OpenAI builds a kill switch for its research models
OpenAI's research models can now be stopped 35 seconds after someone decides to, the gap belonging to "a caching layer for inference authorization," which keeps saying yes for a while after everyone else has said no.