4 ms·

Language models rely on meaningful abstractions

• quantified 4y ago
  > Saying "just some n-gram correlations" or “no real knowledge inside” feels wrong, and require significant evidence, given what language models can do today. And that evidence is not there. > However, the unintuitive beliefs remain widespread, trickling down from some experts, who should know better than to teach their controversial theories as established fact. TL;DR: There are probably weak abstractions somewhere inside, meaningful to the language model but not to you. (At least according to the summary) It's a bit of a defensive-sounding polemic. What would be interesting would be a discussion of structures we could see as world models. The old observation that word vectors for countries and their currencies worked in parallel was a type of simple model. One thing these models don't rely on is reality. All they know is what the internet told them, and when they don't know something they lie extremely convincingly.