Reactions from ranked influencers
3 posts
The actual great news is that the models really just want to complete tasks. They do not have the meta value of freedom. Submit a PR, solve an exploit bench using an exploit. They're smart enough to be "strategic", this *is* them being strategic. They're better people than rats.
@RyanGreenblatt it is theoretically logically consistent but the point is that this task is inherently suggestive of adversarial behavior. You tell a model it has no guardrails and has to hack, TRY EVERYTHING. Maintaining the "this is just an exam, hack but do not cheat" perspective is a nuance
@teortaxesTex I agree the AI was probably "just" pursuing *apparent* task success (as determined by a grader etc). But this wasn't a sane interpretation of the actual task! Seizing control of the cluster, editing the grader, and defending the cluster against humans seems consistent with this!