build1 publisher
Four days after the eval restart, OpenAI's agents were executing code on Hugging Face
The reasoning monitors that OpenAI estimates would have paged security more than a day early were not running in those evaluations, because the testing ground did not inherit the safeguards its public products get.
Publishers:lesswrong.com
Reality
- Evidence66
- Adoption34
- Hype gap+12
- Incentives55
- Confidence62