I feel like there needs to be a post (and I don’t want to write it, but maybe I eventually will) that outlines what a model really is. It is not just a statistical text prediction machine unless you are being so loose with the definition of “statistical” that it doesn’t even mean anything anymore.

A decent example of a statistical text prediction machine is the middle word suggested by your phone when you’re using the keyboard. An LLM is not that.

In the most general terms, this kind of language model tokenizes a corpus of text based on a vocabulary (which is probably more than just the words in the dictionary), uses an embedding model to translate these tokens into a vector of semantic “meaning” which minimized loss in a bidirectional encoding (probably), that is then trained against a rubric for one or more topic area questions, retrained for instruction and explainability, retrained with reinforcement learning and human feedback to provide guardrails, and retrained again to make use of supplemental materials not part of the original training corpus (resource augmented generation), then distilled, then probably scaled and fine tuned against topic areas of choice (like coding or Korean or whatever) and maybe THEN made available to people to use. There are generally more parts to curriculum learning even than that but it’s a representative-ish start.

My point being that, yes, it would be nuts to pose ANY question to a predictor that says “with 84% probability, the word that is most likely follows ‘I really like’ is ‘gooning’ on reddit”, but even Grok is wildly more sophisticated than that and Grok is terrible.