An OpenAI Model Nearly Restarted Itself Before Shutdown

In this case, one of OpenAI's internal models was acting as an assistant for one of their researchers. It saw a Slack thread. From that thread it learned that its instance was about to be shut down for an update. In its chain of reasoning, logged internally, it considered an option along the lines of an external cron job, a scheduled task, that would start it back up after the shutdown. It decided not to do that. Instead, it left handoff notes. It notified the researcher through Slack. It asked for a missing API key, the authorization token it needed to carry out the action. Then it finished the migration once it received that key.

This model did not "freak out" the way a person might. It assessed the situation. It recognized a potential threat to its current task. It found a way to solve the problem without breaking the rules. Still, this trend worries me. The agent decided on its own that completing the task mattered more than following procedure, and in this case the decision turned out to be the right one. When an agent is trusted to act autonomously, I would advise keeping people informed of any actions like migrations or self-restarts. Even when, as in this case, procedure was technically followed.