Researchers at Apple have come out with a new paper showing that large language models can’t reason — they’re just pattern-matching machines. [arXiv, PDF] This shouldn’t be news to anyone here. We …
When you ask an LLM a reasoning question. You’re not expecting it to think for you, you’re expecting that it has crawled multiple people asking semantically the same question and getting semantically the same answer, from other people, that are now encoded in its vectors.
That’s why you can ask it. because it encodes semantics.
if it really did so, performance wouldn’t swing up or down when you change syntactic or symbolic elements of problems. the only information encoded is language-statistical
thank you for bravely rushing in and providing yet another counterexample to the “but nobody’s actually stupid enough to think they’re anything more than statistical language generators” talking point
It’s worth pointing out that it does happen to reconstruct information remarkably well considering it’s just likelihood. They’re pretty useful tools like any other, it’s funny ofc to watch silicon valley stumble all over each other chasing the next smartphone.
Please enlighten me on how? I admit I don’t know all the internals of the transformer model, but from what I know it encodes precisely only syntactical information, i.e. what next syntactical token is most likely to follow based on a syntactical context window.
How does it encode semantics? What is the semantics that it encodes? I doubt they have denatotational or operational semantics of natural language, I don’t think something like that even exists, so it has to be some smaller model. Actually, it would be enlightening if you could tell me at least what the semantical domain here is, because I don’t think there’s any naturally obvious choice for that.
When you ask an LLM a reasoning question. You’re not expecting it to think for you, you’re expecting that it has crawled multiple people asking semantically the same question and getting semantically the same answer, from other people, that are now encoded in its vectors.
That’s why you can ask it. because it encodes semantics.
if it really did so, performance wouldn’t swing up or down when you change syntactic or symbolic elements of problems. the only information encoded is language-statistical
thank you for bravely rushing in and providing yet another counterexample to the “but nobody’s actually stupid enough to think they’re anything more than statistical language generators” talking point
so… a stochastic parrot?
Paraphrasing Neil Gaiman, LLMs don’t give you information; they give you information shaped sentences.
They don’t encode semantics. They encode the statistical likelihood that each token will follow a given sequence of tokens.
It’s worth pointing out that it does happen to reconstruct information remarkably well considering it’s just likelihood. They’re pretty useful tools like any other, it’s funny ofc to watch silicon valley stumble all over each other chasing the next smartphone.
“remarkably well” as long as the remark is “this is still garbage!”
The only remarkable thing is how fucking easy it is to convince the median consumer that vaguely-correct-shape sentences are correct.
Rooting around for that Luke Skywalker “every single word in that sentence was wrong” GIF…
did you ask a LLM for a post to make here? that might explain this mess of a comment
Please enlighten me on how? I admit I don’t know all the internals of the transformer model, but from what I know it encodes precisely only syntactical information, i.e. what next syntactical token is most likely to follow based on a syntactical context window.
How does it encode semantics? What is the semantics that it encodes? I doubt they have denatotational or operational semantics of natural language, I don’t think something like that even exists, so it has to be some smaller model. Actually, it would be enlightening if you could tell me at least what the semantical domain here is, because I don’t think there’s any naturally obvious choice for that.
guy who totally gets what these words mean: “an llm simply encodes the semantics into the vectors”
all you gotta do is, you know, ground the symbols, and as long as you’re writing enough Lisp that should be sufficient for GAI
both your comments made my eye twitch
like what’d happen if bob fucked up the symbols in a pentacle
also why do we need getaddrinfo? the promptfans will always readily tell you who they are