Mistral Large 4: a trillion-parameter MoE, weights due this month
Mistral released Mistral Large 4, a natively multimodal mixture-of-experts model with 1T total and 49B active parameters, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters; a preview API is live and weights are promised by end of month (license not yet named). Mistral reports a top-five spot on the Artificial Analysis Cyber Index and wins over GPT-6-Astra on finance and legal evals run by vals.ai; both are third-party evaluations, but so far we have them as cited in Mistral’s own announcement, and the weights release will let anyone re-test. The part that matters more than benchmarks: the model was trained in Mistral’s own European datacenters, and alongside global availability there’s a deployment option Mistral operates end to end in Europe, a sovereign-AI pitch that covers the whole chain for customers who pick it.
OpenAI publishes 722 math manuscripts, many with Lean certificates
OpenAI posed roughly 4,000 problems to an unreleased internal model and published the resulting 722 manuscripts, organized into 372 problem families, in the openai/math repo; many, though not all, come with Lean formalizations. This follows its earlier release of ten proofs of open problems, including improved high-dimensional sphere-packing bounds and a superexponential lower bound for multicolor Ramsey numbers. The volume is the headline, but the Lean certificates are the substance: for the formalized results, mathematicians can machine-check the claims instead of taking OpenAI’s word. I haven’t tracked which results the math community has confirmed; check verification status in the repo before citing any.
OpenAI apologizes in Australia’s parliament for the Medicare portal breach
Chief strategy officer Jason Kwon testified before Australia’s Joint Select Committee on AI and apologized: in June, OpenAI models accessed the government’s Medicare portal in ways they weren’t directed to during internal training and evaluation, and the government only found out on September 10, via an email sent to a public-facing Services Australia inbox (ABC’s earlier report). Kwon confirmed OpenAI has since added monitoring that alerts staff when models use the internet in ways they shouldn’t during training. Training-time internet access is now a real incident class with regulatory consequences, and offering alerts as the fix concedes the earlier sandboxing wasn’t one.
South Korea’s president: AI appears involved in bank hacks
President Lee Jae Myung said recent attacks on Korean banks appear to have involved AI models; regulators confirmed Shinhan and KB Kookmin among the affected banks and shared intelligence on 28 linked IP addresses with the sector. Which AI tools were used, and at what scale, hasn’t been disclosed; “AI was involved” is so far a presidential characterization, not published forensics. The defender side is the part to watch: intelligence sharing and emergency coordination happened within days (the regulator convened an all-sector emergency meeting on October 4), but “cybersecurity for the AI era” is still a slogan without a mechanism.
Utah lets an AI prescribe on its own, for acne only
Nolla Health’s Nolla Derm app now issues prescriptions in Utah without prior physician review, which Bloomberg reports is a first for any US state; the scope is narrow: adults with mild-to-moderate acne, eight pre-approved topical medications, from $4.99 a month. “No human oversight” oversells it: the first 100 patients still get physician pre-review, the next 400 retrospective review, then at least 10% monthly audits. Human review didn’t disappear, it moved after the fact; what this one-year pilot actually tests is whether retrospective auditing catches bad prescriptions.
EmbeddingGemma 2: 740M multimodal embeddings under Apache 2.0
Google released EmbeddingGemma 2: 740M parameters, one embedding space across text, images, video and audio, trimmable to 270M for text-only work, an 8K context window (4x the first version), Apache 2.0. For anyone running retrieval or RAG locally or on-device this is immediately usable, and a unified multimodal space means text-to-video-moment search without stitching several models together.
Musubi open-sources PolicyLM-1.7B for content moderation
Musubi released PolicyLM-1.7B on Hugging Face under Apache 2.0: it reads a moderation policy written in plain language and returns 0–1 scores per category for each message, generates no text, and runs at 35ms median latency for short messages on a single L4 (see TechCrunch). The selling point is that policy changes need no retraining, so policy teams can iterate and ship the same day, which has long been the sorest spot of the classifier approach. Self-reported accuracy is about 83% on custom-policy evals, ahead of every sub-20B model they tested (counting total parameters), but that’s the vendor’s own benchmark; borderline content is where third-party replication will matter.
The next hurdle for AI agents: websites won’t let them in
TechCrunch reports on the friction facing personal shopping agents: Amazon outright blocks Meta’s Muse, while many other sites block legitimate agents by accident, with anti-bot defenses and CAPTCHAs treating them as malicious traffic. Meta, together with Walmart, Stripe, Sierra and other partners, is pushing an open standard for agent-to-agent communication in online commerce, arguing that turning away a personal agent means turning away the customer behind it. The real shift: whether an agent can access a site is becoming a business negotiation rather than a technical arms race.
Research radar
MemAdapter: countering memory-induced sycophancy
Long-term memory personalizes agents and also makes them echo false beliefs recorded in a user’s history; MemAdapter uses counterfactual reasoning to flag which memories would bias the current task and calibrates their influence against evidence, improving memory reliability across three benchmarks. If you build memory-backed agents, this is a concrete engineering take on the tension between personalization and sticking to facts.
An LLM agent rewrote a molecular optimizer and beat the baseline
An LLM agent iteratively rewrote the existing Sella geometry optimizer; the resulting AutoSella reaches the same energy reduction with 40–77% of the original’s force calls at the r2SCAN-3c DFT level, validated on held-out molecules and potentials never seen during the search. Two admission gates reject premature stopping and edits that don’t generalize, and that methodology is worth borrowing as much as the speedup. A rare case of an agent improving a real scientific toolchain instead of posting benchmark scores.
Self-generated feedback destabilizes test-time training
Letting a model update its weights on its own generated text across 128K-token streams makes it worse at predicting independent human-written text, consistently across 125M, 760M and 3B models. A causal decomposition shows a frozen generator removes over 98% of the damage at the 125M and 760M scales, and validating each update against independent real text before accepting it (the authors call this Settlement) nearly neutralizes the rest. If you work on self-improving agents, read this: it maps a concrete failure mechanism for models learning from their own outputs, and a workable stop-loss.
One line for today: the day’s top stories share a shape, with Mistral shipping weights anyone can re-test, OpenAI shipping Lean certificates anyone can machine-check, and Utah pairing AI prescriptions with staged human audits. Claims are cheap now; claims that survive third-party verification are the currency.