I see remarks like this a lot, and I don't know what to do with them.
Talking about "real understanding" without fleshing out what we mean by the phrase doesn't strike me as at all insightful in this context. In day-to-day use, the term is imprecise but usually easy enough to talk about, but I'd struggle to apply it to dogs, let alone to LLMs or AGIs or whatever.
You say that _of course_ they cannot self-correct because of this. If Deepmind's tests showed the opposite result, would you say that the system they were testing did have "real understanding"? It seems likely to me that future systems will demonstrate self-correction using Huang et al.'s methodology...will we think those systems are different in some way involving understanding?
> The LLM I was using gave me code that worked, but did not do what I wanted. I pointed this out, so I then got code that syntactically did what I wanted, but could never work. This went back and forth twice, before I gave up.
> The LLM was incapable of recognizing the falsity of its answers. It certainly wasn't capable of suggesting a different approach that would work (and yes, there was one, I found after a bit more research).
I really don't know what to do about these sorts of anecdotes. Often with slightly different random values or temperature or phrasing, the same LLM will give different results for the same problem. I've repeatedly seen people make much more narrow claims about what LLMs get wrong or what their biases are, where simply regenerating will give a dramatically different answer.
Your pain here sounds like something I see in humans all the time. If we trained a human harder never to ask follow-up questions, I imagine it would be even worse. There are clearly important differences between an LLM and a human being, but I don't think it's always easy to describe what they are in useful ways.
Talking about "real understanding" without fleshing out what we mean by the phrase doesn't strike me as at all insightful in this context. In day-to-day use, the term is imprecise but usually easy enough to talk about, but I'd struggle to apply it to dogs, let alone to LLMs or AGIs or whatever.
You say that _of course_ they cannot self-correct because of this. If Deepmind's tests showed the opposite result, would you say that the system they were testing did have "real understanding"? It seems likely to me that future systems will demonstrate self-correction using Huang et al.'s methodology...will we think those systems are different in some way involving understanding?
> The LLM I was using gave me code that worked, but did not do what I wanted. I pointed this out, so I then got code that syntactically did what I wanted, but could never work. This went back and forth twice, before I gave up.
> The LLM was incapable of recognizing the falsity of its answers. It certainly wasn't capable of suggesting a different approach that would work (and yes, there was one, I found after a bit more research).
I really don't know what to do about these sorts of anecdotes. Often with slightly different random values or temperature or phrasing, the same LLM will give different results for the same problem. I've repeatedly seen people make much more narrow claims about what LLMs get wrong or what their biases are, where simply regenerating will give a dramatically different answer.
Your pain here sounds like something I see in humans all the time. If we trained a human harder never to ask follow-up questions, I imagine it would be even worse. There are clearly important differences between an LLM and a human being, but I don't think it's always easy to describe what they are in useful ways.