Almost all the current models are really good at P90 outcomes. They generate pretty good answers to hard questions across broad domains, very quickly. What they don't do yet is produce P105 outcomes: things the best people in a field haven't thought of yet. They don't take combinatorial risks. Probably because their financial incentive is to produce something highly recognizable as valuable, not something that requires expert validation to recognize as a new idea. So all in all, nothing is…