logoalt Hacker News

HarHarVeryFunny • today at 1:12 PM • 1 reply • view on HN

I agree that definitions of intelligence should be based on capability rather than implementation, but given that we're discussing LLMs where the intelligence is in the weights, then "the weights don't change" can obviously be taken as pointing out that the model can't learn - the best it can do is memorize and load memorized stuff into context.

Memorizing facts is not the same as learning skills. Implementation-wise memorization can just be RAG, but learning of skills does need to be in the "weights" - whatever is driving the actual behavior of the model.


Replies

kenjackson • today at 6:32 PM

Again, I'd base this on capability. For example, you state: "Memorizing facts is not the same as learning skills." Maybe it's not. I'm not convinced it's not. In the ChatGPT 4 days (or so), I taught the instance I used some chess tactics/strategy that it didn't seem to know. And it appeared to use them correctly after that. Now you might note that this isn't really learning -- they just memorized it and then applied it appropriately.

With a sufficiently large context window, can you make a lot of things look like learning. And if it's not sufficient then point out the capabilities that aren't possible with this approach. To me that much more convincing.