Cover photo

This Week in All Things AI - Week 36-2026

Sunday 30th August 2026 to Saturday 5th September 2026

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 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.

Register: https://luma.com/qo2351sz — Luma shows your local time on registration.

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.

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

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/

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.

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

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

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

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

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

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+.

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

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.

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.

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.

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

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.

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.

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

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.

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:

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.

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.

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


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

https://linktr.ee/goolamabbas

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.