3 ms·
> Does GPT4 not know that a halfling in a tree can't attack a wolf on the ground with a rapier? It’s funny to think about how the parser would even go about de
by zaroth 4y ago
> Does GPT4 not know that a halfling in a tree can't attack a wolf on the ground with a rapier?
It’s funny to think about how the parser would even go about deconstructing that sentence and trying to tie each noun to some world/physics model that would then put constraints on what the character should be allowed to do.
But as I understand it, GPT doesn’t do any of those things. It just a takes the words that have come before, and tries to guess new words to add on that are thematically consistent.
I think it no more “knows” what a “halfling” is than it knows what a “tree” is, the generated words are designed to be read “smoothly” more than any objective correctness.
- stevenhuang 4y agoEvidence is mounting that LLMs may be building world models, ie true understanding. > A pretty hot question right now is whether LLMs are just bundles of statistical correlations or have some real understanding and computation! This gives suggestive evidence that simple objectives to predict the next token can create rich emergent structure (at least in the toy setting of Othello). Rather than just learning surface level statistics about the distribution of moves, it learned to model the underlying process that generated that data. In my opinion, it's already pretty obvious that transformers can do something more than statistical correlations and pattern matching, see eg induction heads, but it's great to have clearer evidence of fully-fledged world models! https://www.lesswrong.com/posts/nmxzr2zsjNtjaHh7x/actually-othello-gpt-has-a-linear-emergent-world https://www.lesswrong.com/posts/nmxzr2zsjNtjaHh7x/actually-o...