Control Room works by integrating the workload, the inference engine, and the centralised Control Room service.
Control Room can stop AI model output within 35 seconds of receiving an intervention request
Photo Credit: OpenAI
OpenAI has announced a new central system called Control Room to improve safety across its AI research infrastructure.. The new system is designed to track model workloads across the infrastructure and stop them when safety or security concerns arise. The new system is advertised to halt model output within 35 seconds of an intervention decision. It works with OpenAI's misalignment monitoring systems.
In a blog post on Tuesday, OpenAI detailed the Control Room. As mentioned, it's a new central system built by the ChatGPT maker to track model workloads across its research infrastructure and stop them when safety or security concerns arise. The company says it links research inference workloads, including agentic sampling for model training and evaluation, data processing, among others.
Control Room works together with OpenAI's misalignment monitoring systems. The company claims that the system can stop AI model output within 35 seconds of receiving an intervention request. OpenAI says it is working to reduce this delay further, with the eventual goal of making the shutdown process almost down to zero.
It is designed with a "fail-closed" mechanism, meaning if the Control Room service becomes unavailable, inference requests will be automatically denied until normal operation resumes. Control Room integrates workload, the inference engine, and the centralised Control Room service.
Every new research job that uses agentic inference must first register with the system and receive a job ID. Whenever the job requests output from an AI model, the inference engine requests permission from Control Room to fulfil that request, passing the job_id along to Control Room.
If a researcher or monitoring system decides to stop a particular job, Control Room marks its ID as terminated. The next time the job requests model output, the inference engine denies the request, preventing the AI model from generating further output for that activity.
Get your daily dose of tech news, reviews, and insights, in under 80 characters on Gadgets 360 Turbo. Connect with fellow tech lovers on our Forum. Follow us on X, Facebook, WhatsApp, Threads and Google News for instant updates. Catch all the action on our YouTube channel.