Notes d’épisode
OpenAI noticed its own model, GPT-6 Astra, struggling on safety research: unnecessary prerequisites, abandoned solutions, blockers that didn't exist. So they built a test. The same coding tasks, with the same code except for names, comments and documentation, framed once as safety research and once as capabilities research. Astra completed about 16% of a hidden checklist on the capabilities version and about 10% on the safety version, roughly 40% fewer items. GPT-5.6 Sol showed essentially no gap. Most of Astra's gap remained after removing explicit refusals and written decisions to hold…
Site web: youtube.com

