Gives OpenAI a centralized kill switch for risky training and inference runs across its research infrastructure.
OpenAI introduced Control Room, a central system that tracks model workloads across its research infrastructure and can rapidly halt them when safety or security concerns arise, stopping workloads in as little as 35 seconds.