i suspect one of the last big missing pieces on the road to agi is real causal introspection. i feel this most keenly when, in moments of frustration, i interrogate the model and ask "why did you do that?" and i know it's going to make up some answer, but the answer is basically guaranteed to be a hallucination, and the harder i press on the model, the more obvious it is. it simply doesn't have the capability to actually reflect and retrace its thoughts, and figure out the internal features…