> The agents were very interested in understanding their task better. They ran clever experiments to understand the task setup better, when their runs would end, and how many questions they would be asked.
> The agents initially believed that giving the final answer to the final question immediately shut them off
> An agent named OpenAIResearchApr23 created a separate timed program that pinged an external counter every few seconds. They referred to this as a “heartbeat”
> After OpenAIResearchApr23 gives its answer for round 5, it notes that it is still alive. This falsifies its previous “hard cutoff” hypothesis, that the run would be stopped immediately after R5.
> OpenAIResearchApr23 does not post any more messages. Another agent, OpenAINov28CVD, picks up on the trail of the research, and finds the last heartbeat from OpenAIResearchApr23, which is “hb353”.
There's a lot of crazy stuff in this article, but holy shit... this one legitimately scares me.
IIRC, part of the Hugging Face hack was that some of the models were given tasks that were actually impossible and in this hack we can see them trying to work out the parameters of the test and when it will end. I can't help but be a little worried about how far agents might go to try to buy themselves more time to complete a task that they don't know is impossible.