Anthropic gives METR employee-level access and drops its publication veto
Anthropic announced that six METR staff now get employee-level access: company hardware, office space, and the internal tools its own risk teams use. METR can publish findings without Anthropic’s approval; Anthropic may only redact narrow categories like customer data and attorney-client material, and METR keeps the right to say publicly whether a redaction was justified. Limited access and publication control are the two places where third-party evaluation most often falls short, and this moves both; watch whether the tiered-access framework (SecureBio is reviewing the bio-chem section of Anthropic’s August risk report, Hugging Face is in early talks) becomes a template for other frontier labs.
Meta releases Muse Spark Open, tuned for Mac Studio
Meta released Muse Spark Open: natively multimodal, 816B total parameters with 42B active, a 1M-token context window, weights on Hugging Face, optimized with Apple’s MLX framework to run locally on a pair of M5 Ultra Mac Studios (see also TechCrunch). Meta calls it the most capable US open-weight model to date, and that framing is the story: Chinese labs have set the pace in open weights since late 2024, when Artificial Analysis marked their frontier overtaking the US, and this is a direct answer. Meta’s sources only say you need two of the machines; my read is that at that price, “local frontier model” means enterprises and serious developers, not laptops.
FT: Anthropic told investors Q3 revenue hit $16.8bn
Per the FT (paywalled; figures corroborated by Yahoo Finance), Anthropic disclosed Q3 revenue of $16.8bn, up about 45% from $11.5bn in Q2, with an adjusted operating margin of 11% for the year so far and $1.70 of revenue per $1 spent on compute in the quarter. That last ratio is the number to remember: self-reported data in the “is inference profitable” debate is rare, and this arrived weeks before a planned November IPO.
NVIDIA opens Nemotron 3.5 Super VL plus edge-scale Bolt models
NVIDIA released Nemotron 3.5 Super VL, a 120B open vision-language model, alongside two small Bolt models for edge devices. Read together with Meta’s same-day release: the open-weight race heated up this week, and this time US companies are doing the heating.
OpenAI launches the Rosalind Academic Plan
OpenAI is offering academic labs sponsored access to GPT-Rosalind for life-science research. The announcement doesn’t spell out the incentive, but my read is that subsidized access buys the lab validation cases in real science, and a published paper that used your model is marketing few ad budgets can match.
Mistral introduces Timestral
Mistral released Timestral, lightweight probabilistic time-series forecasting models in four sizes that run zero-shot on CPUs. Foundation-model economics are spreading beyond language: a model that forecasts zero-shot can cover many of the business metrics that each used to need their own trained model, with fine-tuning still available where domains demand it.
Simons Foundation leads a $15M effort to help mathematicians adapt to AI
The Simons Foundation, together with Anthropic and OpenAI, committed $15M in combined funding and resources to help mathematicians work through what AI tools are doing to research practice; the foundation manages and awards the money. A discipline-level funder now treats “AI changed how research works” as a problem worth dedicated money, which says more than any single benchmark result.
Figma ships Specs to give agents design rules
At Config India, Figma introduced Specs, a Markdown container on the canvas where teams write design rules for both humans and AI agents to follow; code layers entered open beta, and the Figma Agent is now generally available with 20 new MCP integrations (see also the Specs post, TechCrunch). It is the same move as AGENTS.md in the coding world: context for agents becomes a first-class product primitive, now in design tools too.
Research radar
The Thinking Leakage in Reinforcement Learning for LLMs
Hybrid reasoning models like Qwen3 improve after RL training in NO-THINK mode, and this paper traces the gain to “thinking leakage”: the model quietly reactivates its reasoning mode rather than learning to answer without it. If you do RL post-training, the warning is direct: a rising metric does not mean the mechanism you intended is the one doing the work.
Cortex: Consolidating Context into Parametric Memory
Cortex proposes memory consolidation: distill the performance gains a context provides into model parameters, so they survive after the context is removed. It frames model editing, skill learning, and behavior steering as one objective, worth the read if you work on model memory or continual learning.
Sparse Autoencoders for Mechanistic Interpretability in NLP: A Critical Survey
A critical survey of 138 SAE interpretability papers finds that design assumptions, evaluation targets, and experimental setups vary so much across the field that it is hard to draw consistent conclusions about whether SAEs actually work. For interpretability researchers who lean on SAEs, this is useful triage: which validity claims hold up, and which evaluation setups are themselves the problem.
One line for today: Read Anthropic’s two stories together, $16.8bn in quarterly revenue sprinting toward an IPO and publication rights handed to outside evaluators the same day. At that scale, “don’t take our word on safety” stops being a gesture; my bet is the timing is no accident, and outside scrutiny is now part of the IPO story itself.