I don’t think LLMs have theory of mind, but your point is not very strong. You can literally query ChatGPT right now and see that it can figure out intentions (

I don’t think LLMs have theory of mind, but your point is not very strong. You can literally query ChatGPT right now and see that it can figure out intentions (both superficial and deep) of a gun is held to a head quite easily. Because, obviously, training data probably includes a decent amount of motivation breakdowns as a function of coercion. It doesn’t know why, but it knows what to say.