[stavros]

But that's also like saying "humans don't have a memory property, any 'memory' is in the hippocampus". It's not useful to say that "an LLM you don't bother to keep training has no memory". Of course it doesn't, you removed its ability to form new memories!

[datsci_est_2015]

So why then do we stop training LLMs and keep them stored at a specific state? Is it perhaps because the results become terrible and LLMs have a delicate optimal state for general use? This sounds like an even worse case for a model of intelligence.

[stavros]

Nope, it's not that, but it's nice of you to offer a straw man. Makes the argument flow better.

[datsci_est_2015]

Not entirely a straw man. What is the purpose of storing and retrieving LLMs at a fixed state if not to guarantee a specific performance? Wouldn’t a strong model of intelligence be capable of, to extend your analogy, running without having its hippocampus lobotomized?

Given the precariousness of managing LLM context windows, I don’t think it’s particularly unfair to assume that LLMs that learn without limit become very unstable.

To steelman, if it’s possible, it may be prohibitively expensive. But somehow I doubt it’s possible.

[stavros]

It is, indeed, prohibitively expensive. But it's not impossible. The proof is in the fact that you can fine-tune LLMs.