This week’s AI landscape was defined by aggressive model commoditization, agentic workflows maturing, and incumbents pulling key verticals in‑house. DeepSeek and Xiaomi drove frontier‑grade capabilities toward “practically free,” with DeepSeek slashing cached token prices and Xiaomi open‑sourcing trillion‑parameter MiMo models under MIT with huge token grants, while Poolside and xAI pushed specialized coding and agentic models like Laguna and Grok 4.3 into the spotlight.
Major platforms shifted their economics and UX around agents: GitHub Copilot moved to usage‑based billing, Anthropic released deep integrations for creative tools, and OpenAI’s Codex evolved into a “Codex for everything” control layer over the desktop and browser
Microsoft’s new Legal Agent effectively turning legal AI into a Word feature. Strategic and philosophical debates continued in the background: China blocked Meta’s $2B Manus acquisition, Ed Zitron and Shaun Warman questioned AI’s unit economics and “apprenticeship” phase,
Readers who think others in their family, friends and acquaintances who are curious in knowing more about rapidly evolving AI tools/services/use cases and would benefit from being subscribers of this weekly newsletter are encouraged to share this publication link to them and invite them to subscribe
The following messages were posted on the 'All Things AI' Telegram group from Sunday 26th Apr 2026 to Saturday 2nd May 2026
via Valerii Anufriev a link to the Deepseek V4 Pro paper https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf
DeepSeek Slashes AI API Prices to One-Tenth for Cached Inputs
DeepSeek announced Saturday a lasting cut in cached input token costs to one-tenth of original prices, starting at $0.0028 per million on V4-Flash and $0.003625 on V4-Pro, stacking with a 75% discount on V4-Pro until May 5, 2026.
My understanding is that these price drops are only available via DeepSeek's inference endpoint in China. OpenRouter mentions the following about this specific provider input
To our knowledge, this provider may use your prompts and completions to train new models
DeepSeek Input Cache Price Drop!Effective immediately, the price for input cache hits across the ENTIRE DeepSeek API series is reduced to just 1/10th of the original price! Build more efficiently for less.
Reminder: The DeepSeek-V4-Pro 75% OFF promotion is still active
via David James
I encountered an interesting situation last week, the conundrum facing a client: In today's environment of deep learning and AI is signing an NDA necessary? Information acquired by tech is more invasive and sneaky than a human leak. And when this event happens a company is likely to make the employee or partner the scape goat.
China Blocks Meta's $2 Billion AI Startup Acquisition
China's National Development and Reform Commission prohibited Meta's $2 billion purchase of Singapore-based Manus, a developer of autonomous AI agents for tasks like coding and data analysis. The block, announced Monday after a security review, requires unwinding the deal announced in late December 2025; regulators had earlier barred Manus co-founders from leaving China.
PS, Both links behind paywalls
China blocks Meta's $2bn purchase of AI group Manus
Regulators had reviewed whether deal violated Beijing's investment rules
via Alba Chung
Hi, this is Alba with Alibaba cloud.
Alibaba today launched limited beta access to its HappyHorse 1.0, a video generation model designed to help creators produce high-quality, cinematic-style video content. HappyHorse 1.0 is now accessible to creators and enterprise customers globally via HappyHorse official website and through API service on Alibaba Cloud Model Studio
Alibaba Rolls Out HappyHorse 1.0 in Limited Beta
Alibaba today launched limited beta access to its HappyHorse 1.0, a video generation model designed to help creators produce high-quality, cinematic-style
My LinkedIn is https://www.linkedin.com/in/albachung
Feel free to PM me for more details.🙌🏼
Starting June 1st, GitHub Copilot will move to a usage-based billing model as GitHub Copilot supports more agentic and advanced workflows.
GitHub Copilot is moving to usage-based billing
Starting June 1, your Copilot usage will consume GitHub AI Credits.
Xiaomi Open-Sources MiMo-V2.5 AI Models with Top Agent and Coding Performance
MiMo-V2.5-Pro, a 1.02 trillion parameter model leading open-source benchmarks like 1581 Elo on GDPVal-AA and 63.8% on ClawEval. Its sibling, MiMo-V2.5, handles text, images, videos, and audio with 310 billion parameters and supports a 1 million token context for complex workflows. Available under a permissive MIT license on Hugging Face, the models come with instant framework support and a 100 trillion token grant to boost developers.
MIT License, supporting commercial deployment, continued training, and fine-tuning - no additional authorization required.
Two models, both supporting a 1M-token context window :
• MiMo-V2.5-Pro: built for complex agent and
Oh and one more thing — we're giving devs & creators 100T tokens on us. Go build something cool
100T Free Token Grant for Builders
Xiaomi MiMo Orbit-创造者百万亿 Token 激励计划
邀你参加 Xiaomi MiMo Orbit-创造者百万亿 Token 激励计划，100T Credits 面向全球用户限时发放中
MIT License, supporting commercial deployment, continued training, and fine-tuning - no additional authorization required.
Two models, both supporting a 1M-token context window :
• MiMo-V2.5-Pro: built for complex agent and
via Robby Yung, An X article about Ineffable a new AI research lab based in London which is focused on building a superlearner a system that discovers all knowledge directly through its own experience, from elementary motor skills to profound intellectual breakthroughs. No pre-training. No imitation. Just an agent learning endlessly from the consequences of its own actions in a world built to teach it.
via Arnold Concepcion a blog post by Shaun Warman which talks about how Frontier AI is sold at a structural loss because users are still teaching the models.
Three predictions for what happens when the apprenticeship ends — and who gets locked out of the workshop afterward.
Anthropic with a coalition of partners including Blender, Autodesk, Adobe, Ableton, and Splice, are releasing a set of connectors—tools that let Claude work alongside the software creative professionals rely on, so creatives can extend their reach.
Claude for Creative Work
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
US startup Poolside debuts its first open-weight model, Laguna XS.2, a 33B-A3B-parameter MoE model, and Laguna M.1, a proprietary 225B-A23B-parameter MoE model. Poolside founders are Eiso Kant and Jason Warner (former CTO of Github)
We started Poolside because we believed that to build truly capable coding agents, you need to own the full stack: data, training, reinforcement learning, inference.
Khan Academy founder Sal Khan is teaming up with TED to launch a "radically affordable" bachelor’s degree equipping students for an AI-powered workplace. Expected to announce the project in Vancouver at the TED2026 conference
Khan told the SF Standard that unlike degrees from elite universities, the new institute will be “radically affordable. We’re openly talking about sub-$10,000 for the full degree, and I’m hoping to be aggressive with things like financial aid [and] potentially geo-based pricing.”
Sal Khan launching a $10K AI degree, with help from Google, Microsoft, and Replit
The Khan Academy founder is teaming up with TED to offer a bachelor's in AI skills.
via Alex
This is super interesting.
A 13B pre-internet trained model to prove whether LLMs are more than just stochastic parrots:
In short (preliminary results and the approach is flawed in my opinion): LLMs show non-zero learning capabilities from context alone. Obviously python coding problems never touched the training data but it still is able to get some results correct.
The effect is very very small (0.05% success rate) but it shows that LLMs can come to novel (not included in training) conclusions
User
What would an electronic brain be capable of doing?
talkie-1930
An electronic brain would be capable of recording, feeling, and perceiving everything that comes within the range of the senses. It would be capable of knowing colours, sounds, tastes, and smells; it would be capable of pleasure and pain; and it would be capable of thought and volition. In a word, an electronic brain would be a perfectly sensible brain.
User
Would the electronic brain be able to replace human labor?
talkie-1930
No; for, although an intelligent electronic brain would do everything that a human brain can do, it could not do anything more. The utmost that an electronic brain could accomplish would be to think, and to act upon the thoughts; but it could not execute. Muscular labour must still be performed by muscular agency.
The article "AI's Economics Don't Make Sense" by Ed Zitron discusses several interconnected fundamental problems with how generative AI companies operate and monetize their services.
AI's Economics Don't Make Sense
If you liked this piece, please subscribe to my premium newsletter. It's $70 a year, or $7 a month, and in return you get a weekly newsletter that's usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI's finances,
via Po Chu
Cursor + Opus 4.6 broke many safety guardrails and almost destroyed a whole company
PocketOS maker says an AI agent "deleted our production database in 9 seconds."
The details of what happened should be taken with a grain of salt since some of its self-reported by the chatbot, which can be tricky.
original source
Came across this X thread which links to content from OpenAI and Anthropic describing the recommendations in how prompts should be adapted when using their latest models Opus 4.7 and GPT-5.5
Both say the same thing.
Your old prompts don’t work anymore.
But for opposite reasons.
Claude Opus 4.7 stopped guessing what you meant. It does exactly what you type. Nothing more, nothing less.
Vague
A claim from Theo that certain phrasing in commit messages triggers different behaviour from Claude Code Also sharing a link to the quote tweets. Sharing this with no commentary and I think Anthropic deserves their right to respond. Would appreciate if chatter about this is minimal or none and we wait for the dust to settle.
This is an empty repo, I'm just calling Claude Code directly. Insanity.
This is an empty repo, I'm just calling Claude Code directly. Insanity.
via Abdeali Chandanwala
Hey everyone 😊
I am looking for a AI ADs management tool which can create, manage, monitor ad campaigns on linkedin, facebook, instagram and google PPC and give report on what's working and why
Adobe CX Enterprise seems to be an option - but its in Beta I guess - no open access found to it. Any alternatives ?
OpenAI Updates Codex App for Coding and Everyday Tasks
Andrew Ambrosino from OpenAI announced the update, dubbing it 'Codex for everything' with a dynamic UI that shows progress panels, artifacts like Notion docs and GitHub pull requests, and tools for slides, sheets, and annotations. Key gains include 20% faster computer and browser control, role-tailored onboarding for finance, data science, or marketing, and a cleaner design that feels natural across tasks. Users like Sherwin Wu call it their most-used work app, praising the polish, though some note minor lag and request extras like remote connections.
- Dynamic UI for the task at hand
- 20% faster computer & browser use
- Even better slides and sheets
- Annotate in browser, artifacts, and code
- Easier to get started
- Cleaner design across the app
- Performance improvements
- (no clunky
xAI has launched Grok 4.3, achieving 53 on the Artificial Analysis Intelligence Index with improved agentic performance, ~40% lower input price, and ~60% lower output price than Grok 4.20
I was reading the documentation earlier. It's unique that it has the X Search tool
The X Search tool enables Grok to perform keyword search, semantic search, user search, and thread fetch on X (formerly Twitter). This powerful tool allows the model to access real-time social media content, analyze posts, and gather insights from X's vast data.
The release of Grok 4.3 places @xAI just above Muse Spark and Claude Sonnet 4.6 on the
The release of Grok 4.3 places @xAI just above Muse Spark and Claude Sonnet 4.6 on the
X Search | xAI Docs
Learn how to use the X Search tool for searching X posts, users, and threads.
In major news for the legal tech market, Microsoft is specifically targeting that sector with a ‘Legal Agent’ in Word. It is understood that the team that joined from Robin AI has been integral to this product launch.
Learn more: msft.it/6010vyjyA
Harvey raised $200M at an $11B valuation in March on the bet that legal AI is its own surface. The numbers held that up. $190M ARR per TechCrunch's December reporting. 100,000 lawyers across 1,300 organizations
Below is a Youtube playlist of videos from Sequoia Capitals AI Ascent 2026 event including sessions with Nvidia's Jim Fan, OpenAI's Greg Brockman, Andrej Karpathy and Nobel Prize laureate Demis Hassabis of Google Deepmind.
https://www.youtube.com/playlist?list=PLOhHNjZItNnOkkZThzULo1Ygg7JR6T3MG
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 OpenAI's ChatGPT Images 2.0 model within the Freepik tool via the following prompt
a traditional pagoda surrounded by blooming cherry trees, in the style of terraced cityscapes, light red and sky-blue, japanese photography, anime aesthetic, mountainous vistas, silver and crimson, multi-layered