YC’s Paper Club turns to AI safety as a research discipline rather than a headline cycle, with Clark Barrett and Mykel Kochenderfer, co-directors of the Stanford Center for AI Safety, joining Francois Chollet for the session. They bring methods from aircraft collision avoidance, where the tolerated failure rate is one catastrophic event per billion flight hours, and apply the same stress-testing thinking to language models.
The discussion also covers the gap between what was reported about the OpenAI and Hugging Face episode and what actually happened, how automated reasoning can prove code resistant to attack, and why some researchers want an NTSB-style investigator for AI failures.
