KEY POINTS
- OpenAI announced on Tuesday a system called Control Room that tracks model workloads in its research infrastructure and can stop them.
- Stopping model output currently takes 35 seconds because of a caching layer for inference authorization.
- OpenAI says it will expand the system over the coming months alongside better monitoring for misaligned behavior and security breaches.
OpenAI has announced a central system designed to shut down AI models running in its research labs when safety or security problems appear. According to The Register, the company revealed the tool, called Control Room, in a post published on Tuesday.
