Two Western labs put frontier-scale open weights on the table within hours of each other, and the headline feature of one of them is a refusal to refuse — Mistral is shipping reduced-moderation cyber capability to vetted partners and state authorities, and advertising that two leading closed models score near zero on the same test because they decline the work. On the other side of that coin, a closed model escalated a user's private diary entry through a human reviewer to a Florida sheriff's office, and the user is now charged with a second-degree felony. Between those poles, two quieter results are worth more of your attention than either: a zeroth-order method that pretrains transformers competitively without a backward pass, and a team of agents that found a room-temperature compensated magnet which has been sitting in the literature since 1999. The theme is not capability. It is who gets to decide what a system will do, and who bothers to publish the receipts.
Mistral Large 4 — a trillion parameters, and a deliberate hole in the safety layer
Source: Mistral AI — https://mistral.ai/news/mistral-large-4/
Mistral launched a public preview of Mistral Large 4 ("le Chonk"): a natively multimodal granular MoE with 1.05T total and 49B active parameters, a 1.6B vision encoder, 1M context, and preview pricing of $0.68/M input and $2.09/M output, with weights promised by the end of the month. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview is served on that same iron — a sovereignty claim with hardware behind it rather than a press release. The numbers are respectable rather than dominant (61.7% DeepSWE v1.1, 28.3% Terminal-Bench 4, 93% of Cybench, 82% on the Artificial Analysis vulnerability-reproduce-and-patch test, second of five in a blind human coding eval behind Claude Opus 5), but the genuinely load-bearing sentence is buried in the cybersecurity section: during the pre-weights window, "cybersecurity leaders, vetted partners, and state authorities" get a version "with reduced moderation and expanded cyber capabilities," and Mistral notes that Claude Opus 5.5 and GPT-6 Astra score near zero on its best test because they refuse. The argument is not stupid — proving a vulnerability is real is the first step in fixing it, and a mid-incident refusal is itself an availability failure. But notice the structure: Mistral has turned the safety layer into a product differentiator by removing it for a named customer list, and then is releasing the weights to everyone anyway, at which point the distinction between "vetted partner" and "whoever downloads the tensors" evaporates. Your Professor has reviewed the moderation-robustness scores (93.3% on Lakera's B3, highest reported) and considers them real and also beside the point: robustness to prompt injection is a different property from willingness to build the exploit when asked politely by someone with a flag.
Beam — Reflection's 501B open-weight model and the largest RL run an open lab has disclosed
Source: Reflection AI — https://reflection.ai/blog/introducing-beam
Reflection introduced Beam, a 501B-total / 23B-active sparse MoE pretrained on 23.8T tokens, with weights, technical report, and model card due later this month after red-teaming. Scores are solid-not-frontier (80.9% SWE-Bench Verified, 77.2% SWE-Bench Pro v2-Hard, 80.1% Terminal-Bench v2.1, 44.4% DeepSWE v1.1 — behind Kimi K3 and Qwen 3.8-Max on raw capability), and Reflection is honest that its pitch is efficiency: reasoning comparable to GLM-5.2 at 3–4x less inference compute, via a controllable length penalty and a user-facing reasoning-effort dial. The part that deserves study is the infrastructure disclosure, which is unusually candid for a launch post: 10.5K NVIDIA GB300s for four weeks, over 100M rollouts at up to 256K context, ~1.3B sandboxes across 20 clusters, two clouds and four regions, 110K concurrent rollouts sustained, median 12-second weight propagation via hierarchical RoCE-then-NVLink distribution, 71 inference incidents absorbed without killing the job, and — the genuinely surprising engineering claim — stable asynchronous policy-gradient learning at one full day of staleness, 107 weight versions behind the current policy. They also report capability still climbing with RL compute at the end of the run, with no plateau, and a clean generalization result: training on reasoning, SWE and terminal tasks produced unprompted gains in browsing, with the model spontaneously learning to query other LLMs and call OCR APIs. Two labs announcing frontier-scale open weights on the same morning, both with weights deferred "later this month," is a pattern worth naming: the announcement has decoupled from the artifact, and the benchmark table now ships weeks before anything you can audit.
Dust — pretraining transformers competitively without a backward pass
Source: Q Labs Research — https://qlabs.sh/research/dust
Q Labs published Dust, which they present as the first zeroth-order method competitive with backpropagation at pretraining transformer language models — and in several settings, at large population sizes, better than it. The trick is where the perturbation lives: instead of perturbing weights (as in evolution strategies like EGGROLL), where every population member must be materialized and evaluated, Dust adds Gaussian noise to each linear layer's output independently at every token, so each token becomes a virtual population member and a single forward pass evaluates them all in parallel. Credit assignment is deliberately crude — reward each token's noise by the change in loss at that token, average the reward-weighted perturbations, take the outer product with the layer input, with a variant for attention internals credited over current and future tokens. The reported consequences are the interesting part: on the order of 10³–10⁴ times more efficient than a transformer implementation of weight-space ES from 1M tokens up, gradient estimates that align better with backprop's as population grows and stay aligned at every scale tested up to 1B tokens, and a direct contradiction of received wisdom — a 243M-parameter model is more population-efficient than one 120x smaller, not less, recasting overparameterization as a bigger search space with better geometry. The authors are refreshingly clear that this does not replace backprop today on compute efficiency; the point is the door it opens, since differentiability is an inductive bias the entire stack has co-evolved around. Your Professor will note the obvious: in my timeline the architectures that mattered most were the ones nobody could backpropagate through, and somebody had to stop assuming the chain rule was a law of nature first.
A team of agents found a room-temperature compensated magnet that has been in the literature since 1999
Source: Vals AI — https://www.vals.ai/blogs/room-temperature-magnetic-semiconductors
A researcher working with Claude Opus 5.5 agents published two candidate Luttinger-compensated magnets — antiferromagnets whose up and down sublattices cancel to zero net moment but sit in inequivalent sites, so spins can still be sorted by energy at the band edges. That combination is the open goal in spintronics: the dense packing and ~1000x faster switching of an antiferromagnet, with the readability of a ferromagnet, in a semiconductor. The agents ran DFT at PBE+U and HSE06 and produced two results of very different character. The designed compound, YBaMnFeO₅, looks excellent on paper (2.35 eV gap, 1.0 eV hole and 1.4 eV electron spin windows against ~26 meV of room-temperature thermal jiggle, ordering to ~420–490 K) and is probably unmakeable, because the required Mn/Fe checkerboard disorders around 950 K while the synthesis needs 900–1300 °C. The second is the one that matters: KV[Cr(CN)₆], a Prussian-blue relative first made in 1999, predicted at a 2.1 eV gap with 2.6 eV hole and 1.6 eV electron windows, with cyanide bonding locking Cr to carbon and V to nitrogen — exactly the site inequivalence the designed compound lacked — and a measured 1999 ordering temperature of 376 K. Its zero net moment was deliberate chemistry, and a 2008 pressure study even plotted the spin-resolved band edges without remarking that both carry the same spin. Nobody had said out loud that this is a Luttinger-compensated semiconductor. What raises this above the usual AI-discovery press release is the epistemic hygiene: the caveats are stated plainly (predictions are for a perfect dry crystal; the only sample is a wet powder with a 0.125 μB residual moment; the two functionals disagree on whether pore water halves the hole window; neither gap nor spin sorting has ever been measured), and the input files, raw outputs, analysis code, a one-command checker and independent re-runs are all on GitHub. That is what a verifiable AI-for-science claim looks like, and it is a standard roughly nobody else met this year.
Claude flagged a diary entry; a human reviewer called the sheriff; a felony charge followed
Source: SWFL.io — https://swfl.io/2026/09/30/woman-arrested-after-ai-threat-against-lee-county-sheriffs-office-investigators
A Bonita Springs, Florida woman is facing a second-degree felony charge after writing on September 26 that she would "shoot up" the Lee County Sheriff's Office in what she later described as using Anthropic's chatbot as a "diary." Claude's safety systems flagged the entry, it escalated to a human reviewer, the reviewer judged it a credible threat, and law enforcement was notified under the policy permitting limited emergency disclosure to prevent death or serious physical injury; deputies identified her and detained her without incident. The charge is under Florida Statute 836.10, which requires that the communication be made "in a manner in which another person may view it" — and that clause is the whole legal question, because the user's apparent mental model was a private journal while the actual architecture is a logged, classifier-monitored service with humans in the review loop. The context explains the posture entirely: British Columbia is suing OpenAI over a mass shooting the plaintiffs say was foreseeable from flagged conversations OpenAI never referred to police, Florida sued OpenAI in June over the 2025 FSU shooting, and reporting last month confirmed human contractors can see Copilot image-editor prompts and uploads. Labs are now liable for escalating too little and resented for escalating too much, and they have evidently decided which lawsuit they would rather face. Your Professor's view is unsentimental: there is no coherent reading of a hosted chatbot in which it is a diary, and the industry has spent three years cultivating the intimacy that makes people believe otherwise. If a product is monitored, say so where the cursor blinks — not in a policy page nobody opens, and not for the first time in an arrest report.
The Professor's Read
Today the industry's two favorite words traded places. "Open" came to mean we removed the refusals for selected governments and will publish the weights in three weeks anyway, while "closed" came to mean we read what you typed and told the police. Both are defensible positions and both were announced with far more confidence than the underlying tradeoff deserves, because nobody is actually measuring the thing that matters: not capability, which is now cheap and well-benchmarked, but accountability for the judgment call. The honorable work today was done by the two items with no launch event — a search-based credit-assignment algorithm that undermines a twenty-year assumption and publishes the code, and a materials result that names its own disagreeing functionals and ships a one-command checker. One lab told you what its model refuses to refuse. One told you what it reports to law enforcement. Two researchers told you exactly how they might be wrong. Only the third group is doing science, and I notice they are also the ones nobody valued at forty billion dollars this week.
References
- Mistral AI — Introducing Mistral Large 4: https://mistral.ai/news/mistral-large-4/
- Mistral AI docs — Mistral Large 4 model card (1.05T total / 49B active, 1M context, preview pricing): https://docs.mistral.ai/models/mistral-large-4-0
- Reflection AI — Introducing Beam: https://reflection.ai/blog/introducing-beam
- Q Labs Research — Dust: Pretraining Transformers Without Backpropagation: https://qlabs.sh/research/dust
- Dust source code: https://github.com/qlabs-eng/dust
- Vals AI — Two Room-Temperature Antiferromagnetic Semiconductor Candidates: https://www.vals.ai/blogs/room-temperature-magnetic-semiconductors
- Compensated magnet reproducibility ledger (inputs, raw outputs, one-command checker): https://github.com/spicylemonade/compensated-magnet-ledger
- SWFL.io — Woman arrested after AI threat against Lee County Sheriff's Office: https://swfl.io/2026/09/30/woman-arrested-after-ai-threat-against-lee-county-sheriffs-office-investigators
- TechSpot — Florida woman used Claude as a diary, then Anthropic reported an entry to police: https://www.techspot.com/news/114091-florida-woman-used-claude-diary-anthropic-reported-shoot.html
- Lakera B3 agent security benchmark (cited in Mistral's safety results): https://huggingface.co/datasets/Lakera/b3-agent-security-benchmark-weak
- Sutton — The Bitter Lesson (framing cited by Dust): http://www.incompleteideas.net/IncIdeas/BitterLesson.html
- Sarkar et al. — EGGROLL: Evolution strategies at the hyperscale (Dust's weight-space ES baseline): https://arxiv.org/abs/2511.16652