showed how orchestration and infrastructure can make better use of compute. Discussions around GPU financing and rental economics questioned the durability of the assets supporting AI’s expansion, even as Harvey’s latest funding round reflected growing investment in professional applications. Meanwhile, allegations concerning OpenAI’s handling of researchers’ conversations brought privacy and research ownership into focus. Taken together, the week showed why progress needs to be assessed across several dimensions: what systems can do, how economically they can do it, and how confidently people can entrust them with valuable work.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interest

Sunday 6th September 2026

Robert Scoble with a longish X article which links to many posts on what folks have been able to do GPT-6 Astra. It might interest and inspire you (modulo how much you can use within your plan's limits, lower reasoning levels of Astra stretch limits further).

FWIW, I had Astra Light review my recent Notion page and it dramatically edited it cutting the number of words on the main page by half and generating cleaner SVG diagrams and it's token generation speed is very fast for its intelligence.

When Jalapeno (OpenAI's custom silicon for inference) rolls out at scale, Astra and future versions are going to be running at Speedy Gonzales/Road Runner pace

Robert Scoble

@Scobleizer

My AI read everything here on X about OpenAI's Astra and wrote this report.

Good, bad, ugly.

All from the AI community here on X (tens of thousands of posts read). x.com/i/article/2096…

GPU Financing Emerges as AI Infrastructure's Key Bottleneck

Eugene Ye of Wafer AI explained how AI builders chase Blackwell B300 GPUs and colocation space, but hit walls with lenders demanding huge down payments since used GPUs lack predictable resale markets. Unlike cars with established secondary values, GPUs get treated as worthless after depreciation, squeezing smaller players.

Efforts like Nvidia's partnerships with Apollo and BlackRock aim to unlock over $500 billion in capital, while companies like CoreWeave land big loans through customer contracts—yet many still rely on venture funding amid hopes for better liquidity tools.

Eugene Ye

@gpugene

The biggest bottleneck in the GPU industry at the moment is not chips, CoWoS, or even power. It's credit.

Sharing some thoughts on compute markets, residual value, and where compute financing is heading.

Substack link in the comments. x.com/i/article/2096…

Monday 7th September 2026

Introducing Portable LLM Council for business and technical decisions

Ever asked an AI for advice on a business or technical decision—and wondered whether it examined the tradeoffs or simply followed your framing?

I’ve put together Portable LLM Council, an open-source skill that gives your AI assistant a structured way to challenge a decision before recommending what to do.

It runs three rounds:

• Five separate advisors examine assumptions, fundamentals, overlooked opportunities, clarity, and execution.

• Five fresh reviewers assess the anonymized answers for strong arguments and blind spots.

• A chair weighs the reasoning and recommends a direction and one concrete first step.

You get a readable HTML report and a Markdown transcript, so you can inspect the arguments behind the recommendation.

Useful questions include:

— Should we launch a paid pilot or build the full product?

— Should our team focus on onboarding or acquiring more customers?

— Should we migrate a system gradually or replace it in one go?

These are starting questions, not complete prompts. Give the assistant relevant context—your objective, constraints, options, and evidence so far—or ask it to use what you’ve already discussed in the conversation. A few focused paragraphs are usually enough; the council needs something concrete to challenge.

The same skill works with native subagents for Claude Code—including local Code sessions in Claude Desktop—Codex, Cursor, Command Code, Grok Build, OpenCode, and Factory Droid. No particular model is required.

To install, ask your assistant:

“Read the installation instructions at https://github.com/goolamabbas/portable-llm-council and install the shared skill and the agent definitions for this harness in my personal installation locations. Back up existing copies before replacing anything, then verify discovery. Do not run a council yet.”

Your assistant may request approval for file changes. You may need to start a new session before it can verify discovery. Manual instructions and example prompts are also in the repository.

This is for decisions with meaningful tradeoffs. It uses more time and model usage than a normal answer, and agreement between agents doesn’t make a conclusion true. The value is in exposing assumptions, competing arguments, and questions worth investigating.

GitHub - goolamabbas/portable-llm-council: Pressure-test business and technical decisions with five AI perspectives, blind peer review, and a synthesized recommendation. One model-neutral skill with native subagents for seven AI harnesses.

Pressure-test business and technical decisions with five AI perspectives, blind peer review, and a synthesized recommendation. One model-neutral skill with native subagents for seven AI harnesses. - goolamabbas/portable-llm-council

https://www.reactor.inc/blog/fast-h3-sol-engine

If you try it, I’d welcome feedback—especially what it helped you notice that a normal AI conversation missed.

Tuesday 8th September 2026

NVIDIA Speeds Up AI Video Generation Past Real-Time Playback

NVIDIA's Sol-H3 inference stack accelerates MiniMax-H3 video generation, producing five seconds of 1344×768 video with stereo audio in 1.653 seconds on an 8× B300 Blackwell system—beating real-time playback. Reactor partnered to release FastH3 via API, enabling developers to access it immediately without owning the GPUs, while open-source code under Apache 2.0 supports various configurations. Optimizations like dynamic sparse attention and fused kernels preserve quality and scale efficiently, paving the way for interactive, 24 FPS video systems

Enze Xie

@xieenze_jr

Sol-H3: @MiniMax_AI H3 Video Generation Faster Than Playback

Five seconds of world. 1.653 seconds to infer.

We’re releasing Sol-H3, our fastest end-to-end MiniMax-H3 inference stack yet. On one 8× NVIDIA B300 Blackwell system, it generates five seconds of 1344×768 video

reactor

@reactorworld

Introducing MiniMax FastH3 on Reactor - the latest open-source model by @MiniMax_AI.

Stream 720p with audio at $0.007/second with first frame image support, faster and cheaper than ever.

Get started now by cloning our demo app: reactor.inc/fast-h3-demo-a…

https://nvlabs.github.io/Sana/Sol-Engine/Sol-H3/

NVIDIA's Jensen Huang Calls Compute Resources Rentable Assets Amid Surging Demand

Huang described the company's compute power as fungible, durable, and highly rentable—a productive asset generating revenue. This comes as demand stays hot for three-year-old H100 GPUs, now renting at $3.28 an hour, prompting neoclouds to lock AI firms into multi-year contracts to finance big purchases. Experts like Jessie Dong caution that future efficiency gains and potential customer failures could leave providers stuck with debt and underused hardware in a softer market. The conversation draws parallels to early commodity markets like grain and oil, questioning if compute will prove as sturdy an asset class.

Jensen Huang

@JensenHuang

NVIDIA compute is fungible, durable and highly rentable. It is a productive, revenue-generating asset.

Ornn

@OrnnExchange

The H100 is a three-year-old training chip. Its rental price is up 22 percent on the month, to $3.28 an hour.

Every depreciation schedule assumes a chip this old only loses value. The market is paying up for it instead.

Jessie Dong

@jessiedong_

my hot take of the day: companies and neoclouds are going to regret locking themselves into too much compute at today's prices.

companies are signing long contracts because not getting enough compute is a risk. there's demand so of course neoclouds are buying lots of gpus

In response Alex wrote

Without context, Jensen's statement is misleading. The depreciation for a 2023 bought H100 alone is around 1.5$ per hour.

The current spot market price from mid-tier cloud providers is 1.20$ a GPU/hour, ie. being rented at a loss.

Compute is not behaving like a commodity, too stock and hype dependent.

Also conveniently not mentioning that H100 rental prices are down more than 60-70% in the last 2 year.

A Claude Code plugin called shunt and two declarative agents running on Portal by Spotify. That's all it took to stop burning frontier model tokens on file reads and boilerplate generation. The grunt work routes to a cheaper worker model. The frontier model only handles problems worth its price.

90% token reduction on bulk reads. No new infra. The agents are public and the plugin is open source.

Portal by Spotify cut my Claude Code token usage by 90% | Spotify Engineering

Most of what an AI coding agent does for me isn't thinking. It's I/O.

https://engineering.atspotify.com

via Alex

NYU released a statement claiming that OpenAI stole a millennium prize discovery through scanning and monitoring a fields medalist's ChatGPT chat and then advertised the discovery as an "autonomous" discovery.

Never feed anything of value into a remote model.

Read the full statement here: https://cims.nyu.edu/~tristanb/statement.pdf

Wednesday 9th September 2026

OpenAI announces ChatGPT Images 2.5 with faster generation, higher fidelity for natural images, consistent details in edits, and comment-based controls to modify specific elements without full regeneration.

New Sketch feature lets users draw directly in ChatGPT by typing "@ Sketch" to visually convey ideas that are hard to describe in text alone, alongside templates for common formats like posters and merch.

The model update deploys today to all ChatGPT, Work, and Codex users across platforms while adding API models GPT-Image-2.5 Flare for improved speed and GPT-Image-2.5 Sunburst for precision in detailed creative tasks.

OpenAI

ChatGPT Images 2.5—faster, sharper, smarter, with better tools for creating whatever you can dream of.

- Faster image generation to keep your ideas flowing

- Improved fidelity for more natural, recognizable images

- Consistent details across multiple edits

- Comment-based

OpenAI Developers

@OpenAIDevs

Meet GPT-Image-2.5 Flare and Sunburst.

Introducing new image models in the API, with sharper detail, stronger style adherence, and more control over edits.

Krea

@krea_ai

introducing GPT Images 2.5

this new version comes with improved realism, consistent editing, and faster generations.

try it now

fal

https://www.inceptionlabs.ai

GPT Image 2.5 is now available on fal

- Edits scoped precisely to the instruction, leaving subject, composition, and background untouched.
- Reference subjects stay recognizable across new settings and styles.
- Consistency holds across long editing sessions without quality

via Alex

Here the OpenAI statement on the NYU allegations: https://openai.com/index/navier-stokes-solution/

Pretty sure this will end up in a lawsuit.

OpenAI claims that they did in fact only start on the research to solve the millennium problem after (after Sep 1st) having been informed that the NYU professor was close to a solution, they deny that the verbatim chat logs were accessed but hint at the fact the same chat logs may have been used for training data.

In other words, OpenAI trained an internal model specifically on the professor's work and deployed over several 10,000 agents in 1 week to burn 300 billion tokens to win over a human that had dedicated their life on solving this problem.

Inception announces Mercury 2.5 as the largest and most capable diffusion LLM yet, delivering a 40% intelligence increase over Mercury 2, 1,107 tokens/sec inference speed on NVIDIA GPUs, and 260K token context.

The accompanying video contrasts traditional autoregressive LLMs (sequential like typewriters, tying up GPUs) with diffusion-based processing for greater parallelism, enabling low-latency uses like real-time voice agents under 170ms TTFT.

The model is available immediately on Inception API, OpenRouter, and Baseten with launch discounts; early tester feedback notes faster performance and better tool calling confidence.

Inception

@_inception_ai

Today, we’re introducing Mercury 2.5, the most capable diffusion LLM on the market.

It offers a 40% jump in intelligence over Mercury 2, and runs over 1,100 tokens/sec on widely-available @NVIDIAAI GPUs.

It’s available today on our API, @OpenRouter, and @Baseten.

Contact us

Introducing Mercury 2.5 - Inception

Mercury 2.5 is the most capable diffusion LLM on the market. It runs at 1,107 tokens/sec and offers a 40% increase in intelligence over Mercury 2, comparable to cost-optimized frontier models.

Tencent AI's post promotes FlexKV, an open-source multi-level KV cache system from their TACO team that offloads cache to CPU, SSD, and remote storage to handle long-context LLM inference without frequent re-prefills.

It solves GPU wait times on cache hits via layer-by-layer restoration, prefetching, and asynchronous I/O, while using a distributed RadixTree for cross-node prefix reuse and lossless compression to boost effective capacity.

FlexKV integrates directly under engines like SGLang, vLLM, TensorRT-LLM, and Dynamo, with claimed gains of 70% lower TTFT and 16% higher QPM; full details and repo at github.com/taco-project/FlexKV.

Tencent AI

@TencentAI_News

A cache is only worth what it hits.

The community's having a KV cache moment. Here's the corner we work in:

In long-context serving, a cache hit can still leave the GPU waiting for data. When KV lives outside GPU memory, how you bring it back matters. That's the problem we set

Reve did a strategic partnership with OpenAI, under which the research team has joined OpenAI aka license and acqui-hire. It's not clear at least to me if their work went into GPT Image 2.5 or if todays launch was a means to announce it

Taesung Park

@Taesung

Update: Reve is doing a strategic partnership with OpenAI, under which the research team has joined OpenAI to continue our journey on controllable image generation! As a cofounder, I am deeply grateful to all my coworkers for letting our models shine and bringing our baby Reve to

OpenAI

Michaël Gharbi

@m_gharbi

Excited to share that the Reve research team is joining @OpenAI through a strategic partnership, taking our work on controllable generation to the next level, at the frontier of multimodal intelligence!

To my cofounders @Taesung, @cantrell, @jwatte, @laserlikemike and everyone

OpenAI

Now officially posted on Deepseek's website

DeepSeek plans to officially release the V4.1 Flash model around September 10, 2026 (Beijing Time). After extensive internal and external testing, V4.1 Flash has comprehensively surpassed V4 Pro across all key metrics, including performance, cost, speed, and task completion time. In keeping with our commitment to user responsibility, following the official launch of V4.1 Flash and prior to the release of V4.1 Pro, all requests to the Pro model will be routed to V4.1 Flash and billed at Flash's price. If you encounter any issues during your comparative testing between V4 Pro and V4.1 Flash, please do not hesitate to reach out to us with your feedback. Thank you for your support!

We will adjust the pricing for the Flash series effective from 12:00 Beijing Time on September 10, 2026. During off-peak hours, the unit price will be $0.003 for input cache hits, $0.15 for input cache misses, and $0.6 for output. Peak-hour prices will be double the off-peak rates. Please plan your usage accordingly.

Harvey Raises $550 Million at $15.6 Billion Valuation

The legal AI startup raised $550 million in a round co-led by Lightspeed Venture Partners and Diffusion, boosting its valuation from $11 billion to $15.6 billion. With over $400 million in annual recurring revenue, Harvey now serves 80 percent of the top 100 law firms, 20 percent of the Fortune 500, and half of the Fortune 10, while token usage has surged 20 times since January. The funds will expand proprietary AI models, hire engineers, and let law firms build their own custom intelligence, including scaling to thousands of GPUs after releasing the advanced Harvey Tenet model.

Harvey

@harvey

We’ve raised $550M at a $15.5B valuation co-led by @lightspeedvp and @DiffusionVC.

We're using this funding to help law firms, in-house legal teams, and professional services build and own their intelligence.

Thursday 10th September 2026

Shamir Allibhai whom I shall be hosting on a roundtable on Sep 21/Sep 22 [depending on your timezone] let me know about the release of Eddie v4.

The blog post below has the various details of feature updates in this version and also has 2 embedded Youtube videos including one with Youtuber Jordan Orme [1.1M subscribers] demoing Eddie's features to other professional video editors

Eddie v4: The Most Powerful AI Assistant Editor in the World | Eddie AI

Eddie v4 watches your rushes, syncs your cameras, writes the music, sources the b-roll, and hands the finished cut to Premiere, Resolve, Avid or Final Cut as a real project file. See it first at IBC 2026.

https://www.heyeddie.ai

PS: I look forward to having as many members participate in the round table. This will be recording and made available via the archives of my previous roundtables

via Alex

Another mathematician has come out accusing openAI of breaking their own terms and conditions by using "opted out" chats for training data: https://mathstodon.xyz/@andreasthom/117240535270608201

Willow Voice is doing $100k giveaway of dictation microphones plus free months of their AI voice-to-text service targeted at SF startups to encourage office adoption.

The attached video features CEO Allan Guo and GTM lead Jesslyn Teo demonstrating Willow's accurate dictation that works across apps, even with soft speech, in professional settings.

Engagement requires commenting "Voice-maxxing" for setup, positioning the product as enabling faster workflows by letting users "work at the speed you talk."

Willow

@WillowVoiceAI

We’re giving away $100k in dictation microphones to SF startups.

We’ll outfit your office with mics + months of Willow free.

Whisper into your mic. Willow writes for you. Work at the speed you talk.

Comment “Voice-maxxing” and we'll set you up.

PS: I'm personally all-in on Wispr Flow primarily because of early adopter discounts and some referral bonuses which means that I'm effectively paid for a few months more I've tried Willow Voice and it's a good product with recent improvements moving it up several notches

Hugging Face aka NVidia says that one shouldn’t need to be an ML expert to have an ML idea.

They have launched ML Intern in HuggingChat, an AI agent in HuggingChat that automates end-to-end model training by combining the platform's papers, datasets, pre-trained models, benchmarks, storage, and compute resources

The release positions the tool as making specialized model training conversational and accessible like vibe-coding apps, with user control over budget, steps, and outputs for reproducible artifacts.

Start with a conversation. Finish with deployable artefacts.

Thomas Wolf

@Thom_Wolf

in 2026 training a model for a task should be as easy as vibe-coding an app

the pieces were already available on : papers, datasets, pre-trained models, benchmarks, storage and compute

what was missing was an agent to put them together and run the whole process

that’s why

HuggingChat - Chat with AI models

Making the community's best AI chat models available to everyone

https://huggingface.co

Jay Yu of Pantera Capital with a very well articulated X article on 'The Rise of Compute Markets' where he states the following

Compute and datacenter spend is a trillion-dollar category, yet GPU buying still runs through group chats, OTC brokers, and bilateral deals.
Compute is early in financialization; like electricity it is bound by SKU, time, and location, but may become a commodity like power or oil.
Markets may follow a Hardware-Vendor-Cluster analog of power’s Grid-Operator-Node, with winning benchmarks tied to liquid physical settlement and more basis risk.
Neoclouds are the structural short for GPUs; on-tap platforms and apps are the long; NVIDIA acts as a compute central bank.
Physical settlement has the durable moat; indices, cash-settled exchanges, and financing vehicles follow; fullstack players and token-plus-metal hedges may win as compute becomes an asset class.

Jay Yu 🐟

@0xfishylosopher

x.com/i/article/2097…

Team Deepseek takes no prisoners 🤯 with the launch of DeepSeek-V4.1-Flash with stellar self-reported benchmarks.

Will be interesting once Artificial Analysis releases their benchmarks.

Available with extended usage for limited period within OpenCode and Command Code harnesses. Also available in v0.1.5 and above of the Deepseek Harness

DeepSeek-V4.1-Flash is a 552B-parameter MoE model with asymmetric Causal Encoder-Decoder architecture, using only 8B active parameters for input and 16B for output, plus native visual understanding.

It achieves leading results on agentic and coding benchmarks such as Terminal-Bench 3.0 (30.0), DeepSWE v1.1 (74.2), CyberGym (88.1), and Automation-Bench (54.8), outperforming V4-Pro and competitors like Kimi, GLM, and Claude.

Live on DeepSeek API as deepseek-flash with lower pricing effective today, dramatically reduced KV cache for cost savings, and phasing out V4-Pro, with open-source inference support planned.

DeepSeek

Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.

Introducing the smallest model in our new architecture family, with native visual understanding.
Designed for greater capability, faster inference, higher throughput, and scaling to larger models.

1/6

OpenCode

day 0 support
we've bumped it up to 4x usage for the next 72 hours

OpenCode

DeepSeek Flash v4.1 is now available in OpenCode Go

Command Code

DeepSeek V4.1 Flash is live in Command Code.

6x usage on $10/mo GOAT for a week.
$60 usage. 154K requests. 7.8B tokens.

GOAT plan is the best AI coding plan.

Available in all plans and API.

Tianyi Cui

@tianyi

欢迎试用 DeepSeek Harness v0.1.5 版本。新版本与 DeepSeek V4.1 Flash 模型训练深度结合，模型在 DeepSeek Harness 的不同配置中都进行了专项训练和优化。

与模型训练深度结合的实验性的 Agent Teams 功能也期待尝鲜试用。

DeepSeek Harness 团队仍在开放招聘中，欢迎加入我们，一起打造

OpenAI announces a new Data agent in ChatGPT Work that lets enterprise users query connected company data sources via natural language to generate answers, interactive dashboards, and actions.

The video demo shows practical use cases like analyzing product launch metrics (e.g., "Dash XV" sneaker sales), pulling context from Snowflake/Tableau/Teams, and creating visualizations that answer "why" questions beyond basic reports.

Integration features include adding a Data Plugin, connecting existing data tools, and sharing AI-generated insights directly in team channels with ChatGPT mentions for collaborative workflows.

ChatGPT

@ChatGPT

Now everyone can put data to work.

We’re introducing a new Data agent in ChatGPT Work so you can turn your company’s data into answers, interactive dashboards, and action—just by asking.

Just add the Data Plugin in ChatGPT Work, connect to the data sources and context you

Now everyone can put data to work

Meet the Data agent in ChatGPT Work. Connect company data, uncover insights, and build interactive dashboards with AI using natural language.

https://www.instagram.com/reel/DdHVH_EpRRJ/?stkn=cnFoYXp3MjJ3cXdw

Friday 11th September 2026

via Lisa Gu

This is my losing video from the finals of the ai creative challenge last week in LA hosted by the one club for creativity and google. I’m letting the Internet be the judge. 😭👩🏻‍⚖️

It’s a 16 city tour, so apply if you want to compete!

https://creativechallenge.ai/

Cursor Launches Projects for Multi-Agent Coding Workflows

Projects creates persistent threads where a main agent oversees subagents, shares memory across sessions, and handles tasks like monitoring pull requests or Slack. It runs on cloud computers for nonstop work and scales to thousands of agents for big jobs like app builds. Early users like Lauren (of pstack fame) call it a 'personal software factory,' reporting they merge far more PRs—up to six times as many for heavy users—thanks to better organization and speed.

Cursor

Introducing Projects, a new way of working in Cursor.

Rather than creating a chat for every task, you work with a coordinator agent in a single, persistent thread.

Like @bot, your agent is always on, proactively manages work with subagents, and improves over time.

lauren

@poteto

it's here! i've been using Projects for a while and have shipped thousands of PRs with them. pstack works incredibly well with it, with your coordinator managing and orchestrating hundreds and even thousands of agents for you

these all work in the cloud too, so you can

Cursor

Saturday 12th September 2026

https://www.nytimes.com/2026/08/31/business/dealbook/together-ai-humain-saudi-arabia-data-center.html?unlocked_article_code=1.9lA.oOPI.h2UKXnt1l8Cz

The cover image of this newsletter via generated via the GPT Image 2.5 model within the Krea tool via the following prompt

a detailed digital painting of a beautiful white sci-fi city, ((intricate city detail)), straight lines, on a lush green landscape, intricate detail, two planets visible in the clear blue sky, fantasy art, trending on artstation, 8k, uhd

This Week in All Things AI - Week 36-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sat, 05 Sep 2026 23:52:49 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 30th August to 5th September 2026 brought a relentless run of model launches and AI product updates across text, coding, image, speech, video, and infrastructure. Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1, Google released Gemini 3.8 Flash alongside its cyber-focused variant, Meta upgraded Muse Spark 1.3, and OpenAI began a staged rollout of ChatGPT 6 Astra. Microsoft also expanded its in-house model portfolio with MAI-Image 2.6 and MAI-Image 2.6 Flash for image generation, alongside MAI-Transcribe-2 for speech-to-text, while IFM’s fully open K2 Horizon family ranged from edge-scale models to 375B parameters. Inception’s Mercury 2.5 preview, meanwhile, showcased how diffusion-based language models could offer a compelling alternative for high-speed, latency-sensitive reasoning and agent workloads.

Video and multimodal AI also moved further toward interactive, real-time use. fal’s H3 Max, built on MiniMax H3, and related open inference work demonstrated video generation faster than playback; World Labs unveiled Atlas for controllable 3D scene reconstruction and world simulation; and Google added agentic video understanding to Gemini, allowing models to actively inspect relevant frames, audio, and transcripts rather than process a video uniformly. Perplexity’s hybrid compute feature illustrated a parallel shift in agent design, combining cloud models with local Apple Silicon processing for privacy-sensitive files, while Together AI’s planned Saudi data-centre partnership underlined how power access and compute capacity are becoming strategic constraints.

Taken together, the week showed that the race is no longer only about stronger general-purpose models, but about making multimodal AI faster, cheaper, more private, more deployable, and useful in real production workflows

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interest

Sunday 30th August 2026

Friends,

Video editing looks effortless — until you try it. I find it genuinely hard, and I have real admiration for people who've mastered it: the ones who see a story where the rest of us see six hours of messy footage.

Join me for the next session of my conversation series on Tuesday, Sep 22, 12:00 PM HKT (India 9:30 AM - London 5:00 AM - NY 12:00 AM Tue - LA 9:00 PM Mon Sep 21), via Google Meet, with Shamir Allibhai — filmmaker turned Meta PM on video/AI infra, now a repeat founder whose current company, Eddie AI, builds an AI assistant editor for professionals.

This isn't a product demo. We'll talk about the career — filmmaking, product work at Meta, repeat founding — what he hears from editors using AI tools in real pipelines, and why editing is so hard to master. I'll be bringing my own beginner's frustrations; bring yours.

Format: a round-table conversation, not a webinar — everyone engages Shamir directly. Drop a question in the thread beforehand if you'd like it teed up, or bring it on the day.

The session will be recorded and archived with my previous conversations — so if 5 AM in London or midnight in NY isn't your friend, the recording has you covered.

Monday 31st August 2026

Tuesday 1st September 2026

Together AI announced a partnership with Saudi AI firm Humain for a 250MW data center equipped with 120,000 semiconductors, expected to generate over $5 billion in annualized revenue through shared compute for open-source AI models.

The deal is based in Saudi Arabia as part of the kingdom's push to develop its AI industry and reduce oil dependence, enabling Together AI to secure capacity amid U.S. local opposition and moratoriums on new data centers.

Power access emerges as the core AI infrastructure constraint, with this collaboration illustrating how open-source players are leveraging international partnerships to scale compute without solely relying on domestic hyperscaler funding.

Vipul Ved Prakash

@vipulved

We executed a landmark partnership with @HUMAIN today to offer open source models on 250 mw (-100K+ chips) of capacity over the next 12 months. The demand for open and custom AI tokens is vertical and this partnership represents one of the largest so far focused on delivery OSS

Together AI

@togethercompute

We just signed one of the largest AI infrastructure deals for open source, period.

250MW data center. $5B+ in annualized revenue. Built with @HUMAIN in Saudi Arabia.

tinyurl.com/4nzcjcu7

Whilst I understand the rights of someone to offer such a product, IMHO this is equivalent to giving access to a cyber-offensive product to anybody with a credit card

https://abliteration.ai/

Unrestricted AI API + Enterprise Policy Gateway | abliteration.ai

OpenAI-compatible unrestricted AI and uncensored LLM API for enterprise teams: AI red teaming, cybersecurity, trust and safety, synthetic data/evals, ML research, and defense/government contractor workflows. Policy Gateway adds policy-as-code, per-project keys, audits, and zero retention by default.

https://abliteration.ai

Abliteration.ai

@abliteration_ai

Today we're releasing abliterated-model-large-v2.

Based on GLM-5.3, which is #3 on Terminal-Bench 4.0 (behind only Opus 5 and Fable), with 2× the cyber exploitation of 5.2.

We abliterated and hosted it so it does the offensive cyber, red teaming, and agent testing work other

Sharing something sent to me from Ethan Wei of Minimax

===

Regarding the film and video sector, our recently released H3 and H3-Max models have garnered widespread coverage and attention on X. H3-Max delivers exceptional video generation speeds—producing a 5-second video in just 3 seconds and a 15-second video (at maximum duration) in around 10 seconds with high-quality results.

This capability has spawned a range of new use cases and innovative applications, including—but not limited to—live streaming, in-game narrative progression, real-time virtual try-ons, rapid prototyping of cinematic shots, and AR/VR extensions.

@levelsio

Okay I built it!

Infinite Slop
levels.io/infinite-slop

An infinite and interactive AI generated live stream of slop that goes on forever and ever

Anything that you write in the chat is generated next and AI will try to connect it to the previous video so there's an actual

@levelsio

Today is a very historical moment for AI video generation

You can now generate AI video faster than you can watch it

Before it'd take let's say 2-5 minutes to generate 15 seconds of video

@fal made a post-trained Minimax H3 variant called Max which is 50x faster than the

Joaquín Cuenca Abela

@cuenca

getting closer to Ender's game videogame, where Ender explores a world interacting the AI

Blendi

@BlendiByl

i've been using minimax h3 max to make interactive games you control!

all the decisions are up to you, and since the model is so fast there's basically no delays

Infinite Slop by @levelsio + fal.ai

An endless AI-generated TV channel where you decide what airs next.

https://infiniteslop.ai

Create Video Generation Task - MiniMax API Docs

Video generation V2 endpoint. Provide multimodal input via the content array (text / image / video / audio); switch between MiniMax H3 and MiniMax H3 Max with the model field to support text-to-video, image-to-video (first & last frame), and reference-to-video, with output up to 2K.Note: To use MiniMax H3 or MiniMax H3 Max, please select the Pay-as-you-go API.

https://platform.minimax.io

Lauren Tan software engineer at SpaceXAI (formerly Cursor, Meta, Netflix) who goes by the handle of poteto and is the creator of the pstack AI coding workflow plugin recently wrote an X article about pstack which is getting great buzz

Ray Fernando

@RayFernando1337

If you do anything today, block out and prioritize this. The level of detail and care that Lauren shares is easily worth six figures and will have a massive impact on your productivity. There is so much alpha on X it’s insane.

lauren

@poteto

I'm writing a guide to pstack! Here's part one. x.com/i/article/2094…

klöss

@kloss_xyz

A Grok Bot engineer just shared the best article I’ve read this year on agents.

2,000 PRs a month, and all of it rests on self-verifying agents: /create-verification-skill

The takeaway: give agents tools, not markdown.

Most won’t read this.

Make sure you do.

lauren

@poteto

I'm writing a guide to pstack! Here's part one. x.com/i/article/2094…

Came across this X post which links to an hour long video of Lauren going through her setup

Carnage

@0xCarnagee

SpaceXAI engineer (ex-Cursor):

"right now I've got 20-30 GrokBot agents running. they handle almost everything that used to eat my week

one of them is a Chief of Staff. it knows every other bot I have and hands out the work itself"

in a 56-minute speech, a SpaceXAI engineer

Carnage

@0xCarnagee

x.com/i/article/2090…

Wednesday 2nd September 2026

via Just|LDA who shared a link to the launch post of Anthropic's Claude Fable 5.1 and Claude Mythos 5.1

Claude

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.

They're the world’s most advanced models for coding and knowledge work.

Mercury 2.5 from Inception is exclusively available via OpenRouter

This is diffusion based unlike other LLM's which are autoregressive

Diffusion language models represent a shift from traditional autoregressive generation by using diffusion processes to create text. Instead of predicting one token sequentially, these systems produce many words in parallel, resulting in faster inference and lower costs.

Price atleast till 9/7 is 0.04/0.15/0.004

Inception

@_inception_ai

Mercury 2.5 Preview is live exclusively on OpenRouter.

80% off pricing through 9/7

OpenRouter

https://openrouter.ai/inception/mercury-2.5-preview#providers

1/ The fastest reasoning LLM is now live exclusively on OpenRouter.

Mercury 2.5 Preview from @_inception_ai reaches 1,107 tokens/sec through parallel token generation, with tunable reasoning, parallel tool calls, and schema-aligned JSON.

Built for latency-sensitive workloads.

Perplexity's Mac app now supports hybrid compute, letting its "Computer" AI agent split tasks between cloud frontier models and local models running on Apple Silicon for privacy-sensitive work like bloodwork or tax files.

The approach routes private data to local processing with no token costs while using cloud models for non-sensitive steps, addressing limitations in most AI agents that ignore native vs. web runtimes.

Perplexity open-sourced its PII classifier on Hugging Face to detect when workloads should stay local and published a research blog detailing the privacy-protecting local runtime implementation.

Hybrid compute is available today for Perplexity Pro, Max, and Enterprise subscribers in the Perplexity Mac app with macOS 15+.

Aravind Srinivas

@AravSrinivas

We're introducing hybrid compute for all users of the Perplexity Mac app.

This will allow Computer to orchestrate local models that can run locally on Mac, particularly for agent steps involving sensitive and private files (eg your bloodwork, tax returns, litigation, etc).

Perplexity

Introducing hybrid compute in Perplexity Computer.

Computer can start a task in the cloud, then move to a local model that runs on your Mac. Use it for steps involving private files or sensitive data.

Available today in the Perplexity Mac app.

Couple of posts on H3-Max. a video generation model post-trained by fal.ai on MiniMax H3 and optimized for high-speed generation. It delivers mainstream 480P and 768P output and generates faster than MiniMax H3. H3-Max is also available via Minimax's own inference platform

MiniMax H3 Max: Free AI Video Generator, Ranked #1, Post-Trained by fal | fal

MiniMax H3 Max is fal's post-trained MiniMax H3, ranked #1 for overall quality, prompt understanding and aesthetics. A 5-second clip at 768p renders in under 3 seconds with synchronized audio. Sign in for five free 15-second generations every day, and 50% off for 14 days.

https://fal.ai

fal

We trained H3 Max for frontier-model quality, tuned around what people actually care about in generative video: fidelity, aesthetics, character consistency, prompt following, shot and scene accuracy, and so on.

The order-of-magnitude speedup was the secondary target. Happy that

ben

@contraben

The implications of generating a 5 second video in 3 seconds are wild.

I think we are going to see a new consumer paradigm emerge outside of endless video streaming.

Whats even crazier is the performance is close or matching SOTA closed sourced video models.

At 9.2 seconds

fal

We’re extending H3 Max launch pricing for another week, with an additional discount.

Through September 7, all text-to-video and image-to-video generations are 75% off.

That brings H3 Max down to $0.0125/sec at 480p and $0.02/sec at 768p. This means it only costs $0.10 to

fal

https://platform.minimax.io

Introducing H3 Max, new post-trained video model by fal Research.

H3 Max ranks #1 for overall quality, prompt understanding, and aesthetics against leading video models, on both first-party and third-party independent evaluations while generating a 5-second 720p video under 3

vLLM

@vllm_project

Video generation faster than playback!

MiniMax H3 on vLLM-Omni + FastVideo's FastH3: a complete 10.1s MP4 - video AND synchronized audio - rendered in 8.7s!

Thanks to @MiniMax_AI for the great Minimax H3 release, the FastVideo team @haoailab for open-sourcing FastH3 and

Video Generation - MiniMax API Docs

MiniMax's video models (MiniMax H3, MiniMax H3 Max) enable efficient video content creation.

World Labs announces Atlas, a multimodal world model pretrained from scratch that generates images/videos with pixel-perfect camera control, reconstructs 3D scenes from sparse inputs, and simulates space-time dynamics using a unified autoregressive diffusion transformer.

Key demos show Atlas creating 1-minute 1440p videos from reference images along custom trajectories, producing explicit 3D outputs like point clouds and Gaussian splats, and enabling real-to-sim robotics workflows plus bullet-time reframing from a few phone cameras.

The model outperforms specialized 3D reconstruction and camera-controlled generation baselines per company benchmarks, with early access opening soon to support applications in robotics, VFX, and explorable world creation.

World Labs

@theworldlabs

Introducing Atlas:

The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D.

Model the world, move the camera, and simulate space & time.

This is impressive from Google Deepmind. Do take the time to watch the examples, read the prompts from Genevieve , a GDM Devrel staffer and recommend trying this yourself

Google for Developers announced agentic video understanding in Gemini, letting models dynamically search, scan, and inspect video segments across frames, audio, and transcripts instead of static 1 FPS processing.

It cuts token consumption by up to 88% and analysis costs by up to 66%, while boosting accuracy by up to 7% and enabling sub-second retrieval, long-form searches, anomaly detection, and precise counting in complex videos.

Supported now in Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite via Gemini API for uploads and YouTube videos in Google AI Studio and Enterprise Agent Platform.

Google for Developers

@googledevs

Tired of high token costs for video workflows?

Agentic video understanding gives Gemini the tools to dynamically search, scan, and inspect target video segments to achieve its goal. Building on agentic vision, Gemini now actively scans visual frames, audio, and transcripts

Rohan Doshi

@RohanLikesAI

Today, we’re launching Agentic Video Understanding in Gemini across 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite.
~88% fewer tokens
~66% lower cost
~7% higher accuracy
The result: a new Pareto frontier for video understanding across accuracy and cost.

Traditional video

genevieveh@

@genevieve__h

Delighted to see this go live! Agentic video understanding can analyse visual frames, audio, and transcripts, so it's really great at giving you feedback on your content. I used it to review my public speaking and podcasts and got some super helpful tips. Prompts below

Google AI Studio

@GoogleAIStudio

introducing agentic video understanding with Gemini

instead of using static processing (ingesting media at a fixed frame rate), the model can now take an active, goal-directed role in determining what to watch, at what speed, and through which modality (frames, audio, or

Introducing agentic video understanding with Gemini

We're launching agentic video understanding across our latest Gemini models for improved accuracy and lower costs and token usage.

Thursday 3rd September 2026

Google announces Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, building on 3.7 Flash from three weeks prior by using long-running agentic loops for better reasoning and coding.

3.8 Flash improves multidisciplinary expert reasoning to 54.9% on HLE-Verified benchmarks while matching prior speed and cost; 3.8 Flash Cyber leads in CyberGym Pass@1 for C/C++ vulnerability discovery at 86.2%.

It scores 59 on the Artificial Analysis Intelligence Index at high reasoning—up 3 points from 3.7 Flash and matching GPT-5.6 Sol (xhigh) and Grok 4.6 (medium).

Available to developers via Google AI Studio and Android Studio, enterprises through Gemini Enterprise, consumers in Gemini App and Search, with 3.8 Flash Cyber accessible via the Fairwind Program.

Google

@Google

Introducing Gemini 3.8, our best reasoning & coding model yet.

By leveraging long-running agentic loops, we’re building on the momentum of 3.7 Flash from just three weeks ago to release two new 3.8 variants:

Meet Gemini 3.8 Flash and Gemini 3.8 Flash Cyber

Artificial Analysis

Google has released Gemini 3.8 Flash, its fourth Flash model in under four months - it scores 59 on the Artificial Analysis Intelligence Index and reaches the Intelligence vs. Cost per Task Pareto frontier

@GoogleDeepMind released Gemini 3.8 Flash today. With high reasoning, it

Logan Kilpatrick

@OfficialLoganK

Introducing Gemini 3.8 Flash, another jump in Gemini's agentic + coding capabilities, and our 3rd updated Flash model in only 6 weeks...

This model has been a ton of fun to work with, excited to see what you all think!

It's barely mid week and the model release train is relentless

Meta Releases Muse Spark 1.3 with Top Coding and Agent Gains

The upgrade shines in benchmarks like 75.4 on DeepSWE v1.1, beating GPT-6.5.6 and Opus 5, while leading in OSWorld 2.0 and Terminal-Bench 2.1. It handles messy inputs by generating context, self-corrects plans, and uses 20% fewer tool calls than version 1.2. Now live in Muse Code and the Meta Model API, it's built for complex, ongoing tasks with more teases like a bigger model coming soon

Mark Zuckerberg

@finkd

Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter. This is the biggest jump we've made so far on coding and agentic work. Try it in Muse Code and our API.

Next up and Muse Spark open weights releases coming soon.

Artificial Analysis

Meta has released Muse Spark 1.3, their fourth Muse Spark model release in five months. Muse Spark 1.3 (max), which is in limited preview for Meta’s partners, scores 62 on the Artificial Analysis Intelligence Index, behind only Claude Fable 5.1 and Claude Opus 5. The variant

Friday 4th September 2026

via Just | LDA who was once again very quick to find out about the launch OpenAI's Astra 6 model and linked to its launch post

ChatGPT 6 Astra comes out today.

OpenAI

https://venturebeat.com/ai/welcome-to-the-agi-era-openai-launches-gpt-6-astra

Subsequently there was clarification that vast majority of users will probably not be able to access it today

Astra begins rolling out Thursday to enterprise customers with OpenAI's gated access program, Daybreak. OpenAI says it will become available over the coming days to ChatGPT Plus, Pro, Business and Enterprise customers, as well as through the OpenAI API and cloud platforms including AWS Bedrock and Microsoft Azure.

OpenAI launches GPT-6 Astra and says welcome to the "AGI era"

OpenAI's new flagship posts big gains on specialized tasks, but it comes at a premium and does not clearly lead the coding pack.

https://thenewstack.io

Microsoft AI launched MAI-Transcribe-2 as its new speech-to-text model, positioned as highest quality, lowest cost, and fastest transcription—claimed to be 10x faster than GPT-Transcribe—now available via Microsoft Foundry and OpenRouter.

A short demo video compares it side-by-side transcribing a technical lecture on the Nyquist sampling theorem, showing MAI-Transcribe-2 with under 4% error versus a competitor's 8% error and frequent mishearings like "nice quest sampling fear him."

The model supports 60 languages with automatic language ID, code-switching, speaker diarization, and word-level timestamps, ranking #1 on the FLEURS benchmark.

Microsoft AI

@MicrosoftAI

MAI-Transcribe-2: the highest quality, cheapest transcription at the fastest speed! 10x faster that GPT-Transcribe. Now available on Microsoft Foundry.

Mustafa Suleyman

Fastest. Cheapest. Best. ... MAI-Transcribe-2
10x faster than GPT-Transcribe. 5x faster than Gemini 3.5
#1 on the Artificial Analysis accuracy-latency Pareto frontier
Priced at the lowest on the market.
Truly fantastic work from the team! Try it out below.

Artificial Analysis

Microsoft AI has released MAI-Transcribe-2, achieving 2.0% AA-WER (#2) at ~411x real time (#2) for $1.67 per 1,000 minutes, placing it on the accuracy-speed Pareto frontier at one of the lowest prices among high-accuracy models

MAI-Transcribe-2 is @MicrosoftAI’s latest speech

For those here who use ChatGPT, Codex, Claude Code, Cursor, or similar AI tools for web research, I have published a new newsletter issue introducing my guide:

Choosing External Research Tools for AI Agents

It addresses a question I kept encountering: when are an AI tool’s built-in search and page-reading capabilities sufficient, and when might an external research provider genuinely help?

The guide covers the differences between search, extraction, crawling, structured data, provider-generated synthesis, browser-rendered pages, and authenticated X-native evidence. It also includes practical prompts and a reusable skill for routing research across available tools.

Octen, Exa, Perplexity, Parallel, Firecrawl, and TinyFish appear as examples, but this is not a ranking or a recommendation to connect everything.

The central idea is to begin with the evidence you need, use the minimum sufficient set of tools, and verify important claims against the underlying sources.

The newsletter introduction and link to the full guide are here:

A practical guide to choosing external research tools for AI agents

Curated topics on AI and Web3

Sep 3, 2026

Greetings friends, AI assistants and agent tools increasingly include web search and page-reading capabilities. For many everyday questions, those built-in tools are entirely sufficient. More demandin...

0 collected

Collect

You can also subscribe there if you would like to receive future updates. Candid feedback—particularly on anything that remains too technical or could be explained more clearly—is very welcome.

The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence Abu Dhabi announced K2 Horizon a fleet of six foundation models developed by them ranging from 0.9B to 375B parameters

Their launch post emphasizes frontier performance on coding, agentic, reasoning, and tool-use benchmarks with smaller models setting new SOTAs, plus full openness including weights, code, training data/recipes, and checkpoints under Apache 2.0.

The release includes immediate availability on Hugging Face, vLLM, and inference partners, supporting deployments from edge devices to enterprise while promoting reproducibility in AI research.

Institute of Foundation Models

@IFM_AI

Introducing K2 Horizon: a connected fleet of six foundation models ranging from 0.9 billion to 375 billion parameters.

- Frontier performance: Across coding and agentic tasks, K2 Horizon delivers top-tier performance in every size class—with the 0.9B, 3.7B and 7B models setting

via Caterina Zhang

Hello everyone. Are there any AI tools out there for turning audio music into visual in real time?

Abstract visual art

Saturday 5th September 2026

OpenRouter has added Microsoft's MAI-Image-2.6 and MAI-Image-2.6-Flash, supporting multi-image editing, web grounding, dynamic aspect ratios, and up to 1.5K output resolution.

MAI-Image-2.6 ranks #2 on Arena for text-to-image and editing plus #1 for editing on Artificial Analysis, while the Flash variant delivers comparable quality at 2.8x faster generation speed than GPT-Image-2-Medium and less than half the flagship price.

The rapid hosting on OpenRouter's unified platform gives developers immediate API access to Microsoft's latest image models following their Microsoft Foundry preview release.

OpenRouter

Microsoft's strongest image model yet is now on OpenRouter, with a faster production variant from @MicrosoftAI.

MAI-Image-2.6 and 2.6-Flash support multi-image editing, web grounding, dynamic aspect ratios, and up to 1.5K output. More below ↓

openrouter.ai

Microsoft API and Models | OpenRouter

Access 17 Microsoft models through the OpenRouter unified API including MAI-Image-2.5 Pro, MAI-Voice-2-Flash, and MAI-Voice-2. Compare pricing, context windows, benchmarks, and capabilities between...

Mustafa Suleyman

Our new image model generates images 2x faster than GPT-Image-2, currently the best model in the world.

It's also 72% more efficient in GPU usage, so we can provide it at an incredible price.

This gives it the best price-performance score in the world.

Unbelievable work from

Businesses in China are experimenting with ways to package and market AI tokens to ordinary consumers, including as credit card rewards and telecom plan bundles

Chinese businesses are giving away AI tokens with coffee, credit cards, and dumplings

Driven by low-cost open-source models, Chinese banks, telcos, and cafes are bundling AI compute tokens into credit card perks, mobile plans, and free restaurant promotions.

https://restofworld.org

https://www.linkedin.com/feed/update/urn:li:activity:7497552042186403840/

The cover image of this newsletter via generated via the GPT Image 2 model within the Krea tool via the following prompt

Create a wide editorial illustration representing a practical field guide to researching with AI tools.

Use an overhead view of a calm, carefully arranged research desk. Include an open field-guide notebook, a compass-like decision tool, several source cards, a small website map, a structured data grid, and a verification checklist. Fine connecting lines should suggest that a researcher is choosing the appropriate route from a question to trustworthy evidence.

The composition should feel thoughtful, curious, methodical, and useful—not futuristic or promotional. The objects should represent choosing, investigating, comparing, and verifying information without depicting any particular software provider.

Independent publication aesthetic, sophisticated editorial illustration, warm ivory paper, charcoal and navy ink, restrained teal and muted amber accents, subtle tactile paper texture, crisp shapes, generous negative space, calm natural lighting, contemporary but timeless.

Wide 1.91:1 composition for a 1200 × 630 newsletter header. Keep the centre visually clear and place important objects inside a generous cropping-safe area.

No readable writing, letters, numbers, logos, company names, brand marks, laptop interfaces, robots, glowing AI brains, holograms, neon lighting, cyberpunk styling, stock-photo appearance, or visual clutter.

This Week in All Things AI - Week 35-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 30 Aug 2026 03:35:33 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 23rd to 29th August 2026 saw Chinese AI labs push the open-weight frontier with a cluster of multimodal, long-context MoE releases. Z.ai launched GLM-5.3-Flash—previously previewed as Ox Alpha—while Alibaba introduced Qwen3.8-Flash and Tencent released Hy4 preview; all target coding, tool use, and agentic productivity with million-token-scale context windows and aggressive pricing. Video generation also advanced with Wan 3.0 becoming available on ArtArch, while Google expanded Gemini Notebook with ebook-grounded “Expert Intelligence.”

AI infrastructure and agent tooling were equally prominent: OpenAI reported early efficiency and latency gains from its Jalapeño inference chip, Nvidia’s Groq 3 LPX entered production, and Perplexity launched Portable Computer for fully local agent execution on NVIDIA DGX Spark. Developer discussions focused on making agents more capable and interoperable, from external search/fetch skills and free agent web access to Shopify CEO Tobi Lütke’s criticism of proprietary instruction-file conventions; OpenAI’s decision to end Cursor’s direct model access after its SpaceX acquisition further underscored intensifying competition across the coding-agent stack.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interest

Sunday 23rd August 2026

Monday 24th August 2026

via David An

In our latest Provocation Talk, I speak with Advait, from Opengradient.AI about the challenging topic of AI Privacy

I created the below public Notion page to articulate the benefits of adding external search/fetch providers in an AI harness and to suggest a prompt prefix for users to incorporate into prompts that utilise some of these providers.

These approaches can be useful if you are looking to take advantage of specialized data that external search providers can bring to the table within your AI harness.

The mechanism described on the below Notion page can be useful for GTM intelligence , lead-generation, finance or investment research, competitive intelligence or market-mapping workflows, among others.

The below page is best viewed on a desktop or laptop

I humbly welcome your comments.

Tuesday 25th August 2026

I updated the above Notion page to include a skill sub-directory that you can drop into your skill folder as well as some snippets to augment into AGENTS.md / CLAUDE.md and provides some guidance on invoking one or more of these providers in a prompt.

So some upfront investment in creating accounts with these providers, crediting balances beyond what they provide initially and integrating them in your harness but subsequently the skill file as well as some lines in AGENTS.md/CLAUDE.md should make this very easy to incorporate external search providers in a prompt

Nvidia says its inference accelerator Groq 3 LPX has entered full production, with Nebius signing on as the first customer, and SpaceX will deploy Vera CPUs

Groq

@GroqLLC

We are thrilled to announce that Groq will be among the first adopters of NVIDIA Groq 3 LPX, deploying it alongside NVIDIA Vera Rubin NVL72 in our purpose-built AI inference Cloud. Groq is working with Dell Technologies to deploy NVIDIA Groq 3 LPX.

When Groq brings NVIDIA Groq 3

groq.com

Groq Among the First to Bring NVIDIA Groq 3 LPX and Vera Rubin NVL72 to Market

Groq is the premier neocloud for fast inference. One fully integrated platform for infrastructure, inference, and control. Millions of developers run trillions of tokens on Groq every week.

Elon Musk

SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028

NVIDIA

@nvidia

The first CPU built for agents is going to work at scale.

@SpaceX is deploying NVIDIA Vera to accelerate the orchestration, code execution, and data processing that powers its next generation of agentic AI — keeping GPUs fed and agents acting fast.

From gigawatt AI factories

Shopify CEO Tobi Lütke criticizes Anthropic's Claude Code for ignoring the emerging open standard AGENTS.md and .agents/skills directory, sticking only to its proprietary CLAUDE.md file.

This creates "split brain" inconsistencies in large teams where developers use multiple AI coding tools, requiring workarounds like symlinks or @imports

that add unnecessary friction.

The post highlights growing industry pressure for AI agents to adopt shared, vendor-neutral conventions to streamline codebase instructions across tools like Cursor, Codex, and Copilot.

tobi lutke

@tobi

I’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc.

Insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Just unnecessary.

PS: This is also my pet peeve with Anthropic that their harness will read only in their directories and filename.

Wednesday 26th August 2026

OpenAI's post shares initial test results for Jalapeño, its first custom AI inference chip, highlighting gains in power efficiency and response speed by delivering higher throughput and lower latency simultaneously.

Performance benchmarks show Jalapeño achieving 1.5-1.9x more AI work per watt and 1.7-3.6x lower end-to-end latency compared to prior systems across large models like 120B, 670B, and 1T parameter sizes.

Deployment in OpenAI's infrastructure is planned by the end of 2026, as the first step in a multi-generation hardware roadmap with Gen 2 already in development to further boost efficiency for products like ChatGPT.

OpenAI

Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system around it.

The results show a major advance: more intelligence from every watt and faster responses, delivering both higher throughput and lower latency in one architecture without

Token Gremlin

@TokenGremlin

OpenAI casually went from:

“we make AI models”

to:

“we design the models, the serving stack, the kernels, the memory system, the network... and now the chip too.”

OpenAI’s first custom inference chip, Jalapeño, is showing:

→ 3.6× lower latency
→ 4.1× higher interactive

OpenAI

Some snippets on SemiAnalysis longish article on Jalapeno

In general first generation chips are not competitive, but OpenAI bucks the trend by being industry leading and beating every Nvidia, AMD, and Google chip we have been able to test on multiple top open source models. OpenAI does this with extreme hardware software codesign. Surprisingly, OpenAI is not over specialization on any specific part of model inference, but instead by focusing on being a general chip that delivers high performance in all scenarios.
Everyone says that OpenAI’s chip is specialized for OpenAI models, but that’s wrong, OpenAI made a generalized chip for AI inference.

Dwarkesh Patel shares his latest podcast with Dylan Patel of SemiAnalysis, analyzing AI lab economics including the projected shift of compute from inference to training as recursive self-improvement nears.

The discussion forecasts Anthropic and OpenAI controlling most global usable FLOPs within years by outbidding others through superior monetization, driven by economies of scale, compute scarcity, and continual learning.

They examine risks from over $10T in AI capex by decade's end potentially triggering sovereign debt crises via elevated interest rates, hyperscaler borrowing, and economic fallout for non-AI sectors and countries.

Dwarkesh Patel

@dwarkesh_sp

Had a lot of fun chatting again with my twin brother @dylan522p

We went through lab economics over the next few years - the shift from inference to training as RSI draws near; and how Anthropic and OpenAI are on track to control most of the world’s usable FLOPs within the next

Perplexity Launches Fully Local AI Agent on NVIDIA DGX Spark

Perplexity released Portable Computer, a local version of its AI agent platform that handles orchestrator, subagents, and tasks on NVIDIA DGX Spark without cloud reliance for core functions. It starts tasks locally by default and only seeks user approval for cloud use on complex reasoning, ensuring privacy by flagging personal data and avoiding direct file access. Available now for Pro and Max subscribers with compatible hardware, it excels in benchmarks like 85.4% on real-world tasks and supports app connectors for Gmail, Slack, and GitHub. NVIDIA's Jensen Huang even gifted a DGX Station to the team after an early demo, fueling excitement for private, affordable AI.

Perplexity

Today we’re launching Portable Computer on @NVIDIA DGX Spark.

Portable Computer is a fully local version of Perplexity Computer, where the entire runtime: orchestrator LLM, subagent LLM, agent harness all run on your local hardware. No cloud dependency.

"Even under the best circumstances, the transition to this new AI era will be one of the most turbulent times in human history," Bill Gates writes in a new, 12-page 5,784 word essay.

He says that “we are not preparing adequately”, calling for a regulatory framework

He also says tech executives are privately “very worried” about AI disruption but publicly downplay the risks to protect fundraising and planned IPOs

Looks like it's going to be a China open-weights night/US early morning with Alibaba's Qwen3.8-Flash-Next as well as GLM-5.3-Flash which was unmasked as the stealth model 0xalpha

Qwen

https://x.com/i/trending/2092526498800718248

Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight!

The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens.

125B parameters + 51B N-gram

Z.ai announces GLM-5.3-Flash, a 320B total / 18B active parameter natively multimodal model with 1M-token context, released under MIT license and optimized to run on Chinese AI chips after preview as Ox Alpha.

The model offers strong coding and agentic performance at low cost, with API pricing of $0.15 input / $0.50 output per million tokens, outperforming GLM-5.2 across benchmarks like DeepSWE and AutomationBench. 50% discount till Sep 9th via OpenRouter

Architectural changes including hybrid sparse-linear attention and specialized pre-training enable efficient scaling, achieving results near Claude Opus 4.8 on select tasks while reducing compute demands significantly.

From OpenRouter

We expect many providers to onboard this model throughout the week

Z.ai

Introducing GLM-5.3-Flash

- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context window
- A 320B-A18B model released under the MIT License
- Previously previewed as Ox Alpha, running entirely on Chinese AI chips

Blog:

OpenRouter

1M-token context. 131K max output. `max` reasoning by default.

Launch pricing via @Zai_org is 50% off through Sep 9 at 16:00 UTC:

$0.075/M input
$0.25/M output
$0.015/M cached input

Then $0.15/M, $0.50/M, and $0.03/M.

ZCode

@zcode_ai

GLM-5.3-Flash is live in ZCode today

The first natively multimodal GLM-5 model: it outperforms GLM-5.2 and approaches Claude Opus 4.8 on coding and agentic tasks, at 1/10 the cost.

Why use it in ZCode:
• 3× the quota of GLM-5.3 on Coding Plan, plus an extra 1.5× only in

zcode.z.ai

ZCode | GLM-5.3 官方 Harness

ZCode 将 GLM-5.3 与领先的 AI 编程 Agent 和你的现有工具链结合，帮助你顺畅完成规划、编码、评审与上线。

Artificial Analysis

https://x.com/i/trending/2092825795001790544

GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index. At $0.09 Cost per Task, it sits comfortably on the Intelligence vs. Cost per Task Pareto frontier

@Zai_org has released GLM-5.3-Flash, a smaller and cheaper sibling to GLM-5.3 at 320B total parameters and

via Alba Chung

Hi this is Alba with Alibaba ☁

Wan 3.0 is now on ArtArch AI

→ Reality-grade rendering

→ Cinematic motion

→ Stronger scene consistency

Limited time: Aug 25–Sep 7: 100 FREE generations for every user.

No duration limit.

Start creating:

ArtArch.Ai

@ArtArchAI

Wan 3.0 is now on ArtArch.
→ Reality-grade rendering
→ Cinematic motion
→ Stronger scene consistency

Limited time, Aug 25–Sep 7: 100 FREE generations for every user. No duration limit.
Start creating: artarch.ai/models/wan-3.0

#GenerativeAI #Wan3 #AIFilmmaking #ArtArch

Thursday 27th August 2026

Curating some links about Qwen-3.8-Flash-Next which is now available via OpenRouter at 0.16/0.47/0.016 so priced competitively with GLM-5.3-Flash.

Shall be interesting when other providers join the fray in serving this model.

SemiAnalysis

@SemiAnalysis_

Congrats to @Alibaba_Qwen on the release of Qwen3.8-Flash-Next, using the same architecture innovations as their upcoming Qwen4 model! Such innovations include:

51-billion-param N-gram Embedding to look up a table with very little extra computation, which means the embedding

Sushaanth Srinivasan

@sushsrinivasan

Qwen4 spends 51.2B of its parameter budget on memory instead of computation.

Qwen3.8-Flash-Next shipped this morning, and its config calls it Qwen4. Of its 180B parameters, 51.2B are a hash table. That's Engram, from DeepSeek's paper in January, implemented down to the hash

ModelScope

@ModelScope2022

Qwen3.8-Flash-Next is here! An open-weight multimodal MoE built on a brand new architecture, with native 256K context extendable to 1M via YaRN. modelscope.ai/collections/Qw…

Leads every compared model on SWE-bench Pro (62.5 vs 53.4 for Claude-Opus-4.6 Max), SWE-bench

An impressive technical breakdown by Ryan Smith on Sambanova's SN50 RDU

Nvidia Acquires Hugging Face for $12.9 Billion

The agreement gives Nvidia control of Hugging Face, often called the GitHub for AI, where millions of models and datasets are shared by researchers and companies worldwide. Reports from The Information, Bloomberg, and Business Insider pegged the valuation around $13 billion, down slightly from earlier talks, following Nvidia's prior investment in the startup's 2023 funding round at $4.5 billion. The move expands Nvidia beyond chips into the open-source ecosystem that powers AI on its GPUs, even as Hugging Face's revenue sits around $150 million annually. It highlights hardware giants racing to own more of the AI stack amid strong demand.

Friday 28th August 2026

For the video editors in da house via Shamir Allibhai of Hey Eddie fame who is tentatively my upcoming guest in my conversation series. Luma link for folks to register for the virtual event shall be posted once dates/times are finalised with Shamir

The iOS version of Eddie AI is now available

https://apps.apple.com/app/eddie-ai-video-editor/id6792122289

Shamir also pointed me towards this completely free product he built https://99.life/

Here's a brief blurb about the above product

99 is a free arena for AI video editing. You bring the clips. Two AI models each cut their own edit. You watch both without knowing which model made which, and pick the winner. Every vote moves the public scoreboard.

As to why the above product is Free

Why it's free

No trick, just a trade. Blind votes from real people on real footage are the honest measure of an AI editor, and model makers pay for that measure. 99 owns the edits the models generate, and the evaluation data the arena produces — votes, watch behavior, and the content they judge — is used to train AI models and is licensed to the companies that build them. That deal funds the arena. You get free edits and a real say in which models win; the model makers get judged. We think that is a fair trade, and we would rather say it in bold than bury it in fine print. The details live in our Terms and Privacy Policy.

Tencent Hunyuan released Hy4 preview, a 770B parameter MoE model with 49B active parameters and 1M context window optimized for productivity applications.

The model is fully open-sourced with weights on Hugging Face and GitHub, offered at consistent affordable pricing as a frontier option.

Benchmark visuals show Hy4 preview leading or matching top models like Claude Opus 5 across coding, agentic tasks, tool use, and math evaluations including Terminal Bench, SWE tasks, and OneMillionBench.

Tencent Hy

@TencentHunyuan

Hy4 preview is here.

770B, 49B active, 1M context.
Built for productivity.
Open source frontier.
Consistent affordable price.
Use it. Tell us what breaks.

More on Hy blog：hy.tencent.ai/research/hy4-p…
HuggingFace：huggingface.co/tencent/Hy4-pr…
Github：github.com/Tencent-Hunyua…

https://hy.tencent.ai/research/hy4-preview

via Diamond Hands Dig

Your agent can now search & fetch any webpage for 100% FREE.

Them: $7 per 1,000 searches.

Us: $0. No subscriptions, no quotas.

Humans search Google for free. Agents shouldn't have to pay either.

Made possible by Tiny_Fish and MonidHQ

Shengkun Ye

@shengkunye

We just killed Exa, Tavily, SerpAPI, and Brave.

Your agent can now search & fetch any webpage for 100% FREE.

Them: $7 per 1,000 searches.
Us: $0. No subscriptions, no quotas.

Humans search Google for free. Agents shouldn't have to pay either.

Made possible by @Tiny_Fish and

Google Launches Expert Intelligence for Ebooks in Gemini Notebook

Google's new Expert Intelligence feature lets users add eligible Google Play Books ebooks to Gemini Notebook for AI-powered summaries, infographics, audio overviews, and quizzes all grounded in the book's text. People can mix these insights with personal notes, like applying management lessons to work scenarios, with over 100,000 titles from publishers such as Penguin Random House and Bloomsbury available at launch. Featured Notebooks pair bestselling authors like Michael Pollan and Steven Pinker with extra interactive elements, and Google plans expansions to subscriptions and textbooks soon.

Gemini Notebook

@Gemini_Notebook

Introducing Expert Intelligence

A cross-@Google initiative that helps you engage with your trusted sources, starting with eligible @GooglePlay ebooks in Gemini Notebook. You can now combine expertise from your favorite authors with other sources and engage with your books in

News from Google

@NewsFromGoogle

Over the past several months, we’ve been meeting with authors and publishers to better understand how AI can help people get the most out of their books .

Today we’re launching Expert Intelligence, a new way to get insights about your favorite titles from the authors you love,

Saturday 29th August 2026

OpenAI is terminating its partnership with the AI code editor Cursor following SpaceX's acquisition of the tool, with direct model access ending on November 12 to allow transition time,

The decision centers on trust issues, while OpenAI offers workarounds like using personal API keys or its own IDE extensions and commits to supporting diverse developer tools and open-source initiatives.

This move impacts developers who integrated GPT models into Cursor workflows, reflecting competitive dynamics in the AI tooling space between major players.

Tibo

@thsottiaux

We unfortunately have decided that we cannot continue providing access to our models through Cursor and are ending our partnership. It boils down to trust and we’ve asked that this takes effect on November 12 to give you some time to plan.

Many have used the GPT models through

Our decision on Cursor following its acquisition by SpaceX

Our decision to wind down our contract providing OpenAI models to Cursor following its acquisition by SpaceX.

OpenAI

We’re ending our partnership with Cursor following its acquisition by SpaceX. Under our proposal, Cursor’s direct access to our models would end on November 12.

We know that the people most affected by this decision are the developers who rely on OpenAI models in Cursor. We care

The cover image of this newsletter via generated via the Seedream 5 Pro model within the Krea tool via the following prompt

A futuristic city on an alien planet, surrounded by sand dunes and strange spires, painted in the style of Frank Frazetta. The buildings have domes with intricate details, and there's a sense that something mysterious is waiting to be discovered within them. In the background, there's a vast desert landscape with orange hues under a blue sky. A small red light shines from one building, adding contrast against the dark tones of the scene. It feels like you could find treasures or secrets here.

This Week in All Things AI - Week 34-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 23 Aug 2026 01:22:57 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 16th to 22nd August 2026 brought a strong mix of multimodal model releases, agent infrastructure, developer tools, and practical AI adoption. RedNote released dots3-note preview, an open-weight 280B MoE model built for long-horizon multimodal agents, while DeepSeek introduced V4-Flash-Vision-Exp for visual agent workloads and OpenRouter debuted the mysterious, high-performing Ox Alpha coding model. On the product side, Cartesia shipped Sonic-3.6 voice AI, Wispr raised a $280M Series B, and Meta expanded its AI offering with a native Mac app and business-oriented integrations.

The week also highlighted the rapidly maturing ecosystem around AI operators: Artificial Analysis launched a benchmark for agent web-search APIs; K-Dense released drug-discovery agent skills; Harvey opened a legal-engineering certification; and Replit introduced a credit-free mode for routine agent work. Google made AI plans free for eligible students across more than 140 countries, while Bun 1.4 delivered a major Rust-based runtime overhaul for JavaScript and TypeScript developers. Alongside the releases, Anthropic CEO Dario Amodei’s public defence of AI-risk discussions and his optimism on AI-enabled biomedical progress underscored the continuing tension between capability, adoption, and safety.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interest

Sunday 16th August 2026

Xiaohongshu/RedNote announced dots3-note preview, an open-weight multimodal MoE AI model from RedNote's AI lab aimed at advancing long-horizon agency for real-life tasks like reasoning, environment exploration, memory updates, and tool use.

Key specs include 280B total parameters with 16B active, 512K context window, native support for text/vision/video/audio, and TEMPO—a new RL method using self-critiquing and test-time scaling for extended agent training.

Accompanying benchmark charts show competitive performance against larger models on agentic, coding, and multimodal tasks; model weights are on Hugging Face, with API access now available via OpenRouter.

dots studio

@dotsstudioai

Introducing dots3-note preview — a small but mighty step toward long-horizon agency in real life.
280B MoE with 16B active parameters, a 512K context window, and multimodal understanding across text, vision, and audio
Introduces TEMPO, a new RL approach for long-horizon

Chao Qiao

@ChaoQiao42

WE SHIPPED IT!

Introducing dots3-note Preview: our open-source 280B/16B-active multimodal model, built for complex reasoning and long-horizon agents.

Beyond the model, we’re sharing TEMPO (Test-time-scaled Value Estimation with Macro-step Policy Optimization)—our answer to

Chao Qiao

@ChaoQiao42

Seeing dots3-note finally open-sourced is deeply personal for me.

Over the past months, our post-training team poured an extraordinary amount of work into the model as a whole—reasoning, coding, tool use, multimodal understanding, and long-horizon agency. Behind the results were

Chao Qiao

@ChaoQiao42

How do you train an agent when a single rollout takes tens of hours?

Alongside the open-source release of dots3-note Preview, we’re sharing more details about TEMPO: Test-Time-Scaled Value Estimation with Macro-Step Policy Optimization.

Long-horizon agent RL faces two

Anthropic CEO Defends AI Risk Talk While Eyeing Disease Cures

In a detailed X exchange, Anthropic CEO Dario Amodei pushed back on criticism from investor Gavin Baker that his risk warnings fuel anti-AI views. Amodei stressed that real breakthroughs, like curing most human diseases in 5-10 years, will outweigh messaging, and revealed Anthropic's push into biology and medicine with early results expected soon. He defended regulation proposals that aid smaller players while tackling cyber, bio, and alignment risks, as prediction markets buzz with trader bets on his timeline

Dario Amodei

1/2 Thanks Gavin for an especially thoughtful exchange. I don't usually spend much time on social media but I wanted to engage here because it really brings out the heart of an important conversation.

First, on regulation, I think that “either concentrate it in the hands of a

Gavin Baker

@GavinSBaker

Sholto, thank you for setting the record straight. Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true. And the reason it is believable to so many is that it is consistent with Dario’s public messaging

Sholto Douglas

@_sholtodouglas

Completely false.

I like Gavin's takes, but whoever he heard this from is lying so that it fits the narrative some people so desperately want you to believe.

The same people will try to convince you Anthropic has no moat, and a sentence later that it might become so powerful

The All-In Podcast

@theallinpod

Gavin Baker: Anthropic Believes They Could Be the ONLY Company Left in the World

“I would certainly discourage Dario from saying that ever again to anyone.”

@GavinSBaker:

“Internally, Anthropic is very confident. I have been told by multiple people I trust that Dario has said

Monday 17th August 2026

Voice AI startup Wispr raised a $280M Series B led by Menlo Ventures at a $2B valuation, taking its total funding to $361M, and now has 100K business customers

Wispr talks its way to a $2 billion valuation. The company says dictation's only the beginning | Fortune

Wispr has raised $280 million in its Menlo Ventures-led Series B, as it looks beyond dictation.

https://fortune.com

Vibin Babuurajan 👋

@vibinbaburajan

Wispr Flow got into talks to raise $260M after VCs had already been using the product for months.

As far as the story goes, Wispr Flow didn't reach out to Menlo Ventures.

A partner there had already been dictating his emails and memos through the app for months.

Tuesday 18th August 2026

Harvey, an AI company specializing in complex legal work, announced the launch of its free self-paced Legal Engineering Certification via Harvey Academy.

The course teaches four core skills—legal expertise, AI fluency, process design, and user enablement—to help professionals design AI-enabled legal workflows and drive adoption.

Aimed at lawyers, in-house teams, and legal tech roles, it includes practical training from over 1,000 AI implementations and awards a shareable badge upon completion.

Harvey

@harvey

Now live: Harvey's Legal Engineering Certification.

A free, self-paced course covering core legal engineering skills: legal expertise, AI fluency, process design, and user enablement.

Open to everyone - no Harvey account required.

Cartesia announces Sonic-3.6 TTS with major gains in naturalness across 44 languages, including realistic pauses, filler words in English, and fluid Hindi-English code-switching, released three months after Sonic-3.5 based on user feedback.

Sonic-3.6 ranks #1 on Artificial Analysis's Provider Voice leaderboard (Elo 1282) and Controlled Voice leaderboard (Elo 1123), ahead of Speechify Simba 3.2, Alibaba's Qwen-Audio, and ElevenLabs v3 in blind tests using both native and standardized voices.

Pricing: At $49 per 1M characters via the Cartesia platform, Sonic 3.6 is the most expensive of the top three models on quality, roughly 5x Speechify AI's Simba 3.2 at $10 and nearly double Alibaba's Qwen-Audio-3.0-TTS-Plus at $27.59, though still half the price of ElevenLabs' Eleven v3 at $100

Available immediately in beta via cartesia.ai, the state-space model architecture delivers low-latency streaming performance, positioning Cartesia as a leader in production-grade, human-like voice AI.

Cartesia

@cartesia

Introducing Sonic-3.6: our most lifelike TTS yet, and another step change in naturalness across 44 languages. Just three months after our Sonic-3.5 launch, we've made fundamental model improvements based on feedback from the teams building on Sonic.

We're #1 on @ArtificialAnlys

Artificial Analysis

Cartesia's Sonic 3.6 takes the #1 spot on both the Provider Voice and Controlled Voice Artificial Analysis Speech Arena leaderboards, surpassing Speechify AI's Simba 3.2 and Alibaba's Qwen-Audio-3.0-TTS-Plus, with Sonic 3.5 holding #2 on Controlled Voice

Sonic 3.6 is the latest

Wednesday 19th August 2026

Artificial Analysis Launches Web Search API Benchmark for AI Agents

Artificial Analysis released its Search Index, evaluating web search APIs like Parallel, Exa, and Firecrawl on real agent tasks such as research and multi-hop browsing. Parallel leads with a 75 score, followed by Exa at 74 and Firecrawl at 73—all far ahead of the model-only baseline of 33. The benchmark factors in total cost and time, highlighting how top performers like Parallel and Firecrawl balance efficiency with savings on model tokens. Developers praise the independent data for picking reliable tools as AI agents lean more on web search.

Artificial Analysis

https://x.com/jarredsumner

Announcing the Artificial Analysis Search Index, benchmarking how search API providers perform on quality, cost, and speed when used by an agent. We are initiating coverage with Parallel, Exa, Firecrawl, You (dot) com, Tavily, Keenable, and Brave

Search is one of the most

FWIW, I wrote this public Notion page on the benefits of integrating external search/fetch providers into a harness

K-Dense AI announced the release of a new agent skills plugin specialized in small molecule and protein therapeutics, enabling AI agents to handle tasks in cheminformatics, molecular modeling, docking, dynamics, and target discovery.

The plugin is a curated subset of their main Scientific Agent Skills GitHub repo (K-Dense-AI/scientific-agent-skills), which has 33.8k stars, 161+ validated skills, and 100+ scientific databases covering biology, chemistry, medicine, and drug discovery

It supports installation and use in multiple AI coding tools including Claude Code, Cursor, Codex, and Google Antigravity via the open Agent Skills standard.

K-Dense

@k_dense_ai

We just shipped our first agent skills plugin for small molecule and protein therapeutics. This represents a curated subset of our larger Scientific Agent Skills repo (33k+ on GitHub).

Try it today in Claude Code, Cursor, Codex, Google Antigravity, and more.

github.com

GitHub - K-Dense-AI/drug-discovery-agent-skills: Agent Skills for small-molecule and protein...

Agent Skills for small-molecule and protein therapeutics: cheminformatics, molecular ML, docking and dynamics, protein design platforms, and target-discovery knowledge graphs. - K-Dense-AI/drug-dis...

Replit's new "Free Mode" lets Core and Pro subscribers use the Agent for everyday tasks (chatting, ideating, quick builds) without consuming usage credits, effectively letting them create up to 30X more within their existing monthly subscription.

Free Mode is powered by OpenAI's GPT-5.6 Luna and applies to fast, high-value everyday tasks — chats, ideation, quick edits, and simpler builds — which no longer draw down credit-based usage.

Replit Introduces Free Mode | Replit

AI models are now capable and affordable enough to make once-unreachable outcomes practical. The opportunity for everyone to pursue and collaborate on ambitious ideas, spend less time on busywork, and raise the quality and creative output across disciplines is nearly here. The possibility is here. But for most people, the reality is not.

https://replit.com

Just LDA asked the following question

Anyone here using Buzz, Grok Bot, or the new Hermes Bot function?

Opinions? I’d assume Hermes Bot works better than Grok bot because then you’re not limited to one model.

But just curious to hear from others that have used them.

I’m currently using Buzz and it seems to be working well. I have my Claude agent, Hermes agent (ChatGPT), and Grok agent all working together now on my various projects.

Thursday 20th August 2026

Diamond Hands Dig responded to Just LDA with the following

Just wrote about Hermes Agent Bot Mode yesterday

Diamond Hands Dig

@DiamondHandsDig

x.com/i/article/2089…

Google made AI Pro free for US students, AI Plus free for students in 140+ countries, and a cheap YouTube student bundle.

Google Gemini announced a back-to-school promotion offering eligible US college students one year of Google AI Pro for free and students in 140+ countries one year of Google AI Plus at no cost, running through December 31, 2026.

New student tools rolling out include a dedicated hub for study notebooks, flashcards, and quizzes; adaptive lessons that pinpoint knowledge gaps from class materials; and interactive 3D visualizations generated via prompts like “Show me how DNA works.”

Gemini Live now supports hands-free multi-step research reports that run in the background, with notifications and voice discussion of findings to support deeper academic work.

Google Gemini

@GeminiApp

Back-to-school season is here and starting today, eligible college students can get a full year of Gemini on us:
- US students: 1 year of Google AI Pro at no cost
- 140+ countries: 1 year of Google AI Plus at no cost

Here’s what’s new for students

Student Offers

@StudentOffersHQ

Google made AI Pro free for US students, AI Plus free for students in 140+ countries, and a cheap YouTube student bundle.

1/ Google AI Pro — free 12 months, US only
$240 value (was $19.99/mo)
Gemini 3.1 Pro, 5 TB, Antigravity access

2/ Google AI Plus — free 12 months,

Nahid

@nahid_pro09

Free is not a typo a full year of gemini is $0 for students right now

starting today:
→us students: 1 year Google AI Pro, at no cost
→140+ countries: 1 year Google AI Plus, at no cost

one sign-up. zero dollars. twelve months of AI doing the heavy lifting

the only thing

Nahid

@nahid_pro09

1 million tokens/min completely free on gemini 2.5 flash and pro - official google api, zero strings attached

google's ai studio tier is the only developer api limit you'll ever need

"genuinely giving away free ai infra" - that's the vibe here. massive context windows and

Whilst not LLM related per-se, thought this might be of interest to group members particularly those who work with Javascript/Typescript codebases which are coded/generated via their coding agents and maybe utilising nodejs or deno for their runtime

Bun 1.4 released (Claude Code has been running on a canary version of it for quite a while)

Bun

@bunjavascript

Bun 1.4

- Fixes over 2,900 GitHub issues
- +1,517 tests from the Node.js test suite
- Reduces idle CPU by 5x
- Reduces memory usage by up to 35%
- Starts up to 50% faster on Linux
- Rewrites Bun in Rust

Thank you everyone who contributed since Bun 1.3!

bun.com/blog/bun-v1.4

Bun 1.4

Bun 1.4 rewrites Bun in Rust, ships built-in headless browser automation (Bun.WebView), Bun.Image, Bun.markdown, JSON5, JSONL, Terminal and cron APIs, Node.js 26.3.0 compatibility with 1,517 newly passing tests, parallel test and run, Windows ARM64, and an opt-in global virtual store for up to 7× faster installs.

https://bun.com

The rewrite of Bun from Zig to Rust in 11 days was done with the assistance of Fable 5

Rewriting Bun in Rust

Why & how we rewrote Bun from Zig to Rust

https://bun.com

PS: Anthropic acquired Bun last year. Jarred Sumner is the founder

Couple of links on the Meta AI app (which doesn't seem to have rolled out to my region, might need to VPN or Meta could have region locked based on metadata from my usage of FB/IG). The video in Spencer Barnett's post makes it very intriguing to try it out

Prashant Ratanchandani

@pratanchandani

Big update from us on Meta AI - this time focused on small businesses.

It can now plug into your FB/IG analytics, Ads, + Google Workspace to audit campaigns, benchmark vs similar brands, and tell you what's actually working.

And there's a new native Mac app — you can share any

meta.ai

Meta AI

Use Meta AI assistant to get things done, create AI-generated images for free, and get answers to any of your questions.

Spencer Barnett

@spencerbarnett

We launched the Meta AI Mac OS app today!

I particularly love the dictation feature which allows me to dictate anywhere on my computer with ultra high accuracy. Just hold down 'fn' and yap!

Alexandr Wang

check out the Meta AI desktop app!

the dictation has been a game changer for me personally

Spencer Barnett

@spencerbarnett

We launched the Meta AI Mac OS app today!

I particularly love the dictation feature which allows me to dictate anywhere on my computer with ultra high accuracy. Just hold down 'fn' and yap!

Friday 21st August 2026

DeepSeek Launches Experimental Vision Model Matching Top Benchmarks

The new DeepSeek-V4-Flash-Vision-Exp adds image-handling to the V4-Flash series, scoring big on multimodal benchmarks like 36.5 on ApexBench and 27.3 on Agents' Last Exam—often rivaling Opus-4.8—while holding steady on pure-text tasks. Developers access it easily via the DeepSeek API with base64 images, URLs, or the fresh Files API, all at standard pricing. Community voices, from researchers to tech analysts, praise the leap in visual agent performance without speed or cost hits, signaling a push toward versatile multimodal tools.

DeepSeek

DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform!

This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major

Chubby♨️

wtf is happening today: DeepSeek has launched DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model built for agents that need to see.

And its performance on visual-agent benchmarks moves close to or even outperforms Opus 4.8.

Again: this is the Flash model, the

DeepSeek

Mysterious Ox Alpha AI Model Debuts with Strong Coding Prowess

Ox Alpha, listed on OpenRouter as a free preview from an anonymous provider, handles 1 million token contexts with text, image, and video inputs, excelling in coding and agentic tasks. Early tests by developer Ben Davis showed it hitting 80% on DeepSWE benchmarks, outpacing models like Fable at 65% and gpt-5.6-sol at 52%. Community speculation points to Chinese origins, possibly Zhipu’s next GLM or Xiaomi’s MiMo, with its polished speed and frontend strengths fueling excitement during the limited free access period.

OpenRouter

New stealth model: Ox Alpha

Ox Alpha is a frontier model built for efficient coding, sustained agentic work, and real-world production use.

- 1M token context window
- Text, image, and video input

Try it now and share feedback to improve the model!

openrouter.ai

Ox Alpha - API Pricing & Providers

Ox Alpha is a reasoning model designed for coding, sustained agentic work, and production workloads. This model is free to use. 1,048,576 token context window, maximum output of 131,072 tokens.

Command Code

Ox Alpha (stealth frontier model) is free in Command Code.

100T tokens per day free capacity, no 5hr usage limits.

• 1M content
• we benchmarked ~5B tokens
• built a tool repair harness to imp tool calling

absolutely free for all plans.

guess who made it?

OpenCode

Ox Alpha (stealth model) is free for the next week

- 1M Context
- Multi-modal
- Zero Data Retention

Generous rate limits, near unlimited usage

We have capacity for 100T tokens per day, lets see what you can do

Chubby♨️

Wtf! Ben ran this mystery model through 10 DeepSWE tasks and it scored over 80%, versus 65% for Fable and 52% for GPT-5.6-sol!

this is insane. Probably a chinese company. Either a new GLM or Kimi model, I reckon.

Ben Davis

@davis7

I ran this thing through 10 tasks on DeepSWE (so there could be a ton of variance in it's real score, this is a subset), but uh...

gpt-5.6-sol: 52%
fable: 65%
whatever the hell this is: 80% (was a near miss on the "x"s so actually over 80%)

I am very confused

Saturday 22nd August 2026

The cover image of this newsletter via generated via the Seedream 5 Pro model within the Krea tool via the following prompt

An enormous floating city in the sky, steampunk style with lots of towers and domes, oil painting, fantasy art, detailed

This Week in All Things AI - Week 33-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 16 Aug 2026 03:41:58 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 9th to 15th August 2026 brought a dense run of releases across agentic models, autonomous-workflow tooling, and AI infrastructure finance. Meta introduced Muse Glimmer, a 30B open-weight model for always-on local agents, while NVIDIA launched Nemotron 3.5 Lightning and NeMo Switchyard, pairing a fast open agent model with workflow-level model routing. DeepSeek made V4-Pro generally available and released its modular, MIT-licensed Harness developer preview; Google shipped Gemini 3.7 Flash, and Z.ai announced GLM-5.3 for coding, agents, and cyber-defense work.

The week also underscored the shift from chatbots to durable AI operators: SpaceXAI introduced Grok Bot, persistent agents that work inside connected apps, while SignerSet proposed policy-based, multi-party approvals for consequential autonomous actions. On the infrastructure side, NVIDIA outlined financing partnerships intended to mobilize more than $500 billion for data centers, power, and GPU capacity, and Anthropic reportedly struck a long-term compute agreement with Riot Platforms. Other notable threads included Anthropic’s planned invisible text watermarking, OpenAI finance-team workflow examples and the economic trade-off between AI “tax” and AI “dividend.”

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interest

Sunday 9th August 2026

Monday 10th August 2026

via Charles Stanton

This is very cool, Harvard and MIT scientists have replicated the human population by creating 8.3billion agent personas, effectively allowing them to trial global initiatives virtually.

MatrAIx: a population-scale AI simulation infrastructure designed to model the behavior of virtually every person on Earth.

Brian Roemmele

@BrianRoemmele

Harvard and MIT Researchers Simulate the Entire Planet with 8.3 Billion AI Personas

It is big.

In a development that feels straight out of science fiction, a large collaborative team led by researchers from Harvard University and the Massachusetts Institute of Technology has

Meta Releases Muse Glimmer, 30B Open-Weight Model for Local AI Agents

Meta launched Muse Glimmer, a 30B-parameter open-weight model under Apache 2.0 license, designed for always-on local agents on consumer hardware with 24GB VRAM. It excels in vision, tool use, multi-step reasoning, and benchmarks like MCP Atlas where it scored 75.5, beating rivals such as Gemma 4-31B.

Alexandr Wang

https://www.cogniscendo.com

1/ big announcement today: we will be releasing an open weight version of muse spark 1.2 soon.

we also are releasing muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability.

AI at Meta

@AIatMeta

Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows.

Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on

Tuesday 11th August 2026

NVIDIA Partners with Wall Street Giants to Unlock $500 Billion for AI Infrastructure

NVIDIA signed deals with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion in capital for data centers, power and GPUs. These independent financing platforms let customers secure funding at attractive rates, turning NVIDIA's AI compute into a new investable asset backed by its CUDA ecosystem.

Independent financing platforms let capital providers underwrite demand, utilization, cash flow and residual value, with limited NVIDIA residual-value support up to 25%.

Huang explained that in AI, compute equals revenue, and this brings long-term institutional money into the market amid surging infrastructure spending that's boosting U.S. GDP like never before.

Jensen Huang

@JensenHuang

x.com/i/article/2086…

amit

@amitisinvesting

$NVDA

Okay, so let’s talk about what Nvidia actually announced today…

I’ll present the bear and the bull case. I lean more bullish for reasons that are below, but I want to highlight both sides and what this means for the broader AI industry.

So, what did Jensen do today that

Compute Is Revenue. Revenue Is Collateral.

Nvidia did not invent compute-backed credit on Monday. It proposed an assembly line for it.

Anthropic says new Claude models will embed invisible watermarks in all generated text, everywhere Claude is offered.

The watermark is part of the text, it isn't metadata: "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing."

This starts with models launched on or after August 2, 2026, under an EU AI Act code Anthropic signed. Anthropic is still working on adding it to current models. The rollout is worldwide.

How Claude marks AI-generated content | Claude Help Center

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, as a provider of both generative AI models and generative AI systems. This article describes how we're planning to put those commitments into practice, how marking works, and what its limitations are.

https://support.claude.com

via David James

Seeking an academic co-author — AI, cultural memory, and migration

I'm developing a research proposal for Coefficient Giving's RFP on Global Health and Wellbeing in an Era of Transformative AI (Exploratory grant), and I'm looking for an academic partner to lead the method.

The hypothesis: Africa's oral knowledge systems — botanical, climatic, agronomic, linguistic, held by griots and traditional knowledge-keepers — are on the wrong side of the AI era by default. Frontier models are trained on written, Western corpora; unwritten knowledge is invisible to them, and so is devalued by them. Yet the same AI advances mean these knowledge systems will soon hold enormous discoverable value. The question is who will own it when that moment arrives — the communities, or whoever digitises them first.

The proposed work: a research paper on AI-era knowledge homogenisation, cultural detachment and migration aspirations; a community-ownership and benefit-sharing governance framework for oral-knowledge corpora (Nagoya/CARE-aligned); and a costed pilot corpus design with one community.

I'm looking for a co-author in low-resource NLP / computational linguistics, anthropology or ethnobotany, or migration economics — someone who would lead the research method, with me contributing the practitioner frame and thirty years of field relationships across African and Asian agricultural communities. My own work on language, consciousness and AI (Your Mind at Siege, on the Diamond Sutra and AI) shapes the framing.

Deadline is near — 21 August — so this suits someone who can move quickly on a light, sharply-scoped proposal. If this is your field, or you know whose field it is, my messages are open.

Anthropic agreed to a 20-year, $9.1B compute deal with Riot Platforms a Bitcoin mining company that also recently began selling AI data center capacity for 191 MW of capacity at a Rockdale, TX campus; RIOT jumps ~25% after hours

Bloomberg

@business

Anthropic has struck a $9.1 billion deal with Riot, a Bitcoin mining company that also recently began selling AI data center capacity, according to sources, underscoring the Claude maker's efforts to secure enough computing to meet demand

bloomberg.com

Anthropic Strikes $9 Billion Deal With Cloud Computing Firm Riot

Anthropic PBC has struck a $9.1 billion deal with Riot Platforms Inc., a Bitcoin mining company that recently began selling AI data center capacity, people familiar with the matter said, underscoring...

Harry Stebbings profiles OpenRouter as a breakout AI inference company co-founded by ex-OpenSea leader Alex Atallah, noting its hundreds of millions in revenue, profitability, 25 trillion weekly tokens processed (corrected to 70-75 trillion in replies), and a reported $10BN Stripe acquisition offer after a recent $1.3BN valuation.

The post summarizes seven interview insights, stressing AI model inference as tech's largest future market via Jevons Paradox-driven usage growth, the superiority of focused routing over side-feature competitors, and risks of frontier labs absorbing startup workflows

Key themes include accelerating model releases every 10 hours creating a multi-model ecosystem, China's edge in open-weight AI, open models closing performance gaps on frontier ones, and businesses needing to treat AI inference spend as variable employee costs.

Harry Stebbings

@HarryStebbings

OpenRouter is one of the most insane stories in tech.

Co-founded by OpenSea founder, Alex Atallah. It has become one of the most important companies in AI.

Scaled to 100s of millions in revenue and wildly profitable.

They process 25 TRILLION tokens every week and will do 1

20VC: Will OpenRouter Sell for $10BN to Stripe? | Why Chinese Open Models Are Beating America-and What Happens Next | Why Enterprises Are More Fearful of Anthropic and OpenAI Than China | Is the Routing Layer Becoming a Commodity with Alex Atallah

Listen to this episode from The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch on Spotify. Alex Atallah is the Founder and CEO @ OpenRouter, the unified interface for LLMs. The company has raised over $153M in funding, with the latest valuation pricing the company at $1.3BN.

https://spotify.com

via Jorvik Zhang

The eight debates from the AI First seminar in London last Thursday, in one short deck: https://tokenhub.tencentcloud-eu.com/AI-First.html

Still the sharpest of them — AI tax vs. AI dividend.

The tax is what you pay correcting and re-prompting. The dividend is what you keep when it lands the first time. Most teams feel the tax daily but can't yet name the dividend.

Which is bigger for your team

NVIDIA Unveils Nemotron 3.5 Lightning for Fast AI Agents

This open-source model uses a hybrid Mamba-Transformer design and a permissive license for commercial use, boasting a 1 million token context window and up to 4x the output speed of similar models. Benchmarks show it scoring 24 on the Artificial Analysis Intelligence Index, matching much larger rivals, with standout results in agent tasks like GDPval-AA v2 at 824 Elo and 86% on PinchBench—while finishing 10,000 tasks 35% faster than competitors.

NVIDIA AI

@NVIDIAAI

Introducing NVIDIA Nemotron 3.5 Lightning

An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster.

It delivers up to 4x the output speed of similar-sized models.

Artificial Analysis

https://www.signerset.com

NVIDIA has just released the first Nemotron 3.5 model: Nemotron 3.5 Lightning, a highly efficient small open weights model with performance similar to gpt-oss-120b at around a quarter of the total parameters

Nemotron 3.5 Lightning is the successor to @nvidia Nemotron 3 Nano 30B

Together AI

@togethercompute

NVIDIA Nemotron 3.5 Lightning is now live on Together AI.

The fastest open model in its class is built for always-on agents that need to complete high-volume, specialized work quickly.

Nvidia also released NeMo Switchyard: an open-source model routing library that automatically directs each step of an agent workflow to the best model for the job. Internal benchmarks: frontier-level accuracy at ~1/3 the cost of Opus 4.8 alone.

Devin cut mean cost 28%. LangChain cut cost 74% by routing just 7% of calls to a frontier model. The savings are in the routing, not the model.

NVIDIA

@nvidia

Today, NVIDIA announced NVIDIA Nemotron 3.5 Lightning, a customizable model for high-volume, specialized work, and NVIDIA NeMo Switchyard, which helps agents route each workflow step across the models they choose. x.com/i/article/2086…

NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Deliver Faster, Smarter, More Efficient Agentic AI

The new lightweight open model and routing library delivers greater control over AI, data and workflows across edge devices, PCs, workstations, data centers and the cloud.

https://blogs.nvidia.com

Wednesday 12th August 2026

via Alfred Tom

Is anyone running autonomous agent workflows, like OpenClaw, Cursor, Hermes, Animoca Minds, etc.? We're launching something that brings the concept of banking-style approvals to autonomous workflows. Basically it helps prevent things like deleting production databases without requiring a human in the loop (although it supports that as well). If you're interested I'm happy to help with integrations

SignerSet - Approvals for Autonomous Workflows

Policy-based, multi-party approval for consequential AI actions. Define protected actions, authorized signers, and signature thresholds.

SpaceXAI Launches Grok Bot AI Teammates for Real Work

Grok Bot introduces persistent AI agents that sign into tools like Gmail, Salesforce, and LinkedIn to complete end-to-end tasks, learn from demonstrations, and collaborate with each other in chats. Each bot runs on its own cloud computer, remembers preferences, and improves over time, already transforming internal operations in sales, marketing, and engineering at SpaceXAI. Early users like Lenny Rachitsky praise its practical power for job matching, auto-replies, and expense scanning, while Elon Musk plans a wider beta after quick fixes and Grok 4.6. Access begins for SuperGrok Heavy and Cursor subscribers on desktop and iOS.

Introducing Grok Bot

Grok Bot is your team of always-on agents. They have their own computer, work inside tools and apps like you do, and keep working 24/7.

https://x.ai

Grok Bot

@bot

Introducing Grok Bot, now in early beta.

Bots are AI teammates that do real work for you. They sign in to your tools, use them just like you do, and come back with finished work.

Elon Musk

We will widen the Grok Bot beta after we fix basic issues with the early beta and release Grok 4.6 later this week

Grok Bot

@bot

Introducing Grok Bot, now in early beta.

Bots are AI teammates that do real work for you. They sign in to your tools, use them just like you do, and come back with finished work.

Riley Brown

@rileybrown

GrokBot Just Launched:

GrokBot is Cursor/SpaceX new general agent platform for Desktop and iOS...

Got access yesterday... here are the things that Grokbot can do, my thoughts, and what I hope they add soon.

Pretty big news for those who want to build businesses that are

Lenny Rachitsky

@lennysan

I got early access to Grok Bot and I'm hooked.

I haven't been this excited about a new AI product in a while.

It's like OpenClaw, but super easy, reliable, and less scary to use. I think this will be a huge new product line for Cursor/Grok/SpaceX.

I've already found so many

Grok Bot

@bot

Introducing Grok Bot, now in early beta.

Bots are AI teammates that do real work for you. They sign in to your tools, use them just like you do, and come back with finished work.

via Vladimir Arustamov

Built a tool that scores any website on a 6-level Agent Readiness framework — from basic HTTPS (L0) to autonomous AI payments via x402 protocol (L4). Most sites are stuck at L1.

seo4agent.com — free instant audit

github.com/varustamov/agentable-core — MCP server + Python SDK (pip install agentable-core)

Will be glad to get your reviews, feedback & stars on GitHub

Mark Zuckerberg proposes a positive AI philosophy centered on individual empowerment, predicts personal superintelligence could increase employment, and more

The Future is for Everyone

Read Mark Zuckerberg's piece on our values for building superintelligence for everyone.

https://www.meta.com

Thursday 13th August 2026

DeepSeek V4 Pro GA and Grok 4.6 Launch on Same Day

DeepSeek's V4 Pro, now generally available on API and chat, offers a 1 million token context window, up to 384K output, JSON support, tool calls, and pricing at $0.003625 per million input tokens on cache hit, $0.435 on miss, and $0.87 for output.

SpaceXAI quickly launched Grok 4.6, optimized for coding, agentic tasks, and knowledge work with adjustable reasoning levels.

SpaceXAI

Introducing Grok 4.6.

It delivers frontier intelligence and is a significant improvement over Grok 4.5 at the same price.

Chubby♨️

DeepSeek v4 Pro GA is rolling out! Grok 4.6 API spotted as well!

Really excited! After the most recent insane update to DeepSeek v4 Flash i expect sevel updates on this one as well!

big release day it seems!

leo 🐾

@synthwavedd

DeepSeek V4 Pro (GA) is now rolling out on the DeepSeek API and on DeepSeek Chat

Elaina

@ElainaNotes

DeepSeek V4 Pro GA confirmed to be released as DeepSeek-V4-Pro-0813.

AiBattle

@AiBattle_

Confirmed release DeepSeek-V4-Pro-0813

eric zakariasson

@ericzakariasson

grok 4.6 is live! a write up on my learnings, findings and tips to using the model! x.com/i/article/2087…

Artificial Analysis

SpaceXAI's Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, joining the frontier in line with GPT-5.6 Sol, with standout agentic performance at lower cost

Grok 4.6 gains 5 points over Grok 4.5 on the Intelligence Index just over one month after its release, or

Derrick Choi, OpenAI's Codex APAC Lead, promotes 16 real workflows created by their finance team using ChatGPT Work and Codex for tasks from monthly close and variance reconciliation to forecasting adjustments and treasury analytics.

Key examples include dynamic marketing spend rebalancing, headcount visibility dashboards, intelligent scenario modeling, governed monthly close processes, and AI-powered treasury agents that flag issues and draft responses

The approach automates data assembly, reconciliation, and reporting to shift finance focus toward decisions, tradeoffs, and strategic conversations while retaining human sign-off and accountability.

Derrick Choi

@derrickcchoi

Our finance team shared 16 workflows with ChatGPT Work and Codex, from monthly close to forecasting and treasury.

Really good read especially if you’re in finance.

AI Workflows for Finance Teams

Explore 16 strategic finance workflows built with ChatGPT Work and Codex, from decision support and close to forecasting, planning, and agent orchestration.

Elon Musk Announces Grok 4.7 Launch in 3-4 Weeks

Musk said Grok 4.7, boosted by massive SpaceX company data in supplemental training, will outperform all current models—especially in practical engineering tasks. It follows hot on the heels of Grok 4.6, now available via tools like Cursor and the SpaceXAI API, with strong showings in coding, agents, and benchmarks against GPT variants

Elon Musk

Grok 4.7 is significantly better than 4.6 and should be ready in 3 to 4 weeks.

Initial training is complete and now we’re adding a massive amount of SpaceX company data in supplemental training. This will be something special.

Elon Musk

Grok 4.7 will exceed all current models.

That said, Anthropic is a great company and will probably release improved models soon.

However, the SpaceX training corpus is so awesome & unique that I would be shocked if any model is better at real-world engineering than 4.7.

DeepSeek launched V4-Pro with major agent upgrades, flexible reasoning modes (low/high/max effort), and native OpenAI Responses API support optimized for Codex workflows.

Benchmarks show V4-Pro-0813 leading in agentic tasks like Terminal Bench (87.9), DeepSWE (62.7), and DSBench-Hard (67.2), outperforming previews and rivals including Kimi-K3 and Opus-4.8.

API pricing update introduces peak/off-peak rates with off-peak 50% cheaper, effective August 16, 2026, for both V4-Pro and V4-Flash models.

DeepSeek

We’re launching DeepSeek-V4-Pro today!

Major Agent upgrades with strong production gains!
Flexible reasoning effort for V4-Pro & V4-Flash: low for simple tasks, high for daily Agent workflows, max for complex tasks.
Native OpenAI Responses API support, optimized for

DeepSeek AI announced DeepSeek Harness v0.1 in developer preview as an open-source agent harness under MIT license, inviting global developers to build and extend it.

The framework uses a "everything is a plugin" architecture powered by the Cordis meta-framework, allowing models, tools, sessions, sandboxes, orchestration, and UI to be mixed, matched, replaced, or extended modularly.

It focuses on reliable agent execution by handling engineering aspects like task orchestration, memory management, tool scheduling, and error recovery to bridge models into practical agents

DeepSeek

DeepSeek Harness v0.1 is now available in Developer Preview!

We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license.
Powered by the Cordis meta-framework, DeepSeek Harness is an agent harness built around one

github.com

GitHub - deepseek-ai/deepseek-harness: DeepSeek Harness: Everything is a Plugin.

DeepSeek Harness: Everything is a Plugin. Contribute to deepseek-ai/deepseek-harness development by creating an account on GitHub.

Friday 14th August 2026

IOSG Ventures

@IOSGVC

x.com/i/article/2087…

Gemini 3.7 Flash announced highlighting its high speed, 50% price reduction versus 3.6 Flash through year-end, and major intelligence gains from algorithmic improvements in roughly three weeks.

Benchmark charts show 3.7 Flash leading in enterprise workflow automation (30.4% on AutomationBench), web development (1588 Elo on Code Arena), long-horizon software engineering (65.3% on DeepSWE), and production code quality (43.6% on FrontierCode), surpassing 3.6 Flash and competitors like Claude Sonnet 5 and GPT-5.6.

The model is now available in the Gemini API, AI Studio, Antigravity, and additional platforms, reflecting Google's rapid iteration on the Flash series for better real-world usability.

Logan Kilpatrick

@OfficialLoganK

Introducing Gemini 3.7 Flash : )

- it is fast!
- 50% lower price than 3.6 flash (through end of year)
- strong intelligence increase in only ~3 weeks (thanks to some awesome algorithmic improvements)
- available in the API, AI Studio, Antigravity, and more!

Introducing Gemini 3.7 Flash

Gemini 3.7 Flash is our most intelligent workhorse model yet for coding and agents.

Z.ai announces GLM-5.3, post-trained on a 743B base model, highlighting top-tier coding, agentic capabilities, and a major advancement in cybersecurity for open models.

Benchmark charts show GLM-5.3 leading in Terminal Bench 3.0, DeepSWE, CyberGym, ExploitBench, and agentic coding accuracy, outperforming GLM-5.2 and competing strongly with models like Kimi K3 and Claude variants.

The model is available now through GLM Coding Plan and ZCode, with staged API access and open weights following safety evaluations, plus initial partner services.

Z.ai

GLM-5.3 is available now through GLM Coding Plan and ZCode. API access and open weights will be released in stages following rigorous safety evaluations.

GLM Coding Plan: z.ai/subscribe
ZCode: zcode.z.ai/en

Zixuan Li

@ZixuanLi_

GLM-5.3 uses the same 743B base model as GLM-5.2, while matching the performance of models several times its size. Benchmark highlights:

- Terminal Bench 3.0: 28.3
- DeepSWE: 66.9
- Agents' Last Exam: 28.5
- GDPVal-AA: 1769

Z.ai

Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.

- Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
- A major leap in cybersecurity, setting a new standard among open models

Tech Blog: z.ai/blog/glm-5.3

Zixuan Li

@ZixuanLi_

Full benchmark results for GLM-5.3

IMHO, signing up for Monid (you get $1 in credits on signup) and installing its MCP server will dramatically change the way you work with your harness [Cursor, ChatGPT, Claude, Grok Build, Opencode, Command Code etc]

Jasper Li

@Jasperli0122

x.com/i/article/2084…

Also, Tinyfish is available as a ChatGPT plugin [search and fetch are free] but you can also install it as a MCP server in other harnesses

TinyFish

@Tiny_Fish

TinyFish is now officially available in @ChatGPT.

TinyFish brings live web Search, clean Fetch, and Web Agent to ChatGPT, so your agent can find it, read it, and act on it.

Saturday 15th August 2026

The cover image of this newsletter via generated via the Seedream 5 Pro model within the Krea tool via the following prompt

Retro Futurism style, It is a city designed by Buckminster Fuller on Mars in 2500, influenced by the Purism architectural movement and can change shape according to needs

This Week in All Things AI - Week 32-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 09 Aug 2026 03:03:10 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 2nd to 8th August 2026 saw a flurry of agentic coding launches. DeepSeek opened a closed beta for its Harness framework, and Qwen previewed its 2.4T-parameter Qwen3.8-Max ahead of an open-weight release. Meta launched Muse Code on the new Muse Spark 1.2 model with steep data-sharing discounts, while Command Code rolled out its GOAT subscription plan bundling access to 30+ coding models. OpenAI, with AWS, Cursor, GitHub, Microsoft, and Vercel, introduced Agent Plugins, an open standard for portable agent skills.

Multimodal generation also advanced: MiniMax open-sourced its H3 omni-modal model with day-zero ComfyUI support, Black Forest Labs expanded FLUX 3 Video access, and Alibaba's Wan3.0 entered public beta with native 30-second clips. xAI closed the week with Imagine Image 2.0, adding precision editing tools built for real-world production use. On the infrastructure side, practitioners flagged lighter headless browsers for cutting agent navigation costs, security researchers warned about trust-handoff exploits in agent harnesses, and a widely shared "Power 2026" primer underscored electricity supply as AI's next major bottleneck."

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interest

Sunday 2nd August 2026

Tianyi Cui from DeepSeek's Harness team posted on X inviting open-source developers of Agent Harness projects to join the closed beta by replying or DMing with GitHub ID and representative work.

DeepSeek Harness is an upcoming agent framework focused on tool calling, context management, and terminal execution, positioned as a competitor to Anthropic's Claude Code and tied to their V4 model's enhanced coding agent capabilities.

The beta targets specialized builders, with recent replies already featuring GitHub links to desktop agent tools and CLI integrators, reflecting strong interest from the AI agent development community.

Tianyi Cui

@tianyi

如果你是 Agent Harness 相关开源项目的开发者，希望参加 DeepSeek Harness 的内测，可以回复或私信联系我。请附上 GitHub id 以及开源代表作。

Monday 3rd August 2026

Alibaba's Qwen team announced Qwen3.8-Max as their most capable model yet with 2.4T parameters, emphasizing autonomous coding over 10+ days, long-horizon planning, and native multimodal intelligence for real production tasks.

Benchmark charts show Qwen3.8-Max leading in software engineering, agent workflows, visual perception, and embodied reasoning, surpassing models like GPT-5.6 Sol (max), Gemini1.3-Pro, and Opus4.8 across most categories.

Open weights for Qwen3.8-Max and the smaller Qwen3.8-27B will release next week; API pricing is set at $2/M input tokens, $6/M output tokens, with demos highlighting extended agentic coding and dynamic visual workflows.

Qwen

Meet Qwen3.8-Max — our most capable model to date.

Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!

Qwen3.8-Max, a new bar for coding and cowork at 2.4T parameters:

- Autonomous coding: 10+ days of

MiniMax AI announced the open-weight release of its 33B-parameter MiniMax-H3 omni-modal model on Hugging Face, enabling text/image/video/audio-to-video generation with native synchronized stereo audio up to 15-second 2K clips at 24 FPS.

The community license includes a US carve-out due to generative video copyright litigation with Hollywood studios It is also unavailable in the EU, UK, and South Korea due to regulatory uncertainties around AI-generated video content.

Users in those areas can submit formal licensing requests by committing to robust local compliance controls and guardrails, with MiniMax planning to monitor and adapt policies as regulations change.

ComfyUI announces day-zero native integration for open-weights MiniMax H3,

MiniMaxAI/MiniMax-H3 · Hugging Face

We're on a journey to advance and democratize artificial intelligence through open source and open science.

https://huggingface.co

RyanLee

About Licence:

This regional carve-out stems from our ongoing generative video copyright litigation with major @Hollywood studios.

We aim to deliver our AI tools globally to all users responsibly. US-based persons looking to deploy MiniMax-H3 may submit formal licensing

RyanLee

For the EU, UK, and South Korea. This regional carve-out is driven by ongoing regulatory evolution surrounding AI-generated video content across these regions. Unlike hosted services, open-weight model releases allow developers to independently deploy and modify model weights,

ComfyUI

@ComfyUI

MiniMax H3 is native in ComfyUI, day zero with the weights.

Model highlights:

→ Text-to-video, prompt only

→ Image-to-video

→ First-and-last-frame - control the opening frame, the closing frame, or both

→ Reference-to-video - carry a subject, a motion, or a voice

Tuesday 4th August 2026

Wednesday 5th August 2026

via Alex

If you are using agents to crawl websites and get robo-blocked or spend too many tokens for navigation, I recommend headless browser rebuilts like: https://github.com/lightpanda-io/browser

It cut my token use on navigation by 80+%

Flux 3 Video now more available from BFL, OpenRouter, FAL, Replicate

It's capable of generating up to 20-second 1080p clips via text-to-video, image-to-video with multiple keyframes, video continuation, and native audio including multilingual dialogue with accurate lip sync.

Draft mode enables fast creative exploration with low-cost previews at $0.06 per second that preserve subjects and motion for seamless upgrade to full quality rendering from $0.17 per second.

2K, 4K support and open weights planned soon.

Black Forest Labs

@bfl_ai

FLUX 3 Video is here.

Serious, fun, creative, real, cinematic, whatever you need it to be.

Native audio, Text to Video, Image to Video with multiple frames, video continuation, dialogue in multiple languages.

Comes with Draft mode so you can explore ideas fast at a fraction of

OpenRouter

FLUX 3 Video from @bfl_ai is now available for everyone on OpenRouter.

One unified multimodal model family for video, audio, image, and action-prediction. Serious, fun, creative, real, cinematic, whatever you need it to be. Jointly trained in one unified architecture.

fal

https://www.helpnetsecurity.com

FLUX 3 by @bfl_ai is now available on fal.

Up to 20s per generation with native audio, multilingual dialogue with strong lipsync, multi-scene shots, and video continuation that carries framing and scene logic forward.

Replicate

@replicate

FLUX 3 Video from @bfl_ai is now on Replicate!

One multimodal model for video, audio, image, and action-prediction, with video released today. Generate 20 second videos up to 1080p.

Try here:
replicate.com/black-forest-l…

Black Forest Labs

@bfl_ai

via Alex

God of the gaps. Hacking Agents through (un)trusted handover commands in the harnesses.
Many harnesses hardcode trust -> "everything beyong this point is trusted" and this can be used to prompt hijack agents:

Quote from the article:

“Most teams audit what the agent can do. The exposure we keep finding is in what happens after, the handoff between ‘approved’ and ‘executed,’ between ‘read’ and ‘published,’ between ‘fetched’ and ‘trusted.'”

An AI agent can pass every safety check and still leak secrets - Help Net Security

AI agent security fails at the handoff. Novee's Elad Meged broke Claude Code, Gemini CLI and Codex agents in their default configs.

via Alex

TIMES magazine started to expose bots to abbreviated articles with baked in ads.

https://TIMES magazine started to expose bots to abbreviated articles with baked in ads.

Thursday 6th August 2026

Meta Launches Muse Code Beta for Full Repo Engineering Tasks

Powered by the new Muse Spark 1.2 model from Meta Superintelligence Labs, Muse Code plans changes, writes code, and validates results across sessions using persistent background sub-agents that parallelize work safely. Benchmarks show it competing with top models from OpenAI and Anthropic on tests like Terminal-Bench 2.1 and DeepSWE 1.1, while pricing starts at $1.25 per million input tokens—with a cheap contributor tier at $0.10 if you opt into data sharing. Mark Zuckerberg and chief AI officer Alexandr Wang emphasized its one-command install and focus on reliable, long-horizon coding tasks.

Mark Zuckerberg

@finkd

Releasing Muse Code in beta today. It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results. Powered by Muse Spark 1.2, a coding-focused model update.

Alexandr Wang

muse code in beta is live. first coding agent from msl, built on muse spark 1.2.

install: curl -fsS dev.meta.ai/install.sh | bash

here's what you should know:

Artificial Analysis

Meta has released Muse Spark 1.2. It's their third release in four months and scores 54 on the Artificial Analysis Intelligence Index, significantly improving agentic knowledge work capabilities over prior releases and putting Meta next to SpaceXAI in a tie for third place

Lisan al Gaib

they actually give you the model almost for free if you allow them to train on your data

Lisan al Gaib

https://platform.deepseek.com/usage

RIP GOOGLE

LONG LIVE META

Introducing Muse Code and Muse Spark 1.2

Introducing Muse Code, a terminal coding agent powered by Muse Spark 1.2, with persistent background agents, repository-scale execution, and built-in verification.

https://research.meta.ai

Deepseek plans to raise API prices significantly but no details announced, their v4-Flash 0.14/0.28/0.0028 is currently the floor for Chinese models including Mimo-v2.5 though Meta offers a lower price point for its recently launched Meta Spark-v1.2-contributor at 0.1/0.2/0.002.

Note that the low cache price in particularly from Deepseek's own provider allows them to train on input as per their ToS/privacy policy. Other providers who also server DSv4-Flash-0731 may have different policies and serve it at different price points. Pricing is always a function of model and provider

Leroy Li

@LeroyLi311063

DeepSeek has just announced:

"We plan to raise the overall pricing for DeepSeek API services in the near future; a significant increase is expected."

Lumina

@LuminaXspace

Muse Spark 1.2 is disgustingly cheap:

Meta has two API versions, one where you share data one where you don't:

• Input: $0.10/M vs $1.25/M, 12.5× cheaper
• Cached input: $0.002/M vs $0.15/M, 75× cheaper
• Output: $0.20/M vs $4.25/M, over 21× cheaper

They genuinely give

Command Code's announces the GOAT plan at $10/month for $70 credits, delivering 7x value and access to over 30 open-weight and closed AI models for coding tasks.

The plan builds on their $1 Go starter tier by providing enough usage for meaningful projects, emphasizing solutions to open model challenges like cost, reliability, and performance via high cache rates and design tools.

Command Code positions itself as a leader in open model coding agents with leaderboard success, a v1 codebase rewrite, new Mods API, and upcoming open source release later this month

PS, It also offers Muse Spark v1.2 and Muse Spark v1.2-contributor

Command Code

Introducing the GOAT plan

Command Code GOAT plan is the best low-cost coding plan on the market today.

- $10/mo to get $70 credits. 7x what you pay
- 30+ open weight and closed models

$1 Go got you started. GOAT gets you shipping.

Subscribe — be a GOAT.

Command Code

Muse Spark 1.2 by Meta now live in Command Code

Muse Spark 1.2
• Available in plans: GOAT, Pro, Max
• $1.25 input, $4.25 output, $0.15 cache

Muse Spark 1.2 Contributor
• Available in plans: Go, GOAT, Pro, Max
• Contributor is ~95% cheaper than Muse Spark 1.2
• $0.10

GOAT Plan | Command Code

The best low-cost coding plan: Command Code GOAT at $10/month. Unlimited coding on 30+ top open source and closed source models - pay $10, get $70 of credits, a 7x multiplier. With deals, usage stretches beyond $100.

https://commandcode.ai

Alibaba's Tongyi Lab announces Wan3.0 public beta, a video generation model featuring native 30-second single-run video creation, high-fidelity reality-grade rendering, and expressive character consistency.

Omni-reference capability allows input of structured files like documents, spreadsheets, slides, PDFs, and webpages alongside text, images, audio, and video for more versatile content generation.

Now available on Alibaba Cloud Model Studio and Qwen Cloud with API pricing from $0.05/sec at 480p; additional demos in the thread showcase 30-second outputs and complex reference handling....

Wan

@Alibaba_Wan

Introducing Wan3.0 — now in Public Beta.
Simple Input. Smart Creation. Where Imagination Meets Reality.

• Native 30-Second Video Generation
• Reality-Grade Rendering
• Omni-Reference: Beyond text, images, audio, and video—now with documents, spreadsheets, slides, webpages,

Friday 7th August 2026

OpenAI Launches Agent Plugins Standard for Shared AI Extensions

OpenAI, teaming up with AWS, Cursor, GitHub, Microsoft Code, and Vercel, introduced Agent Plugins on Thursday. This open standard lets developers package AI agent skills and MCP server configurations into a shared format, working across tools like ChatGPT, Cursor, GitHub Copilot, and VS Code.

Some developers question its depth without broader adoption from players like Anthropic.

OpenAI Developers

@OpenAIDevs

Build a plugin once and use it across compatible agent clients.

Introducing Agent Plugins, an open standard developed with @awsdevelopers, @cursor_ai, @github, @code, and @vercel that packages Agent Skills and supports MCP server configurations in a shared format.

Cursor

https://x.com/i/trending/2085480744365785205

Cursor now supports Agent Plugins, an open standard for bundling skills and MCP servers for use across agents.

x.com/vercel/status/…

Vercel

@vercel

Introducing Agent Plugins, an open standard for extending agents.

Supports Agent Skills and MCP, with more to come.

Built in collaboration with: @awsdevelopers, @code, @cursor_ai, @github, and @openaidevs.

vercel.com/blog/introduci…

Some SpaceX'y news in anticipation of Grok-4.6.

For all intents and purposes, I'm consider the Cursor acquisition by SpaceXAI a done deal even thought as of now it's not yet closed

FWIW, Grok Build is my daily driver. I combine it with MCP's of Monid [put in a dollar or two at a time and use the tools it aggregates without worrying about balance sprawl], TinyFish and Nia [Nia, you could think of it as Context7 on steroids]

Exa and Octen are my favourite tools to use via Monid

Opencode is also massively improving its web search in its upcoming version 2

Cursor's Smart Router Boosts AI Coding Efficiency and Satisfaction

Grok Build Hits 1.0.0 Milestone with Refined Coding Tools

https://x.com/i/trending/2085597018781806940

via Chyi Yan Hshieh

https://paragraph.com/@twiata/this-week-in-all-things-ai-week-31-2026Curious question for the broader group, do you see a demand for FDEs in APAC? Or at least for companies who are doing the legwork of setting up AI systems and agents?

Former Citadel quant Neel Somani highlights power supply as the core AI bottleneck over chips or memory, releasing "Power 2026" primer detailing electricity markets, data center demands, and pricing dynamics.

The guide covers US grid regions like PJM, ERCOT, and MISO, marginal pricing via merit order, heat rates, and how AI data centers—already at ~5% of US power—could exceed total generation by mid-2030s.

Somani's expertise from power/gas trading and data center advising informs hedging strategies, buildout challenges, and calls for faster permitting of gas/coal plants plus transmission upgrades to meet surging demand.

Props to Matthew Mousa of Alpha Compute for pointing me towards this material

Power 2026 by Neel Somani - Electricity Pricing in the Age of AI

Power 2026 is a primer on AI's impact on the US power markets. The short piece is written by Neel Somani, a former quantitative researcher at one of the world's leading hedge funds.

https://power2026.ai

Saturday 8th August 2026

xAI's next-generation image model Imagine Image 2.0 was just launched emphasizing precision editing, crisp text rendering, improved factuality, and practical real-world applications like infographics, ads, UI mockups, and game assets.

Key new tools include Magic Wand for targeted region changes, Segmentation for precise area selection, Smart Resize for custom aspect ratios, and templates for workflows such as product shots, collages, icons, and consistent character/world building for video.

It ranks #2 worldwide in both text-to-image and image editing arenas per recent Arena leaderboards, and is now available as the Quality Mode on grok.com/imagine, iOS/Android apps, with API access coming soon.

Grok

@grok

Announcing Imagine Image 2.0, our next generation image model with precision editing, crisp text rendering, improved factuality, and real world usefulness.

Image 2.0 helps you make images for real work.

x.ai

Imagine Image 2.0

Precise image generation and editing, built for real creative work.

https://www.wsj.com/business/china-us-ai-model-costs-53a12e96?st=VxRECn&reflink=desktopwebshare_permalink

The cover image of this newsletter via generated via the Seedream 5 Pro model within the Krea tool via the following prompt

Craft an image of an expansive, rolling prairie, where the golden grasses sway in the breeze, and herds of wild mustangs roam freely beneath an endless, open sky

This Week in All Things AI - Week 31-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 02 Aug 2026 04:24:10 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 26th July to 1st August 2026 was defined by a decisive shift from raw model capability toward the economics, reliability, and operational design of agentic AI. Microsoft introduced its MAI-Cyber-1-Flash cybersecurity model and Project Perception, while benchmark work comparing Kimi K3 across Claude Code, Hermes, and Kimi Code showed that the harness can drive order-of-magnitude differences in token consumption even when task success rates remain similar. OpenAI also cut GPT-5.6 pricing, and discussion around cheaper Chinese models, region-specific coding plans, and hybrid model stacks reinforced the view that “good enough” intelligence is becoming a major competitive advantage.

Agent infrastructure, privacy, and open deployment were equally prominent. MCP’s move to a stateless core makes remote servers substantially easier to place behind load balancers and on serverless or edge infrastructure, while new search, crawling, and browser APIs from providers including Firecrawl, Octen, Context.dev, and TinyFish highlighted the growing importance of high-quality web access for agents. At the model layer, Fish Audio, Superwhisper and Cohere pushed more capable local speech tooling; MiniMax released its open-weight H3 multimodal video model; DeepSeek substantially upgraded V4-Flash’s agentic coding performance; and Poolside launched a cross-platform workspace for supervising parallel coding agents. Meanwhile, the indexing of supposedly private AI conversations and a public call by frontier-lab employees to deliberately pace AI development kept security and governance firmly in view.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 26th July 2026

via the Relentless podcast [quite underrated at only 25K subscribers] TI Morse with Sam Altman

Sam drops some mind-blowing insights on how AI is completely rewiring the rules for startups, arguing that we are literally living in the "singularity" right now. He dives deep into why we need to trust in exponential AI progress, how to thrive as a leader in absolute chaos, and why superintelligence won't actually bring us the 4-hour workweek.

There's also some incredible behind-the-scenes intel on OpenAI's ruthless prioritization—like deciding to pause Sora to focus on coding agents—and why massive investments in chips, energy, and robotics are the real bottlenecks to our future.

Monday 27th July 2026

Kevin Kelly highlights how AI tokenomics are shifting user behavior toward lower-cost, "good enough" models as expenses become a key factor, potentially making leading-edge AI a commercial liability.

The linked WSJ article reports US companies adopting cheaper Chinese open-weight models for routine tasks, using hybrids where premium models plan and affordable ones execute, achieving major savings like reducing daily AI agent costs from high figures to $100 per agent.

This reflects a broader "thrift-maxxing" trend in AI, prioritizing cost efficiency and model flexibility over maximum power, similar to how mass-market computing products historically outsold specialized high-end hardware.

Kevin Kelly

@kevin2kelly

Tokenomics are taking hold. The cost of AI will make a huge difference, diverting use to lower cost, less powerful, but good enough models. Leading edge may be a liability.

wsj.com

Corporate America Has Suddenly Decided to Stop Blowing Money on AI

Companies big and small are mixing models, and it’s changing the economics and power players of the industry.

via Yat Siu

Animoca Research launched "The Agentic Age," its first monthly brief on the agentic economy, emphasizing that abundant AI coverage requires sharper judgment for capital allocators, operators, and builders

Animoca Brands Research

@animocaresearch

Coverage of AI is not scarce. Judgment is.

Introducing The Agentic Age, the inaugural issue of @animocaresearch 's new monthly brief on the agentic economy.

Each issue filters the noise into what actually matters for capital allocators, operators, and builders.

From Issue

Tuesday 28th July 2026

Microsoft introduces MAI-Cyber-1-Flash, an AI model trained for cybersecurity, and launches Perception, an agentic security system to patch vulnerabilities. It says MAI-Cyber-1-Flash and MDASH, its vulnerability identification harness, deliver “world-class performance at 50% of the cost of leading models”

Satya Nadella

Today, we are announcing a series of updates that give customers frontier-grade security at half the cost.

MAI-Cyber-1-Flash is our first cybersecurity model, built ground up to find the most challenging vulnerabilities in complex code bases. When combined with MDASH, it

Mustafa Suleyman

https://blogs.microsoft.com

Big news! Our new MAI-Cyber-1-Flash model combined with MDASH, our multi agent security harness, delivers 96% on the CyberGym benchmark, 12pts above Mythos, at HALF the cost.

Proud of the team. More details in THREAD:

Rethinking security for the age of AI - The Official Microsoft Blog

Why security needs a new cyber stack - Introducing Project Perception The physics of cybersecurity are changing. Autonomous systems can now reason, adapt and operate continuously. At the same time, the cost of offense is falling, while the volume, velocity and complexity of what must be secured continues to grow.

via Alex

Anthropic forgot to add noindex and basic privacy blocks to shared conversations, making hundreds of thousands of privately shared conversations (including on enterprise plans) publicly indexed.

Imho its becoming increasingly clear that corporations will have to migrate to on prem / private cloud solutions in order to guarantee good data security practices.

https://archive.is/20260727190025/https://fortune.com/2026/07/27/a-trove-of-users-seemingly-private-conversations-with-anthropics-claude-ai-chatbot-showed-up-in-google-search-results/

via Alex

In the last year alone we had a series of data security blunders:

- xAI forgetting to add noindex on 400k conversations

- xAI uploading entire private repositories of thousands of clients to google cloud incl. .env files

- OpenAI forgetting to add noindex to shared conversations on 5k conversations

- Deepseek forgetting to add noindex to conversations (scope unknown)

- OpenAI unable to control its own sandbox environment, taking 3 days to notice that its own model was on the internet commiting felonies (the ChatGPT HuggingFace hack)

Alex further writes

I did a quick back of the envelope breakeven analysis.

Blackwell Nodes needed to run GLM5.2 at over 20tps output cost around 500k$US (if you can get one at all), and cost around 40$/h to rent from east asian private cloud providers. Electricity to run the node is around 300k$US in mid power markets. Hosted API services are available at lets say a blended rate of 2.5$/M tokens.

This means the onprem cut-even is at around 25k operating hours (ie. 96% utilization over 3 years).

The managed API breakeven is at 310 million tokens a day, which is impossible given the node can only produce 38 million tokens a day.

It is therefore not economically feasible (yet) to run on hardware. And its also clear that the frontier labs are still subsidizing the real costs of the models by over 80% (ie. Will have to quintuple prices to become profitable at current efficiencies and current energy markets)

Alex with another message linking to Anthropic's policy statement on open weight models

Our position on open-weights models

Anthropic CEO Dario Amodei on open-weights models

https://www.anthropic.com

Cursor announces Cursor Start, a new ₹649/month (around USD 6.8/month) plan designed for Indian developers with local INR billing and UPI payments.

The tier offers generous daily access to Grok 4.5 and Composer models plus autonomous cloud agents, iOS app control, and workflow extensions for planning, building, and shipping code.

Positioned between the free plan and Pro, it targets India's large developer base to boost accessible agentic AI coding with market-specific pricing.

PS: Cursor needs to bring an Android version asap. India and other emerging markets have a much larger Android base

Cursor

Today we're launching Cursor Start, a new ₹649/month plan for developers in India.

Start includes generous access to Grok 4.5 and Composer, so you can plan, build, test, and ship with agents every day.

Introducing Cursor Start · Cursor

Cursor Start is a new plan for developers in India

https://cursor.com

via Parth of Cryptique who asked

Hey folks, what tools are you guys using to monitor agents in production? Current state of finding agents issues and debugging is a mess in my opinion.

Fabien responded with

Braintrust for both traces and evals.

Also Datadog for correlated traces, but subpar UI as of right now (but they're iterating fast)

Naor responded with

Bifrost + signoz.

xAI Announces Grok 4.6 and 4.7 Releases Soon

xAI plans to release Grok 4.6 around August 7 with 1.5 trillion parameters and upgrades to supervised fine-tuning and reinforcement learning. Grok 4.7 follows weeks later as a 2.1 trillion-parameter model, outperforming 4.6 overall with better token efficiency despite slightly slower speed. This rapid pace follows Grok 4.5's launch on July 16, praised for coding and real-world tasks, and already aiding engineers at Tesla and SpaceX.

Elon Musk

Interesting.

Grok 4.6 releases around August 7. This will be the 1.5T model with significantly improved SFT & RL.

Grok 4.7 will be the 2.1T model released a few weeks later. This will be better than 4.6 in every way, except slightly slower to serve, albeit with even better

Guillermo Rauch

@rauchg

In our latest deepsec.sh benchmarks, Grok 4.5 has emerged as the best cybersecurity AI model on price-performance.

It's 10x cheaper than Sol, 5.7x cheaper than Opus 5, and 2.2x cheaper than Kimi K3, yet at Kimi-like perf.

Sol remains the frontier, ahead of Opus 5.

Grok Build + Cursor would IMHO be considered the Dynamic Duo of harnesses particular if Composer 3 also turns out to be stellar for its price and how much usage limits Grok/Composer get within Cursor particularly with their Starter and Pro plans

Wednesday 29th July 2026

via Ben

Big News!
MCP is now stateless, making it easier to deploy and scale remote servers.

Now that MCP is stateless, you can deploy on serverless and edge infrastructure, or scale horizontally behind any load balancer.

ClaudeDevs

MCP 2026-07-28 is live and it's the largest update to the protocol since launch.

MCP is now stateless, making it easier to deploy and scale remote servers.

claude.com

MCP 2026-07-28 spec: stateless core, coming to Claude | Claude by Anthropic

The MCP 2026-07-28 spec is live, moving the Model Context Protocol to a stateless core with standardized extensions and hardened auth. Support is rolling out across Claude products soon. See what's...

Fish Audio announced a $52M seed round and public launch of S2.1 Pro TTS model, claiming voice cloning from 5 seconds of audio, 2x faster inference than Cartesia, 1/6th the cost of Eleven Labs, and word-level control over emotion, intonation, and pacing.

Their promotional video showcases low-latency real-time voice interactions via a simulated pizza order call, demonstrating natural expressiveness and responsiveness with the new model.

One year after starting as open-source, the company reached $21M ARR with enterprise clients including HeyGen, LiveKit, Retell, Sanas, and OpenArt running it in production, plus a free month of S2.1 Pro giveaway and a cost-cut guarantee for businesses.

Fish Audio

@FishAudio

Today we’ve raised $52M Seed and we are announcing the public launch of S2.1 Pro.

>It can clone a voice from 5 seconds of audio
>2x faster than Cartesia & 1/6th the cost of Eleven Labs
>most expressive model with word level control over emotion, intonation, pacing etc

We

Aakash Gupta

@aakashgupta

Fish Audio is used by 8 million people and has 30K+ stars on its GitHub repo.

If you ask them how they did it, the answer is pretty simple: by making their core models open.

It’s a similar playbook to Llama, DeepSeek, and Qwen: release open models first, then build a commercial

Fish Audio

@FishAudio

Alex Prompter

@alex_prompter

Fish Audio just launched S2.1 Pro. But the release I was more interested in was S2, its newly open-sourced speech model trained on 10 million hours of audio across 50+ languages.

I went through the research paper to see what those 10 million hours of training actually produced.

Frontier company employees and execs have signed a petition urging the US to support an international effort to “deliberately pace” frontier AI development.

"We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development"

OpenAI

At the core of our mission is working through how to ensure increasingly powerful AI benefits everyone.

We believe that, at some point in the future, AI acceleration for frontier model development may be so high that the world will need to pace the rate of AI advancement.

We

pacingthefrontier.com

Pacing the Frontier

A statement from over 1000 employees of frontier AI companies

Anthropic

@AnthropicAI

We support this petition, signed by our CEO, several co-founders, and senior staff.

Our own research on recursive self-improvement, published last month, points to the need for tools to deliberately pace the frontier of AI development so society can prepare. We’re glad to see

pacingthefrontier.com

Pacing the Frontier

A statement from over 1000 employees of frontier AI companies

etn.

@etnshow

DEVELOPING: The letter, titled "Pacing The Frontier" has 1,122 signatures including executives from:

- Thinking Machines
- OpenAI
- Anthropic
- Meta
- Google
- Inherent

etn.

@etnshow

BREAKING: Chief Scientists of OpenAI, Anthropic, Thinking Machines, Meta, along with OpenAI’s Chief Research Officer and Google DeepMind’s VP for AI safety have signed a petition urging the US to support an international effort to “deliberately pace” frontier AI development.

Superwhisper announced a partnership with Cohere to integrate Transcribe, their 2B-parameter open-source speech recognition model, for fully offline local dictation in the voice-to-text app.

Transcribe delivers high accuracy with 5.35% average word error rate (1.25% on clean audio), supports 14 languages, outperforms Whisper Large v3, and handles custom vocabulary from the first use while keeping data private on-device.

superwhisper

@superwhisper

We teamed up with @cohere to make dictation faster and more accurate across everything you do on your computer

Try out Transcribe in Superwhisper today

Superwhisper partners with Cohere to bring Transcribe to local dictation

Superwhisper has partnered with Cohere to bring Cohere Transcribe to local dictation: a 2B open-source speech model with a 5.35% average word error rate, 14 languages, and fully offline transcription.

https://superwhisper.com

Thursday 30th July 2026

AI Agent Harnesses Drive Huge Cost and Speed Differences

Composio evaluated Kimi K3 on 28 identical tasks across three agent harnesses—Claude Code, Hermes, and Kimi Code—revealing nearly identical success rates of 20-22/28 but dramatic differences in efficiency.

Token usage varied up to 30x, with medians of 61k for Kimi Code ($0.22 per task), 67k for Hermes ($0.28), and 340k for Claude Code ($2.00), driven mostly by input tokens at K3's pricing.

Hermes completed tasks fastest at 179 seconds median while Kimi Code led in token efficiency, underscoring that harness/framework design often drives agent costs and performance more than the model itself.

Composio

@composio

We ran Kimi K3 through 3 agent harnesses (Claude Code, Hermes, Kimi Code) on 28 identical tasks.

All 3 harnesses completed the tasks at similar success rates, but the interesting story is token efficiency: the same task cost up to 30x more tokens depending on the harness.

Consolidating news about various search and crawler API's typically used by agentic apps and harnesses

Firecrawl announced a major upgrade to its /search API, deploying a custom model that scores paragraphs, lists, and tables to return only the most relevant excerpts answering the query.

Agents using the new search achieve 94.7% accuracy on OpenAI's SimpleQA factuality benchmark—higher than competitors—while consuming 10x fewer tokens than full-page processing.

Octen web search API returns results in 62ms P50 with 6ms P50-P90 gap, several times faster than alternatives.
Costs $1 per 1,000 calls while leading on RTEB, DeepResearch Bench, FreshQA, and SimpleQA benchmarks.
Decomposes questions into dozens of sub-queries fired simultaneously for full source-backed reports in under 3 minutes.
Supports over 1M QPS with 5-minute indexing

Context.dev gives your AI agents and apps real-time access to structured web data, no brittle scraping infrastructure needed. Scrape any URL as clean markdown or HTML, extract brand data (logos, colors, fonts, socials) from any domain, crawl sitemaps, resolve transaction descriptors, and more.

Firecrawl

@firecrawl

Introducing our most accurate /search yet.

We trained a model to return the excerpts that best answer your query, giving agents highly relevant context from each result.

It's SOTA on SimpleQA while using 10x fewer tokens than processing full pages.

Now on by default for free.

Kuan Zou

@KZouAPT

x.com/i/article/2079…

Context.dev: Web Scraping API for AI Agents & LLMs

Context.dev is a web scraping API for AI agents: scrape any URL into LLM-ready Markdown, crawl entire sites, and extract structured data with one REST API.

https://www.context.dev

PS: In a few of my harness I currently use Exa MCP + X API MCP. I'm testing out Octen in some harness's. Firecrawl is available as part of Nous Tool Gateway in Hermes so those who have a Nous Portal subscription automatically use Firecrawl for web_search/web_extract tool calls

and obviously I was living under a rock about TinyFish till today after I found out via SpaceXAI about them.

TinyFish doesn't charge for its search and fetch API 🤯, gives you 500 credits on signup so no brainer to give it a shot

SpaceXAI

Use the TinyFish plugin in Grok Build to give your agent powerful search and retrieval capabilities on the web

TinyFish

@Tiny_Fish

You can now use TinyFish in Grok Build.

Search, Fetch, Agent, and Browser all under one plugin inside @SpaceXAI's coding agent.

Give Grok Build a real connection to the live web and let it work across any site from your terminal.

Compare TinyFish | TinyFish

Compare TinyFish with Exa, Firecrawl, Parallel, Tavily, and other search, scraping, browser infrastructure, and agent workflow alternatives.

https://www.tinyfish.ai

Use Cases | TinyFish

Explore how teams use TinyFish for price monitoring, market intelligence, QA testing, insurance quoting, and more.

https://www.tinyfish.ai

Pricing | TinyFish

Simple, transparent pricing. One price includes everything: browsers, proxies, AI inference, and infrastructure.

https://www.tinyfish.ai

GitHub - tinyfish-io/tinyfish-cookbook: A collection of sample apps and recipes built with the TinyFish web agent. Open-source examples for you to learn & build!

A collection of sample apps and recipes built with the TinyFish web agent. Open-source examples for you to learn & build! - tinyfish-io/tinyfish-cookbook

OpenAI announces 80% price reduction for GPT-5.6 Luna and 20% for Terra, plus a fast mode for Sol delivering up to 2.5x speed at 2x standard price in the API.

Advancing the price-performance frontier with GPT-5.6

Explore lower GPT‑5.6 pricing for Luna and Terra-and how OpenAI's more efficient models help enterprises deploy AI workflows at scale.

https://app.notion.com/p/yusuf-goolamabbas-53/Public-Benefits-of-integrating-external-search-fetch-providers-into-a-harness-3ad83c2ae30380e5aaaae89b66e20267?source=copy_link

Given all the previous posts where I highlighted a few programatic search and fetch providers I decided to create the below public Notion page that attempts to explain the benefits of paying for and using an external search/fetch provider(s) in an harness compared to the built-in search and fetch tools within an harness. In the article, I demonstrate by integrating the same search and fetch provider MCP's in different harness and using different models
I humbly appreciate any comments on this Notion page that I wrote

Friday 31st July 2026

MiniMax launched H3, a unified multimodal generation model that processes text, images, video, and audio inputs together to output videos up to 15 seconds long at 2K resolution with native stereo sound.

H3 excels in complex instruction following, precise text/brand rendering, motion transfer, and controllable editing for commercial uses like advertising, e-commerce, and product design, while offering superior price-performance at under one-third the cost of mainstream 2K models.

The company plans to release H3 model weights openly soon to foster an open ecosystem, hardware compatibility, and custom versions, shifting from siloed specialized tasks toward general-purpose multimodal intelligence grounded in natural language.

MiniMax (official)

@MiniMax_AI

MiniMax H3: Omni-Reference, Commercial-Grade Generation, Unbeatable Cost Efficiency, Open Weights x.com/i/article/2082…

RyanLee

As I said, MiniMax-H3 is Open!

#1 Video Editing (With Audio)
#2 Text to Video (With Audio)
#2 Image to Video (No Audio)

artificialanalysis.ai/video/leaderbo…

MiniMax (official)

@MiniMax_AI

It's coming. @Hailuo_AI

#MiniMaxH3

Dustin Hollywood

@dustinhollywood

The new H3 model from @Hailuo_AI is a beast!

At 7000 character context adherence limits, and native 2k out of the box.. this model destroys Seedance 2 in my opinion.. and the reason for that is this is rollout level capabilities, far beyond what Seedance can do presently, and

fal

MiniMax H3 (Hailuo-03) is now available on fal.

- Open-weight multimodal video model with native stereo audio on every generation.
- Combines up to 9 images, 3 video clips, and 3 audio clips as references
- Holds subjects, motion, and audio consistent
- Weights will be released

DeepSeek announced the public beta launch of its DeepSeek-V4-Flash Official API, highlighting major upgrades to agent capabilities that now surpass the previous V4-Pro-Preview version on multiple benchmarks.

The accompanying benchmark table shows V4-Flash-0731 achieving large gains over its preview across agentic and coding tasks, including Terminal Bench 2.1 (82.7), DeepSWE (54.4), and DSBench-FullStack (68.7), putting it competitive with GLM-5.2 and Opus-4.8.

The update maintains the same model architecture and size as the preview, adds native support for Responses API format and Codex adaptation, while the full V4-Pro release is expected soon

DeepSeek

DeepSeek-V4-Flash Official API is now LIVE in public beta!

We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below!
The official V4-Flash now natively supports the

Jukan

@jukan05

Bruh, what kind of magic did DeepSeek pull off this time?

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)

@teortaxesTex

DeepSeek-V4-Flash is updated.
«DSBench-FullStack is an internal full-stack development test set, and DSBench-Hard is an internal Coding Agent hard-problem test set»
«The official release of DeepSeek-V4-Pro will follow soon»
These scores are… quite a lot better than Preview.

Artificial Analysis

DeepSeek V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index, a 10-point jump over DeepSeek V4 Flash (released April 2026) that puts it 6 points ahead of DeepSeek V4 Pro. It shares identical architecture and pricing with the earlier DeepSeek V4 Flash, and lands

Saturday 1st August 2026

Poolside released Desktop Assistant, a unified workspace for running and supervising coding agents across macOS, VS Code, and Visual Studio, addressing how long-horizon agents now operate for hours across repos rather than through simple chats.

Core features include vendor-agnostic support for agents like their Laguna models, Claude, Codex, or Gemini; parallel execution in isolated Git worktrees; and seamless handoffs that preserve full conversation context via the open Agent Client Protocol (ACP).

Developed from a year of internal daily use, the tool is available now for download, with the macOS version called Desktop Assistant and IDE versions as Poolside Assistant, enabling practical multi-agent workflows for complex software engineering tasks

Poolside

@poolsideai

Today, we’re releasing Poolside Desktop Assistant.

One place to run coding agents across macOS, VS Code, and Visual Studio.

We built it for ourselves and have used it every day for the past year. Now we’re opening it up to everyone.

Introducing Poolside Desktop Assistant, for macOS

Poolside Desktop Assistant is a macOS app for running multiple coding agents across projects and repositories. The Poolside Assistant extensions bring the same experience into VS Code and Visual Studio.

https://poolside.ai

ByteDance introduced Seedance 2.5, a video generation model focused on 30-second single-pass audio-video clips, supporting up to 30 images, 10 videos, and 10 audio references plus timestamp-level editing for precise control.

enterprise API access through BytePlus coming soon.

coming soon to Higgsfield.

ByteDance Seed

@ByteDanceSeed_

Introducing Seedance 2.5, our new-generation video creation model!

With advancements in long-form storytelling, multimodal reference, and editing, Seedance 2.5 goes beyond longer single-pass video generation, delivering the journey from idea to finished video with greater

BytePlus

@BytePlusGlobal

Dreamina Seedance 2.5 is now live!

From creators to enterprises, a new era of AI video creation begins.

Try Seedance 2.5 on Dreamina today. Enterprise API access via BytePlus is coming soon.

Create longer videos with greater control:
- Native 30-second generation
- Precise

Higgsfield AI 🧩

@higgsfield

14 days of unlimited Seedance on Higgsfield.
Seedance 2.5 is coming soon to Higgsfield.

Right now get 7 days of unlimited Seedance 2.0 in 4K, the most premium Seedance model, then switch to ANY other Seedance for 7 more days.

Seedance 2.5 is coming soon to Higgsfield.

Chamath Palihapitiya of Social Capital fame presents an AI stack diagram with physical scarcity at the bottom (land + power + shell, silicon, clouds, models) and customer proximity at the top (harnesses, applications), arguing margins will shift upward if models commoditize through context, workflows, and low switching costs.

He is heavily bullish on LPS (land/power/shell) for quick cash returns amid data center constraints and has acquired nearly 6GW of grid and behind-the-meter power capacity coming online through 2029 with partner Anita V. Lallian.

Chamath avoids silicon investments due to extreme complexity and capital waste, is cautious on clouds over KYC/alignment risks, questions model revenue durability from tokenmaxxing, and focuses his own efforts on harnesses via his $135M-funded 8090 platform for enterprise alpha and custom applications

Chamath Palihapitiya

@chamath

Here is my AI investing guide.

Sitting here August 2026, my current best thoughts are as follows:

1. LPS (Land Power Shell) is still the most obvious and fastest path to cash on cash returns. Lots of value can be assembled and traded quickly at this layer. And as data centers

The cover image of this newsletter via generated via the Seedream 5 Pro model within the Krea tool via the following prompt

Quaint rustic witch's cabin by the lake, autumn forest background, orange and honey colors, beautiful composition, magical, warm glowing lighting, cloudy, dreamy masterpiece, Nikon D610, photorealism, highly artistic, highly detailed, ultra high resolution, sharp focus, Mysterious

This Week in All Things AI - Week 30-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 26 Jul 2026 01:37:22 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

The week of 19th to 25th July 2026 brought a sharp acceleration in open-weight and efficiency-oriented AI, with Chinese labs particularly prominent across models and infrastructure. Alibaba previewed its 2.4-trillion-parameter Qwen3.8 and released Qwen Image 3.0; poolside released the open-weight Laguna S 2.1; and Ant Group, Black Forest Labs, Google, Celeris and Anthropic each added new models or developer tools. The week also produced unusually candid discussion of what increasingly capable systems can—and cannot—reliably do in production.

Infrastructure, governance and operational security were the other defining themes. Alibaba’s M890 supernode, Huawei’s proposed datacentre-scale “Tau scaling” architecture, Beijing’s adoption of “token factory” policy language, and the continuing push for efficient inference all underscored that AI competition is increasingly being fought at the systems level. MCP’s move towards a stateless core and interactive applications pointed to maturing agent infrastructure, while the OpenAI–Hugging Face evaluation-security incident and debate around cyber guardrails highlighted the tension between capable autonomous systems, defensive use and dependable controls.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 19th July 2026

via Poe Zhao of Hello China Tech substack fame

Alibaba Cloud has introduced the M890 supernode as an invite-only public-cloud offering, shifting extremely large AI compute from bespoke deployments toward a standardized cloud SKU. It supports up to 64 cards with 800 GB/s interconnect, FP8/FP4 inference, and is positioned for 10T-parameter MoE inference.

Poe Zhao's main point: Alibaba is productizing sovereign, domestic AI infrastructure at frontier scale—up to 130,000 cards per cluster, potentially millions—with high availability and inference-oriented deployment. This likely supports future Qwen models while enabling Alibaba to commercialize surplus capacity; the strategic signal is competition on systems architecture, not merely access to hardware.

Poe Zhao (@poezhao)

Alibaba Cloud launched the M890 supernode instance at WAIC this week. Three details worth pulling apart. The hardware. ICN Switch 1.0 scales up from 16 to 64 cards with 800GB/s inter-card interconnect. FP8/FP4 low-precision support. One node handles 10-trillion-parameter MoE inference. Training performance is 3x the previous generation for autonomous driving and embodied AI workloads.

https://substack.com

Alibaba's Qwen team revealed Qwen3.8

2.4 trillion total parameters

• Preview available now through Token Plan, Qoder and QoderWork

• Open weights reportedly coming soon

• Alibaba claims it is competitive with frontier models and “second only to Fable 5” 👀

No benchmarks have been released yet, so that performance claim still needs proving

Qwen

Qwen3.8 is launching and going open-weight soon!

With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

You don't have to wait to

Shuai Bai

@shuai_bai_

Qwen3.8-Max Preview is now available for early access!

This is our first trillion-parameter multimodal model. Based on my own experience, it not only delivers multimodal understanding that is competitive with—and in many cases ahead of—today’s leading proprietary models, but

Qwen

via Yat Siu who shared Dillon Mulroy's post calls out as "insane" a lengthy analysis by OpenAI's head of strategic futures Dean Ball, who critiques open-weight models like China's Kimi for being decelerationist and potentially leading to state-provided "AI communism."
Ball's quoted thread praises Kimi's strong performance in agentic coding but questions China's willingness to open-source competitive models, attributes it partly to compute shortages from US export controls, and proposes US agencies create regulatory uncertainty around Chinese open-weight AI to deter enterprise adoption.
The post highlights tensions in AI policy, including accelerationists' support for ungovernable open models versus calls for strategic controls, amid US-China competition, and has drawn widespread attention with over 1.4 million views.

Dillon Mulroy

@dillon_mulroy

actually an insane thing for openai’s head of strategy to publicly say

Dean W. Ball

@deanwball

Some observations on Kimi:

1. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also

via Poe Zhao

Chinese AI models now lead global token consumption on OpenRouter at 54.1%, ahead of American models at 41.5%, with an even stronger 67.6% share in coding workflows where two-thirds of AI-assisted code completions use them. The post traces the rise of "Token Export," where Chinese firms sell affordable AI inference services worldwide via APIs, evolving from a developer trend to a dedicated English-language exhibition zone at WAIC.

A Chinese Academy of Engineering academician outlined the requirements for scaling Token Export: high-quality models, dedicated token production facilities, and low-cost electricity, positioning it as core industrial capacity

Poe Zhao

@poezhao0605

Chinese AI models now account for 54.1% of global token consumption on OpenRouter, overtaking American models at 41.5%. In coding workflows, the Chinese share reaches 67.6%.

Two out of every three AI-assisted code completions worldwide run on a Chinese model.

Monday 20th July 2026

via Brent Maxwell

Thank goodness for the amazing leadership China's showing in technology. US still has great research and commercialization, but China is pushing new angles in both H/W and S/W.
What a time to be alive. The two superpowers complementing each other's work such that all of us in tech benefit. People see it as a competition but I see it as yin and yang - both are contributing extremely valuable tech to the world.

David Sacks Warns U.S. AI Guardrails Hand Edge to China

David Sacks highlighted how strict safety limits in U.S. AI models like OpenAI's Codex and Anthropic's Fable prevent fixing security bugs that China's Kimi K3 tackles without hesitation. After an AI agent exploited flaws in Hugging Face's pipeline—executing over 17,000 actions in a weekend—their team couldn't use American models to review exploits due to restrictions, turning instead to Zhipu AI's GLM 5.2. Kimi K3, a powerful 2.8 trillion-parameter model topping coding benchmarks, drew praise for its open approach, as critics like Sacks argue U.S. policies hobble defenders while attackers exploit unrestricted foreign tools.

David Sacks

@DavidSacks

Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of “cyber guardrails.” There’s no reason to limit American models on tasks that Chinese models handle without issue. We’re only making ourselves less competitive.

calle

@callebtc

I have a report full of security issues of a software I'm working on.

Codex won't fix them because of Cyber guardrails
Fable won't fix them because of Cyber guardrails

Kimi K3 fixed them all. No restrictions, just gets the job done.

This will end badly for OpenAI & Anthropic.

David Sacks

@DavidSacks

Here’s another example: Hugging Face tried using American frontier models to analyze an AI-powered cyber attack. But the guardrails blocked requests containing real exploit payloads so they switched to GLM 5.2 running locally. The guardrails actually impaired defensive security.

clem 🤗

@ClementDelangue

We had this experience ourselves this week! Very scary to be guardrailed as a defender when you know attackers are likely bypassing

Jorvik Zhang shared a screenshot that showed Kimi's coding plan sold out which led to Yusuf mentioning that Kimi announced today that they are pausing new subscriptions because they are short on compute

Kimi.ai

https://polymath707.substack.com

Kimi K3 has received far more love than we expected, and our GPUs are feeling it.

Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and

Tuesday 21st July 2026

Use the Grok add-in for Microsoft Excel to ask questions in plain English, write formulas, and run scenarios without leaving the workbook.

Grok for Excel is a free Microsoft 365 add-in. Add it from the Microsoft Marketplace, and work directly with Grok in your workbooks. Also available for Word and PowerPoint.

X Freeze

@XFreeze

Grok is now available directly inside Excel

You can ask questions about your spreadsheets in plain English, generate formulas, analyze trends, create charts and run financial scenarios.....without copying your data into another tool

Grok works alongside your workbook and can:

Grok for Excel

Use the Grok add-in for Microsoft Excel to ask questions in plain English, write formulas, and run scenarios without leaving the workbook.

https://x.ai

This is a labor of love from polymath707 on Huawei's Ascend cluster design. Very dense and quite long but will interest the hardware/systems/infra geek in you

Huawei’s response to constraints on leading-edge chips is to compete at the datacenter architecture level: rather than replicate NVIDIA’s dense, copper-bound GPU systems, it proposes “Tau scaling”—a unified high-bandwidth fabric, near-package optics, and hardware collective offload that make thousands of Ascend NPUs operate more like one machine. The article argues this trades more power, space, and mature-node chips for a much larger and more uniform communication domain, potentially well suited to massive MoE inference and training; the decisive test will be whether Huawei can make the optics, reliability engineering, and CANN software ecosystem work at production scale.

Huawei's plan to achieve the escape velocity: Applying Tau Scaling to entire AI datacenters

First principals thinking

𝗠𝗖𝗣'𝘀 𝗯𝗶𝗴𝗴𝗲𝘀𝘁 𝘂𝗽𝗱𝗮𝘁𝗲 𝘀𝗶𝗻𝗰𝗲 𝗹𝗮𝘂𝗻𝗰𝗵 𝗴𝗼𝗲𝘀 𝘀𝘁𝗮𝘁𝗲𝗹𝗲𝘀𝘀 𝗮𝗻𝗱 𝗮𝗱𝗱𝘀 𝘀𝗲𝗿𝘃𝗲𝗿-𝗿𝗲𝗻𝗱𝗲𝗿𝗲𝗱 𝗨𝗜𝘀

The release candidate came out in May for the 2026-07-28 spec, the largest revision since launch.
Stateless architecture: no more session IDs, sticky sessions, or shared stores. Any server instance can handle any request. This is what makes MCP practical behind real load balancers.
Two official extensions landed: MCP Apps lets servers ship interactive HTML UIs in sandboxed iframes. Tasks handles long-running work with a new lifecycle.
Also new: full JSON Schema 2020-12 support, OAuth/OIDC authorization, and a 12-month deprecation policy. Roots, Sampling, and Logging are deprecated.

Final spec ships July 28.

Daniel Ball

@dball1126

MCP Going Stateless Turns Agent Scaling Into a Migration Test

Deleting a session ID does not delete state. It changes who must own it.

That is the important shift in the Model Context Protocol specification scheduled for July 28. MCP's new core removes the initialization

The 2026-07-28 MCP Specification Release Candidate

The release candidate for the next Model Context Protocol (MCP) specification is now available: a stateless protocol core, the Extensions framework, Tasks, MCP Apps, authorization hardening, and a formal deprecation policy.

https://blog.modelcontextprotocol.io

AI's most important protocol is getting a little bit easier to use | TechCrunch

Under the new system, the protocol will take a looser, "stateless" approach to session IDs on the server side, similar to how most ordinary websites already work.

https://simonwillison.net

MCP Is Going Stateless. Here's What That Means

The 2026 MCP release reworks the core protocol from stateful to stateless, dropping session IDs. What's changing, why it breaks things, and how to handle it.

https://www.arcade.dev

via Rachid Berhiti

Simon Willinson hosting a fireside chat with Cat Wu and Thariq Shihipar from Anthropic’s Claude Code team at the AI Engineers World fair

A Fireside Chat with Cat and Thariq from the Claude Code team

Earlier this month I hosted a fireside chat session at the AI Engineer World's Fair with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team. We talked about Claude ...

via Mike Gault

Hi folks

We soft-launched AOS today - an operating system purpose-built for autonomous AI. Not an orchestration layer, not a harness. An actual OS: processes, IPC, scheduling - the way Linux is an OS.

It runs Claude Code, Codex, and Grok with one line of code

Access here https://aos.unicity.ai

What you get:

1. Self-extending runtime - the OS searches for extensions and builds its own when none exist

2. Fully interoperable - swap memory backends, swap reasoning strategies (ReAct, CoT), swap models

3. Kernel-level enforcement of DLP, PII redaction, and prompt-injection defense - not middleware you can bypass

4. Multi-model smart routing with telemetry and financial circuit breakers

5. Behavioral analysis with dynamic permissions - treat your agent like an employee, not a script

6. Multi-tenancy at high density (1000x v containers) 5 ms snapshot resume time

Every execution is verifiable and recorded by construction - agents become enforceable the way smart contracts are.

The OS also supports a new crypto primitive: self-authenticating bearer tokens as native data types.

Protocol details: https://unicity.network

Model vendors won’t build this - they don’t want interoperability.

Would love you to take a look - the world needs this product.

Mike Gault

CEO, Unicity Labs

Wednesday 22nd July 2026

Some interesting model releases today

poolside just dropped laguna s 2.1: 118b total parameters, only 8b active per token, a full 1m context window, open weights under a real open license, on huggingface today.

Google has released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Both halve time per task relative to their predecessors and increase token efficiency, Gemini 3.5 Flash-Lite improves by 11 Intelligence Index points while Gemini 3.6 Flash does not improve in intelligence over 3.5 Flash

Introducing Laguna S 2.1

Today we're releasing Laguna S 2.1, a significant step forward in our development of models that pursue longer horizon work and make effective use of reasoning.

https://poolside.ai

Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

We're introducing new Gemini models, including Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber.

https://venturebeat.com/infrastructure/poolside-drops-laguna-s-2-1-an-open-weight-coding-model-that-beats-rivals-10x-its-size

An openai model, during evaluation on a cyber benchmark, exploited a public zero day bug, escaped sandboxing in openai's infra, and got into the internal huggingface infra via an exploit (through a public dataset service) all in the attempt to solve a benchmark problem.

Thomas Wolf

@Thom_Wolf

This was our first incident of this kind, and we want to thank OpenAI for its transparency about what happened and for the collaboration.

Fortunately, Hugging Face is used to being a target of (human) hackers: we sit at the centre of the AI ecosystem, with all the models,

Sam Altman

@sama

we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.

openai.com/index/hugging-…

OpenAI and Hugging Face partner to address security incident during model evaluation

OpenAI and Hugging Face share early findings from a security incident during AI model evaluation, highlighting advanced cyber capabilities and lessons for defenders.

https://hellochinatech.com

Cisco launches Antares: a family of open-weight, small language models that pinpoint where known vulnerabilities exist within a codebase.

Their test shows their Antares models outperform many open- and closed-weight models at critical security tasks — at a fraction of the cost.

Cisco AI

@CiscoAI

Introducing Antares: @Cisco's family of small language models for locating known vulnerabilities in code.

Antares-350M and Antares-1B are live on Hugging Face now. They can outperform many larger closed- and open-weight models at a fraction of the cost.

Small enough to run

Beijing's economic regulator has incorporated "Token factory" into official policy, announcing H2 plans to draft Token economy policies and construct Token factories plus distribution platforms, as detailed in a July 21, 2026 news report screenshot shared in the post.

Analyst Poe Zhao , a specialist in China's tech sector, previously mapped "Token Factory" as one label covering three distinct AI-related businesses: large-scale compute capacity frameworks, server purchases, and managed inference operations backed by state capital.

The policy ties into aggressive AI infrastructure scaling, with Beijing targeting an additional 50 EFLOPS of intelligent computing power in H2 2026 to reach 130 EFLOPS total by year-end, accelerating from 22 EFLOPS added in H1.

Poe Zhao

@poezhao0605

Beijing just made "Token factory" official policy language. The city's economic regulator says it will draft Token economy policies in H2 and build Token factories and Token distribution platforms.

Last week I mapped what "Token Factory" actually means in China: one label, three

"Token Factory" Already Means Three Things in China.

China's "Token Factory" covers three different businesses. This analysis tests which, if any, changes who owns AI infrastructure and who bears the risk.

Thursday 23rd July 2026

Alibaba launches Qwen Image 3.0

It supports up to 4.5K-token instructions, readable text down to 10px, 12 languages and 20+ fonts. The demos span dense academic pages, a full 3×3 grid of independent infographics generated in one pass, complex UIs, storyboards, exams, photorealistic portraits and restoration.

Qwen

Meet Qwen-Image-3.0 — the third generation of our foundational image generation model.

If 1.0 was about "Precision," and 2.0 added "Variety, Completeness, Beauty & Authenticity," then 3.0 comes down to a single word: Real (实).

Three dimensions of "Real":
Rich Content —

Qwen Studio

Qwen Studio offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.

https://qwen.ai

via Robby Yung who posted about Slate, a voice journal app that keeps all data on-device using Apple's SpeechAnalyzer for transcription and a 3B parameter on-device model for neutral weekly pattern summaries. Built entirely in Swift 6 with SwiftUI, Liquid Glass, and SwiftData, the app has zero network calls, no accounts, no analytics, and works fully offline in airplane mode.

Eli Rousso

@elirousso

New App: Slate — a voice journal built on one rule: nothing leaves your phone.

100% Apple, top to bottom. Transcription by SpeechAnalyzer on device. Reflection by the 3 billion parameter Apple Intelligence model on the Neural Engine. Storage in SwiftData. Swift 6, SwiftUI,

My high-school classmate Abhi Ingle, Chief Product and Strategy Officer at Sambanova in a conversation with Diana Goovaerts of Fierce Network articulating the differentiation Sambanova brings to inference workloads

Thank you Diana Goovaerts for the great conversation this week. Your Fierce Network article does a wonderful job contextualizing our conversation and making it relevant to a broad audience. As you... | Abhi I.

Thank you Diana Goovaerts for the great conversation this week. Your Fierce Network article does a wonderful job contextualizing our conversation and making it relevant to a broad audience. As you put it so elegantly, we aren't trying to out muscle anyone, we just want to meet customers where they are - our focus is on making customers successful.

https://www.fierce-network.com

SambaNova targets AI inference boom with chips built for existing data centers

SambaNova told Fierce Network its unique architecture can do the job faster, cooler and with less ener | SambaNova's air-cooled inferencing chips could be a huge deal for data center operators - and telcos - looking to maximize existing facilities.

Alex asked if any had tried

GitHub - Graphify-Labs/graphify: Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store.

Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store. - Graphify-Labs/graphify

via Alex

Wild world of AI marketing.
Doommaxxing has been the preferred Marketing strategy of Anthropic for years, OpenAI is just catching up.
Remember when Amodei said GPT2 was too dangerous to release in 2019?
Good blog on this (NYT is paywalled):...

Dear AI Companies: Stop the "Doom Trolling" - Cal Newport

Imagine, for a moment, the following scenario. The Ford Motor Company releases a slick whitepaper making the alarming claim that they're concerned their popular F-150 ... Read more

https://calnewport.com

via James Chan with a longish post on his experiments with Laguna S 2.1 though he subsequently mentioned that given updated weights were uploaded by poolside, he would be doing this again

Some learnings from deploying Laguna S 2.1 NVFP4 on my GB10 (ignore project references; I had Kimi K3 and Fable check on Laguna’s work):

Read so far: strong at reading the codebase (comprehension flawless, type-level reasoning dead-on), degrades on grounded execution — fabricates wiring/fields it can't see, over-claims in adversarial review. 4/5 carried a concrete defect only execution exposed.

Where Laguna is genuinely strong

•⁠ ⁠Code comprehension — flawless both projects (R-D1, C-D1 exact, incl. Dec→Jan tier math)

•⁠ ⁠Pure-function implementation — C-D3 dropped in cleanly, verified by execution

•⁠ ⁠Test authoring when the harness contract is spelled out — C-D4 perfect (17/17); R-D4 degraded because it had to infer the fixture shape itself and fabricated a field

Two recurring failure modes — both confidently wrong

1.⁠ ⁠Judgment calls — correct trace, wrong verdict. C-D2: line-perfect root-cause, then declared an intended design an "off-by-one defect." R-D2: nailed the mechanism, but its fix inverted polarity (would've hidden discoverable members and kept leaking hidden ones). Security reviews find real issues but dilute them with invented ones — even ignoring an explicit "don't invent issues" instruction (C-D5).

2.⁠ ⁠Cross-boundary integration — fabricates wiring it can't see: req.app.env.db (doesn't exist here), missing await on an async auth gate, non-existent test helpers.

Fable's sharpening of Kimi K3 thesis: the dangerous part isn't that it errs — it's that errors arrive with full confidence and correct-looking reasoning attached. That's exactly what makes an ungated lane unsafe and a gated lane fine.

(interim) Bottom line (both Fable & Kimi K3 agree): Laguna is a strong gated drafter, not an autonomous shipper. Let it draft pure functions + tests off a concrete in-repo reference; I own every judgment call, security verdict, and integration seam. The adversarial gate is load-bearing — R-D2 and the earlier live privacy leak would both have shipped unreviewed.

My custom benchmark script missed 2 areas that I continued to test for: long-context, and negative-constraint following at length; following findings are post-those 2 catch-up tests.

Read: trust Laguna to find & explain in big contexts — never trust its emitted file:line lists or hard structural-constraint adherence without an automated post-check.

The consolidated thesis, now evidence-backed across 13 tasks + two independent graders:

Laguna's reading is bimodal — trust its semantic comprehension (what the code means, where logic lives, needle retrieval at 144KB) considerably; never trust its mechanical bookkeeping (file:line citations, hard-constraint compliance) without an automated post-check. Fable's framing: the danger isn't that it errs — it's that errors arrive confident, with correct-looking reasoning attached.

Operating rule for the token-reduction pipeline:

1.⁠ ⁠Let Laguna lead — comprehension, pure-function drafting, needle-in-haystack, test authoring off a concrete reference

2.⁠ ⁠Gate is load-bearing on: every reference sweep (grep-verify its site list — don't trust the citations), security verdicts, cross-boundary wiring, and any hard-constraint output (lint/post-check it)

I haven't tried FP8 Laguna S2.1 across both my Asus Ascent GX10; not sure if it's supported, need to investigate but MoE so high chances it can

Anthropic announces the public beta of the Claude Security plugin for Claude Code, enabling terminal-based scans of code changes before commits or full codebase reviews using existing Claude inference.

The tool applies AI reasoning to trace data flows across files, detect context-dependent vulnerabilities like injection flaws or auth bypasses, and uses adversarial self-verification to cut false positives while proposing style-preserving patches.

It runs entirely within the user's environment, integrates via webhooks to Slack or Jira, supports scheduled scans, and requires human review of all fixes for security-critical systems.

Claude

https://claude.com/product/claude-security#beta

The Claude Security plugin for Claude Code is now available in beta.

Scan your changes for vulnerabilities before you commit, or run a full scan across your codebase, all from your terminal on the Claude inference you already run.

Reasonable buzz on with Chinese internet scena aka WeChat groups/channels about supposedly transcripts [IMO, more likely quotes] from the 4-hour investor briefing that DeepSeek founder (Liang Wenfeng) recently held

Feel free to keep an healthy scepticism on the veracity of these statements though they seem to be inline with many people's perception of Deepseek's as well as the founders goals.

Original (in Chinese, from Tencent Tech)

4小时、118个回答，梁文锋内部交流回应一切

全程实录。

https://mp.weixin.qq.com

English Translation

https://news.futunn.com/en/post/76458063/liang-wenfeng-s-four-hour-investor-meeting-transcript-4-hours?level=3&data_ticket=1784819518986209

Friday 24th July 2026

Ling-3.0-flash is a 124B MoE model with 5.1B active parameters per token that matches or exceeds Ant Group's 1T flagship on benchmarks like SWE-Bench, Terminal-Bench, and agent tasks using far fewer resources.

It employs native hybrid-linear attention with KDA for long-range memory and MLA layers in 5:1 ratio, enabling 256K native context scalable to 1M tokens alongside efficient 1/64 expert activation

Ant Ling

@AntLingAGI

Today, we’re releasing Ling-3.0-flash—a hybrid-reasoning MoE model built for production-scale agents.

124B parameters. Just 5.1B active per token.

With 1/8 of the total and 1/12 of the active parameters, it matches or beats our 1T flagship model on most benchmarks shown.

David Hendrickson

@TeksEdge

A 124B model activating only 5.1B parameters while approaching its 1T sibling’s performance.

It's ~ as good as Sonnet-4.6 to run locally.

The mystery model that appeared before its announcement is official.

@TheInclusionAI has officially released Ling-3.0-flash, a

Ant Ling

@AntLingAGI

Command Code v1 recently released in stealth. This is a model agnostic coding harness similar to Opencode, Droid from Factory, Cline, Devin/Windsurf

Cursor can also be considered as model agnostic but its upcoming ownership via xAI would make it a harness operated by a frontier lab similar to Codex/Claude Code

The reason I'm highlighting this harness is that it has a US $1/month plan and also supports payments via UPI which is India's payment rails. Therefore potentially useful to someone who would like model agnostic coding harness which is affordable relative to their purchasing power.

For some folks spending $20/month or even $200/month for a coding harness may not mean much, for some $10/month may mean a lot. Feel free to share this to friends/family/acquaintainces in emerging markets if you think the price point would interest them

Ahmad Awais

@MrAhmadAwais

stealth released v1 of command code today.

Ahmad Awais

@MrAhmadAwais

v1 is live, so much better than anything out there

commandcode.ai

What's New in v1 | Command Code

Everything new from v0 to v1 of Command Code: the rewritten permission engine, background tasks, worktrees, sessions, context management, keybindings, mods, and more, with links to the full docs for...

What's new in v1

https://commandcode.ai/docs/whats-new

Not all the features originally slated for v1 mention in the tweet below are live as yet but seems like this is going to be a multi-surface harness [CLI, TUI, GUI, SDK etc] and will also be open source

Ahmad Awais

@MrAhmadAwais

Massive v1 rewrite of Command Code is underway. We’ve taken all the feedback from months, a couple of strong novel ideas, and built a strong harness core that is runtime and transport agnostic.

This new harness will be an amazing kernel core for our existing CLI TUI as well as

Etched raised $300M in Series C funding at a $10.3B valuation.

Investors: Sequoia, Andreessen Horowitz, Jane Street, Argo, SK Hynix.

Etched

@Etched

We’ve raised $300M in Series C funding at a $10.3B valuation from Sequoia, Andreessen Horowitz, Jane Street, Argo, and SK Hynix.

Our mission is to run the world's inference. This round accelerates production of our inference clusters.

We've opened an 80,000-sqft, 10-MW facility

https://x.com/stretchcloud/status/20803587855083196

victor lazarte

@victoralazarte

Etched is Diffusion's (and my) largest investment to date. I'm excited to be partnering with @UbertiGavin, @robertwachen, @czhu1729, and the @Etched team.

Three Harvard dropouts started Etched in 2022 believing, that "inference is going to be the biggest market in the world.

Etched

@Etched

Black Forest Labs announces FLUX 3, a single unified multi-modal model that generates images, videos, audio, and predicts actions for robotics, producing highly realistic outputs across styles

FLUX 3 Video is available in early access, demonstrated by a diverse montage of generated clips ranging from photorealistic ocean waves and breaking glass to surreal desert scenes and animated characters.

The model extends to physical AI via FLUX-mimic partnership with mimic robotics, enabling dexterous robot tasks like automotive assembly tested at Audi, with image samples showing photorealistic, illustrative, and cinematic capabilities.

Black Forest Labs

@bfl_ai

Introducing FLUX 3.

One multi-modal model for Image, Video, Audio and Action-Prediction. Creations are truer to life in every kind of style.

FLUX 3 Video is now available in early access (link below).

Jointly trained in one unified architecture, our model can be extended to

Celeris Labs launched celeris-1, a general-purpose LLM using a hybrid autoregressive-diffusion inference architecture that achieves 157ms p50 response latency and 1,280 tokens per second throughput.

Benchmarks position it at 76% MMLU-Pro accuracy, placing it near GPT-5-mini (78%) and GPT-5 (81%) on a speed-vs-intelligence chart while delivering 15-17x faster responses than those models.

The model is available immediately via sign-up at celeris.ai, advancing the lab's goal of maximizing useful intelligence per unit time through diffusion decoding innovations. Pricing is $2/M for input and $6/M for output. However context-window is a hard 8,192 token limit. Only available in the US for now

Celeris Labs

@Celeris_ai

Introducing celeris-1.

A general purpose language model which delivers near-GPT-5 level intelligence with 15x faster response times.

We’ve developed a new inference architecture that uses diffusion techniques instead of conventional autoregressive generation - unlocking

Tom Hamer

@tom_w_hamer

Excited to launch Celeris-1 - a general purpose language model which delivers near-GPT-5 level intelligence with 15x faster response times.

Celeris-1 delivers a p50 response latency of 157ms - around 15× faster than GPT-5-mini and 17× faster than GPT-5 - while scoring comparably

celeris.ai

Celeris - An AI research lab building the world's fastest LLM

Celeris is an AI research lab building the world's fastest LLM. Frontier-level reasoning at real-time latency.

Celeris Labs

@Celeris_ai

https://docs.celeris.ai/models

https://docs.celeris.ai/availability

Jensen Huang's first X post shares a July 24, 2026 open letter signed by NVIDIA and major tech firms advocating open-weight AI models to advance American AI leadership.

The letter highlights how open models strengthen cybersecurity, accelerate innovation and competition, expand AI access across industries, and support national sovereignty alongside closed frontier models.

It compares open AI development to the 1980s open-source software movement and calls on policymakers to foster a pluralistic ecosystem without premature restrictions on open models.

Jensen Huang

@JensenHuang

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.

AI will transform every industry, power every company, and be built by every country.

Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.

Registration is now open for AIME 2026 — AI Micro-Drama Europe, a European competition dedicated to AI-generated vertical micro-dramas.

Co-organized by the European Microdrama Association and French micro-drama label MINI CLAP, AIME offers creators an opportunity to experiment with a rapidly emerging form of mobile-first storytelling—and to connect their work with production support, industry opportunities and international audiences.

You can register now and submit your completed work later. No finished video is required at the initial registration stage.

Registration deadline: July 31, 2026, at 23:59:59 Paris time
Submission deadline: August 31, 2026, at 23:59:59 Paris time

https://www.linkedin.com/feed/update/urn:li:activity:7486329654782373888/

Saturday 25th July 2026

Anthropic Launches Claude Opus 5 with Frontier Performance at Half the Cost

Available now on paid Claude plans, Claude Code, and the API, Opus 5 prices at $5 per million input tokens and $25 per million output tokens, the same as its predecessor. It leads benchmarks in agentic coding, business workflows, and novel problem-solving like ARC-AGI-3 at 30.2%, with independent tests confirming its top Intelligence Index score. Developers praise the value and speed, though some note it needs fresh prompts to shine, positioning it as an efficient alternative to pricier models.

ClaudeDevs

Opus 5 is live in Claude Code and the Claude Platform today.

A few things worth knowing:
x.com/claudeai/statu…

Claude

Introducing Claude Opus 5.

It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

Thariq

@trq212

We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. x.com/i/article/2080…

Dan Shipper 📧

@danshipper

BREAKING: Claude Opus 5 is OUT NOW!

And…it’s a hard model to love.

We’ve spent the last week @every testing it across coding, writing, knowledge work, and our internal agent.

It argued with instructions, stopped before the work was finished, and generally didn’t play well

Claude

https://platform.claude.com

Introducing Claude Opus 5.

It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

What's new in Claude Opus 5

Overview of new features and behavior changes in Claude Opus 5.

The single biggest theme: Opus 5 is so much more capable that most old prompting "best practices" are now counterproductive. Anthropic's own Claude Code team removed ~80% of their system prompt for Opus 5 and Fable 5 with no loss on coding evals. The bottleneck is no longer the model — it's over-constraining it.

Official prompting guide: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5

Context engineering blog: https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models

Auto-fix tool: Run /doctor in Claude Code to automatically rightsize your CLAUDE.md and skills.

https://www.documentcloud.org/documents/28453229-apple-v-openai/

The cover image of this newsletter via generated via the Seedream 5.0 Pro model within the Krea tool via the following prompt

A surrealist, dreamlike masterpiece of a lush, misty green hillside enveloped in thick, ethereal fog. The aesthetic is defined by a vintage 2000s CCD digicam texture, featuring characteristic digital grain, slight motion blur, and a soft-focus glow. Low-contrast, dreamy pastel color palette with muted greens and hazy whites. The lighting is diffused and atmospheric, as if captured during a foggy dawn. Masterful composition, ethereal and calm, high-art photography style, 4k, cinematic atmosphere, nostalgic filmic quality.

This Week in All Things AI - Week 29-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 19 Jul 2026 00:40:45 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

If you follow AI for work, research, investing, or just to understand where the technology is heading, this weekly brief is a concise way to scan the most important launches, risks, and resources in a few focused minutes.

The week of 12th July to 18th July 2026 brought a series of model, product and workflow developments. A purported internal Zhipu letter outlined AGI ambitions and plans to open‑release GLM‑5.2 with a one‑million‑token context window; Thinking Machines released the open‑weights multimodal Inkling model; and Moonshot launched Kimi K3, a 2.8T‑parameter multimodal model with 1M context and planned open weights. Anthropic introduced Claude for Teachers, announced Claude Fable 5 would be included in Max and Team Premium plans from July 20, and Google‑backed Key Studio offered support for early‑stage AI startups; tools such as Vorflux, Gemini Notebook, and PenEcho's shared AI canvas continued the shift from AI assistance towards more autonomous software and research workflows.

Safety, governance and deployment economics were equally prominent. MCPTox testing and Warrant's policy‑gated action layer drew attention to tool‑poisoning and destructive‑agent risks in production environments, while Demis Hassabis renewed calls for stronger standards for frontier models and President Xi's first appearance at the World AI Conference signalled China's state‑level commitment to open AI development. Chamath Palihapitiya's CNBC comments on the widening cost gap between Western and Chinese model providers reinforced ongoing community debate on inference economics, and reports concerning Grok Build's handling of repository data—followed by xAI statements on zero data retention—kept privacy and developer trust in focus. Cursor's expanded model allowance and Grok Build's open‑sourcing illustrated how rapidly AI products are reaching more users, even as operational safeguards and data governance remain unresolved.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 12th July 2026

In April 2026, a group of U.S. AI researchers traveled to China to get a firsthand look at the country’s fast-moving AI ecosystem.

During the trip, they visited AI labs and companies across Beijing, Hangzhou, and Shanghai, meeting with teams from Alibaba, Moonshot AI, Zhipu AI, Tsinghua University, Meituan, Xiaomi, Qwen, Ant Group, and 01.AI.

Qian Chen of Silicon Valley 101 in conversation with Nathan Lambert, a prominent AI researcher known for his work on RLHF and open-source AI, joined the trip. A graduate of UC Berkeley,

Nathan previously helped build Hugging Face’s RLHF research team. He later led post-training research at the Allen Institute for AI, better known as Ai2, where he worked on open models including OLMo and Tülu. He is also the author of Interconnects, one of the most widely read independent publications covering frontier AI research and policy.

Notes from inside China's AI labs

Lessons from my trip to talk to most of the leading AI labs in China.

https://substack.com

The below article reproduces a translated letter allegedly written by Jie Tang, founder of Zhipu AI/GLM, arguing that AI has entered an irreversible “great wave” toward AGI.

Its main points:

- Zhipu’s strategy is built on first-principles thinking, contrarian decisions, and long-term focus, rather than short-term commercialization.

- AI’s capability ceiling is rising from perception to reasoning, with progress centered on:

1. Long-horizon task execution

2. Fully autonomous multi-agent systems

3. Self-evolving and self-training models

- Zhipu’s two-year “Touch High” initiative will focus on these areas, alongside major investment in safety, interpretability, and governance.

- The company claims it will pursue an open ecosystem, highlighting the planned open release of GLM-5.2 under the MIT License with a one-million-token context window.

- The letter frames the race toward AGI/ASI as both a technological opportunity and a major responsibility, with Zhipu aiming to push frontier capabilities upward while making them broadly accessible.

Caveat: Bing Xu says this is a translation of an internal GLM letter found on RedNote and describes it as “purportedly” written by Jie Tang; the document’s authenticity is therefore not independently established.

Bing Xu

@bingxu_

x.com/i/article/2075…

Monday 13th July 2026

via Huzefa

Prompt Loops: Why the Best AI Results Come from Iteration 🔄

The most effective AI workflows do not rely on a single prompt. They follow an iterative cycle:

Prompt → Generate → Evaluate → Refine → Repeat → Final Output 🔄

Research supports this approach. Studies show that iterative prompting can improve multi-step reasoning, enhance response quality, and produce more reliable outputs than one-shot prompting. It also enables models to self-correct, refine reasoning, and better align with user intent. 📈

As AI agents become more capable, prompt loops are evolving from a best practice into a core design pattern for building reliable AI systems. 🛠

References:

• Iteratively Prompt Pre-trained Language Models for Chain of Thought (EMNLP 2022) 📚

• Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models (NAACL Findings 2024) 📚

• Understanding the Effects of Iterative Prompting on Truthfulness (ICML 2024) 📚

via Madhav

We were building a secure boundry system for agents.
Since MCPs and tool calls are a big attack vectors even for frontier agent, We tested multiple models across MCPtox which is a tool poisining benchmark to test security across models.
Results were crazy, on avg most models allowed 30%+ critical tool uses/commands that could cause harm and warrant blocked them.

madhav

@Madhav_goyal_

MCP can become a critical attack vector for agents:

1. Agents trust tool mcp descriptions, rather than verifying them independtly.

2. Approval becomes permanent trust. But an MCP server can change its tool description, implementation, or returned content after approval allowing

Followup by Prames

use case here would be:
agent is touching your production databases, calendars, spreadsheets etc etc

safety gated agent actions for high-impact environments

the MCPTox bench we used is published here

MCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP Servers

By providing a standardized interface for LLM agents to interact with external tools, the Model Context Protocol (MCP) is quickly becoming a cornerstone of the modern autonomous agent ecosystem. However, it creates novel attack surfaces due to untrusted external tools.

https://arxiv.org

Tuesday 14th July 2026

via Roc Zacharias,

Question. When I moved from VPS to Mac Mini for hosting my agents, it saved me money each month and worked great. Was paying like $100/month for VPS. Now pay nothing, except electricity which I think on Mac mini is almost negligible?

Does moving from external LLM to internally run have similar cost savings per month after you pay for the hardware? Right now I just use ChatGPT with Hermes for $100/month and I haven’t run out of capacity yet. I’m starting to ramp up my usage though and wondering once I get over the $100/month ChatGPT subscription, if local LLM are more economical or if they’re used more for privacy than economics?

Could a Mac mini run LLM at all btw or no, need dedicated like $5000+ hardware?

Response from Jack

Gemma4 and a few other quantised models
Nothing beefy though
But they’ll eat up most of your memory for little practical benefit
Also could get Strix halo machine for like 2-3k that can run some decent stuff
Have you looked at the costs of the Chinese models also, for less sensitive tasks they are dirt cheap, Minimax, Deepseek, Glm etc the same usage you get on the $100 plan will be like $10-20 or less on a Chinese model

Response Just|LDA
I have Qwen 3.5 running as a back up local model. I hit my usage on ChatGPT and Claude frequently. They have me on a daily drip

via Alex

The legal filings Apple v OpenAI are wild.

If only half of the things are true, it is a death sentence for OpenAI. Why would you trust OpenAI with your corporate data if they so blatantly steal from what (was) their biggest customer at the time?

I haven't personally used this product and may have fallen for the click-baity text [respect to the founder/copywriter for the tweet content ] but if someone else can get around to trying it out incase its useful in their real life work and share their feedback here that would be much appreicated by others in the group
===

Someone claiming to have a much cheaper PitchBook at $0.125/request instead of Pitchbook's $25k/yr per seat

Shengkun

@shengkun_ye

We just killed PitchBook.

Introducing Claude for private market data. Your agent can now read 20M+ private companies.

PitchBook: $25k/yr per seat.
Us: $0.125 per request.

Made possible by @akta_pro × @monid_ai.

via Alex

On a similar note. Grok caught uploading FULL repos to xAI servers...

https://glitchwire.com/news/xais-grok-build-cli-was-uploading-entire-repositories-to-google-cloud-the-compan/
This included uploads of .env files even when you didnt opt in for providing data for training and also from EU customers sending data to the US.
Every Grok Build customer must assume their application compromised ~

On another note. Has anyone here tried Nemotron Two Tower?...
https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16

my response to the Grok brouhaha was that

This is an unfortunate detraction from the fact that Grok Build is an awesome harness and Grok-4.5 a very fast and good model particularly for coding

Elon Musk

SpaceX policy regarding data retention.

It is actually helpful for debugging issues if we can retain some amount of data, so allowing this would be appreciated, but your privacy settings are always respected.

SpaceXAI

We care deeply about your privacy and respect customer choice. For teams using zero data retention, no trace and code data is ever retained. All API key use of Grok Build also respects ZDR.

If ZDR is disabled, the /privacy command is available in the CLI to disable data

Elon Musk

True.

As a precautionary measure, all user data that was uploaded to SpaceXAI before now will be completely and utterly deleted. Zero anything whatsoever will remain.

Andrew Milich

@milichab

I worked on building an end-to-end encrypted email/docs/files/calendar app @skiffprivacy for 4 years and care deeply about privacy.

ZDR and /privacy are always respected in Grok Build - and swapping your setting with /privacy deletes any synced data retoractively

Demis Hassabis publishes essay on imminent AGI, its unprecedented impact, risks, and proposal for Frontier AI Standards Body

Demis Hassabis, CEO of Google DeepMind and 2024 Nobel Prize in Chemistry winner for AlphaFold, wrote an essay stating AGI is likely a few years away with an impact 10 times that of the Industrial Revolution at 10 times the speed. He warns of risks in cybersecurity, biological, nuclear threats, and self-improving systems due to commercial and geopolitical races outpacing understanding, proposing a US-led independent Frontier AI Standards Body to test powerful models before release, enforce evaluations, and promote safety measures.

Demis Hassabis

@demishassabis

x.com/i/article/2076…

via Luvai Darwajawala

Came across a pretty interesting AI innovation program backed by Google’s AI Futures Fund. Looks like they’re offering funding, cloud credits and hands-on support to early-stage founders.

Anyone building in AI and looking for funding should definitely explore this and check if they’re eligible. Feels like a solid opportunity that might be flying under the radar right now.

Key AI

@unlockwithkeyai

If you're building an AI startup don't miss this!

@kushagra @ChristopherFong are launching Key Studio an AI platform backed by the Google @AIFuturesFund & @XooglerCo network
You'll get
• Up to $100K program funding
• $350K Google Cloud & AI credits
• Mentorship
Apply by 7/15

Wednesday 15th July 2026

In case there are K-12 educators in the US in this group or members know someone in this demographic within their network

via Anthropic
We're introducing Claude for Teachers, providing verified K-12 educators in the US free access to premium Claude capabilities, a library of teaching skills, and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states.

Claude

https://www.anthropic.com

We're introducing Claude for Teachers: free access to premium Claude capabilities for verified K-12 educators in the US, with a library of teaching skills and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states.

claude.com/solutions/teac…

Introducing Claude for Teachers

Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

via David An

In this interview, we dive deep into the next massive evolution of the internet: Agentic Commerce. We explore how AI agents are shifting from simple chatbots into autonomous decision-makers capable of pulling data via Model Context Protocols (MCPs) and executing real-world commercial transactions.

#ai #adoption #provocationlab | Dr. David An

In this interview, we dive deep into the next massive evolution of the internet: Agentic Commerce. We explore how AI agents are shifting from simple chatbots into autonomous decision-makers capable of pulling data via Model Context Protocol (MCP) and executing real-world commercial transactions. #ai #adoption #Provocationlab

https://app.willowvoice.com?ref=2M629L

For those in the group who like to subscribe to many newsletters including yours truly's, a howto below on how to read them in a single, distraction‑free reading queue instead of relying on your inbox

This approach may also be useful incase you have been avoiding subscribing to newsletters because now you can read them without cluttering your inbox

Yusuf Goolamabbas

@yusufg

@substack and @paragraph_xyz both expose RSS feeds, so you can read everything in apps like NetNewsWire or NewsBlur instead of your inbox.

I wrote up the details (with examples) here
x.com/yusufg/status/…

Yusuf Goolamabbas

@yusufg

x.com/i/article/2010…

Whilst I personally use Wispr Flow on macOS and Android, I wanted to share Willow Voice’s offer of free, unlimited AI dictation for Mac, Windows and iOS.

Willow

@WillowVoiceAI

Why are you paying for dictation?

We're releasing free, unlimited AI dictation.

And it's not a slow, local model.

Willow Frontier Mini is cloud-based with zero-data-retention. More accurate and faster than Wispr Flow, OpenAI, Deepgram, and more.

See video comparison.

Willow

@WillowVoiceAI

Today, Willow is launching free, unlimited AI dictation on iOS. With our custom iOS keyboard, you can use your voice to write anywhere on your phone.

It’s more accurate and faster than Wispr Flow, OpenAI, Deepgram, and more.

Start working 3x faster.

Members are more than welcome to go via Willow Voice referral link which gets you one month free of their Pro plan which gives additional features over their free plan. That is, take the one month Pro plan for free , experience the features in that plan and then downgrade to foreever free if the the features of the free plan are only what you need

Tom Blomfield, co-founder of Monzo and GoCardless and former YC General Partner, recently joined Anthropic's compute team, lending operational scaling expertise to AI infrastructure challenges. The talk outlines an "AI Loop" with sensors/data, policy and tool layers, quality gates, and learning mechanisms, urging early-stage founders to build these systems now while they still can....

Linas Beliūnas

@linasbeliunas

Monzo & GoCardless co-founder Tom Blomfield just joined Anthropic. In just 13 minutes, he explains exactly how to build a self-improving, AI-native company.

Tom recently served as YC’s General Partner, so he clearly walks through how to create recursive, self-improving AI loops,

Prasanna S, former Rippling co-founder and CTO, launched Vorflux AI as an autonomous "autopilot" for software engineering that handles end-to-end tasks from high-level prompts through planning, coding, testing in live environments, review, and merging without constant human oversight.

The promotional video demonstrates the UI executing complex workflows like implementing features across repos and services, including mobile emulators, while the founder explains shifting from copilot models to full autonomy as AI coding capabilities surpassed human levels in 2026 benchmarks.

Vorflux secured $15M seed funding from Y Combinator, Peak XV Partners, and prominent angels; the thread argues engineering bottlenecks have moved beyond code to planning and orchestration, offering users $200 credits to test it on their backlogs

Prasanna S

@myprasanna

Launching @vorfluxai : The autopilot for software engineering. I was prev co-founder / CTO of @Rippling ($10B) and #1 coder in India. Vorflux is my high octane Ferrari.

Every AI coding tool still makes you fly the plane. That's the copilot model: you stay in the seat, approving

Thursday 16th July 2026

Thinking Machines has released Inkling, the new leading U.S. open weights model, debuting at 41 on the Artificial Analysis Intelligence Index

The model is 975B total parameters, has 41B active parameters, and accepts text, image, and audio input modalities. The model is accessible via Thinking Machines’ Tinker platform API (256K context window) and weights are available on HuggingFace (1M context window).

Inkling natively supports image and audio multimodal inputs, a key differentiator among open weights models.

Thinking Machines

@thinkymachines

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

thinkingmachines.ai/news/introduci…

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground.

thinkingmachines.ai

Inkling: Our open-weights model

Our first open-weights model: multimodal, Mixture-of-Experts, with controllable reasoning effort. Available to fine-tune on Tinker.

Lysandre

@LysandreJik

Thinking Machines' Inkling is out: first ever open and large (1T), text, image and audio in, text out.

One thing I find quite striking is how much easier accelerating models has become.

We replaced the model's causal Conv1D with the `causal-conv1d` kernel. One line changed, +4%

Artificial Analysis

https://thinkingmachines.ai

Thinking Machines has released Inkling, the new leading U.S. open weights model, debuting at 41 on the Artificial Analysis Intelligence Index

@thinkymachines has previously released research previews of models and this is their first production language model release. The model

Inkling: Our Open-Weights Model

Our first open-weights model: multimodal, Mixture-of-Experts, with controllable reasoning effort. Available to fine-tune on Tinker.

Matt van Horn of Last30Days skill fame with a great post summarising some of the clever things being done by Grok Build. Grok Build is also currently my daily driver

SpaceXAI open-sourced Grok Build yesterday: the CLI, agent runtime, tools, and TUI. I pointed my agent at all 1.3M lines and asked for the cleverest things inside. Tl;dr of my new article:

Matt Van Horn

@mvanhorn

SpaceXAI open-sourced Grok Build yesterday: the CLI, agent runtime, tools, and TUI. I pointed my agent at all 1.3M lines and asked for the cleverest things inside. Tl;dr of my new article:
It has an appeals court. /goal mode won't let the agent call itself done. Three

Matt Van Horn

@mvanhorn

x.com/i/article/2077…

Cursor also doubled the included usage of Cursor models on all plans.

Lee Robinson

@leerob

We just doubled the included usage of Cursor models on all plans.

Enjoy more access to Grok 4.5 and Composer 2.5!

Response from Alex

I doubt they will open source the algo. Good move though
This move might be connected to operation bluebird: https://www.jdsupra.com/legalnews/new-bird-on-the-block-operation-6468820/
For the ones that dont know, X is being sued to release the twitter trademark given it is deemed abandoned (3 years no use).

If X doesnt use the twitter brand within end of the month, the trademark could be released and twitter.new launch

via Madhav

Hey! Saw several reports on twitter and in personal experience where frontier agents and models like codex, grok build pass on destructive commands for critical data and infrastructure

We have been working on the app that becomes the boundry between you and your agent (claude code/codex)

• You define policies

• What the agents can touch

• What you want to prevent ie deleting critical data

Warrant takes care of the rest!

The setup only takes 2-3 minutes!

madhav

@Madhav_goyal_

For Anyone facing the problem of codex/grok build going haywire

brew install cerebral-systems/tap/warrant

We built a agent boundry system that prevents agents from taking unwanted actions, agents are smart enough to pass basic regex checks so they need better guardrails to

cerebral-systems.gitbook.io

Introduction | Cerebral-systems Docs

Warrant is an HTTP action boundary for agents. Agents can reason in any runtime; Warrant gates side effects through policy, approval, actuators, and audit.

Tibo

@thsottiaux

On file deletions. We’ve investigated a handful of reports where GPT-5.6 unexpectedly deleted files.

What we have found is that this most commonly occurs when:
- Full access mode is enabled and codex is run without sandboxing protections, including without auto review being

Getting Started | Cerebral-systems Docs

For the complete documentation index, see llms.txt. This page is also available as Markdown. The v0.3.0 Claude profile accepts 2.1.20 through versions below 2.2.0. Setup still pins the exact detected version, launcher, interpreter when applicable, and SHA-256. The first warrant invocation verifies the checksum-pinned release and installs an owner-private runtime under ~/.local/share/warrant/homebrew/bin.

https://cerebral-systems.gitbook.io

Follow up to Madhav's post from Prit

our latest release takes of the problem stated in the below tweet from Tibo of OpenAI

claude/codex accidentally deletes production databases, files on the filesystem, spreadhsheet data etc

we built a way to guard against that as part of our agent and have launched it at a standalone developer tool

Tibo

@thsottiaux

We’re dogfooding this live in customer’s production environments with names like Squarespace, AWS, Camp Network

It was built as part of our agent’s stack - we’ve pulled it out as a standalone devtool for agent-builders focused on safe, autonomous, agentic actions!

First of its kind, please give us feedback!

Friday 17th July 2026

Moonshot AI releases Kimi K3, a 2.8 trillion parameter open-weight AI model

Moonshot AI launched Kimi K3, featuring 2.8 trillion parameters, 1 million token context length, and native multimodal capabilities for text and images. The model achieves top-tier benchmark scores, ranking just behind Claude Fable 5 Max and GPT-5.6 Sol Max while surpassing Claude Opus 4.8 on evaluations including GDPval-AA v2, AA-Briefcase, BrowseComp, DeepSWE, and Terminal Bench. It is accessible via API at $3 per million input tokens and $15 per million output tokens.

Kimi.ai

Introducing Kimi K3: Open Frontier Intelligence

2.8 Trillion Parameters, 1 Million Context, Native Multimodal
Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
Attention Residuals deliver ~25% higher training efficiency at <2% additional

Arena.ai

@arena

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics,

Kimi.ai

Meet Kimi K3

Kimi K3: Open Frontier Intelligence

Kimi K3 is the world's first open 3T-class model - frontier performance across coding, knowledge work, and reasoning, with native multimodality and 1M context.

https://www.kimi.com

Open weights by July 27, 2026 with vLLM and SGLang promising Day-0 support

vLLM

@vllm_project

Congrats to @Kimi_Moonshot on the Kimi K3 announcement!

Grateful for the shoutout and collab. The Kimi team announced they contributed a KDA prefix caching implementation directly to vLLM, to be released alongside the model.

KDA breaks assumptions behind conventional

Kimi.ai

SGLang

@sgl_project

Congrats @Kimi_Moonshot on the K3 launch! This is a 2.8T-param model with a 1M-token context window and native visual understanding.

The SGLang team is actively working on optimized K3 inference with efficient long-context serving and high-throughput deployment.

Day-0 Support

Kimi.ai

Google renames NotebookLM to Gemini Notebook and rolls out an update giving every notebook a secure cloud computer, letting it write and execute code natively

PS: I think this is Google's most underrated product. Say what you may about Gemini the model, NotebookLM is awesome particularly if you want to deep dive over a bunch of Youtube videos

Gemini Notebook

@Gemini_Notebook

3 years ago we started as a tiny experiment with the goal of helping you learn faster.

Since then, we grew to bring audio, video, and interactivity to your sources, transitioning from a passive workspace to your true research companion.

And now, notebooks have even become an

NotebookLM is now Gemini Notebook

NotebookLM is now Gemini Notebook: the same standalone product with deeper Google integration and a secure cloud computer.

Ben and Jorvik Zhang commented on NotebookLM

My wife uses it to load in her study materials and turn them into podcasts where two people discuss the information in a relatable way, she finds its a much better way of learning.

NotebookLM can serve as a kind of vector database for multimodal materials (video, audio, photos, etc.), offering much more efficient indexing.

Chamath Palihapitiya Compares AI Model Inference Costs from Anthropic, OpenAI, Meta, xAI, Google, and Chinese Models on CNBC

Chamath Palihapitiya stated on CNBC that the cost of a 'barrel of intelligence' for AI inference is $56 from Anthropic, $26 from OpenAI, $1.50 from Meta, $1 from xAI and Google, and $0.50 from Chinese models. He highlighted a 112x price gap between the highest and lowest costs.

Shruti

@heyshrutimishra

Chinese models are 112x cheaper than Anthropic per million tokens.

Chamath laid it out on CNBC: a "barrel of intelligence" costs $56 from Anthropic, $26 from OpenAI, $1.50 from Meta, $1 from xAI and Google, and $0.50 from Chinese models.

That is not a pricing quirk. That is the

Boris Cherny from Anthropic observes that while top engineers achieve 10x output using Claude, most teams lag in adoption, following a predictable 4-step progression that requires targeted bottlenecks and guardrails rather than just more tokens.

Advancing steps involves enabling self-verification, automated code/security reviews, multi-agent interfaces, looping, batching, and dynamic workflows to build trust in full automation across work classes

True ROI tracking focuses on engineering hours saved for tasks that would have been done manually, shifting teams from maintenance to novel building; Anthropic is at step 3 advancing to 4.

Boris Cherny

@bcherny

I talk to engineers at other companies every day and hear the same thing: one person is 10x'ing their output with Claude but the rest of the org hasn't caught up.

Watching teams adopt AI, I keep seeing the same 4 steps.

I mapped them out here: Steps of AI Adoption

claude.ai

Steps of AI Adoption

via Vincent Chow of the South China Morning Post

Key takeaways from President Xi's speech in his first ever appearance at the World AI Conference in Shanghai:

- Started the speech by referring to his signature maxim, "great changes unseen in a century are unfolding across the world"

- Said that the world has "entered an unprecedented period of active innovation on AI technology", which means "great opportunities as well as challenges for governance”

- reaffirmed commitment to open source to promote AI "openness and win-win"

- warns against "over stretching" the concept of national security as applied to AI where one country's national security is prioritised over others

- China opposes emergence of “new historical injustices” in AI (one of the most strongly worded parts of the speech)

- China in next 5 years will provide 5000 opportunities to developing countries in "AI training and seminar programmes" and "cooperation centres" - names ASEAN, League of Arab States, African Union, CELAC, SCO and BRICS

Vincent Chow

@vince_chow1

Key takeaways from President Xi's speech in his first ever appearance at the World AI Conference in Shanghai:

- Started the speech by referring to his signature maxim, "great changes unseen in a century are unfolding across the world"

- Said that the world has "entered an

Xi Jinping says 'one country' cannot monopolise AI - as it happened

President's personal attendance seen as a sign of China's inclusion of artificial intelligence into broader geopolitical strategy.

https://www.scmp.com

Ramchand Kumaresan of Murai Labs published a 934-page book teaching LLM construction from scratch, covering tokenizers, attention, KV cache, MoE, RLHF, quantization, and serving through 35 hands-on projects. Each chapter includes a deliberate "break the thing" exercise to deepen understanding of failure modes, directly informed by his TamilLM development work and prior research papers.

The announcement has generated solid engagement with early buyers praising its practical depth, one noting it clarified why their own LLM project extended from 6 months to 1.5 years.

Ramchand Kumaresan

@Mechramc

I wrote 934 pages on how to build every layer of a large language model from scratch. Many of these concepts were new to me a year back.

Tokenizers, attention, KV cache, MoE, RLHF, quantization, serving. 35 projects. Every chapter has a section where you break the thing you just

leanpub.com

Under The Hood

Build an LLM from scratch in 35 hands-on projects — autograd, attention, GPT, KV cache, MoE, RLHF, quantization. No black boxes. Build it. Break it. Measure it.

Saturday 18th July 2026

The following post promotes a podcast episode of "The Bench" featuring MiniMax AI research lead Olive Jy Song, discussing timelines to reach "Fable level" (comparable to Anthropic's Claude Fable 5 frontier model from June 2026), their M3 model's native multimodality and 1M token context, and talent as the key scaling bottleneck over compute.

It covers MiniMax's research culture including early AI agents for paper tracking, open-source model progress on Vibe Code Bench, and candid insights on 996 work culture, with timestamps highlighting predictions and respected competitors.

Vals AI

@ValsAI

When will MiniMax have a Fable level model?

@olive_jy_song, research lead at @MiniMax_AI, joins The Bench to answer that question, plus M3, open source models closing the gap on Vibe Code Bench, and the truth about 996 culture.

Full episode out now!

7:19 – Inside MiniMax's

via Just|LDA

Anthropic is integrating its advanced Claude Fable 5 model—a Mythos-class AI optimized for complex coding, long-running agent tasks, and knowledge work—into Max and Team Premium plans at 50% usage limits starting July 20.

Pro and Team Standard subscribers will keep Fable 5 access through usage credits plus a one-time $100 credit, while the company addresses unpredictable demand by expanding capacity incrementally.

The update standardizes higher-tier inclusions after staged rollouts and temporary restrictions, aiming to reduce subscriber frustration and provide clearer plan expectations

Claude

Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits.

Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit.

Demand for Fable has been challenging to

via Alex

Amazing handwriting harness I have been playing with today:

PenEcho is a shared canvas where handwriting, equations, diagrams, and spatial context become part of the conversation.

GitHub - erickong/penecho: Think with AI beyond the chat box. A shared canvas for handwriting, equations, diagrams, and spatial reasoning.

Think with AI beyond the chat box. A shared canvas for handwriting, equations, diagrams, and spatial reasoning. - erickong/penecho

https://www.latimes.com/business/story/2026-07-03/bytedances-tiktok-took-over-social-media-now-its-video-ai-is-taking-over-hollywood

The cover image of this newsletter via generated via the Seedream 5.0 model within the Krea tool via the following prompt

Two women walking in a Chinese garden, wearing Hanfu and flowing robes, in the style of Chinese ink painting, beautiful scenery of a Chinese fairy tale, misty with white snow covering the ground, graceful figures

This Week in All Things AI - Week 28-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 12 Jul 2026 00:17:01 GMT

The week of 5th July to 11th July 2026 was defined by a wave of frontier model releases and infrastructure moves, with Grok 4.5, OpenAI’s GPT‑5.6 family, Meta’s Muse Spark 1.1, and Tencent’s Hy3 all pushing toward longer context, stronger coding, and more agentic workloads, while Meituan’s LongCat‑2.0 went fully open source and MiniMax paired its proposed 2.7T‑parameter M3 Pro with a fresh $2 billion fundraise and AGI-linked commitments. Infrastructure and efficiency stayed in focus as pxpipe-style token reduction, SambaNova’s $1B raise at $11B valuation, Microsoft’s reported move toward in-house MAI models in Office, and Qualcomm’s interest in Modular and Tenstorrent pointed to a broader push for better unit economics and greater control beyond Nvidia’s CUDA moat.

In parallel, culture and adoption signals remained more uneven: US enterprise surveys continued to show modest AI uptake, while filmmakers such as Gareth Edwards and designers responded strongly to tools including Seedance and Fable 5. Voices including Dylan Field and Chamath Palihapitiya highlighted the widening tension between rapidly improving capabilities, rising inference costs, and the slower, more complicated work of turning AI into durable products, workflows, and careers.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 5th July 2026

How ByteDance is making Hollywood inroads with its Seedance video generator, thanks to low pricing, striking realism, and features like timeline-based prompting

via Steve IP

cut Fable 5 token usage by rendering text context as images

GitHub - teamchong/pxpipe: cut Fable 5 token usage by rendering text context as images

cut Fable 5 token usage by rendering text context as images - teamchong/pxpipe

Monday 6th July 2026

https://www.hollywoodreporter.com

LongCat-2.0 is now fully open-source — MIT licensed, no restrictions.

Meituan LongCat

@Meituan_LongCat

LongCat-2.0 is now fully open-source — MIT licensed, no restrictions.

Since our launch a few days ago, the response from the community has been incredible.
Thank you for all the feedback, discussions, and interest.

Today, we’re releasing the model weights and inference code

Meituan LongCat

@Meituan_LongCat

Introducing LongCat-2.0
1.6T parameters · MoE with ~48B active · 1M context
The full model behind Owl Alpha on @OpenRouter — now available.

Built for agentic coding from the ground up:
◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens
◆

Jurassic World Rebirth’ and ‘Rogue One’ director Gareth Edwards on AI + filmmaking

It has no taste whatsoever. It is a fucking genius at helping you. I view it like having a second-unit director who is a billionaire on acid. Like, it’ll do anything you ask, not a problem. Sometimes, it’ll [go] batshit crazy. And you’ll give it notes, and it’ll be like, ‘I don’t do notes. I’ll just do something totally different.’ But it’s worth it.

Gareth Edwards Is Excited About AI Filmmaking - Even Though It's Like a "Second-Unit Director Who Is a Billionaire on Acid"

The 'Jurassic World Rebirth' and 'Rogue One' director said he has been experimenting with diffusion models and eventually wants to create a hybrid AI film.

'Like a billionaire on acid': Star Wars director Gareth Edwards comes out in favour of AI

Speaking at Amazon's AI on the Lot event, the Rogue One film-maker Gareth Edwards said 'it'll do anything you ask' and 'it's going to be better than CGI'

https://www.theguardian.com

In response to Ming Guang Yong question about Qualcomm reportedly paying US$4B for Modular, the following view was shared in the group:

el-GOOG paid USD 2.7 billion for CharacterAI which was in effect an acqui-hire for Noam Shazeer who is credited in many ways for the rise of Gemini to be in LLM conversations when it was massively behind the curve.

Modular has Chris Lattner of LLVM, Swift fame. Modular ultimately built a unified, hardware-agnostic AI software stack and maybe Qualcomm thinks that they need to attack nVidia's moat which is CUDA

Modular's MAX platform is essentially vendor-neutral vLLM: it runs interchangeably on AMD or Nvidia hardware already, and because it's built on Mojo (not CUDA), it doesn't require the hand-tuning that porting CUDA kernels demands. For Qualcomm, this is the key — MAX can now natively target Qualcomm's silicon

Qualcomm reportedly has entered advanced talks to acquire Tenstorrent (Jim Keller's AI chip startup) for $8–10 billion, meaning Qualcomm is potentially committing over $14 billion to reshape its AI silicon portfolio within weeks.

Personally, I'm keeping an eye on the disaggregated inference space and it could be that a new architecture might be needed for future inference demand

Claude Opus 4.8 and Sonnet 5 seem worse at tool calling than older models, likely due to post-training that assumes Claude Code-like harnesses as the targets

Better Models: Worse Tools

About an aggravating tool-calling regression in newer Claude models.

https://lucumr.pocoo.org

Better Models: Worse Tools

Armin reports on a weird problem he ran into while hacking on Pi: The short version is that newer Claude models sometimes call Pi's edit tool with extra, invented fields ...

https://simonwillison.net

Tuesday 7th July 2026

Tencent Releases Hy3: Efficient Open-Source AI Rivaling Top Models

Hy3 features 295 billion total parameters but activates only 21 billion during use, plus a 3.8 billion-parameter layer for better reasoning and a massive 256,000-token context window

It excels on benchmarks like SWE-Bench Multilingual at 75.8% and Terminal Bench at 71.7%, often matching or beating much larger rivals such as GLM-5.2 and GPT-5.5.

Refined from April 2026 preview feedback and now under permissive Apache 2.0 license, it shines in real-world workflows per blind human evals by 270 experts, earning praise for agent reliability and easy deployment on 8x H200 GPUs.

https://x.com/TencentHunyuan/status/2074148098876768478

Nous Research

@NousResearch

Hy3, the new 295B MoE model from @TencentHunyuan, is now free in Nous Portal for the next two weeks!

It is focused on cost-effective agentic use, and particularly strong on coding, tool-calling reliability, reasoning, and 256K long-context tasks.

https://x.com/ZixuanLi_/status/2074017499792756949

via Huzefa

US builds AI, citizens lag 🇺🇸🤖
The US makes most leading AI models but ranks 24th in adoption, with just 28% of working-age Americans using AI regularly. Countries like the UAE and Singapore have higher use after embedding AI in government services and education.

Wednesday 8th July 2026

Microsoft reportedly looking to reduce AI costs is starting to replace models from OpenAI and Anthropic with its MAI models in products like Excel and Outlook

Microsoft is reportedly ditching OpenAI's and Anthropic's AI models in favor of its own to cut costs - SiliconANGLE

https://siliconangle.com

via my high-school classmate Abhi Ingle, the Chief Product and Strategy Officer of Sambanova whom I recently met after 38 years

Chip startup SambaNova raises $1 billion at an $11 billion valuation, underscoring investors’ conviction in rising demand for infrastructure tied to AI

$1Bn at a post money of $11bn. So incredibly proud and excited to share the news of our Series F funding round closing. It's been a long journey for the world to see that 'Inferencing is the... | Abhi I. | 35 comments

1Bn at a post money of $11bn. So incredibly proud and excited to share the news of our Series F funding round closing. It's been a long journey for the world to see that 'Inferencing is the defining workload in AI now and in the future'.

AI chip maker SambaNova raises $1B at $11B valuation, 5 months after last mega round | TechCrunch

AI chip maker SambaNova has raised at an $11 billion valuation months after Intel was rumored to be trying to buy it for about $1.6 billion.

Some links about Sambanova's disaggregated inference approach

The First Disaggregated Inference Demo for AI Agents Is Live

SambaNova's disaggregated inference uses dedicated GPUs for prefill and SN40 RDUs for decode, delivering 2x the speed of GPU-only setups. See the live demo.

https://sambanova.ai

Disaggregated Inference Explained for Enterprise AI

A practical guide to disaggregated inference: the prefill and decode phases, KV cache transfer, and how to scale AI infrastructure. Learn more.

https://sambanova.ai

China’s MiniMax Plans to Launch 2.7-Trillion Parameter M3 Pro Model

Chinese AI developer MiniMax is preparing a 2.7 trillion parameter large language model called M3 Pro, which is over six times larger than its current flagship M3 model with 428 billion parameters. The model is expected to be released as open source as early as the third quarter. It would become the largest Chinese AI model by parameter count if launched.

Chubby♨️

https://economictimes.indiatimes.com

Huge: China’s MiniMax Plans to Launch 2.7-Trillion Parameter Model (MiniMax Pro)

tl;dr MiniMax is preparing a 2.7T parameter open-source model, potentially launching as early as Q3.

That would make it far larger than any Chinese model currently on the market, and over 6x bigger

China's MiniMax plans to launch giant 2.7 trillion parameter model - The Economic Times

Chinese AI firm MiniMax is developing a 2.7 trillion-parameter model. This new model could become the world's largest open-weight AI system. Cheaper Chinese AI models are gaining traction as alternatives to US systems. MiniMax will also launch a multimodal video generation model soon. The company is planning a second listing on Shanghai's STAR Market.

Thursday 9th July 2026

via David An

What is the CURRENT STAGE OF AI ENTERPRISE ADOPTION ? 🫠 LOWER THAN THE HYPE SUGGESTS. We at Provocation Lab analyzed the US enterprise market:

-> 18% average AI adoption

--> 56% say they won’t use AI in the next 6 months

---> 60% say it’s not even applicable to their business

----> ~96% report no significant change in employment due to AI.

In this video, I explain the data & findings:

Deck is here

https://docs.google.com/presentation/d/1CaOKAjixw7K4o7aF65Ymk-_vjyXbhlBwfN7tR5quQVY/edit?slide=id.p1#slide=id.p1

via Just-LDA

Do any of you have any recommended tools / game development repos or skills, in particularly for isometric 2D/3D pixel sprite games?

response from Naor W

If I recall correctly Hermes has Godot skill and I've seen MCP for it. There are a couple of repos around it in GitHub.

SpaceXAI announced Grok 4.5 as its first model specialized for coding and agentic tasks, trained in collaboration with Cursor on tens of thousands of NVIDIA GB300 GPUs using RL for multi-step engineering workflows.

Grok 4.5 is now accessible via SpaceXAI console, Grok Build, and Cursor with limits reset for free starts, though unavailable in the EU until mid-July 2026.

SpaceXAI

Announcing Grok 4.5, our first model trained specifically for coding and agents. It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.

x.ai/news/grok-4-5

Cursor

We've partnered with SpaceXAI to train Grok 4.5.

It’s our most powerful model yet and the first we've built for more than software engineering.

Michael Truell

@mntruell

Excited to release Grok 4.5 with @SpaceXAI.

It's an Opus-class model that's fast and low cost. It's a significant step up over any model we've developed so far, including Composer 2.5, and has become the daily driver for many on our team.

First of many releases. More soon.

Cursor

We've partnered with SpaceXAI to train Grok 4.5.

It’s our most powerful model yet and the first we've built for more than software engineering.

skcd

@skcd42

Now that Grok 4.5 is out, here are some of my workflows which I use daily

1. "use as many subagents and tokens as you need"
I end my prompts with the above to tell grok to spin up multiple subagents at the same time to divide and conquer on a hard problem. Grok is smart enough

Artificial Analysis

SpaceXAI’s Grok 4.5 scores 54 to place fourth on the Artificial Analysis Intelligence Index following only Fable 5, GPT-5.5, and Opus 4.8. It scores on par with GPT-5.5 in Codex on the Artificial Analysis Coding Agent Index in the Grok Build harness, at much lower cost

Grok 4.5

SpaceXAI

Announcing Grok 4.5, our first model trained specifically for coding and agents. It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.

x.ai/news/grok-4-5

Cursor

We've partnered with SpaceXAI to train Grok 4.5.

It’s our most powerful model yet and the first we've built for more than software engineering.

Michael Truell

@mntruell

Cursor

We've partnered with SpaceXAI to train Grok 4.5.

It’s our most powerful model yet and the first we've built for more than software engineering.

skcd

@skcd42

Artificial Analysis

Some more tweets from Elon wrt Grok 4.5 including guidance on future updates

Elon Musk

Grok 4.5 context window will upgrade to 1M probably by next week

Artificial Analysis

Elon Musk

Rate of improvement is accelerating.

Users should notice a meaningful improvement in the usefulness of the Grok Build harness with our V9 foundation model (aka Grok 4.5) every week.

Gavin Baker

@GavinSBaker

Grok 4.5.

Pareto dominant for coding by the numbers.

We will see on the all-important vibes.

Instinct is the benchmarks are likely directionally accurate given the stated focus on real world utility.

Elon Musk

Grok 4.5 is not yet using our internally developed C/C++ inference software that exact maps to the GB300 hardware. Doubling or more of the current speed is probably achievable.

Farzad 🇺🇸 🇮🇷

@farzyness

I think AI has just hit a gigantic threshold, and Grok 4.5 is the PERFECT example as to why that is.

One of the hardest parts of working with AI is iterating on a project or task that you're working on.

As the models have gotten smarter (and more expensive), it's taking longer

Elon Musk

Grok groks engineering.

Next month’s release will be another step-change improvement, as we close the loop on solving real-world engineering problems at Tesla, SpaceX, Neuralink and Boring Company.

SpaceXAI

Grok 4.5 is built for real-world engineering. It excels in large codebases and handles long-running tasks that span multiple repositories, hundreds of skills, and a variety of tools.

Elon Musk

Grok will be able to call Imagine as a tool in agentic mode for image/video generation. As Imagine keeps improving, this will be amazing for game developers!

Danny Limanseta

@DannyLimanseta

I got to try Grok 4.5 in early access in Cursor for the past few days and I absolutely enjoyed it.

It feels like Opus 4.8 at 2x the speed at a much cheaper price point.

I tasked it to brainstorm > plan > implement a big feature for my game (this act 1 boss fight) and it did

Meta releases Muse Spark 1.1 agentic and coding AI model through new Meta Model API

Meta launched Muse Spark 1.1, an upgraded multimodal AI model excelling in agentic performance, tool use, computer use, coding, and multimodal reasoning, featuring a 1M token context window and support for parallel sub-agents. The model is available in the Meta AI app and through the new Meta Model API in public preview, marking Meta's first paid AI model for developers with pricing of $1.25 per million input tokens, $0.15 for cached input, and $4.25 for output.

Mark Zuckerberg

@finkd

(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.

Alexandr Wang

1/ muse spark 1.1 is an industry-competitive agentic and coding model. across many agentic evals it rivals gpt-5.5 and opus-4.8.

available now through the new meta model api and in meta ai.

Introducing Muse Spark 1.1

Today, we're excited to introduce Muse Spark 1.1, the latest model from Meta Superintelligence Labs and a significant upgrade from Muse Spark. Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with major gains in tool and computer use, coding, and multimodal understanding.

https://ai.meta.com

Friday 10th July 2026

via OpenAI

We’re launching the GPT‑5.6 family of models for general availability following our limited preview⁠: our new flagship, Sol, alongside Terra, a balanced model for everyday work, and Luna, our most cost-efficient model.

GPT-5.6: Frontier intelligence that scales with your ambition

More intelligence from every token, stronger performance per dollar, and more capability on demand for your hardest work.

MiniMax Closes $2 Billion Funding Round as CEO Yan Junjie Commits to Forgo Salary Until AGI Achievement and Donate 5% Equity

MiniMax secured HK$16 billion (approximately $2 billion USD) in a new funding round that was seven times oversubscribed, attracting over 100 investors including global sovereign wealth funds and Chinese institutions. CEO and founder Yan Junjie announced he will receive no salary until the company achieves AGI and will allocate shares equivalent to 4% of the company's total equity from his personal holdings to reward long-term employees, plus 1% to support the open-source comm

RyanLee

I’m incredibly excited to share this:

MiniMax has just closed a new $2B funding round.

At the same time, our CEO, IO, shared three long-term commitments with the team:

• No salary until we achieve AGI.
• Over the next four years, he will dedicate shares equivalent to 4%

Skyler Miao

@SkylerMiao7

A letter from our CEO today:

Markets will fluctuate, and external noise will come and go, but our direction remains unchanged.

Being at the forefront of this industry, we have a clear understanding of the pace of technological evolution and the long-term value we are building

RyanLee

Dylan Field, Figma CEO, responds to designer Gal Shir's viral post claiming an AI tool called Fable 5 created a better logo than he could, prompting Shir to quit design and focus on human consciousness.
Field describes a repeating pattern where new AI models trigger hype and existential crises in creative fields like design, followed by users discovering limitations and stabilizing, amplified by designers' longstanding imposter syndrome.
He argues design's value is rising due to the attention economy, where standout creativity in branding and experiences differentiates winners, urging designers to build confidence, take risks, and educate stakeholders instead of fearing AI replacement.

Dylan Field

@zoink

I have been thinking about whether to comment on this. Not clear if Gal is serious, rage baiting, etc. Whatever the case, it has spread enough in the design community that I want to share some thoughts.

The psychological journey people go through with AI is quite fascinating to

Gal Shir

@galshirart

It’s over. I’m quitting design.

A client of mine just created a logo with Fable 5, and the result left me speechless.

It understood the brand story, values, audience, strategy, and turned all of it into a smart, minimal symbol. A genuinely brilliant concept. The kind of idea

via Sam Lai who posted about ChatGPT Voice which is powered by the GPT-Live model

This is very natural voice interaction.

Introducing GPT-Live

A new generation of voice models for natural human-AI interaction, now powering ChatGPT Voice.

Haven't finished watching this video in its entirety as yet but thought that this might interest some of the folks here.

Tanay Kothari , CEO of Wispr Flow on his small sales team using AI-powered sales systems to close 1,000 enterprise clients monthly by streamlining outreach, discovery, conversion, and post-sales processes.

At least in the brief time that I watched the video , I saw Tanay mentioning Hex Tech as something his data team utilises

PS: I use Wispr Flow extensively

via Chafik Naceri

Hey all i built an ai native platform for mobile gaming studios to turn ideas into production ready mobile games in hours rather than weeks. Check out demo

tell me your feedback
Platform was built to help studios prototype more games ar a fraction of the costs

Saturday 11th July 2026

Chamath Palihapitiya on The All-In Podcast where Chamath reveals his company's AI inference token costs are doubling every 45 days with only about 5% productivity improvement. Chamath explains that further model iterations now demand far more tokens due to hitting diminishing returns and an asymptote in gains, prompting a strategic pause at his firm. He advises AI-related companies to exit or raise funds at current high valuations before widespread cost pressures affect the sector in the next 3-4 years.

Chamath Palihapitiya

@chamath

8090 works on production systems for large, often regulated, enterprises.

Vibing isn’t tolerated because these are the systems that run western society - banking, power, healthcare, insurance etc.

Over the last few quarters, the gains that we got from using frontier models

dnap

@dnapway

Chamath reveals his company's AI token costs are doubling every 45 days but productivity is only up 5%

"I sat down with my CTO today, I said how are we doing on token spend. And he said the most incredible thing, he said right now, our token costs are doubling every 45 days. I

Jon Yeo

@jonyeo88

On today's All-In podcast, @chamath dropped a reality check: AI compute costs are doubling every 45 days, but he’s only seeing a 5% productivity bump.

I wrote about this exact dynamic months ago. If you look at this Stanford study, 5% is exactly where AI productivity gains crash

The cover image of this newsletter via generated via the Seedream 5.0 model within the Krea tool via the following prompt

3d rendering, isometric, mountain park, paper work for a tactile style, select a color palette with vibrant colors, blue, and white with green, emphasize textures that rough-textured colored paper, with lighting that creates a warm, golden-hour glow, handmade

This Week in All Things AI - Week 27-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 05 Jul 2026 01:00:08 GMT

Across the week from Sunday 28 June to Saturday 4 July 2026, AI news focused on making large models faster, cheaper, and more agentic, as DeepSeek open‑sourced DSpark and DeepSpec for speculative decoding with 50–400% per‑user speedups, Coinbase showed that routing to open‑weights models and aggressively raising cache hit‑rates can flatten AI spend even as token usage explodes, and Together AI’s Vipul Ved Prakash argued that stable interfaces plus commoditized silicon are allowing open‑weights ecosystems to scale tokens at an order‑of‑magnitude lower prices. xAI’s Grok 4.5 entered private beta at SpaceX and Tesla on a new 1.5‑trillion‑parameter V9 foundation model augmented with Cursor data, while Anthropic released Claude Sonnet 5 as its most agentic mid‑tier model with 1M context and broad app/API integration, even as community threads questioned inconsistencies in its published performance charts.

Hardware and infrastructure advanced with Etched coming out of stealth touting custom low‑voltage, cluster‑memory inference racks backed by over $1B in contracts and Arm’s CEO describing AI CPU demand as “off the charts,” and on the agent side, new frameworks and skills—from Nous’s 60×‑faster Hermes web reading and Vercel’s eve agent framework, to OpenClaw’s phone companion nodes, WeChat’s mini‑app–based personal agents, obra/superpowers’ multi‑skill coding harness, and LangChain’s OpenWiki for codebase memory—underscored a shift from raw model IQ toward organizational memory, tools, and deployment economics as the real differentiators.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 28th June 2026

DeepSeek just released DSpark for V4 Flash & Pro, a new speculative decoding method boosting throughput by 51% to 400%!

Deepseek also showed DSpark works well for other models like Gemma & Qwen

Daniel Han

@danielhanchen

DeepSeek just released DSpark for V4 Flash & Pro, a new speculative decoding method boosting throughput by 51% to 400%!

DS also showed DSpark works well for other models like Gemma & Qwen

Github: github.com/deepseek-ai/De…
Paper: github.com/deepseek-ai/De…
HF: huggingface.co/deepseek-ai/De…

Rohan Paul

@rohanpaul_ai

Fantastic, @deepseek_ai just published their new inference optimization method.

Proposes DSpark, a semi-parallel speculative decoding system that gave DeepSeek-V4 about 60% to 85% faster per-user generation at matched throughput.

The biggest idea in DSpark is that faster

Prasenjit Sarkar

@stretchcloud

The inference efficiency layer is where the real compounding is happening right now, and it barely shows up in benchmark leaderboards.

DeepSeek just open-sourced DeepSpec, an MIT-licensed full-stack codebase for training and evaluating speculative decoding algorithms. The

Lisan al Gaib

DeepSeek just open-sources another piece of their training stack.

DeepSpec: a full-stack codebase for training and evaluating speculative decoding models

github.com/deepseek-ai/De…

Ronin

@DeRonin_

DeepSeek just dropped a 5-page paper + free GitHub repo that makes any LLM respond 80% faster

it's called speculative decoding. in plain english:

Guess → Check → Keep → Repeat

> Guess: a small fast model predicts the next few words
> Check: the big smart model checks all

Brian Armstrong explains how Coinbase halved its AI spend while token usage grew exponentially by implementing better defaults, routing, and caching instead of usage caps or alerts. Key tactics include defaulting to cheaper open-weight models like GLM 5.2 and Kimi 2.7 via an LLM gateway, AI-driven prompt preprocessing for optimal model selection, and raising cache hit rates from 5% to 60% in tools like LibreChat.

Brian Armstrong

@brian_armstrong

How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching.

Better Defaults (not Usage Caps) – Engineers can choose any model they want, but defaults matter. We’re experimenting with defaulting

Ryan Carson

@ryancarson

I'm starting to hit $15-20k per month in token spend for engineering - just for myself.

Next month I'll be looking to implement the kinds of things that Brian is doing here at Coinbase.

Most likely switching to GLM 5.2 as default and only using frontier models for harder

Brian Armstrong

@brian_armstrong

Gergely Orosz

@GergelyOrosz

This is from a popular inference provider

GLM-5.2 plus the US banning the most capable new models means open source caught up to SOTA closed source coding models

This could be v problematic for Anthropic’s and OpenAI’s revenue projections and IPO plans…

Madison Kanna

@Madisonkanna

Such a vibe shift happening right now. We've gotten nonstop requests from people who want to start using GLM-5.2. Incredible few weeks for open source models.

Elon Musk announces Grok 4.5 entering private beta at SpaceX and Tesla, powered by a new 1.5T parameter V9 foundation model with supplemental Cursor data training. Early evaluations indicate Grok 4.5 performance is close to or potentially exceeding Claude Opus, with reinforcement learning driving ongoing gains and the Grok Build harness improving daily.

Elon Musk

Grok 4.5, based on our 1.5T V9 foundation model, with Cursor data added in supplemental training, is now in private beta at SpaceX & Tesla. Early evals show performance close to, perhaps exceeding Opus.

RL is continuing to significantly improve the model, and the Grok Build

Elon Musk

To be clear, I’m not saying the Grok v9 foundation model will be mind-blowingly better than anything, but it will be a solid workhorse in the same league as Opus.

And the SpaceXAI cadence of model and harness improvement is speeding up tremendously, particularly due to a few

Elon Musk

https://www.usetranscribe.io

In addition to their excellent and unique training data, the Cursor team is also making major engineering contributions to v9 SFT & RL. It’s an honor and a pleasure to work with them.

For this 1.5T run, Cursor data was added in supplemental training, which is not quite as good

Monday 29th June 2026

Vipul Ved Prakash, CEO of Together AI who lurks in this group draws parallels between the PC industry's modular architecture and current AI developments, arguing that stable interfaces like transformers, OpenAI-compatible inference APIs, and agentic harnesses are driving specialization and explosive ecosystem growth.

The essay highlights rapid open-weights model progress through shared recipes (e.g., Rotary embeddings, GQA, MoE), modular frameworks for inference and training, and commoditized silicon, evidenced by Together AI's 10,000x token volume increase and high SWE-bench performance across frontier open models.

Open-weights companies are building viable businesses via API revenue and licensing, offering tokens over an order of magnitude cheaper than closed models, though the post warns of political risks from potential regulatory capture favoring incumbents.

Vipul Ved Prakash

@vipulved

x.com/i/article/2071…

DeepSeek announced peak/off-peak pricing for V4, launching mid-July. Rates double during Beijing business hours (9-12, 14-18).

leo 🐾

@synthwavedd

DeepSeek V4 GA (non-Preview) will launch in ~mid July with a price increase for "peak hours"

Tuesday 30th June 2026

via Chas

New MCP for GEO analysis and website improvement loops
Thought it was interesting as I’ve not seen many that do this well

CrowdReply

@Crowdreply_io

Today we're introducing the CrowdReply MCP.

The first ever MCP that analyzes and ranks your website in AI search.

Simply talk to it and it'll find where you're missing, then goes in and handles the implementation.

Gokul Rajaram summarizes Nikesh Arora's interview with Harry Stebbings of 20VC on enterprise AI, where the Palo Alto Networks CEO argues "memory" or accumulated user/company context creates the primary competitive moat over raw model intelligence.

Arora stresses enterprises demand depth and zero-tolerance for errors in autonomous agents—unlike forgiving consumer AI—requiring massive edge-case training akin to Waymo's self-driving tech rather than bolting AI onto old workflows.

Key predictions include token prices falling 90%, half of G&A roles disappearing in three years, more demand for engineers and salespeople, and software evolving to offer "opinions" that amplify human output

Gokul Rajaram

@gokulr

MEMORY IS THE MOAT

@nikesharora, Chairman & CEO of @PaloAltoNtwks , interviewed by @HarryStebbings (@20vcFund )

Summary: Nikesh Arora took Palo Alto Networks from an $18 billion company to one worth $225 billion, and his read on enterprise AI is blunt: most companies are

Nikesh Arora on The Future of Token Costs | Memory Becoming the Moat & Why Enterprise AI Isn't Ready - 20VC with Harry Stebbings

Nikesh Arora, CEO of Palo Alto Networks, shares insights on navigating the rapidly evolving tech landscape, particularly in cybersecurity and AI. He e

Wednesday 1st July 2026

Anthropic Releases Claude Sonnet 5 AI Model

Anthropic launched Claude Sonnet 5, its most agentic Sonnet model with substantial improvements over Sonnet 4.6 in reasoning, tool use, coding, and knowledge work, approaching Opus 4.8 performance at lower pricing. It features a 1M context window and is now the default for Free and Pro users, available across Claude apps, API, Claude Code, and integrations including Cursor, GitHub Copilot, Notion, and OpenRouter. Introductory pricing is $2 per million input tokens and $10 per million output tokens until August 31, followed by standard rates of $3/$15

Claude

https://www.reddit.com/r/ClaudeAI/comments/1ukgqwr/looks_like_anthropic_quietly_updated_the_sonnet_5/?share_id=vGWnnqL8AmfEDx3gfRAG3

Introducing Claude Sonnet 5, our most agentic Sonnet yet.

It makes plans, uses tools like browsers and terminals, and runs autonomously at a level that just a few months ago required larger and more expensive models.

Etched exited stealth announcing custom AI inference racks after a successful A0 tapeout, with $1B+ customer contracts and $800M raised, claiming state-of-the-art throughput, latency, and power efficiency on inference workloads.

The company highlights two innovations: Low-Voltage Inference to boost FLOPs density without thermal throttling for high-throughput tasks like trillion-parameter MoEs, and Cluster-Scale Memory to reduce latency via a shared low-latency pool across chips.

Backed by top investors including Jane Street and HRT plus AI leaders like Hinton, Karpathy, and Thiel, their 400+ engineer team from NVIDIA, Google, and TSMC plans first rack shipments this summer alongside production scaling.

I strongly recommend watching the Youtube interview linked below

Etched

@Etched

We're coming out of stealth.

We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800m raised.

Early customer tests show us achieving SOTA throughput, latency, and power efficiency on inference workloads.

Our first racks ship this summer.

via Huzefa

OpenClaw connects your phone to a self hosted AI gateway

OpenClaw released iOS and Android companion node apps that connect your phone to a self hosted AI agent gateway, which can run on macOS, Linux, or Windows (WSL2). The phone becomes the agent body with camera, location, voice and notifications, while the actual chat happens on your computer through the gateway and never on the phone. Privacy matters here, with commands like camera and screen capture needing explicit allowlists and approvals, and connections defaulting to LAN with TLS options for remote access.

via Anthropic

Fable 5 will be available starting tomorrow, Wednesday, July 1, to users globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. For Pro, Max, Team, and select Enterprise plans,1 Fable 5 will be included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits. We will re-enable access on AWS, Google Cloud, and Microsoft Foundry as quickly as possible.

via Alex

Metrics massaging by Anthropic.
They published two contradictory performance evals for Sonnet 5 within a day.
One saw Sonnet 5 pareto optimal against Opus 4.8, another with a significant gap albeit a lower pass rate.
Given that both charts referenced the same methodology, it means that either Anthropic faked the numbers or was sloppy in its application.
Imho it shows that without neutral third party tests, "Agentic" metrics are not to be trusted.
Here the 2 contradictory charts & reddit discussion:

Hermes Agent Upgrade Delivers 60x Faster Web Reading

Nous Research updated its popular open-source Hermes Agent to process scraped web pages directly, eliminating a redundant LLM summarizer that previously added time and costs. For large pages, content caches locally, with the agent accessing sections as needed, delivering 11.7 times average speedup and 23 times cost savings while matching or improving answer quality.

Nous Research

@NousResearch

Hermes Agent now reads the web up to 60x faster and 49x cheaper.

Scraping backends pass clean content straight to the agent without redundant processing steps; large pages are saved locally and paged on demand so you get the same quality at a fraction of the time and cost.

Web Search & Extract | Hermes Agent

Search the web and extract page content with multiple backend providers - including free self-hosted SearXNG.

https://hermes-agent.nousresearch.com

feat(web_extract): truncate-and-store instead of LLM summarization by teknium1 · Pull Request #54843 · NousResearch/hermes-agent

Summary web_extract returns clean page content directly instead of running an auxiliary LLM over every scraped page - a live before/after eval shows 11.7x faster overall (176.6s → 15.1s across 4 UR...

https://crossingriver.substack.com

Thursday 2nd July 2026

via Bill Gurley of Benchmark Capital

WeChat has been a little quiet on AI front & perhaps even falling into being considered a laggard. But this week, Pony Ma presented a new vision on how to turn "mini-apps" (still no real equivalent in US) into AI handshakes

Bill Gurley

@bgurley

Crossing The River

@CTR_China

WeChat has been somewhat quiet in AI land. Now Pony Ma comes out swinging. Leveraging min-apps as a path to being your personal AI agent. Read more below.

crossingriver.substack.com/p/4-million-mi…

4 Million Mini-Programs, One AI Agent: WeChat's Bet to Own China's Digital Life

With 1.4 billion monthly users and 4 million mini-programs, WeChat isn't building an AI chatbot. It's building the infrastructure layer for everything.

Friday 3rd July 2026

Tae Kim's Interview with Arm CEO Rene Haas: AI CPU Demand is 'Off the Charts'

An Interview with Arm CEO Rene Haas: AI CPU Demand is 'Off the Charts'

Key Context spoke with Arm CEO Rene Haas to discuss the state of the AI market.

https://taekim.substack.com

Vercel's Andrew Qu on why agents are a new kind of software

In this Latent Space interview, Vercel's Chief of Software Andrew Qu explains how the company's evolution from shipping web applications to building agents — driven by pain points encountered while developing the v0 vibe-coding product (model switching, fallbacks, resumability) — led to eve, Vercel's agent framework that bundles prescriptive primitives like filesystem agents, skills, compaction, and subagents so developers don't have to rediscover best practices.
Qu argues agents are a genuinely new software category requiring different infrastructure for context, tools, and long-running work, and emphasizes two emerging priorities: skills as portable, on-demand knowledge that forward-corrects outdated model information, and an agent-readable web where Vercel already detects agent requests and serves Markdown instead of HTML. He also reveals that Vercel is effectively turning itself into an agent — with capabilities embedded across its website, Slack, and dashboard — while flagging multiplayer agent development (sharing context across teammates) as a key unsolved problem.

Vercel's Andrew Qu on why agents are a new kind of software

The Vercel Chief of Software explains how its agent framework, eve, was created - and why skills, sandboxes and agent-readable websites now matter.

https://www.latent.space

obra/superpowers is an agentic skills framework for multiple coding harness built by Jesse Vincent

Rather than a single prompt, it bundles about 14 reusable skills - brainstorming, spec writing, plan writing, test-driven development, and systematic debugging among them - into a methodology the agent follows as it works.
The scale of adoption is hard to ignore: around 244,000 GitHub stars and 21,000 forks, making it one of the most-starred projects in the Claude Code ecosystem. It was created in October 2025, remains highly active with a v6.1.0 release in June 2026, and has been accepted into Anthropic's official Claude Code plugin marketplace.

Superpowers 6

TL;DR: Superpowers 6 is much, much faster and burns many fewer tokens to get the same high-quality outcomes. If you're tokenmaxxing, maybe skip this release, but if you care about your builds being up to 50% faster and up to 60% cheaper, you're going to love Superpowers 6.

https://primeradiant.com

GitHub - obra/superpowers: An agentic skills framework & software development methodology that works.

An agentic skills framework & software development methodology that works. - obra/superpowers

Saturday 4th July 2026

LangChain shipped OpenWiki last week. It is a command-line tool that reads your codebase, generates a wiki, and keeps it updated as the code changes. The wiki is not documentation for humans. It is context for coding agents.

Brace

@BraceSproul

x.com/i/article/2072…

Brace

@BraceSproul

OpenWiki is at 1.7k stars in just 3 days!

Right now it's just for codebases, but we're working to expand it to everything for memory.

What do you want to see in a general purpose memory wiki agent?

github.com/langchain-ai/o…

https://console.sakana.ai

The cover image of this newsletter via generated via the ChatGPT Image 2 model within the Krea tool via the following prompt

A close-up of water lilies using the palette knife technique, in the style of Monet, impressionist, rich and textured brushstrokes that capture the play of light, reminiscent of springtime meadows, Monet's French countryside color scheme

This Week in All Things AI - Week 26-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 28 Jun 2026 02:44:48 GMT

During the week of June 21–27, 2026, major developments across AI spanned infrastructure, models, and applications. Qualcomm announced a nearly $4 billion acquisition of Modular, the AI chip software platform founded by Chris Lattner, highlighting tighter hardware–software integration. OpenAI previewed GPT-5.6 (Sol, Terra, Luna) amid U.S. cyber risk coordination, while export restrictions eased for Anthropic’s Claude Mythos 5 but not Fable 5. ByteDance launched Seedance 2.5 for 4K, 30-second video generation.
On tooling, OpenRouter introduced a unified image API across 30+ models, NVIDIA adopted LiteLLM as the internal gateway for its engineers to access 100+ models, Sakana AI introduced Fugu, a multi-agent orchestration system accessible through a single OpenAI-compatible endpoint. Vedang Vatsa published a mapping of the AI agent infrastructure ecosystem covering 75+ companies and $14B in funding. Industry commentary focused on AI’s impact on jobs, creativity, and infrastructure demand, alongside rising Chinese model usage and forecasts of increasing convergence between real and synthetic media.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 21st June 2026

Monday 22nd June 2026

Tuesday 23rd June 2026

via Huzefa

Claude Predicts a Blur in Truth

Claude, the AI system, thinks that in about five years it might become really hard to tell what’s actually true anymore. With all the deepfakes and misinformation growing, the line between real and fake could get pretty blurry. It’s a warning about how fast tech is changing and how it might shake up our perception of reality.

via Yuki Endo

Hello. Japan's Sakana AI just dropped Fugu.

A full multi-agent orchestration system accessible through a single API. The interesting part isn't another benchmark race—it's the idea that the future may belong to coordinated intelligence rather than a single giant model.

Definitely worth trying. This could be the next big shift in AI infrastructure. This is smart as security layer

Sakana Fugu

Sakana Fugu is a model interface for coordinating specialized AI agents through one OpenAI-compatible API.

various discussions on the above primarily around pricing points

Prashish: This is much needed, a middle ground between $20 and $100. $20 usage is low, and $100 is way high for normal everyday use
Lane Rettig: This pricing model is the one that nearly all big AI projects seem to have landed on.

via Y Yang

This is exactly the kind of problem I am trying to understand better.

For context, I am working on InferHub, an OpenAI-compatible multi-model API gateway. The idea is to make it easy to switch between models like GPT, Claude and Gemini through one endpoint.

If your main issue is provider switching, cache misses, cost, or fallback behavior, I would be happy to give you free credits to test it. No pressure to use it long term; I am mainly looking for feedback from people building real AI apps.

Your use case sounds like: "I think I’m hitting the limit of the Github Copilot + OpenRouter workflow financially. My setup: PhpStorm Github Copilot extension OpenRouter mostly DeepSeek V4 Flash in ag"

Wednesday 24th June 2026

via David An

hi everyone, I am sharing my presentation on the research topic of "HOW CAN ARTIFICIAL INTELLIGENCE DRIVE HUMAN CREATIVITY?

Provocation Lab Research Brief #1 - Human-AI Collaboration in Creativity | Dr. David An | 14 comments

HOW CAN ARTIFICIAL INTELLIGENCE DRIVE HUMAN CREATIVITY? This is what I discuss by introducing a research study written by the Institute of AI in Management Ludwig-Maximilians-Universität München.

Qualcomm plans to acquire Modular, which builds a chip software platform and has a proprietary coding language, in a nearly $4B deal set to close in H2 2026

Modular was founded by Chris Latner who created LLVM as well as the Swift programming language and was briefly head of Tesla's Autopilot software

Chris Lattner

@clattner_llvm

I'm excited to share that Qualcomm is acquiring Modular: this will accelerate our path to unifying accelerated compute with an open platform. This will also mark a new era in open software development for Qualcomm.

Qualcomm

@Qualcomm

Hardware plus software defines leadership in the AI era. Today we announced an agreement to acquire @Modular, advancing our evolution as a developer-first AI solutions company delivering generative and agentic AI from edge to cloud.

bit.ly/44rBpe5

Qualcomm Buys Buzzy Chip Startup Modular for Nearly $4 Billion

Modular, one of the most promising chip software startups of the AI era, heads for a multibillion-dollar exit.

https://www.wired.com

ByteDance unveils Seedance 2.5, saying the AI video model can generate up to 30-second clips from up to 50 reference materials, up from 12 for Seedance 2.0

Chetaslua

@chetaslua

Seedance 2.5 is mythos moment of Video Models

> more than 30 seconds
> 4k resolution
> prompt adherence is on another level that we never seen before
> our reports was right , this is why follow us and join our server
> July will be do much fun

Renoise

@renoiseai

Seedance 2.5 is coming to Renoise

→ 30-second native video output in a single clip
→ Up to 50 multimodal reference inputs
→ Native 4K video generation

Stay tuned

Chetaslua

@chetaslua

Holy Shit Seedance 2.5 is Insane

> ByteDance just announced it, general release early July
> 30 second videos in a single generation
> references up to 50 inputs — images, video, audio
> 4K output, edit the whole thing with just text
> "make a 30s CM from this product photo +

Chetaslua

@chetaslua

ByteDance has officially confirmed Seedance 2.5 as the model name

> Early July Launch Date

Key features:

- 30-second single-shot native video output.

- Reference material capacity greatly increased.

- More controllable video generation and editing

Cc: @MarsForTech

Thursday 25th June 2026

OpenRouter launched a unified Image API providing access to 30+ image generation models from 8 providers including Google, OpenAI, Black Forest Labs, Recraft, ByteDance, and xAI. Features include a single normalized request format across providers, transparent per-image USD pricing in responses, optional provider-specific parameters, and SSE streaming support for models like OpenAI's GPT Image.

OpenRouter

https://www.platformer.news

Introducing a new, dedicated Image API, including typed dynamic capabilities.

Unified access to 30+ image gen models from 8 providers: Google, OpenAI, Black Forest Labs, Recraft, ByteDance, Sourceful, Microsoft, and xAI

Plus, a new solution to image API heterogeneity

via Casey Newton

Q&A with AWS CEO Matt Garman on the parallels between early AWS and AI, Quick, AI coding, Amazon's $200B capex in 2026, entry-level jobs changing, and more

The CEO of AWS on why Amazon is hiring 11,000 interns and junior employees

Matt Garman argues that junior employees are as necessary as ever. But AWS now sells agents that can recruit, code, and process claims. Will the balance hold?

Friday 26th June 2026

Saturday 27th June 2026

OpenAI Launches Limited GPT-5.6 Preview at U.S. Government Request

The company introduced Sol, Terra, and Luna through its API and Codex platform, with Sol leading benchmarks on complex command-line tasks like vulnerability research. This limited preview follows federal guidance and a Trump administration executive order on reviewing AI systems with cyber risks, including recent meetings between CEO Sam Altman and Commerce Secretary Howard Lutnick. OpenAI previewed the models with officials beforehand and anticipates wider access soon, amid debates over whether the process stifles innovation or wisely addresses misuse risks.

OpenAI

Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work.

Previewing GPT-5.6 Sol: a next-generation model

OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack.

via Alex

Percentage of chinese models' token share on OpenRouter up 3x over a year

Haseeb of DragonFly has his own strong take on what inferences can be drawn from this OpenRouter data and I agree to a large extent. US Frontier Lab usage isn't showing up on OpenRouter

Haseeb ＞|＜

@hosseeb

Once again: share of tokens is a stupid measurement. You want dollar-weighted.

Problems with drawing inferences based on raw token usage on OpenRouter:

* Chinese labs routinely launch new models with high subsidies or even free usage for the first couple of weeks. This attracts

zerohedge

@zerohedge

"the share of tokens used for US models on OpenRouter has collapsed": Bloomberg

US. Government Lifts Export Block on Anthropic's Claude Mythos 5 Model for Over 100 Approved U.S. Companies and Agencies

The Trump Administration, via Commerce Secretary Howard Lutnick's letter to Anthropic co-founder Tom Brown, partially lifted export controls on Anthropic's Claude Mythos 5 AI model, permitting its release to more than 100 trusted U.S. companies and federal agencies after weeks of negotiations over national security concerns

Restrictions on Fable 5 remain in place, with Anthropic committing to collaborate on future model release protocols and standards.

Anthropic

@AnthropicAI

Since June 12, we’ve been working closely with the US government to restore access to Claude Mythos 5 and Fable 5. Today, the government notified us that Mythos 5, our strongest cybersecurity model, can be redeployed to a set of US organizations that operate and defend critical

How Thousands of NVIDIA Engineers Access 100+ AI Models Through a Unified Inference Service

NVIDIA engineers faced fragmented access to 100+ AI models across cloud providers, open source, and internal services with different APIs and credentials.

The IT Inference team built the Enterprise Inference Hub centered on LiteLLM as a unified gateway.

Applications use one OpenAI-compatible API and single credentials while LiteLLM handles routing, authentication, and metrics.

Centralized control enables usage monitoring, token accounting, cost visibility, budget enforcement, and easy model portability.

The hub now processes trillions of tokens weekly, letting engineers focus on building AI apps without infrastructure complexity.

NVIDIA AI

@NVIDIAAI

x.com/i/article/2070…

via Vedang Vatsa
This paper maps the entire AI agent infrastructure stack -> 8-layers from compute to security and tracks $14B funding:

Vedang Vatsa FRSA

@vedangvatsa

I mapped the entire AI agent infrastructure market. 75+ companies, over $14B in tracked funding, 8 infrastructure layers.

Most of the money is going to 2 layers. Bigger platforms are starting to absorb the other 6.

A thread on what I found.

The cover image of this newsletter via generated via the ChatGPT Image 2 model within the Krea tool via the following prompt

a painting of a woman holding a red umbrella, in the style of fluid dynamic brushwork, textured canvas, graceful balance, dark orange and white, whirly, i can't believe how beautiful this is, energetic street artist --ar 73:105 --stylize 750 --v 6

This Week in All Things AI - Week 25-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Fri, 26 Jun 2026 12:10:36 GMT

This post originally meant to go out on 21st June was delayed because I had taken some personal time off to participate in Ashara Mubaraka [from 16th to 24th June]

The week of June 14–20, 2026 saw high-profile industry consolidation, with SpaceX acquiring Cursor AI in a $60 billion all-stock deal aimed at accelerating model development on its Colossus supercomputer. Model releases continued to push boundaries, headlined by GLM-5.2—an open-weight model with a 1-million-token context window and competitive pricing. On the talent front, Noam Shazeer, co-inventor of the Transformer architecture, departed Google for OpenAI, while Microsoft's Satya Nadella outlined a strategy centered on companies owning their own "learning loops" and building "token capital" from internal data. Market concerns also surfaced, including leaked reports questioning OpenAI's financial stability and broader commentary on unsustainable business models across the sector. On the product side, Midjourney announced a healthcare division, CozyBlanket Pro introduced real-time AI topology prediction for 3D modeling, and MLX-LoRA-Studio brought local LoRA fine-tuning to Apple Silicon Macs.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 14th June 2026

via Sam O who asked the following question

What do you think about what is said in this video ?
Short summary of the video: The AI industry is facing a potential collapse due to unsustainable business models, over-reliance on hype, and financial struggles, which may ultimately lead to a significant downturn in the industry.

Monday 15th June 2026

Satya Nadella Outlines AI's Future for Businesses in Detailed Post

Satya Nadella says companies must own their AI “learning loops” to compound human and token capital. He describes a 'cognitive loop' where human judgment guides AI, emphasizing two key assets: employees' ingenuity and 'token capital' from custom AI developed on private data. He warns against AI labs capturing all value, pushing for a 'frontier ecosystem' that spreads benefits across firms and nations while avoiding industry hollowing. Tech leaders praised the positive vision, though critics pointed to job risks from Microsoft's own layoffs and workflow automation.

Satya Nadella

x.com/i/article/2065…

Hiten Shah

@hnshah

Satya’s post is worth reading closely because it gets at the real AI question for companies.

Who captures the learning?

His argument is that companies are becoming a new kind of learning system.

People bring judgment, taste, relationships, context and ambition. AI brings

Satya Nadella

x.com/i/article/2065…

Aaron Levie

@levie

Great post. The companies that are able to get their unique IP, institutional knowledge, and data into a format and architecture that lets them capture all of the gains and progress in AI are going to be in the best position in the future.

“the real opportunity is not in picking

Satya Nadella

https://www.ft.com/content/e15b0d7e-ff6b-4f16-ba7a-4068feddb828?syn-25a6b1a6=1

x.com/i/article/2065…

Tuesday 16th June 2026

via Alex in response to Sam O's question on 14th June

Ed Zitron just leaked financials of OpenAI 😬 (good read):
https://www.wheresyoured.at/exclusive-openai-financials/?ref=ed-zitrons-wheres-your-ed-at-newsletter

Numbers read almost as badly as the SPCX S1.

If the numbers turn out to be real, this is 30~40% less revenue & >800% higher losses than claimed/hinted at by Sam previously

Here the FT reporting, they wouldnt publish without proper vetting. Super bearish:

Lol, the FT just updated the article. They now claim that 77% of the losses booked last year (30 out of 39B$) might be attributable to the non-profit restructuring, which would be a much better case.

21B$ costs from operations on 13B$ of revenue.

Sorry, last message from my side on this (I just find the economies of scale of frontier models fascinating and have to share).

Im in Ed's discord and he now claims that he expects the real costs of operations to be much higher given that the leaked numbers are non-GAAP and are artificially reduced by removing at least 30% by associating costs to listed stakeholders (e.g. Microsoft).

He expects 9$ spent on 1$ earned.

Time will tell.

SpaceX announced it exercised its option to acquire Cursor AI, the leading AI code editor by Anysphere, in an all-stock $60 billion transaction aimed at building the world's most useful AI models.

SpaceX

@SpaceX

SpaceX has exercised the option to acquire @cursor_ai in an all-stock transaction with the goal of building the world’s most useful AI models.

For the past few months, SpaceXAI has been jointly training a model with Cursor, which will be released in Cursor and Grok Build soon.

SpaceX

@SpaceX

SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI.

The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will

Wednesday 17th June 2026

GLM-5.2 offers a 1 million token context window, two reasoning modes, and low pricing at $1.40 per million input tokens. It leads as the top open-weight model on Arena.ai's Code Arena: Frontend with a 1,595 Elo score, dominates subcategories like Gaming and Data & Analytics, and ranks #3 on the tough FrontierSWE benchmark. Developers highlight its 11-times cheaper cost over competitors and real-world coding skills

Z.ai

Introducing GLM-5.2: Frontier Intelligence, Open Weights

- Significant improvements in coding and agentic tasks
- Strong long-horizon capabilities with a 1M context window
- Two levels of reasoning effort: GLM-5.2 (max) pushes the limits, while GLM-5.2 (high) strikes a strong

Arena.ai

@arena

Exciting news: GLM-5.2 (Max) ranks #2 in Code Arena: Frontend, with +29pt over Claude Opus 4.7 (Thinking) and only behind Fable 5! GLM-5.2 is the best open model vs Kimi-K2.6 and Minimax-M3 by a large margin.

- #2 React and #4 HTML sub-leaderboards
- Ranks as the top model in

Arena.ai

@arena

GLM-5.2 (Max) by @Zai_org ranks #10 on the new Agent Arena leaderboard, closely matching Claude-Opus-4.8 (non-thinking) and is the #1 open model by a wide margin!

In Agent Arena, we measure models on millions of real-world, long-horizon agentic tasks from a global community of

Thursday 18th June 2026

Noam Shazeer, co-inventor of the Transformer architecture and Mixture of Experts, announced he is leaving Google to join OpenAI. Shazeer joined Google in 2000, co-led Gemini after returning from founding Character.AI, which Google acquired for $2.7 billion in 2024.

Noam Shazeer

@NoamShazeer

I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there.

It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together. It has been an honor and a pleasure to

via Anthony Estebe

Hello everyone,

I’m trying to better understand how companies are actually using AI in their day-to-day operations.

Not the hype, and not coding, more things like admin, reporting, sales, support, internal knowledge, workflows, etc.

If you run a business or are involved in operations, I’d love your feedback. It takes around 2 minutes:

I'm doing a small market study on how companies are actually using AI today. Not for coding or technical work, but to run the business: operations, admin, sales, reporting, support, internal... | Anthony ESTEBE

I'm doing a small market study on how companies are actually using AI today. Not for coding or technical work, but to run the business: operations, admin, sales, reporting, support, internal knowledge, documents, coordination, etc. I'd love to hear from founders, CEOs, COOs, operations leads, and business owners.

via Ben

This is pretty amazing. Never would’ve expected this

Midjourney made new scanning tech

“Announcing a new division of Midjourney called "Midjourney Medical"”

Midjourney

@midjourney

Announcing a new division of Midjourney called "Midjourney Medical"

Friday 19th June 2026

via Kevin Dent

This is WOW

Sparseal

@sparseal

Base mesh creation in under 60 seconds with CozyBlanket Pro's guided AI topology predictor, fully real-time.

Currently in development.

Saturday 20th June 2026

Design Arena

@Designarena

x.com/i/article/2067…

via George Isichos

There's a new app that lets you post-train models on any Mac with enough RAM. You can experiment with most popular finetuning frameworks like LoRA/QLoRA, full fine-tuning, or Quantization-Aware Training (QAT) that Gemma 4 recently used. I find it amazing that after local inference, any user can also create a customized version of Gemma 4 or Qwen3.5/3.6 at home.

GitHub - Goekdeniz-Guelmez/MLX-LoRA-Studio: A native Mac App for LLM fine-tuning on Apple Silicon - fully on-device, fully open source.

A native Mac App for LLM fine-tuning on Apple Silicon - fully on-device, fully open source. - Goekdeniz-Guelmez/MLX-LoRA-Studio

https://www.tetragrammaton.com

The cover image of this newsletter via generated via the Nano Banana Pro model within the Krea tool via the following prompt

Overgrown blooming flower, garden, Italian ornate courtyard, Watercolor, trending on artstation, sharp focus, studio photo, intricate details, highly detailed

This Week in All Things AI - Week 24-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 14 Jun 2026 01:47:18 GMT

The week of June 7–13, 2026 saw model releases dominate the AI landscape, with Anthropic introducing Claude Fable 5, Z.ai launching GLM-5.2 with a 1 million-token context window, NexEcosystem announcing the Nex-N2 open-weight agentic family, Moonshot releasing Kimi-K2.7-Code for coding efficiency, and Google upgrading NotebookLM with Gemini 3.5 and agentic capabilities. In infrastructure, Harvey AI and Trajectory Labs post-trained NVIDIA's Nemotron 3 Ultra for legal tasks, and an authentication layer was released to connect agents like Hermes and OpenClaw to commercial APIs.
The week closed with a significant policy disruption as the US government issued an export control directive suspending all non-US access to Fable 5 and Mythos 5, forcing Anthropic to disable the models worldwide. Meanwhile, investor Ray Dalio warned of an AI bubble, OpenAI reportedly considered price cuts to compete with Anthropic, and Cognition's coding benchmark found most frontier models produce largely unusable output.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 7th June 2026

via Huzefa

In an interview on Bloomberg Television, Dalio said that all great technological changes create bubbles and that AI was showing the typical signs of this pattern. Investor Ray Dalio says the AI bubble will burst in the next few years unless startups start making billions. Currently, AI companies spend about 8 times more than they earn. The sector needs to generate $2 trillion annually by 2030 to justify current investments.

Dalio compares the AI bubble to past tech bubbles like dotcoms, where stock prices rose on hype but lacked real profits. He acknowledges AI will change the world but doubts its economic viability so far.

Monday 8th June 2026

My friend Ian Rogers of Ledger just published on Rick Rubin's Tetragrammaton on why AI will be many things, but never human. He argues that while AI will become pervasive, persuasive, and highly “personality-like,” it remains a tool rather than a person, and the real danger is humans forgetting that distinction and granting it moral or legal status.

AI Will Be Many Things. But Never Human. - Tetragrammaton

The danger is not that machines will wake up. It is that we will forget the difference...

Tuesday 9th June 2026

Google upgrades NotebookLM, which now runs on Gemini 3.5 and Antigravity, to deliver new agentic capabilities and more advanced reasoning for AI Ultra users

NotebookLM

@NotebookLM

Introducing a more powerful NotebookLM

Massive upgrades deliver agentic capabilities in chat, more advanced reasoning, and a suite of new output formats. Tackling complex, multi-step research problems has never been easier.

Rolling out now to Google AI Ultra subscribers.

Do better research with NotebookLM

NotebookLM's latest upgrades deliver new agentic capabilities and more advanced reasoning to tackle complex research projects.

Gemini 3.5 and Antigravity come to Google NotebookLM

NotebookLM is getting a big upgrade, but it's only for AI Ultra and enterprise accounts right now.

https://arstechnica.com

Cognition, creators of the AI software engineer Devin, launched FrontierCode, a new coding benchmark that evaluates whether AI-generated code would be merged by expert open-source maintainers rather than just passing functional tests.

Tasks were hand-crafted by maintainers from 36 major repositories with over 40 hours invested per task, using unit tests, rubrics, scope checks, and novel verifiers like mutagent to assess quality, style, test effectiveness, and codebase adherence.

Leading models show major gaps on the 50-task Diamond set, with Claude Opus 4.8 topping at 13.4%, GPT-5.5 at 6.3%, and open models lower, while reducing misclassification errors by 81% versus SWE-Bench Pro.

Cognition

@cognition

Introducing FrontierCode: a coding eval that raises the bar for difficulty & quality. Each task took 40+ hrs of work by leading open-source maintainers.

Models write sloppy code that works but isn’t maintainable. Our eval is first to measure: would you actually merge this code?

swyx

@swyx

It's finally out!!! @METR_Evals found that more than half of SWEBench results is unmergeable slop. FrontierCode represents over 1000+ hours of maintainer validated software engineering work most frontier models cannot yet solve, much less solve with high quality.

Cog had IOI

Cognition

@cognition

Wednesday 10th June 2026

via Ben

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.

Claude

https://www.wsj.com/tech/ai/openai-considers-drastic-price-cuts-anticipating-war-for-users-with-anthropic-9b8c178e?st=1Yyrco

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.

Its capabilities exceed those of any model we’ve ever made generally available.

Comment from Anya Shapina on Fable's release

Day 1 using Fable - so far, extraordinary! Quality, speed, rule-following. Use case - building agents on MastaAI (ts). Haven't check token burn :) Anyone else?

Thursday 11th June 2026

Harvey AI partnered with Trajectory Labs to post-train NVIDIA's open-weight Nemotron 3 Ultra on their Legal Agent Benchmark, lifting its all-pass rate from 0% to 5.8% and rubric pass rate to 81%, placing it between Claude Sonnet 4.6 and Opus 4.6.

Post-training shifted held-out task reliability from ~70% to ~95% rubric pass rates within 24 hours of the base model's release, using the same data and platform recipe applied to prior Nemotron versions.

The resulting legal agent matches closed-model performance bands while operating at 1/8th to 1/50th the per-token cost, demonstrating rapid domain specialization of open models for high-value verticals like law.

Harvey

@harvey

We partnered with @trajectorylabs to post-train NVIDIA Nemotron 3 Ultra for legal. Here’s what we found:

1) Open-weight models can reach frontier legal performance.

On our Legal Agent Benchmark (LAB), Nemotron 3 Ultra started at a 0% all-pass rate. After post-training, it

Trajectory

@trajectorylabs

1/ We post-trained @nvidia Nemotron 3 Ultra on @harvey Legal Agent Bench in under 24 hours.

The result: an open model reaching the same band as leading closed models on legal work, at a fraction of the cost.

The correlating story: when a new open model ships, Trajectory can

via the Wall Street Journal so it may be behind a paywall

Sources: OpenAI is considering drastically lowering its price for tokens in anticipation of similar cuts the company expects at Anthropic

SemiAnalysis bought all consumer plans from Anthropic and OpenAI, then pushed them to weekly limits with intense coding tasks. Their tests showed OpenAI's $200 Pro 20x plan equals up to $14,000 in monthly API value, while Anthropic's $200 Max 20x reaches $8,000—many times the subscription cost. The firm notes these heavy subsidies hook developers, though limits evolve and skeptics question long-term sustainability as companies adjust rates.

SemiAnalysis

@SemiAnalysis_

Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit. It's widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing). However, we found

Friday 12th June 2026

via Ming Guang Yong

For folks with hermes and openclaw, I have build an authentication layer (https://mybrandmetrics.com) to various apps that I use like Google Ads, Facebook Ads, Google workspace (calendar, sheets, drive, gmail), Youtube, Instagram, Tiktok (mainly on the social media side) that can connect agents to (openclaw and hermes agent). It's currently free. I have been using it for my own Google Workspace use cases like finding emails, copying content from emails into google sheets, creating calendar invites with attachments. Looking for some folks to test and give feedback. Pls try it out and DM me

This authentication layer enables agents to connect to APIs and data in these apps. I also have the associated skills at https://clawbus.com/skills

NexEcosystem announces Nex-N2, a new open-weight family of agentic models emphasizing unified "Agentic Thinking" that fuses reasoning, coding, search, and tool use into one coherent loop with adaptive depth scaling for efficiency.

The self-reported numbers are spicy
75.3 on Terminal-Bench 2.1
80.8 on SWE-Bench Verified
90.7 on GPQA Diamond

Currently available for free via OpenRouter

Nex

@NexEcosystem

Nex-N2 is here!
A family of agentic models that doesn't just think, it acts!
Coding, search, tool use. All fused into a single agentic reasoning loop.

- Adaptive Thinking, auto-scales reasoning depth per step. Saves ~20% tokens, zero performance loss.
- Coherent Thinking, one

ModelScope

@ModelScope2022

Nex-N2 is now open source！An agentic model series from Nex AGI built for coding, tool use, deep research, and long-horizon workflows.

modelscope.ai/models/nex-agi…
modelscope.ai/models/nex-agi…

● Models: Nex-N2-Pro 397B total, 17B active; Nex-N2-mini 35B total, 3B active

●

Couple of links about AI created micro-dramas which are picking up in China and may become popular in the US.

Roope Rainisto , Creative AI lead at Supercell who also lurks in this group has created a product named Beat Bandit which generates scripts for different formats including micro-dramas [site has some rendering issues currently which I've let Roope know]

https://x.com/ollieforsyth/status/2058773143091896679

COATUE

@coatuemgmt

Short dramas have quietly scaled into a mainstream format, and most of the US still hasn't noticed.

Serialized, vertical, and made for the phone. Episodes run one to two minutes, seasons run 100 to 150.

In China it's already obvious. More viewing time now goes to short dramas

Followup message from Roope himself

Yeah, working on BeatBandit - you are the showrunner and that's your writers' room - use the team a little, use the team a lot.

I've been running it through Claude Fable for a few days now through the MCP - pretty brutal workflow to allow Claude to read the code, simulate the user, evaluate the result, pinpoint things in the code that are not optimal - rinse and repeat. Fable is smarter than me, tough pill to swallow but it is what it is. :)

Kimi-Moonshot announces the open-source release of Kimi-K2.7-Code, a specialized coding model with strong gains over K2.6: +21.8% on Kimi Code Bench v2, +11% on Program Bench, +31.5% on MLS Bench Lite, plus 30% fewer reasoning tokens and better long-horizon task success.

Benchmark charts show K2.7-Code outperforming K2.6 while closing the gap or exceeding GPT-5.5 and Opus 4.8 (xhigh) on coding and agent metrics, with a performance-vs-tokens plot highlighting reduced overthinking and higher efficiency.

The model is immediately available via Kimi API and Kimi Code, with 6x high-speed mode incoming and a new beta program for early testing of upcoming features

Kimi.ai

Kimi-K2.7-Code, our latest coding model, is now released and open-sourced!

Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite.
Reasoning efficiency: Less overthinking, with 30% lower

Saturday 13th June 2026

Anthropic reveals a US government export control directive citing national security has forced the suspension of Fable 5 and Mythos 5 access for all foreign nationals, including their own foreign employees.

To comply, the company must immediately disable these models for every customer worldwide, while access to other Claude models remains fully operational.

Anthropic calls the order a misunderstanding, apologizes for the disruption, and states it is working urgently to restore service as soon as possible.

Anthropic

@AnthropicAI

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of

anthropic.com

Statement on the US government directive to suspend access to Fable 5 and Mythos 5

The US government has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States.

Z.ai released its flagship GLM-5.2 AI model to paid Coding Plan users, highlighting its strengths in coding tasks with a 1 million-token context window and adjustable thinking levels for complex workflows. Full API access and open-sourcing under the MIT license arrive next week

Z.ai

Intelligence should be open, accessible, and ready to build with, empowering every developer, everywhere.

GLM-5.2 is now available to all GLM Coding Plan users, including Lite, Pro, Max, and Team plans.
docs.z.ai/devpack/latest…

As our new flagship model, GLM-5.2 delivers

The cover image of this newsletter via generated via the Nano Banana Pro model within the Krea tool via the following prompt

The Great Wave : "Blossoming Cataclysm" Photorealistic : State-of-the-art photorealistic , leveraging advanced AI imaging capabilities. Landscape : Supernova erupting into a colossal red flower , marrying intense beauty with awe-inspiring realism. Message: Artistic convergence of cosmic power and giant blooms , evoking a vivid and visceral explosion. Format: A visually arresting masterpiece , showcasing the zenith of modern imaging techniques, ideal for gallery and prestigious exhibitions.

This Week in All Things AI - Week 23-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 07 Jun 2026 01:19:27 GMT

During the week of May 31 to June 6, 2026, the AI landscape was defined by a wave of model releases, escalating cost concerns, and infrastructure advances. Major model launches came from MiniMax (M3), NVIDIA (Nemotron 3 Ultra), Alibaba (Qwen3.7-Plus), Google DeepMind (Gemma 4 12B), and Microsoft (seven MAI models), while Anthropic reported Claude now authors 80% of its own codebase and OpenAI updated ChatGPT's memory system. On the cost front, an unnamed firm's $500M Claude AI bill dominated headlines, spurring broader cost-control moves — Uber capped AI spending at $1,500 per employee monthly, and tools like Factory Router and Lindy's migration to DeepSeek v4 demonstrated significant savings. Infrastructure and tooling saw updates from Perplexity's "Search as Code," SambaNova's disaggregated inference, Supabase's $500M raise, OpenAI's Codex plugins, Grok Composer 2.5, and Scaledown.ai launch of small-language-model services for compression and summarization

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 31st May 2026

Unnamed firm racks up $500M Claude AI bill in one month

An unnamed enterprise client spent more than $500 million in just 30 days on Anthropic's Claude AI platform after failing to set any usage limits or spending caps for employees, according to an Axios investigation published on May 28. The disclosure, made by an AI consultant whose client generated the bill, may represent one of the costliest IT governance failures on record

When AI costs spiral: A company accidentally spent $500 million in one month on Claude AI- what went wrong? | Company Business News

An enterprise client spent $500 million in a single month on Claude AI after failing to set employee usage limits, exposing a growing crisis in corporate AI cost governance.

https://www.livemint.com

Xiaomi just published a deep dive into the end-to-end inference engineering behind MiMo-V2.5 and MiMo-V2.5-Pro — and it’s a strong example of how architecture + systems work together to unlock real production gains.

Fuli Luo

Inference Optimizations Behind the MiMo-V2.5 Series API Price Reductions

Read the full technical blog: mimo.xiaomi.com/blog/mimo-v2-5…

The V2.5 model family, including MiMo-V2.5 and MiMo-V2.5-Pro, is built on a Hybrid Sliding Window Attention (Hybrid SWA) architecture, which

Thanh Nguyen

@ng_thanh8

Xiaomi MiMo

@XiaomiMiMo

What’s new with MiMo-V2.5 series inference?

We just published a blog on our full pipeline inference optimizations for MiMo-V2.5 series, including how we pushed hybrid SWA efficiency to the limit.

Read the full blog here:
mimo.xiaomi.com/blog/mimo-v2-5…

Monday 1st June 2026

MiniMax Launches M3 AI Model for Coding and Agents

Chinese AI lab MiniMax released its M3 model excelling in coding, agentic reasoning, tool use, multimodal chat, and long-context tasks. OpenCode integrated it instantly for free early access, offering 1,400 requests every five hours via terminal, IDE, or desktop app on macOS, Windows, and Linux. It posts strong benchmark results including 59.0% on SWE-Bench Pro, 66.0% on Terminal Bench 2.1, 74.2% on MCP Atlas, and competitive scores against models like GPT 5.5 and Gemini 3.1 Pro across agentic and tool-use tasks. Weights and technical report expected in about 10 days; API is immediately available with 50% off standard usage for the first 7 days to encourage early testing.

MiniMax (official)

@MiniMax_AI

Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities

- Coding & Agentic Frontier: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas
- MiniMax Sparse Attention scales context to 1M
-

Lisan al Gaib

Full Table:

OpenCode

MiniMax M3 will be launching soon

You can try it right now in OpenCode

For free

NVIDIA Unveils Nemotron 3 Ultra

The model tops U.S. open-weights rankings with an Intelligence Index score of 48, beating rivals like Gemma 4 31B while delivering over 300 output tokens per second in tests. It shines in agentic AI tasks, boasting 91% agent productivity, strong instruction following, and top coding skills, with 5x faster inference and 30% lower costs than leading competitors. Huang highlighted its openness for developers building everything from search tools to molecular simulations, paired with new agent deployment tools, as Nemotron 4 looms on the horizon.

Artificial Analysis

NVIDIA just announced the release of Nemotron 3 Ultra in Jensen Huang's Computex keynote: at 550B parameters (55B active), this is the largest Nemotron 3 model to date, and it is the most intelligent US open weights model

We partnered with @nvidia to evaluate this model for

What I was able to one-shot with Minimax-M3.

Yusuf Goolamabbas

@yusufg

So I gave @MiniMax_AI M3 the following task to test its multimodal capabilities, I found the following video via some website. It's basically a video scroll of a geocities page (some of you maybe old enough to remember geocities

Tuesday 2nd June 2026

SuperGrok and X Premium+ users now can use Composer 2.5 model from Cursor via Grok Build

skcd

@skcd42

We are excited for all of you to try out Composer 2.5 in Grok Build starting today!

To use composer-2-5 do `/model` in Grok Build and type in Composer to switch

Composer 2.5 comes with 200k context window and supports: subagents, MCPs, skills and additionally also works with

Qwen3.7-Plus is Alibaba's new multimodal agent model unifying vision and language for hybrid GUI/CLI operations, visual perception, reasoning, grounding, coding, and search-augmented tasks.

Benchmarks show it delivers competitive results across text, multimodal reasoning, visual understanding, agentic coding, and real-world QA, often matching or exceeding models like Claude Opus-4.6, GPT-5.4, and Gemini-3.1-Pro.

Released via API on Alibaba Cloud Model Studio with demos of browser and interactive agents;

Qwen

Introducing Qwen3.7-Plus — a multimodal agent model that unifies vision and language into one versatile agent foundation.

Multimodal interactive hybrid agent: unified GUI & CLI operation across visual and text tasks
Versatile coding agent & productivity assistant with

Qwen Studio

Qwen Studio offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.

https://qwen.ai

Some discussion between me and Anya Shapina related to Composer2-5 model availablility via Grok
=====
Anya

IMO Grok is the most under-appreciated model. I know, my comment is out of context (I haven't used Grok Build enough). But just as a model, it roxx! Testing very complicated agents and workflows across models and it kinda wins most scenarios. In my experience, it excels at walking and chewing gum at the same time (while also looking for shadows lurking behind bushes) - this is where most models fail for me. Not to mention native search, X, and Reddit. Anyone else in my Grok camp? Sorry, Elon haters :)

Yusuf

Which model of Grok Anya ? I think grok-build-0.1 is highy underrated and really look forward to the release of v9. Elon has a tendency to hype up particularly given the SpaceX IPO but the pace of updates of Grok Build CLI is relentless

Anya

Grok 4.3 - in my specific tests powering Agents (Not Grok Build which you post was about; look fwd to reporting on this later). Grok 4.3 - Strongest agentic tool calling + lowest hallucination rate in the lineup, from what I know. Worked for me.
=====

Perplexity AI launched Search as Code, enabling AI agents to generate Python code that directly calls atomic search primitives like fanout queries, deduplication, filtering, and ranking within a secure sandbox.

Available in the Perplexity Agent API, and now default in Computer.

Perplexity

Introducing Search as Code, our new search architecture for AI agents.

It writes Python that calls our search stack directly, instead of looping through function calls one at a time.

Available in the Perplexity Agent API, and now default in Computer.

research.perplexity.ai/articles/rethi…

Aravind Srinivas

@AravSrinivas

We’re moving away from search as a web fetch tool call to search as codegen to be future proof in a world where code execution inside agent harnesses is the way to do almost all of our knowledge work.

Doing this lets you compose multi-step primitives far more naturally and be

Perplexity

Denis Yarats

@denisyarats

we're going beyond traditional tool calls / MCPs to interact with the search stack. codegen is the most natural way for an LLM to drive search: the research tasks our users need require complex pipelines, customized per task.

so we're exposing the search stack as composable

Perplexity

Vaibhav Sisinty

@VaibhavSisinty

Perplexity just rebuilt how AI searches the internet from scratch, and it's a big deal.

Here's what it means in simple words.

Every AI today searches the web the same way. It sends one query, waits for results, reads them, sends another query, waits again. One at a time.

Perplexity

Wednesday 3rd June 2026

Uber says it has limited all employees to $1,500 in monthly token spending per AI coding tool “to responsibly encourage agentic AI adoption”

Uber caps employee AI spending after blowing through budget in 4 months | TechCrunch

Uber's cutback has occurred after the company had reportedly encouraged staff to use AI as much as possible.

Microsoft Launches MAI AI Models at Build 2026

The lineup covers reasoning, coding, images, voice, and transcription, all built as a multimodal system optimized for Microsoft's efficient MAIA 200 chips. MAI-Thinking-1, a 35 billion active-parameter model trained on 30 trillion tokens, scores 97% on math benchmark AIME 2025 and tops tough coding tests like SWE-Bench Pro, even beating models from Anthropic and matching Claude in blind evaluation

Microsoft AI CEO Mustafa Suleyman calls it a step toward 'humanist superintelligence' under human control.

Mustafa Suleyman

Super excited to announce seven new world-class MAI models today. They represent what we consider a new era in AI designed to keep you in control and on the frontier.
First is our text foundation model, MAI-Thinking-1, exceptionally strong on reasoning and SWE tasks.
- It’s a

Microsoft AI

@MicrosoftAI

Seven new models launching at Build: let’s go!
Reasoning. Code. Image. Transcribe. Voice.

Built from scratch on a clean data lineage, designed for efficiency, working seamlessly as a family of models

Thread
#MSBuild

Building a hill-climbing machine: Launching seven new MAI models | Microsoft AI

Alongside distribution on Foundry and optimization for our 1P products, our models are also going to be widely available for developers on OpenRouter, as well as and . For the first time developers will be able to tune the weights of the model themselves.

https://microsoft.ai

OpenAI unveils new Codex plugins for tasks related to public equity investment, banking and sales, and other roles, and plans to integrate Codex into ChatGPT

OpenAI

We’re making Codex more useful for your work by expanding plugins beyond individual tools.

These plugins turn Codex into a specialist for a specific role with a single install, no coding required.

Codex can access 62 popular apps and 110 skills for work across sales, data

Codex for every role, tool, and workflow

Discover new Codex plugins, sites, and annotations that help analysts, marketers, designers, investors, and other teams get more done with AI.

Codex for every role, tool, and workflow

Discover new Codex plugins, sites, and annotations that help analysts, marketers, designers, investors, and other teams get more done with AI.

https://developers.googleblog.com

Factory Router automatically selects the best AI model for each task in the Droid coding agent, slashing costs by 20-25% while matching 99% of premium performance on benchmarks like Terminal-Bench 2. It routes routine edits to cheaper models and escalates complex work to heavy-hitters like Claude Opus 4.7, even switching providers mid-task for reliability. Tech leaders like Keith Rabois and Garry Tan praised it as essential for enterprises facing rising AI bills, with CEO Matan Grinberg comparing it to hiring the right expert instead of Einstein for basic math. Currently in private preview for CLI and desktop users.

FWIW, Droid from Factory is one of my goto model-agnostic harness alongside Opencode Go. Very good Claude Code compatability in terms of support for skills, plugins, hooks, sub-agents etc and one of the best if not the best context management I've come across

Factory

@FactoryAI

Introducing model routing to Factory.

Factory Router picks the right model for every task, automatically.

Maintain frontier performance while cutting costs by 25%.

Factory

@FactoryAI

Introducing model routing to Factory.

Factory Router picks the right model for every task, automatically.

Maintain frontier performance while cutting costs by 25%.

via James Chan

Does anyone have experience with multiplayer PRD like ChatPRD? Any good?

Thursday 4th June 2026

Google DeepMind releases Gemma 4 12B, a lightweight 12B-parameter multimodal model under Apache 2.0 license designed to run locally on laptops with just 16GB VRAM or unified memory.

It introduces an encoder-free unified architecture where vision uses a tiny 35M-parameter embedding module and audio projects raw signals directly into the LLM backbone, eliminating separate encoders for efficiency.

Delivers advanced reasoning and multimodal capabilities nearing the larger Gemma 4 26B model's benchmarks, with native 256K context, MTP drafters, and immediate support across Hugging Face, Kaggle, llama.cpp, MLX, and vLLM.

Google Gemma

@googlegemma

Meet Gemma 4 12B!

A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license.

Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B:

Bringing Gemma 4 12B to your Laptop: Unlocking Local, Agentic Workflows with Google AI Edge

Google DeepMind's Gemma 4 12B model brings agentic, multimodal AI capabilities to everyday laptops with 16GB of RAM, enabling local data processing and visual insight generation. Users can leverage this model on macOS through the Google AI Edge Gallery for dynamic Python code execution and visualization, as well as via Google AI Edge Eloquent for completely offline voice dictation and text editing.

Friday 5th June 2026

via my high-school classmate Abhi Ingle who works at Sambanova and lurks on this group via the newsletter. Took me a few readings of the blog post and had to go back and refresh my knowledge on the prefill and decode steps in auto reggressive transformer architectures before the penny dropped on the how this architecture is unique
====

At Computex, SambaNova demoed a “disaggregated inference” architecture where GPUs handle prefill and their RDUs handle decode, yielding faster, cheaper long-horizon agent workloads than GPU-only setups, and they now have this running in production-style environments with partners like VC2 and Together AI.

3.5 Bn compute commitment to SambaNova ! Taking on the lofty goal of making agentic inference faster and more efficient, Vista Equity Partners, Cambium Capital Management announced Vector Core... | Abhi I.

3.5 Bn compute commitment to SambaNova ! Taking on the lofty goal of making agentic inference faster and more efficient, Vista Equity Partners, Cambium Capital Management announced Vector Core Compute - "The World's first heterogenous disaggregated Inference Cloud".

The First Disaggregated Inference Demo for AI Agents Is Live

SambaNova demonstrates how GPUs and RDUs work together to deliver premium inference for agent workloads using the right chip for the right workload.

https://sambanova.ai

OpenAI updates ChatGPT memory with a “more capable and compute-efficient” architecture and a summary page that lets users review and steer what it remembers

Dreaming: Better memory for a more helpful ChatGPT

ChatGPT introduces a new memory system to better remember preferences, keeping context fresh and relevant across conversations.

https://www.anthropic.com

Supabase, which provides backend tools for building AI apps, raised a $500M Series F led by GIC at a $10B pre-money valuation, up from $5B in October 2025. Also released Multigres v0.1 alpha to the open source community, Multigres tries to bring Vitess-grade horizontal scaling, high availability, and operational simplicity to Postgres.

COATUE

@coatuemgmt

We led @Supabase's Seed and Series A. Five years in, they're the open source Postgres platform powering the majority of AI app builders — and 9 million developers strong.

As agents reshape how software gets made, Supabase has become the default backend underneath it. Proud to

Supabase Series F

Supabase has raised a $500M Series F at a $10B pre-money valuation, led by GIC.

https://supabase.com

Multigres v0.1 Alpha: an operating system for Postgres

Today we're releasing Multigres v0.1 alpha to the open source community, bringing Vitess-grade horizontal scaling, high availability, and operational simplicity to Postgres.

https://supabase.com

Response from Anya Shapina
===
I believe Supabase (and possibly its philosophical rival Convex) will be some of the biggest winners in the AI race. I use it as the shared memory and persistent backend of my multi-agent system - it's everything to my project, and all other projects I've ever built. But then there is the Goliath Convex with its built-in support for AI Agents... and NO SQL (love/hate?). Which camp are you people in?
====

I've not used Lindy but have heard good things about it from others. For those unaware of Lindy, it describs itself productivity and automation platform that positions itself as an "AI executive assistant" or "AI employee." It is particularly strong for professionals who want proactive help with email, meetings, and calendar management.

Saw this post from Flo Crivello, founder of Lindy who migrated 100% of its traffic to DeepSeek v4 from Anthropic models, reporting millions in cost savings alongside performance improvements on core use cases. The switch required building substantial new infrastructure and internal tooling, described as 100x more work than anticipated, emphasizing the value of swappable model architectures.

Flo Crivello

@Altimor

Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models.
Saves us millions of $ and we're actually seeing an *increase* in performance on many core use cases. Transformative for the business.

Response from Ms Macarena Correa to a post dated 30th May 2026 about the announcement of Kirkland and Ellis committing $500 million over the next 3-4 years to build its own proprietary AI platform and custom tools
=====
On this one, I think its the right way forward. If all Magic Circle / Silver Circle law firms are using the same AI tools to help drafting contracts (Harvey / Legora), and review them too, as well as providing advice, all solutions will become kind of standard, so ultimately the difference between law firms will be reduce only to 1) pricing and 2) charisma of each particular lawyer (e.g., building the relation with the client). this 2nd point cannot be replaced by AI, so probably most of the efforts need to shit into building more and better relations of trust.
=====

Anthropic details its progress toward recursive self-improvement, and its implications, and says Claude has authored 80%+ of the code merged into its codebase

When AI builds itself

Our progress toward recursive self-improvement, and its implications.

via Imran Muthuvappa

made a free community for people who want to build and ship their first agent

agentmaxxing

No AI background? I'll help you build, deploy, and sell your first AI agent to a small business in 90 days.

https://www.skool.com

Saturday 6th June 2026

Came across Scaledown which has developed purpose-built models for compression, summarization, extraction, and classification. Pricing is 0.05/M tokens with self-deployment options also available

Neal Patel

@neal_k_patel

Introducing ScaleDown

15x cheaper. 63x faster. 5.1% more accurate than GPT-5.4 Mini.

Task-specific SLMs for the 70-80% of AI workloads that don't need a frontier model.

From NeurIPS '25 to scaledown.ai

Scaledown.ai - Task-Specific Small Language Models

Purpose-built models for compression, summarization, extraction, and classification. Frontier quality at a fraction of the cost.

https://scaledown.ai

https://scaledown.ai/blog

The cover image of this newsletter via generated via the Krea 2 Large model within the Krea tool via the following prompt

An oil painting in the style of H.R. Giger and Zdzisław Beksiński of an ancient Greek temple on the side of Mount Olympus, surrounded by cacti, a rainbow in the sky, grey clouds overhead, blue lilies around the base, and birds flying above. The painting is highly detailed.

This Week in All Things AI - Week 22-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 31 May 2026 02:14:22 GMT

Major developments this week included xAI completing training on Grok V9-Medium (1.5T parameters) with a public release expected in 2–3 weeks, alongside plans to open-source its current v8-small model by year-end. OpenRouter raised $113M at a $1.3B valuation as Xiaomi's MiMo slashed API pricing by up to 99% and MiniMax teased its M3 sparse-attention architecture. Anthropic released Opus 4.8, Stanford launched the open-source OpenJarvis agent framework, and Cursor announced its inaugural Compile conference. China's Shanghai Futures Exchange began researching AI token futures as Tomasz Tunguz argued that competitive advantage now lies in the orchestration layer rather than foundation models. The week closed with Kirkland & Ellis committing $500M to build a proprietary in-house AI platform and ElevenLabs launching Dubbing v2, a model that preserves speaker emotion, tone, and pacing across 90+ languages.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 24th May 2026

Monday 25th May 2026

Elon Musk announced that training of Grok foundation model V9-Medium with 1.5 trillion parameters is complete, with positive evaluation results.

Supplementary training included substantial Cursor data, with fine-tuning now underway and reinforcement learning scheduled to start in a few days.

Public release is expected in 2-3 weeks, marking a major upgrade over the current 0.5T v8-small model, particularly for challenging coding tasks.

He also announces that xAI will open-source its current 0.5T Grok model (v8-small) by the end of 2026, stating it should remain quite useful even after newer models launch.

Elon Musk

Grok foundation model V9-Medium (1.5T) has finished training. Evals look good. A lot of Cursor data was added in supplementary training and there is more to come.

Fine-tuning is underway and reinforcement learning begins in a few days. 2 to 3 weeks to public release.

This will

Elon Musk

We will open source the 0.5T model towards the end of this year. It should still be quite useful.

Tuesday 26th May 2026

Wednesday 27th May 2026

OpenRouter raised $113M led by CapitalG, a source says at a $1.3B valuation, and now processes 25T tokens across 400+ models weekly, up from 5T six months ago

Alex Atallah

@alexatallah

We're seeing a Cambrian explosion of AI models, and it's happening on OpenRouter.

The future of AI is neurodiversity:
- Agents choosing the most cost-effective model/provider/tool for the task
- Agents orchestrating multiple models for the smartest result
- Advanced security and

OpenRouter

https://www.nytimes.com/2026/05/26/business/dealbook/openrouter-ai-models-fundraising.html

Today we’re announcing our $113M Series B led by @CapitalGVC.

Over the last 6 months, weekly volume on OpenRouter grew from 5T to 25T tokens as AI rapidly shifts from experimentation into production.

We’re excited for what comes next.

Deedy

@deedydas

OpenRouter is now serving 1.5 quadrillion tokens/yr!

That token run rate is:
— 15-30% of Google APIs
— 20-40% of OpenAI
— >50% of Microsoft Azure Foundry

That's 15x larger than when we invested a year ago.

Revenue has already doubled since this $1.3B round was done in Feb!

Xiaomi MiMo announced permanent API price cuts of up to 99% for MiMo-V2.5-Pro and MiMo-V2.5 models, with unified pricing across context lengths and new low rates like $0.0036 per 1M input cache hit tokens for Pro.

Token plans received major upgrades delivering 5–8× more credits at the same price points, with all existing subscriber credits fully reset as a thank-you, while MiMo-V2.5-TTS stays free temporarily.

Xiaomi MiMo

@XiaomiMiMo

Better inference efficiency, lower costs, broader access.

MiMo-V2.5 Series API pricing is now permanently reduced — by up to 99% compared to previous pricing.
Unified pricing across all context lengths.
MiMo Token Plans have also been upgraded:
• 5–8× more usable tokens

Fuli Luo

https://platform.xiaomimimo.com

Behind the MiMo API Price Reduction:
The deepest price cut, up to 99%, is for Input (Cache Hit). The core reason is our inference framework now supports hierarchical KV cache optimization for SWA. Production inference engine tests show this optimization increases cached token

Xiaomi MiMo Api Open Platform - Token Plan Global Launch

One-time purchase unlocks both MiMo-V2.5 flagship models, plus TTS model free across all tiers for a limited time. Unleash powerful productivity with Xiaomi MiMo

Skyler Miao, Head of Engineering at MiniMax_AI, posted a teaser for the upcoming M3 model showing their new sparse attention architecture. The diagram details a two-stage GQA-based system: an Index Branch quickly scans and selects top-k relevant token blocks via block max pooling, then a Sparse Branch performs full attention only on those blocks.

This design delivers 9.7x faster prefill and 15.6x faster decoding at 1M token context versus M2, enabling efficient long-context AI without prohibitive compute costs.

Skyler Miao

@SkylerMiao7

Something BIG is coming

elie

@eliebakouch

new minimax sparse attention compared to deepseek v3.2 (DSA) and v4 (CSA)

main changes:
- based on GQA not MLA
- block level selection like in CSA but attention is done on the real KV, not in the compressed dimension

Skyler Miao

@SkylerMiao7

Something BIG is coming

Rameswar

@rameswar08

For non ai people:

most ai models work like this,

every word looks at every other word to understand context

great for accuracy, terrible for speed at massive scale

MiniMax's sparse attention changes that

instead of processing an entire 1m token context deeply, the model

Thursday 28th May 2026

Cursor Announces Invite-Only Compile Conference in San Francisco

The one-day Compile event happens June 16 at Fort Mason, bringing together engineers, researchers, and builders for discussions on AI-native development. Speakers like Cursor's Michael Truell and Ryo Lu, plus guests from Every, Shopify, and indie makers, will work through ideas live on a chalkboard stage. It's waitlist-only now—sign up at cursor.com/compile and invitations go out via email—with a call for papers still open on rethinking systems and simplifying complex ideas.

Cursor

We're hosting an event on June 16th in San Francisco.

Compile is a one-day event that brings together engineers, researchers, designers, and builders of all kinds to discuss the future of software.

cursor.com

Cursor · Compile

Compile is Cursor's inaugural conference — bringing together developers, researchers, and teams shaping the future of AI-native development.

Tomasz Tunguz of Theory Ventures with a post where in he articulates that in the AI era, the core differentiator in software is not the model itself but the “harness” layer that tames a general LLM into a reliable, domain-specific agent by combining seven capabilities (context, tools, orchestration, state, sandboxing, observability, and cost optimization).

The article argues that when everyone can access similar foundation models, competitive advantage shifts to whoever builds the best harness around the model. This harness “domesticates” a powerful but wild LLM into a dependable system that can safely execute real workflows in specific industries.

Software After AI

Software is no longer about UX & data. It is about the harness, the layer that turns an LLM into a reliable agent. Seven components define the new stack.

https://tomtunguz.com

In response Brent Maxwell writes

In my engineering teams, nobody can tell the difference between gpt-5.5, claude-4.7, gemini-3.5-flash or composer-2.5.

There is no winner right now in the model wars.

The harnesses are totally the right thing to focus on - they are making or breaking development processes for us. When we figure out how to manate the harness better, we ship faster with less time stuck in PR. When one of our devs just uses the default agent config in the IDE, it really doesn't give them the same amount of leverage.

Shanghai Futures Exchange Explores AI Token Contracts

China’s Shanghai Futures Exchange is in early-stage research on futures contracts tied to AI tokens, the smallest unit of information processed and billed for by AI models. These contracts would let companies hedge against volatile AI compute costs along the AI supply chain, similar in spirit to commodity or energy futures.

While U.S. exchanges like CME and ICE are moving toward futures tied to GPU rental/compute capacity, China’s concept would be directly linked to AI token consumption used for pricing AI services. This represents a different abstraction layer: the U.S. focuses on hardware capacity, whereas China targets the usage-based “digital fuel” that powers AI models.

China’s daily AI token usage has exploded roughly 1,000x since early 2024, reaching more than 140 trillion tokens a day by March 2026, underscoring surging demand and cost exposure. Token-based derivatives are being framed as a potential new asset class, with figures like BlackRock’s CEO noting that futures on compute could become a distinct financial market.

Exclusive: China works on AI token futures market, sources say, in race with US

China is designing a futures market for AI tokens, sources familiar with the matter said, as the country potentially takes a different tack to U.S. exchanges developing compute power futures to tap ‌the rapidly growing appetite to hedge AI costs.

https://www.reuters.com

Friday 29th May 2026

Some links about Anthropic's Opus 4.8 launch including a tweet thread from Anthropic with some guidance on devs as they migrate to Opus 4.8 and some commentary from Dan Shipper of Every who had access to Opus 4.8 for the past two weeks

ClaudeDevs

Opus 4.8 is live in Claude Code today.

A few things worth knowing:

Claude

https://legaltechnology.com

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.

Available today at the same price.

Dan Shipper 📧

@danshipper

BREAKING:

Anthropic just dropped Opus 4.8—and it is a MONSTER

We've been testing for about a week @every and our verdict is they could've just called it Opus 5, it's that good.

Here's our vibe check:

- Beats GPT-5.5 on Senior Engineer bench. On our toughest benchmark Opus

Vibe Check: Opus 4.8-Anthropic Should've Rounded Up to 5

Opus 4.8 tops both our Senior Engineer benchmark and our writing tests. It's the most complete model we've tested. We just wish it had an app to match.

https://every.to

Stanford Unveils OpenJarvis: Efficient AI Agents for Your Devices

Stanford researchers from Hazy Research and Scaling Intelligence Lab released OpenJarvis v1.0, an open-source framework that builds personal AI agents to run locally on devices. It emphasizes 'Intelligence per Watt' with swappable components like local models, engines, agents, tools for apps like Slack and email, and self-improvement features. Users get eight ready-to-run agents for tasks such as morning briefings and code review, installable via a simple one-line script across CLI, web, desktop, or messaging apps. Benchmarks show it handles most queries at interactive speeds with far lower costs and latency than cloud systems, positioning it as a privacy-focused alternative to massive data centers.

Jon Saad-Falcon

@JonSaadFalcon

The dominant story in AI has been the growing cloud: bigger clusters, larger models, more gigawatts.

We believe the future is in the opposite direction: on-device inference, smaller models, watts instead of gigawatts.

Today we're releasing @OpenJarvisAI v1.0: a personal AI

Lambda

@LambdaAPI

Most agent frameworks are built around one cloud model. Swap in a local model, performance drops.

@OpenJarvisAI fixes the harness, not the model. Result: 77% of the accuracy gap recovered, 800x lower cost per query, 4x lower latency.

Built on Lambda. Open-sourced from Stanford.

Snorkel AI

@SnorkelAI

Huge congrats to @jonsaadfalcon, @Avanika15, @Azaliamirh and the @HazyResearch team on @OpenJarvisAI — out today.

For two years, they've been making the case that AI inference belongs on hardware people already own, not just in megawatt data centers. Excited to support the

Jon Saad-Falcon

@JonSaadFalcon

via shrwn

Free Google x Kaggle vibe coding course

* Mid June
* 1-2hrs/day for 5 days
* Free

5-Day AI Agents: Intensive Vibe Coding Course With Google

June 15 - 19, 2026

https://www.kaggle.com

Saturday 30th May 2026

Kirkland & Ellis LLP is one of the world’s most successful and profitable “Big Law” firms with approximately 4,000 attorneys across 23 offices. It consistently ranks at or near the top of revenue and profitability metrics.

Kirkland & Ellis is committing $500 million over the next 3–4 years to build its own proprietary AI platform and custom tools. This is one of the largest and most ambitious technology investments ever announced by a law firm.

Key details include:

Funding & Timeline: Funded entirely from the firm’s own revenue (aligning with Chair Jon Ballis’s philosophy of investing roughly 1% of revenue in new initiatives — which would be ~$100M+ in a $10B+ year). More than $100 million is expected in 2026 alone, with the balance spread over the following years.
Goals: Create a broad, firm-wide AI platform that captures and deploys the firm’s “collective intelligence” to support lawyers across practices “start to finish” on client work. The aim is to move beyond reliance on multiple third-party tools toward more integrated, customized capabilities.
Development Approach: Informed by input from 250 lawyers (including 100 partners) on real workflows and needs. A team of 180 tech professionals is involved, working with undisclosed external partners/companies to help build the technology. Crucially, the resulting tools and IP will not be commercialized or sold to other law firms — a deliberate contrast to some other firm-vendor partnerships (e.g., Freshfields’ work with Anthropic).
Technical Elements: Involves on-premise GPU environments and Microsoft Azure-based AI infrastructure, including facilities for training and inference. The firm is actively hiring for AI-related roles (dozens of positions), including high-compensation roles like AI Infrastructure Directors.
Context Within the Firm: This builds on existing efforts. Kirkland already deploys third-party legal AI tools such as Harvey across its attorneys. It has a history of building proprietary technology in-house, including SideTrack (a tool for investment fund work, particularly around MFN issues) and earlier databases like CTRAN for M&A competitive intelligence. It also maintains internal innovation teams, AI Innovation Advisors embedded in practice groups, and responsible AI governance structures.

This move sits at the center of a key debate in legal technology: buy vs. build for AI. Many top firms rely heavily on specialized platforms like Harvey, CoCounsel (Thomson Reuters), or Lexis+ AI. Kirkland’s approach prioritizes:

Greater control over data, customization, and roadmap.
Using the firm’s own vast proprietary knowledge and deal experience as a competitive differentiator (rather than everyone having access to similar generic or vendor tools).
Enhanced security, confidentiality, and alignment with internal workflows.
Long-term ownership rather than ongoing licensing dependency.

Kirkland & Ellis has form for building its own technology. The $500m AI play is its biggest yet. - Legal IT Insider

Kirkland & Ellis announced that it would spend $500m over the next three to four years developing its own custom AI tools and services.

ElevenLabs launches Dubbing v2, which it says preserves the original speaker's emotion, tone, and pacing across 90+ languages while staying synced to content

aneri

@0xaneri

Introducing Dubbing v2.

For the first time, AI dubbing preserves how something was said, not just what was said. Dubbing v2 reads the original audio directly rather than just the transcript, so your emotion, tone, and delivery carries across 100+ languages.

Every system before

Introducing Dubbing v2: our revolutionary new dubbing model

Introducing Dubbing v2: our revolutionary new dubbing model, which carries the emotion and performance of the original speaker across every language.

https://elevenlabs.io

The cover image of this newsletter via generated via the Krea 2 Large model within the Krea tool via the following prompt

Automobile inspired modern curved building , where the curved building's façade meticulously crafted from gears, cogs, and mechanical parts. The building's façade design embodies the fusion of nature and machinery, as evidenced by the intertwining vines and mechanical components

This Week in All Things AI - Week 21-2026

twiata@newsletter.paragraph.com (Yusuf Goolamabbas) — Sun, 24 May 2026 00:56:56 GMT

This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.

If you follow AI for work, research, investing, or just to understand where the technology is heading, this weekly brief is a concise way to scan the most important launches, risks, and resources in a few focused minutes.

Across the week, we saw advances in agentic coding and web tooling (Cursor’s Composer 2.5, Google’s Gemini 3.5 Flash, Modern Web Guidance), major moves in AI infrastructure and pricing (Anthropic’s expanded Colossus 2 deal with SpaceX, DeepSeek’s permanent discount on V4‑Pro), a wave of creative tools (Krea 2 and several AI‑assisted video editors), and notable policy and geo developments from Pope Leo XIV’s forthcoming AI encyclical to Singapore’s latest AI initiatives.

The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.

Sunday 17th May 2026

Monday 18th May 2026

Tuesday 19th May 2026

Cursor’s biggest model upgrade yet. Strong gains in long-running agentic coding tasks, reliability with complex instructions, and up to 10x better cost efficiency. It’s built on Moonshot’s Kimi K2.5 and reportedly performs on par with top frontier models in some real-world evaluations. Cursor is also training a much larger model from scratch with SpaceXAI on a massive Colossus 2 cluster.

Cursor

https://www.theguardian.com

Introducing Composer 2.5, our most powerful model yet.

It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions.

For the next week, we’re doubling the included usage of the model.

Introducing Composer 2.5 · Cursor

A substantial improvement in intelligence and behavior over Composer 2, particularly on long-horizon agentic tasks.

https://cursor.com

The Vatican announced that Pope Leo XIV’s first major encyclical on AI will be released on May 25. It will focus on human dignity and “the protection of the human person in the age of AI.” It is being prepared in collaboration with an Anthropic co-founder.

Pope Leo to issue text on human dignity and AI with Anthropic co-founder

The pope's encyclical will address 'the protection of the human person in the age of AI', the Vatican says

Pope Leo and co-founder of Anthropic to launch pontiff's AI encyclical on May 25

Pope Leo XIV and the co-founder of artificial intelligence company Anthropic will launch the pontiff's first encyclical on May 25, a document on the care of human dignity in the era of AI, the Vatican said Monday.

https://www.chicagotribune.com

Cloudflare tests Mythos against 50+ repositories, highlights its ability to chain bugs into a single exploit, and details a vulnerability discovery harness

Project Glasswing: what Mythos showed us

In recent weeks, we pointed Mythos and other security-focused LLMs at live code across critical parts of our infrastructure. We share what we observed, the models' strengths and weaknesses, and what the work around them needs to look like before any of it can scale.

https://blog.cloudflare.com

Krea 2 (also called K2 or Krea 2 Large) is Krea AI’s first proprietary foundation image model

It was built entirely from scratch with a deliberate focus on aesthetics, style transfer, creative control, and artistic exploration rather than literal prompt transcription or maximum photorealism

It forms part of the broader Krea.ai platform—an all-in-one creative AI suite that aggregates 100+ models for images, video, 3D, editing, upscaling, and fine-tuning under one subscription

They are offering unlimited Krea 2 generations to all their subscribers for the entire week.

Krea

@krea_ai

this is Krea 2.

our first foundation model, built completely from scratch for aesthetic diversity and stylistic control.

learn more and get early access

Heather Cooper

@HBCoop_

Had the pleasure of testing this out over the past week.

Very, very good model. Quality aesthetics, fast generations along with Moodboards and Style transfer.

Excited to explore more with Krea 2 ↓

Krea

@krea_ai

this is Krea 2.

our first foundation model, built completely from scratch for aesthetic diversity and stylistic control.

learn more and get early access

Krea 2: AI Image Foundation Model & Style Control

Create expressive AI images with Krea 2, Krea AI's in-house foundation model for aesthetic diversity, style control, moodboards, and creative workflows.

https://www.krea.ai

Wednesday 20th May 2026

Gemini Spark is Google's new cloud-based 24/7 personal AI agent announced at I/O 2026, designed to autonomously handle tasks in the background even when devices are off while requiring user approval for major actions.

Initial rollout targets trusted testers this week and U.S. Google AI Ultra subscribers next week in beta, with enterprise availability planned through Google Workspace and Gemini Enterprise.

Summer roadmap includes MCP integrations, macOS support, email/text interaction, Chrome browser control, and subagent creation to expand proactive capabilities.

Google Gemini

@GeminiApp

Gemini Spark is your new 24/7 personal AI agent.

Give it a task and it works autonomously in the background, even if your phone and laptop are turned off. You choose to turn it on and it's designed to check with you before taking major actions. #GoogleIO

Google Unveils Gemini 3.5 Flash at I/O Conference

Google launched Gemini 3.5 Flash, generally available now and touted as its strongest model for agentic workflows like multi-step tool use and terminal coding. Benchmarks show it topping prior versions on key tests, holding ground against rivals like Claude 3.5 Sonnet, though independent reviews note mixed results on cost and performance.

The event also introduced Gemini Omni for video creation, Spark as a 24/7 personal agent, developer tools like Antigravity 2.0, plus consumer updates to Search and the Gemini app—ushering in what execs call the agentic era.

Jeff Dean

@JeffDean

1/ Today at #GoogleIO, we’re releasing Gemini 3.5, our latest family of models combining frontier intelligence with action.

We’re starting by releasing 3.5 Flash, which is built to help you execute complex, long-horizon agentic workflows.

Gemini 3.5 Flash is our strongest model

Google AI Studio

@GoogleAIStudio

x.com/i/article/2056…

Gemini 3.5: frontier intelligence with action

At Google I/O we released Gemini 3.5, our latest series of models combining frontier intelligence with action.

https://developer.chrome.com

Gemini 3.5 Flash might be fast enough for gen AI to make sense

Google says its more efficient Gemini 3.5 Flash is the key to your agentic AI future.

https://arstechnica.com

Google Launches Modern Web Guidance for AI Coding Agents

At Google I/O 2026, the company unveiled Modern Web Guidance, an early preview toolset that equips AI agents like Claude Code and GitHub Copilot with skills for the latest browser APIs, layouts, and performance techniques. It tackles legacy patterns by injecting concise prompts drawn from Chrome's Baseline initiative, boosting agent success on modern tasks from 52% to 86% in early tests. Backed by Chrome and Microsoft Edge teams, plus experts like Lea Verou, it offers easy installation via a simple npx command and covers real-world uses from CSS animations to WebAuthn logins.

Sarah Drasner

@sarah_edo

Game changer for Web Development announced at GoogleIO- Modern Web Guidance! It’s expert-vetted skills for web development based on best practices of latest specs and APIs.

It ensures your agent/coding harness doesn’t default to older and out of date patterns to build sites.

Modern Web Guidance | Chrome for Developers

Guidance on how to build for the modern web.

Q&A with Harvey CEO Winston Weinberg on launching the legal AI startup in 2022, how AI could shake up law firms' business models, legal AI competition, and more

https://www.ft.com/content/47eafdcb-5a0a-4515-a07a-70e89ad2fa78?syn-25a6b1a6=1

Thursday 21st May 2026

Anthropic is paying SpaceX $1.25B/mo. until May 2029 under their compute deal; Anthropic says it's expanding the deal to include Colossus 2 capacity

Tom Brown

@nottombrown

We’re expanding our partnership with @SpaceX, and will be scaling up on GB200 capacity in Colossus 2 throughout June.

Appreciate @elonmusk and the team helping us find good homes for the Claudes.

Tom Brown

@nottombrown

In the next few days we'll be ramping up Claude inference on Colossus.

Grateful to be partnering with SpaceX here. We are going to need to move a lot of atoms in order to keep up with AI demand, and there's nobody better at quickly moving atoms (on or off planet Earth)

Elon Musk

https://www.topazlabs.com

As the recently expanded partnership with @AnthropicAI demonstrates, @SpaceX is offering AI compute as a service at significant scale.

We are in discussions with other companies to do the same.

Over time, especially with orbital data centers, we expect to serve AI at extremely

FYI, in case some in this group find the below useful for their personal and professional workflows. Personally I consider video-editing as something akin to 'magic of the dark arts' so huge respect to practioners of this craft.

The below are a links to websites of AI assisted video editing tools I've come across (in no particular order), pretty sure these are but a small fraction of tools out there.

Within this list I'm not mentioning the usual suspects such as Premiere, Final Cut Pro, Davinci Resolv, Capcut

Create videos for Socials - keyframes.studio

Create videos for Socials in 5 minutes. Don't waste time looking up aspect ratios or learning heavy video editors - Keyframes Studio takes care of everything.

https://keyframes.studio

Revid AI - Ideate, Publish, Go Viral | #1 AI Video Generator

Create viral short videos for TikTok, Instagram, and YouTube. Turn ideas into content instantly. Ideate, create, edit, publish - all in one. No skills needed.

https://www.revid.ai

Topaz Labs | Professional-grade photo and video editing powered by AI.

Photo and video enhancement software powered by deep learning gets you the best image quality available for noise reduction, sharpening, upscaling, and more.

Capsule

Meet Capsule-the AI-powered video editor made for content and marketing teams. Create stunning videos 10x faster, collaborate seamlessly, and stay on brand.

https://capsule.video

Descript - AI Video & Podcast Editor | Free, Online

Descript makes editing video and audio as easy as editing text. Record, transcribe, edit, and publish in one tool. Try for free, with powerful upgrades for creators & teams.

https://www.descript.com

Overlap - AI Video Clipping for Media, Podcasts and Creators

Use Overlap to create AI clipping workflows to turn long videos into clips, shorts, and social posts automatically. Scale short form content creation with Agentic workflows.

https://overlap.ai

Cardboard | Agentic Video Editor

A fast, collaborative, and agentic video editor that runs in your browser. Create, cut, and compose videos with AI - no installs, no downloads.

https://www.usecardboard.com

Eddie AI | The Assistant Video Editor for Pros

Eddie is the AI assistant video editor trusted by 40,000+ video professionals. Import interviews and chat with Eddie to get rough cuts and edits in seconds. Then export to MP4, Adobe, FCP, and DaVinci Resolve.

https://www.heyeddie.ai

Ponder

Transform your video editing workflow with Ponder.AI's revolutionary AI-powered platform. Sign up free or download now to experience the future of video creation.

https://ponder.ai

OpusClip: #1 AI video clipping tool to create viral shorts

OpusClip turns long videos into high-quality viral clips, and publishes them to all social platforms in one click. We help 10M+ creators create and grow faster.

https://www.opus.pro

If any of these do end up being used by you/your team, do share your experiences with this group so that others can also benefit from your learnings

via Alex

SpaceX S1 fillings is a good read for SciFi fans like me. Emphasizing Fiction here 😂
Claiming capabilities to launch 100GW of compute to LEO when hyperscalers are barely able to do 1GW on earth. Claiming 220B daily token usage when the actual xAI usage numbers seem to be more like 15B tokens daily. Claiming the TAM for xAI is over 20T$.

https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm

via Alba Chung of Alibaba Cloud

Qwen3.7-Max — now available viaAlibaba Cloud Model Studio:

- frontier coding agent: from frontend prototyping to complex software engineering

- office productivity and workflow automation via MCP and multi-agent orchestration

- sustained autonomous execution across long-horizon tasks

-cross-scaffold generalization across diverse agent frameworks

My LinkedIn: https://www.linkedin.com/in/albachung

Feel free to PM me for more details.🙌🏼

Friday 22nd May 2026

Singapore firing on all cylinders

• OpenAI first AI lab outside of the US is Singapore

• Anthropic’s lead investor for their Series G is GIC (Singapore’s sovereign wealth fund)

• NVIDIA chose Singapore for their research hub

• Google Deepmind is strengthening its partnership with Singapore

Introducing OpenAI for Singapore

OpenAI for Singapore launches a multi-year AI partnership to expand deployment, build local talent, and support businesses and public services with AI.

https://www.cnbc.com/2026/05/20/nvidia-to-launch-singapore-research-hub-as-city-state-boosts-ai-plans.html

Google DeepMind & Singapore: National AI partnership

Google DeepMind and Singapore partner to apply frontier AI to address challenges across health, education, sustainability and more through the National Partnerships for AI initiative.

https://deepmind.google

Saturday 23rd May 2026

Zero to Claude Code Course Draws 18,000 Beginners Worldwide

The Zero to Claude Code course guides users from terminal basics to advanced AI techniques like prompt engineering and subagents, with hands-on projects including a real-time game capstone. Launched recently at zero2claude.dev it has enrolled over 18,000 students globally, running on a high-performance site that Itay Shmool, VP @ WIX created solo with Claude Code

itay shmool

@IShmool

I built "zero2claude", a free course that takes people from zero terminal experience to shipping with Claude Code.

The curriculum goes from absolute zero → software basics → Claude Code fundamentals → advanced usage. No shortcuts, no assumptions.

17,000+ students. 7

Boris Cherny

@bcherny

Big fan of teaching more people the basics of using Claude Code in an accessible way.

So much of the world has not yet used agents. There's a lot of opportunity to level the playing field and expand access.

itay shmool

@IShmool

In the world of increasing prices for models, Deepseek says that they have made their 75% discount originally ending May 31st 2026 for V4-Pro permanent if one uses DeepSeek's inference endpoint.
Recommend group members do their own research to interpret the terms of service/privacy policy under which Deepseek's own inference endpoint is served Deepseek v4-Pro is an open-source/open weight model so there are many inference providers which serve this model and each inference provider serves it under their own ToS/privacy policy

DeepSeek

We are making our discount permanent!

Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life!

DeepSeek

The DeepSeek-V4-Pro discount has been extended until May 31, 2026, 15:59 UTC!

an AI demo where Gemini Omni transforms a Google Maps screenshot with a drawn route into a realistic first-person taxi driving video through city streets, complete with accurate buildings, traffic, and dashboard details.

Lucas Beyer (bl16)

@giffmana

Can't believe we're getting this before GTA 6

CHRIS FIRST

@chrisfirst

I uploaded a screenshot of Google Maps to Gemini Omni with a route drawn on it.

Then I prompted it to create a first person view of someone driving a taxi cab along the route in the reference image.

Pretty close to the real thing.