Hacker News

points by edstarch 2 days ago | 0 comments

I'm by no means anti-anything but I think what the comments are getting at is that these "world models" do not have a world model. It's a fundamental qualitative gap that no amount of compute has been able to close, just mask by everything else getting better.

This doesn't show up as much where the verifiable reward is "code compiles and tests pass" but you can still absolutely see it in things like movies, prose, or slide decks. Even frontier models are pretty bad at keeping track of narrative threads even as the technical quality drastically improves.

This begs the question of whether the gap can be closed just with more compute (presumably yes, or at least errors can be minimized beyond human capacity to see them) and is that going to be worth it with current tech (possibly no).

[–] exit 1 day ago

does your model of models admit the possibility of partial models?

the average human cannot vividly recollect a walk through a familiar neighborhood, much less invoke a coherent fantasy of going on such a walk. do they contain a world model?

[–] tyg13 1 day ago

Must we, every time the limitations of AI are discussed, always imagine some simpleton human to defend it? It's trite and rarely adds anything to the conversation. The point is to judge the model on its merits. If this is a good world model, then it's a good world model. It doesn't matter whether or not a human can recollect every detail about a familiar neighborhood. It's irrelevant

[–] handoflixue 1 day ago

It matters whether humans can pass the test or not.

Humans also get details wrong when they do stuff like this. So right away, we learn this isn't a limitation of AI. If you want to say that getting that wrong is "proof they don't have a world model", that cuts both ways - it means humans do just fine despite not having one of those either.

But that also lends us insight into fixing the issue on merit! If humans also obviously lack a world model, how do we cope with this? Sometimes we use "live action" and reality forces consistency. Sometimes we use editors and QA. Sometimes it ends up on a website listing every little shot-to-shot inconsistency in every single TV show and movie ever made.

And now you have a few clear ideas on how to train the next generation of image models to do better! And you can probably dramatically improve your own prompting by providing reference images, using consistent descriptions, etc..

So, hardly irrelevant.

[–] cman1444 1 day ago

Exactly. Also, for practical purposes, an AI's capability to meet or exceed human performance is a very important benchmark. It makes the decision between using AI vs paying a human creator very easy.

[–] shoelessone 1 day ago

I think what we're talking about though here in reality is our perception of the model's output.

Our perception in realtime needs to be better than our ability to remember things from the past.

[–] handoflixue 1 day ago

I strongly suspect that if you fed this video through an LLM, it would spot all of the mistakes - the problem is that parsing and generating video is still pretty expensive, so that sort of automatic multi-pass process isn't showing up yet. Better to let the human decide whether they mind enough to pay for edits.

[–] yreg 1 day ago

The point is that humans obviously have world models, just imperfect ones.