Comment by mitxela
8 hours ago
Some concrete facts about LLMs are explained by their next token predictor nature. Every time it says "wait, that's wrong." instead of generating the correct thing the first time.
8 hours ago
Some concrete facts about LLMs are explained by their next token predictor nature. Every time it says "wait, that's wrong." instead of generating the correct thing the first time.
I think that's relatively emergent too though! BERT never really did that (at least to my recollection), presumably because its training was never sufficient for it to develop corrective reasoning in a chain of thought.
What exactly do you mean with "emergent" here?
I don't know what the parent poster means, but if you look at how LLMs are trained, and how it's trained on human communication, you should find it interesting that humans often correct themselves. They also often create facts out of nowhere, and do this crazy thing called dreaming.
If you're an LLM trying to symbolically emulate human behavior, especially when coding, you're going to see git commits, where people write a bunch of code and then go, oh wait, I should fix that!
From the perspective of an outside observer, making mistakes, and then correcting the mistakes, is how you code.
In fact, an outside observer might think that this is how you teach people to code. You make a mistake, then you say, oh, here's the mistake I made, and here's why, and then you fix it.
So I think true emergent behavior will be demonstrated, when an LLM says "Hey wait a minute! They're actually making real mistakes, and they're actually correcting, it's them not on purpose?!", and then stops hallucinating, and stops making mistakes as it codes.
The same is true of bug trackers. You submit a bug request and you often have a patch. Then people discuss it. From a MLM's perspective, you're supposed to write poor code, or code that's not perfect, and then have a conversation.
If I had the resources to train a large-scale LLM, I would clone, for example, GitHub. I would then remove everything where people are fixing broken code.
The outcome would be very interesting.
I've been looking at ways to make enhanced long-term memory stores for LLMs, and there's lots of problems with shifting symbolic relationships if you do it wrong, but definitely once there's true long-term memory, and adaptive behavio,r I think that's the only way you're going to get true real emergent behavior.
BERT isn't a next token predictor. It predicts a single token based on the whole surrounding context in both directions.
I mean sure, but BERT can be and often is used as a next token predictor / generator.
I could have used any series of NLP examples, the point is this CoT behavior only emerges when you get to a certain scale (and training style, presumably).