4 ms·
> Often with slightly different random values or temperature or phrasing, the same LLM will give different results for the same problem. You're fundamentally d
by klyrs 3y ago
> Often with slightly different random values or temperature or phrasing, the same LLM will give different results for the same problem.
You're fundamentally describing a stochastic parrot here, and not an intelligence. When you say that humans can also make the same mistake, you're ignoring the fact that every human testing these systems and finding it lacking are also comparing their experience to a lifetime of interacting with human intelligence. Every anecdote of this sort is an example, typically based on repeatedly trying an assortment of prompts (eliminating two of your variables -- random values & varying prompts), on a large variety of tasks, which is curated down to a single example for the sake of brevity.
To say that "oh maybe it would have gotten the right answer if you got lucky or tried harder or twiddled a knob that you don't have access to" is simply nowhere near the extraordinary evidence that is required to prove the extraordinary claim of "real understanding." The reason that you don't know what to do about these anecdotes is that you lack the evidence to properly rebut them. You can only wave your hands at ill-defined properties and gaslight about the user holding it wrong. Perhaps you should ask your parrot buddy what to do.
But if you were really serious about this, you wouldn't be going after the strawfolk down in the comments, you'd rebut the paper itself.
- yawnxyz 3y agoThis is super interesting because you’ll end up with different people with different opinions on “what is the best answer” if they all have different experiences. Most of us when using LLMs feel it should only produce the one correct answer; sometimes that’s the right way to think about it, but other times we might want an opinionated answer, when there is no true answer. So I think if we get better learning and self correcting machines, we also get more opinionated and sometimes incorrect machines. Kind of like people. And sometimes they’re stubborn or confidently incorrect, but that I think actually points to more intelligence (in people at least), since it points to people deciding based on their experiences.
- lossolo 3y ago> This is super interesting because you’ll end up with different people with different opinions on “what is the best answer” if they all have different experiences. In context of grandparent comment it's code, so while we can have different solutions to the same problem, we can verify if the solution is correct because it's not an opinion on some subject.
- aarong11 3y agoThere's no such thing as "one correct answer" to most tasks in any language. In mathematics, maybe. If you're asking it to recall a specific quote verbatim, yeah. But other than that there are many (infinite?) different ways you can paraphrase differently to get the same point across (at least in English).
- caturopath 3y ago> Every anecdote of this sort is an example, typically based on repeatedly trying an assortment of prompts "Data" isn't the plural of "anecdote" -- you can systematically test things, as the authors of the study the article reports did, or you can say "I tried something and got result ~X, therefore LLMs cannot X", which is my honest understanding of what I've heard people say repeatedly. > But if you were really serious about this, you wouldn't be going after the strawfolk down in the comments, you'd rebut the paper itself. I don't know the details of the paper, but I suspect it's right. My actual problem was with someone explaining that the result was trivial and offering what I thought was a non-helpful explanation of why.
- soco 3y agoAnd I assume there's no understanding of what "real understanding" means. In this so limited context "real understanding" looks like "did what I wanted without me really explaining what I wanted". Although in real life I seldom know from the start what I really want either and takes me a few iterations to clarify it even to myself - so "real understanding" seems to be something even I am missing (and I shouldn't blame poor AI for lacking it). Or maybe I'm an AI myself just unaware of it?