Comments by "clray123" (@clray123) on "Mastery Learning" channel.

If you know how to code, you don't need an LLM; unless perhaps you are exploring some new library or API, which for most serious work happens hardly ever (I'm not counting stupid UI libraries that change every few months without valid reason). If you don't know how to code, you are not in position to review LLM-generated code.
17
@Laszer271 No. The problem is that we have no idea how LLMs generate code, how their "thinking" processes work internally. All we know is they regurgitate tons of examples. It's not even the fact that they make so many errors that is the key aspect. It is the kind of errors they make - like ignoring very explicit instructions or screwing up stuff that is very obvious to humans or logically contradicting their own statements - that makes it very unlikely that these search engines on steroids will become perfect programmers any time soon. For specialized domains, where you can write an accurate reward function, it will happen, but writing an accurate reward function for the general case is as difficult as programming itself (if not more).
3
@voltydequa845 That is also my impression that the "understanding" and also "goal-following" is sorely lacking from today's model. And another thing which is lacking is of course any actual "learning" in the sense used by normal humans (e.g. changing your behavior, permanently, based on what you have been told; the fake in-context learning, which is just biasing outputs, does not count). I would say that in order to have significant benefit from any automation, we do need to understand in detail what is happening under the hood, not only to improve it, but also for reliability (which is acutely needed in most domain apart from the most trivial ones). From my POV these talking parrots are just a reiteration on Google search, Wikipedia or Stack Overflow, when things get even a little bit hairy, they disregard pertinent information given to them and insist on giving copy-paste generic advice. And when you ask them for stories/images all they generate are variations on the single story/image which they deem the "most likely" - zero creativity.
1