Astra system card admits lower chain-of-thought monitorability as three safety staff are fired
A timeline: Astra's system card concedes a substantial decrease in chain-of-thought monitorability, METR testifies to the Senate the same day California's AG subpoenas OpenAI, and then three safety-related staff are fired verbally.
After Hugginface incident , outside auditors (METR) were let in to investigate, and tomek korbak was openai's main contact for them sep 3: openai's own astra system card admits "a substantial decrease in chain-of-thought monitorability". basically it's getting harder to read what the model is thinking. tomek, mikita balesni and jasmine wang were pushing labs to commit to keeping that reasoning monitorable sep 30: METR testifies to the senate on "rogue AI", citing tomek's paper. same day california's AG subpoenas openai over cyber incidents oct 1: news breaks that all three got fired. three different reasons, all verbal, nothing in writing (per them): tomek: "the way I communicated with METR" mikita: "speaking too much to third party safety organizations" jasmine: "that I accessed an executive's email" openai says they broke "our policies on accessing and handling sensitive company information" and that it wasn't about raising safety concerns
This post quotes another post: x.com/balesni/status/2108262814003687745