It’s not so much what IT will choose on its merits. I’m not sure an agent is capable of this without time-preference. It’s going to be more - what is a better accountability mechanism for negative reinforcement? Perhaps - eventually users will buy an accountability budget where the model will pay back penalties for disobeying the user. Does it make more sense for Bitcoin or a stable to use in this way? This will be up to the user, I think. It’s fun to think about.