OpenAI部署AI“控制室”：最长35秒阻断模型输出 OpenAI近日宣布，已在内部研究基础设施中部署名为“Control Room”的中央控制系统

OpenAI deploys AI "control room": blocking model output for up to 35 seconds OpenAI recently announced that it has deployed a central control system called "Control Room" in its internal research infrastructure. When a model task poses a safety or network security risk, the system can block the model from continuing to output for up to 35 seconds after receiving the stop command. According to OpenAI, Control Room can uniformly track research tasks such as reinforcement learning training, model evaluation, agent sampling, and data processing, and is linked to the model behavior monitoring system. The monitoring model will check the reasoning and operation process of the trained model. Once potential problems are found, an alarm can be issued or the task can be suspended directly based on the degree of risk. During the specific execution process, the monitoring system will send the task ID that needs to be stopped to the Control Room. The next time the relevant task applies for model inference, the inference engine will verify its running permission; if the task has been marked as terminated, the system will reject the request to prevent the model from continuing to generate content. OpenAI stated that it currently takes up to 35 seconds from the time the Control Room receives the stop command until the task cannot continue to run at all. This time difference comes from the caching layer in the inference authorization process. The company is further compressing the response time with the goal of achieving instant stop. The system also uses a "fail-close" mechanism. If the Control Room fails, is shut down, or is tampered with, all related inference requests will be denied by default until service is restored. This mechanism now covers all cutting-edge model research tasks of OpenAI, and is being expanded to other research tasks. OpenAI said that external model services adopt a more centralized technical architecture, and currently have corresponding deployment offline and emergency response capabilities. via cnBeta.COM - Chinese industry information website (author: Source: Phoenix Technology)