That depends. As an analyst it is already vastly superior to any human. Yet there appears to be an inventiveness gap. As an inventor -- or at the frontiers o

That depends. As an analyst it is already vastly superior to any human. Yet there appears to be an inventiveness gap. As an inventor -- or at the frontiers of scientific discovery -- it's hopeless and still vastly inferior to the average practitioner. (Hence Dwarkesh's question: https://marginalrevolution.com/marginalrevolution/2025/02/dwarkeshs-question.html https://marginalrevolution.com/marginalrevolution/2025/02/dw... ) I've tried prompting o1-Pro to invent something genuinely novel in materials science, and it resists. It appears incapable. On that note, there was this AI materials science project that was run by Deepmind. They spent a ton of money to identify a bunch of "new compounds" -- which, it turns out, were not "novel, credible, or useful." (https://www.theregister.com/2024/04/11/google_deepmind_material_study/ https://www.theregister.com/2024/04/11/google_deepmind_mater...) ...Most interestingly, as I recall, the ones that were "credible" and "useful" were not novel, they were derivative of existing compounds already known to science. So the AI really achieved nothing at all, and, pointedly, this AI didn't have a gimped set of instructions. What we're seeing is that AIs and SotA LLMs excel at certain things -- analysis, linear prediction, and linguistic processing -- but fail utterly at other things such as multidisciplinary logical inference. So what I predict is not that AI will divide the best from the rest, but that it will exceed the best in some tasks, and prove almost entirely incapable of others. Bullish for science and advanced engineering, bad for financial analysts...