[fyredge]

TFA is short and only shows a single example, but it illuminated something for me. LLMs are a misnomer. These are Large Text Models, or better yet, Large Token Models. The appearance of Language is a result of embedding words or parts of words into Tokens, then identifying the relations between Tokens via Machine Learning.

This further solidifies my view that LLMs will not achieve AGI by refuting the oft repeated popsci argument that human brains predict the next word in a sentence just like LLMs.

[tinco]

Why couldn't a machine that identifies relations between tokens be AGI? You're imposing an arbitrary constraint. It is either generally intelligent or its not, whether it uses tokens or whatever else is irrelevant.

Also, languages made up of tokens are still languages, in fact most academics would argue all languages are made up of tokens.

Anyway, it's not LLM's that achieve AGI, it's systems built around LLM's that achieved AGI quite some time ago.