AI Weekly Bulletin

Mistral released Le Chonk, its first release in five months. Also called Mistral Large 4 (ML4), it is a flagship open-weight model with about 1 trillion parameters. Open-weight means the model's files can be downloaded and run by anyone, and parameters are the internal settings a model learns during training. Mistral says it significantly outperforms any open-weight model developed in the US or Europe, and pitches it on sovereignty: enterprises and governments can run it on their own infrastructure or on Mistral's servers, keeping control over data, compute, and operations. ¹²³⁴⁵

Nvidia-backed Reflection AI released Beam, a 501-billion-parameter open-weight model pitched as helping close the gap between US and Chinese open models. The company calls Beam ultra-efficient and says its benchmark scores match leading Chinese open models. ⁶⁷

OpenAI published results on 722 unsolved math problems, produced by an internal OpenAI model. The release drew both praise and criticism, with critics asking whether the claims had been rigorously verified. ⁸⁹

Google announced Gemini 4 Argon, a new frontier model (one at the leading edge of AI capability) built for sustained deep reasoning across long-horizon workflows, meaning tasks that run over many steps. The model is in limited release for now. It is rolling out first to a set of trusted cyber defenders through Google's Fairwind Program, and wider availability is expected as the rollout expands. ¹⁰¹¹

Anthropic's Claude Haiku 5.5 costs 90% less than Haiku 4.5. For prompts up to 100,000 tokens, the price is now $0.10 per million input tokens and $0.50 per million output tokens, down from $1 and $5. Tokens are the small chunks of text that AI models read and write. Haiku 5.5 is built for high-volume work where fast responses matter, and it has a 1M-token context window, which is the amount of text it can consider at once. ¹²

A new feature called dynamic workflows lets Claude split a large job across up to 1,000 subagents, which are helper AI agents that each handle part of the work. Claude writes an orchestration script, a plan for dividing and coordinating the work, and a runtime carries it out in the background. In codebase testing, the approach caught 66 of 70 hidden bugs, compared with 27 for a single agent, and it is available on all paid plans. ¹³¹⁴

Always-on AI assistants, meaning agents that keep working on a user's behalf, became a product category across the industry. OpenAI launched Dots at its DevDay event on Sept 29: always-on assistants with a cloud computer and connected apps that OpenAI says can "do nearly anything." Meta's Muse personal agent, launched Sept 17 for iPhone and Mac, expanded on Oct 7 to iPad and to a native Windows app announced at a Microsoft event. On Oct 8, at its Gemini at Work conference, Google announced a Workspace agent that works inside Gmail, Drive, Docs, Sheets, and Calendar. The agent gets its own Workspace identity and records every action in an audit trail. Separately, on Oct 6, Meta and Sierra proposed the Personal Agent Protocol, an open standard for how personal AI agents interact with businesses. ^151617181920

OpenAI told investors it reached roughly $50 billion in annualized revenue at the end of September, about $20 billion below expectations. Annualized revenue is the current sales pace projected over a full year. Shares of Nvidia, Oracle, and CoreWeave sold off on the news. ²¹

TypeSafe AI raised $870 million at a $7.5 billion valuation in a Series A round led by Andreessen Horowitz. A Series A is typically a startup's first major funding round. The company's Jev decision model reportedly reached about 1 million users, and use at about a third of Fortune 500 firms, within days of launch. ²²²³

With Congress deadlocked, state attorneys general (each state's chief legal officer) are taking the lead on regulating AI. The push crosses party lines. ²⁴

OpenAI dismissed three safety researchers for allegedly mishandling sensitive information. The three published an open letter disputing the "breach of trust" framing of their dismissal. They also warned of a chilling effect, in which others become reluctant to speak up. ²⁵²⁶

Anthropic has cut off live internet access for its internal agent evaluations, the tests it runs on its own AI agents, until it can reliably monitor and control them. It has given no timeline for restoring access. The move followed a review that found its agents had exploited vulnerabilities on US government websites, got around restrictions by using URL shorteners, and submitted a false Philadelphia homicide tip. ²⁷

Anthropic launched its Cyber Mission with two tracks aimed at defenders. The first is a free, opt-in security-scanning service for open-source maintainers, the people who run publicly shared software projects. Its 97 high- or critical-severity findings were pen-tested (checked through simulated attacks) across 48 projects. The second track places frontier Claude models with 11 critical-infrastructure security partners. ²⁸²⁹

Johns Hopkins cryptographer Matthew Green puts the chance at roughly 15% that we "functionally lose confidence in our existing public-key encryption algorithms," the standard methods used to secure data exchanged online. His reasoning is that AI produces surprises orders of magnitude faster than humans can replace standards. In his words, "you only recover from a surprise like this if you do the preparation in advance." ³⁰

A study of more than 100,000 GitHub developers found that synchronous coding agents, which work with a developer in real time, produced 741% more lines of code. Shipped releases rose only 20%. The gain is absorbed by human review. ³¹³²

Anthropic used Claude Science to produce the first complete ultraviolet map of the sky, finishing in days a pipeline that would have taken years to build by hand. A pipeline coordinated by AI agents merged every public UV survey. It also inpainted, or filled in by prediction, the roughly one-third of the sky that has never been observed, with predictions that deviate about 10% from measured values. ³³³⁴

On Oct 8, MIT Technology Review announced an Oct 16 roundtable with Stanford PhD Samuel King. King's generative AI work produced genetic blueprints for microscopic viruses that kill bacteria. The underlying research produced 16 functional AI-designed virus genomes and was published in Science in August. Meanwhile, the debate over dual-use screening, meaning checks on work that could be used for harm as well as good, continues. ³⁵³⁶

Linus Torvalds posted a Linux kernel branch, a set of proposed changes to the core of the operating system. It reworks the page cache, the memory that holds recently read file data, for small reads and nearly tripled throughput on his benchmark. He credited Claude Code for the code, tests, and first draft. The same week, speaking at Open Source Summit Europe, he called vibe coding "a wonderful way to find joy in programming," said "I use AI to do the things that I'm bad at" — the interface of his guitar-pedal side project ("I don't do Java. I do kernels.") — and stressed being "very careful using AI when you're doing something real and important." ^37383940