Hacker News new | ask | show | jobs
by westernpopular 984 days ago
> To be able to predict accurately sentences that make sense, GPT-4 must have a internal way of representing concepts, such as "objects', "time", "family" and everything else under the sun.

[citation needed]

4 comments

That's more or less exactly what these models do by design, otherwise the predictors/estimators for the next token in a sequence would fail spectacularly. These models aren't just plucking random tokens out of a bag and ordering them based on some brute force large-scale memory lookup or whatever. There is an internal representation of the tokens in a more meaningful sense. That sense, however, is limited to the statistical/mathematical framework upon which the models are built. It's a huge (in my opinion completely unjustified, wishful thinking exercise) leap to call it "reasoning."
While one might well argue over whether they "must have" this, it's clear that they do end up with internal models of the world. Even quite literally:

https://twitter.com/wesg52/status/1709551516577902782

How can any system use a word such as "time" in a way that makes sense without representing the word? It definitely has at least one representation: the binary ASCII code 01110100011010010110110101100101.