Why do models task game?
TLDRHow can we study misalignment with todays models as proxies? Theyre clearly not paperclip maximizers, but they also often do things the user doesnt want. A strong contender for a real misaligned propensity is task gaming: taking actions that dont complete a task but superficially seem like they do, such as hardcoding tests or falsely claiming a task is fully complete. But maybe task gaming is just a crude heuristic, or the model mistakenly trying to achieve the users intent? In this post we…
Read the full story at LessWrong ↗