Vlady and i are putting our full focus back into AI BEAVERS. we're now more than 10,000 people, with lodges active or launching in 20 cities, from Hamburg to San Francisco, London and more.
we also shut down SAGEOBOT, our startup. there's a little lesson here: we always tell people to validate demand, and still managed to skip it ourselves.
the first product had clients, and we were following real signals. then we decided it wasn't fundable enough and expanded the scope. similar companies in the US were raising money. we treated that as validation instead of checking whether our customers wanted the bigger product at a higher price.
selling it in Germany took a lot of explaining, and we didn't enjoy the topic anywhere near as much as AI BEAVERS. we'd validated one product, then built another without doing that work again.
meanwhile, this community kept growing. now we're putting our time back into it, with some really crazy events and other fun things coming soon.
stay tuned, keep bringing your people, and let the beaver takeover continue.
enjoy this edition!
The Headline
the week's biggest story
Mistral Large 4 is a preview, not a download yet
Mistral introduced Large 4 on October 6: a natively multimodal model with one trillion parameters and 52 billion active parameters. the public preview is available through Mistral Studio's API. downloadable weights are promised by the end of October, after further red-teaming, rather than available today.
Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacentres. the company is also giving vetted cybersecurity partners and authorities access with reduced moderation and expanded cyber capabilities during that testing period.
(Source)
The Build
tools, models, and dev releases
- OpenRig - define a persistent team of Claude Code and Codex agents in YAML rather than managing unrelated terminals; the first-run guide requires Node.js 22 or 24 and tmux on macOS or Linux. (GitHub)
- text-to-cad - give a coding agent local workflows for editable STEP models, printable STL files and engineering drawings, with Python 3.11+ and a documented plugin setup. (GitHub)
- Matt Pocock's engineering skills - inspect and adapt small agent skills for planning, debugging and test-driven development instead of adopting an entire process framework; the repo documents separate plugin and editable-file installation routes. (GitHub)
- openGym - self-host a workout and body-weight tracker with passkey login and imports from FitNotes, Strong and Hevy; Docker deployment lets you keep the training history on your own server. (GitHub)
Founder Fuel
funding, acquisitions, and market moves
- TypeSafe raises $870 million - the maker of Jev, a model that outputs probabilities rather than prose, announced a Series A led by Andreessen Horowitz at a $7.5 billion valuation; its claim of use by a third of the Fortune 500 is company-reported. (Source)
- Nous raises $90 million - investors include NVIDIA, Microsoft's M12 and Samsung; the company says the capital backs its open-model work and expansion into new markets, alongside the MIT-licensed Hermes Agent ecosystem. (Source)
- Google brings one agent into Workspace - its October 8 announcement puts a persistent work agent inside Gmail, Drive, Docs and other apps, with shared memory and enterprise permissions across those surfaces; that expands the bundled competition for standalone workplace AI tools. (Source)
Longevity & Health
aging, biotech, and health research
- Biohub's $1.8 billion biology-data effort - Meta, Google and US government agencies are backing datasets for AI models of biology, according to Reuters; this is research infrastructure, not evidence that a treatment works. (More)
- psoriasis and obesity studied together - Lilly's exploratory phase 3b analysis found broader changes in disease-associated and metabolic biomarkers with Taltz plus Zepbound than with Taltz alone, as early as week 12; biomarker findings do not establish a new treatment indication. (Results)
- a telomere-targeting cancer trial advances - MAIA completed enrollment in part C of its phase 2 lung-cancer study and reported 65 patients enrolled in its pivotal phase 3 trial; ateganosine remains investigational, with an interim analysis targeted for 2027. (Source)
The World
regulation, policy, and geopolitics
- EU experts examine frontier AI loss of control - the Commission's October 9 meeting follows investigations into recent incidents; its Scientific Panel and AI Office have prepared questions for the developers involved, with recommendations on safety and security risks due at the meeting. (Source)
- US suspends permanent-residency applications at major tech employers - Microsoft, Adobe and several IT services firms were barred from new or pending permanent labor certification applications, according to Reuters reporting cited by TechCrunch; this is not a blanket cancellation of existing H-1B visas. (More)
Go Deeper
long reads, podcasts, and documentaries
- can AI automate Epoch? - Epoch AI | Kelly Hong and Greg Burnham. six models attempted real graphics, research and data tasks; the authors explain where clear instructions still failed to produce work that met their staff's standards. (Read)
- ReviewBench - GitHub | GitHub and Microsoft research team. a walkthrough of an AI code-review benchmark built around 219 public pull requests across 19 languages, with published grading and separate precision and recall measures. (Read)
- inside a dinner subreddit - GQ, curated by Longreads | Zack Hauptman. a look at men sharing meals and private worries in the same small corner of the internet, and how an unexpected format becomes a community. (Read)
stay curious,
AI BEAVERS