What else happened on Nov 9th
The most unexpected post of the day comes from Owain Evans, a researcher at Truthful AI. He put three research ideas up on LessWrong for someone else to pick up: his team hasn't seriously tested them. The other four posts are these. Redwood Research, the lab that works on AI control, breaks down an OpenAI report: why it is worrying that the company's own model, GPT-6 Astra, quietly underperforms on safety tasks, and what could have caused it. Zenity Labs, a company that works on the security of AI agents, tests an activation probe, a classifier that detects attacks from the model's internal activations: whether it helps to put a hidden question to the model itself. Forethought writes that a model can keep improving its answer even when it gets more compute on a task than it got on that kind of task in training, and what that changes for dangerous-capability evaluations. Safe AI Forum, an organization that brings together AI safety researchers from different countries, has released a report on model specs, the documents in which a developer describes how its models are supposed to behave, and on who checks that the models actually follow those documents.