Skip to content
VietVance

Site search

Science & Technology

OpenAI Deploys AI Emergency Stop System That Can Halt Tasks Within 35 Seconds

OpenAI has deployed an internal emergency stop system called Control Room to monitor research tasks performed by AI models. After researchers decide to pause or terminate a task, the system can prevent it from continuing within 35 seconds. OpenAI is working to shorten that time.

VietVance

1 min read

OpenAI has deployed an internal emergency stop system called Control Room to monitor tasks performed by AI models. When researchers decide to pause or terminate a task, the system prevents the model from generating further output. OpenAI says it takes no more than 35 seconds from receiving a stop instruction until the task can no longer continue.

Illustration of an artificial intelligence system.(VietVance)

The system consists of task workloads, inference systems that run the models, and a centrally managed Control Room service. Monitored tasks include research work such as reinforcement learning experiments, model evaluations, large-scale data sampling and data processing. A task can use one or more models, run by the corresponding inference systems.

When a task starts, Control Room assigns it a unique identifier. Thereafter, the workload must include that identifier when sending a model processing request to an inference system. Before processing the request, the inference system checks the task’s status with Control Room. If the task has been terminated, the request is rejected, preventing the model from generating further output.

In addition to checking each request, Control Room connects to OpenAI’s existing monitoring systems to detect anomalies or potential failures and can stop affected tasks depending on their severity. OpenAI says it is seeking to reduce the time between receiving a stop instruction and terminating a task to zero, so tasks can stop immediately.

Source: VietVance

Related