[IAmGraydon]

>An LLM predicts the next word based on all the words before it. That is the whole story. There is no idea sitting underneath. The words are everything. For an LLM, words are the source, and any meaning is just a byproduct that falls out by accident.

I don't think that's right. What makes LLMs (and other forms of "AI") work is the very fact that they hold a statistical model underneath and put words on top of it. That statistical model, at least to me, appears to be somewhat akin to what a human would call a "concept".

Like a stable diffusion model has a statistical model of what a "chair" is. Not one exact chair, but all chairs. They all have backs, 4 legs, a flat area to sit on. It can then carve an image of a chair out of noise using this concept. Take a music model like Suno as another example. If you tell it to make a rock song, it has a statistical idea of what this would looks like. The harmonic progression, the tempo, the types of instruments used. It then attaches sounds to this statistical model and you get a "song". LLMs also appear to do this. If you ask it for a paragraph of a specific flavor of prose, it has that as a statistical model, and it attaches words to it.

I am not a big believer in LLMs delivering on even half the hype they've generated. However, this very idea of human concepts vs AI's statistical models is the one thing makes me wonder, sometimes, if they're on the right track, even if we are a long way off from AGI. It's kind of funny that the author of the above article landed on the opposite conclusion.