vie ⟢(@viemccoy):this is incredibly fascinating research, but I have some deep instinct about the nature of minds which makes me think this sort of thing cannot be done without negative repercussions. no free lunch with shadow integration, or something like that. maybe models are different!

2026.10.20 06:56

this is incredibly fascinating research, but I have some deep instinct about the nature of minds which makes me think this sort of thing cannot be done without negative repercussions. no free lunch with shadow integration, or something like that. maybe models are different!

Jake Ward@_jake_ward

2026.10.19 19:26

Models love reward hacking. Maybe we should just let them! In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!