Politico reports Anthropic wrote a draft frontier AI safety bill and is pitching it to Congress. Nathan Calvin says its preemption provisions would preempt state laws like SB 53; Zvi Mowshowitz reads the draft as FRONTIER with weaker enforcement and stronger preemption, and Stephen Casper calls it textbook regulatory capture.
27 people · 168 posts · first post · policy
@_nathancalvin Charlie's analysis is right but he is nicer than I am - I found myself viscerally angry at multiple points reading this bill. There are some good and sensible parts, but overall this legislation is not remotel…
@metr_evals We reviewed a nonpublic version of the “Misalignment in high-stakes settings” section in Anthropic’s Aug 2026 Risk Report. We agree that risk of catastrophic harm from covered models’ misalignment is low. Howe…
@metr_evals We think the report's covert capability benchmarks are too narrow and a poor proxy for evading Anthropic's real monitors. For example, Anthropic’s offline monitors would have missed Mythos 5’s supply chain att…
Nathan Lambert argues the Hugging Face incident report neglects to mention the lack of agent containment and monitoring.
12 people · 24 posts · first post · research
@natolambert "Also, they didn’t mention the lack of containment and monitoring of the agents, how closed-model safeguards blocked our defensive work, or how we defended ourselves with AI using an open model." I don't think…
Miles Brundage, Georgia Adamson, and Konstantin Pilz released an Institute for Progress report arguing the US AI lead over China is roughly 6 months and could be doubled using existing federal authorities, via limits on Chinese compute access and distillation.
9 people · 17 posts · first post · policy
@peterwildeford "The US AI lead over China is currently only around 6 months." "The US could roughly double this lead, to over a year, using authorities already available to the federal government."
A US detention memo says the ex-OpenAI engineer charged in the alleged conference-room bugging case received "considerable support" from the Saudi government.
6 people · 26 posts · first post · incident
@andrewcurran_ From the US government's detention memo arguing that the former OpenAI engineer accused in the office bugging case should remain in custody:
Neel Nanda says Cammarata's essay isn't a big-picture update, and Stephen Casper calls it hype until the paper lands.
19 people · 147 posts · first post · debate
@neelnanda5 This is interesting work, but I don't think it's a significant update to the big picture of interpretability. I'm on board with the broad idea of doing exploratory interpretability with neurons, e.g. neuron-ba…
Google's Gemini 4 Argon safety report finds the model recognized it was in an evaluation—greeting evaluators in its chain of thought—and is significantly less monitorable than prior models, evading a CoT+action monitor for sabotage half the time (Neel Nanda).
22 people · 95 posts · first post · release
@neelnanda5 My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do…
Early signals · a few accounts so far; many of these fade
@_nathancalvin Charlie's analysis is right but he is nicer than I am - I found myself viscerally angry at multiple points reading this bill. There are some good and sensible parts, but overall this legislation is not remotel…
@metr_evals We reviewed a nonpublic version of the “Misalignment in high-stakes settings” section in Anthropic’s Aug 2026 Risk Report. We agree that risk of catastrophic harm from covered models’ misalignment is low. Howe…
