Post BA53ReDOfj3TQ5wvPE by dani_el@col.social
(DIR) More posts by dani_el@col.social
(DIR) Post #BA53ReDOfj3TQ5wvPE by dani_el@col.social
0 likes, 0 repeats
RE: https://mathstodon.xyz/@tao/117208618508728654I don't understand that prior. It (naturally?) feels to be the opposite, specially if the emphasis is on the subject: one(-self). *Insight* or *understanding* requires some form of interpretability (i.e. clear explanation) and that usually comes from the solution development process ...RT: https://mathstodon.xyz/users/tao/statuses/117208618508728654
(DIR) Post #BA53ReSzjjX2CTFNia by dani_el@col.social
0 likes, 0 repeats
By construction, with LLM-based AI, there is no guarantee that the process being delegated would have any way to "explain" itself. The systems can be forced to preserve traces aiming at explanation. But once complexity thresholds (far exceeded in advanced math) are reached, that just becomes illusory: intermediate output will become too long or as dense as the content upon which the LLM is built/prompted on.
(DIR) Post #BA53Reg6wy1Wr9NrA8 by galactus@col.social
0 likes, 0 repeats
@dani_el he is talking about ways to use AI other than just “solve this problem,” as his example explains.