3 ms·
People who think they are smart often think they can get by in any domain without knowledge and context, but at a certain point this fails and knowledge is requ
by grey-area 2mo ago
People who think they are smart often think they can get by in any domain without knowledge and context, but at a certain point this fails and knowledge is required, not just intelligence.
- skydhash 2mo agoPretty much this. With the internet and my websearch-fu, I can find some facts on a lots of things. And while I do for some questions I have, I try to never interject answers when I don’t have a modicum of mastery in the domain. Mastery is the formation of an internal model that guide your actions and thinking. It’s not merely knowledge and intelligence.
- grey-area 2mo agoYes sorry for the shorthand, an internal model is a good way to describe mastery. It’s something LLMs don’t have and can’t transmit.
- ndriscoll 2mo agoWe're talking about learning here, i.e., building that knowledge and context. People also refer to more than one source, and I think you overstate the degree to which LLMs make mistakes in current year. Human teachers also make mistakes, and it's typical for a bright student who's paying attention to correct even small details.
- grey-area 2mo agoLLMs are not a shortcut to mastery, in fact the opposite most of the time and their errors are qualitatively different from human errors.
- ndriscoll 2mo agoI don't think anyone is claiming they're a shortcut to mastery. Personal tutors aren't a shortcut to mastery either. Both simply have the advantage of being able to deeply know where the student's knowledge frontier is, so they can tailor their answers to that, make connections and draw analogies to things the student has mastered (while pointing out distinctions), etc. The student still needs to put in the work.
- grey-area 2mo agoLLMs are not tutors and do not deeply know anything.
- ndriscoll 2mo agoIn the sense that they are not programmers and do not deeply know e.g. performance engineering, sure. But I can still give one a program and ask it to optimize it, and it will turn it into a zero allocation streaming SIMD implementation that's orders of magnitude faster than where it started. It may not "know" how to do any of that, but it does it anyway.
- grey-area 2mo agoIt may or may not do that and it may introduce other problems as it does so. Checking the work is not a trivial problem.