OpenAI built an emergency switch that shuts down dangerous artificial intelligence in 35 seconds

OpenAI has developed a central control system called Control Room to halt potentially risky AI models in emergencies. This infrastructure can stop problematic systems within 35 seconds of intervention. The system aims to enhance safety as AI models become more effective and pose greater risks. It connects research workloads, inference engines, and the Control Room itself, ensuring that tasks can be halted if necessary. The mechanism also prevents requests from being processed if the Control Room is unavailable. Currently, the delay for stopping tasks is up to 35 seconds, but OpenAI is working to reduce this to zero.