5 ms·

There are no "real-world concepts" or "semantic meaning" in LLMs, there are only syntactic relationships among text tokens.

• lblume 1y ago
  That really stretches the meaning of "syntactic". Humans have thoroughly evaluated LLMs and discovered many patterns that very cleanly map to what they would consider real-world concepts. Semantic properties do not require any human-level understanding; a Python script has specific semantics one may use to discuss its properties, and it has become increasingly clear that LLMs can reason (as in derive knowable facts, extract logical conclusions, compare it to different alternatives; not having a conscious thought process involving them) about these scripts not just by their syntactic but also semantic properties (of course bounded and limited by Rice's theorem).
  • jibal 1y ago
    > Humans have thoroughly evaluated LLMs and discovered many patterns that very cleanly map to what they would consider real-world concepts Well yes, humans have real-world concepts. > Semantic properties do not require any human-level understanding Strawman. > a Python script has specific semantics one may use to discuss its properties These are human-attributed semantics. To say that a static script "has" semantics is a category mistake--certainly it doesn't "have" them the way LLMs are purported by the OP to have concepts. > it has become increasingly clear that LLMs can reason (as in derive knowable facts, extract logical conclusions, compare it to different alternatives These are highly controversial claims. LLMs present conclusions textually that are implicit in the training data. To get from there to the claim that they can reason is a huge leap. Certainly we know (from studies by Anthropic and elsewhere) that the reasoning steps that LLMs claim to go through are not actual states of the LLM. I'm not going to say more about this ... it has been discussed at length in the academic literature.
• empath75 1y ago
  Do you learn anything from reading books or is everything you know entirely derived from personal experience.
  • jibal 1y ago
    Non sequitur. I learn things from LLMs because the LLM training data contains a vast amount of information. This has nothing to do with whether LLMs have concepts.
• int_19h 1y ago
  If that were true, homonyms would be an intractable challenge to LLMs, yet they can handle them just fine, and do so in tasks that require understanding of their semantics (e.g. give LLM a long text and require it to catalog all uses of the word "right" segregated into buckets according to meaning).
  • jibal 1y ago
    That's a syntactic pattern matching operation.
    • int_19h 1y ago
      How so? Syntactic pattern matching can distinguish "right" used as noun from "right" used as adjective. It cannot distinguish "right hand" from "right way", though - that's semantics. LLMs can reliably do the latter.
    • jibal 1y ago
      > It cannot distinguish "right hand" from "right way", though Of course it can -- pattern matching matches on context.