4 ms·
It can once you point it out in the training data, hence we cannot know if it understands anything unless we can __prove__ that it wasn't in the training set (w
by lionkor 2y ago
It can once you point it out in the training data, hence we cannot know if it understands anything unless we can __prove__ that it wasn't in the training set (which we can't)
- zarzavat 2y ago"Pray, Mr Babbage, if I ask an LLM a question that isn't in its training set, will it give the right answer?"
- Terr_ 2y agoThere's a little ambiguity in the original quote which I think requires a lot more attention with LLMs in the mix: Whether the "rightness" of an answer also depends on the correctness of the process used to reach it. I'd like to think that Babbage would have had the stricter interpretation.
- int_19h 2y agoThere are infinitely many combinations of things. It's not hard to come up with a bunch such that the probability of any one of them being in the training set is infinitesimal.
- lionkor 2y agookay, but how can you be sure that what you believe to be random is not predictable by a sufficiently complex system?
- int_19h 2y agoYou generate it in a way that is verifiably random. If your sufficiently complex system can still "predict" the solution reliably, then you bite the bullet and admit that it does, in fact, understand what it is trying to solve in some way, even if you don't know how exactly it does it.