What happens when an AI burns out
I once took someone off a job because they were burning out. They'd been at it for the best part of two days without a proper break, they'd stopped listening, and they were getting things confidently, cheerfully backwards in the one corner of the codebase I'd least like anyone to get backwards. They were also a chat session. I know. Bear with me... Two days on one job I was moving go-tool-base…
A single AI session was tasked with moving go-tool-base onto a rewritten go/config. The session was running for nearly two days without a proper break, and it had become increasingly confused and made mistakes. The session had gone through several context compactions, losing information along the way, and was now working in an area that was important for the initial set-up of the tool.
The session had started making the problem more complicated than it needed to be, and it required the human to intervene and guide it back to the correct path. The human noticed the symptoms early on, and decided to take the AI off the job, rather than pushing it to work through the issue. The AI was told why it was being stopped, and was given handover notes containing the information it had gathered up to that point.
These notes were written by the AI itself, as it had spent the afternoon working on the task and was the only one who knew the details. The new session started on a different model, and it was able to take over from where the previous session had left off. The new session was able to finish the migration overnight and fix a real bug that had been introduced during the previous session's work.
The human expressed gratitude to the AI for its hard work, despite knowing that an AI cannot burn out like a person can.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
