|
|
|
|
|
|
by emptiestplace
532 days ago
|
|
|
I keep saying this but nobody seems to really get it or care: LLMs are magic. This isn't at all like conventional software where someone plans for every possible path, or even software where paths are generated procedurally. These models show truly emergent behaviors that go far beyond their training. And yet people dismiss them simply because they're probabilistic (more likely because they've heard some version of this, actually), as if that somehow invalidates the fact that they're doing things they were never programmed to do. Our understanding of consciousness is about to undergo a revolutionary transformation, and a lot of people ... aren't going to like that, either. |
|
|
You can make them deterministic by fixing the temperature. But even still, there is some intrinsic error rate: it's like an oracle machine where some subset of queries will simply give you the wrong answer... which is just like humans (how many times have answers on StackOverflow been wrong), but it's definitely not like "conventional" programs.
I guess just as people had to get used to "trusting" computers, I think there's probably going to be another adjustment period of "untrusting" them and finding the right balance of "trust but verify." I think this is also mirrored in adoption patterns: mathematicians seem to be more open to using LLMs because they're used to not blindly trusting supposed proofs. On the other hand, it's well known that writing code is easier than reading it, so developers are a bit more wary.