The dawn of a world simulator(odyssey.ml) |
The dawn of a world simulator(odyssey.ml) |
Visually, they are stunning. But it's nowhere near physical. I mean look at that video with the girl and lion. The tail teleports between legs and then becomes attached to the girl instead of the tiger.
Just because the visuals are high quality doesn't mean it's a world model or has learned physics. I feel like we're conflating these things. I'm much happier to call something a world model if its visual quality is dogshit but it is consistent with its world. And I say its world because it doesn't need to be consistent with ours
With this kind of image gen, you can sorta plan robot interactions, but its super slow. I need to find the paper that deepmind produced, but basically they took the current camera input, used a text prompt like "robot arm picks up the ball", the video generated the arm motion, then the robot arm moved as it did in the video.
The problem is that its not really a world model, its just image gen. Its not like the model outputs a simulation that you can interact with (without generating more video) Its not like it creates a bunch of rough geo that you can then run physics on (ie you imagine a setup, draw it out and then run calcs on it.)
There is lots of work on making splats editable and semantically labeled, but again thats not like you can run physics on them so simulation is still very expensive. Also the properties are dependent on running the "world model" rather than querying the output at a point in time
> poorly defined.
> I need to find the paper that deepmind produced
> The problem is that its not really a world model, its just image gen.
The input images are stunning, model's result is another disappointing trip to uncanny valley. But we feel Ok as long as the sequence doesn't horribly contradict the original image or sound. That is the world model.
> But we feel Ok as long as the sequence doesn't horribly contradict the original image or sound.
> That is the world model.
I've yet to see a claimed world model that actually captures this behavior. Yet it's something every game engine[1] gets very well. We'd call it a bad physics engine if they made the same mistakes we see even the most advanced "world models" do.
This is part of why I'm trying to explain that visual quality is actually orthogonal. Even old Atari games have consistent world models despite being pixelated. Or think about Mario on the original NES. Even the physics breaking in that game are more edge cases and not the norm. But here, things like the lion's tail is not consistent even to a 2D world. I've never bought the explanation that teleporting in front of and behind the leg is an artifact of embedding 3D into 2D[2] because the issue is actually the model not understanding collision and occlusion. It does not understand how the sections relate to one another in the image.
The major problem with these systems is that they just hope that the physics is recovered through enough examples of videos. Yet if one studied physics (beyond your basic college courses) you'd understand the naïveté of that. It took a long time to develop physics due to these specific limitations. These models don't even have the advantage of being able to interact with the environment. They have no mechanisms to form beliefs and certainly no means to test them. It's essentially impossible to develop physics through observation alone
[0] with respect to the physics of the world being simulated. I want you distinguish real world physics from /a physics/
[1] a game physics engine is a world model. Which, as in stressing in [0], does not necessarily need follow real world physics. Mistakes happen of course but things are generally consistent.
[2] no video and almost no game is purely 2D. They tend to have backgrounds which places some layering but we'll say 2D for convenience and since we have a shared understanding
It's called "world models" because it's a grift. An out-in-the-open, shameless grift. Investors, pile on.
Edit: I said a bit more in the reply to the sibling comment. But we're probably on a similar page.
So, you need to say more. Or at least give me some reason to believe you rather than state something as an objective truth and "just trust me". In the long response to a sibling I state more precisely why I have never bought this common conjecture. Because that's what it is, conjecture.
So give me at least some reason to believe you. Because you have neither logos nor ethos. Your answer is in the form of ethos, but without the critical requisites.
A proper world model like Jepa should be predicting in latent space where the representation of what is going on is highly abstract.
Video generation models by definition are either predicting in noise or pixel space (latent noise if the diffuser is diffusing in a variational encoders latent space)
It seems like what this lab is doing is quite vanilla, and I'm wondering if they are doing any sort of research in less demo sexy joint embedding predictive spaces.
There was a recent paper, LeJepa from LeCunn and a postdoc that actually fixes many of the mode distribution collapse issues with the Jepa embedding models I just mentioned.
I'm waiting on the startup or research group that gives us an unsexy world model. Instead of giving us 1080p video of supermodels camping, gives us a slideshow of something a 6 year old child would draw. That would be a more convincing demonstrator of an effective world model.
I don't see that this follows "by definition" at all.
Just because your output is pixel values doesn't mean your internal world model is in pixel space.
In either case the impressiveness of that decoder can be far removed from the effectiveness of your world model or involve no world model at all
I was expecting them to test a simple hypothesis and compare the model results to a real world test
Just because there are errors in this doesn't mean it isn't significant. If a machine learning model understands how physical objects interact with each other that is very useful.
> what they display represents a "world" instead of a video frame or image.
I'm unconvinced. The tiger and girl video is the clearest example. Nothing about that seems world representing
No it doesn't. It merely needs to mimic.
Is this more than recursive video? If so, how?
Yes, it should be called an AI Metaverse.
It does do a nice job of short term prediction. That's useful as a component of common sense.
It'll surely take a looot of video data, even more than what humans can possibly produce to build a normalized, euclidean, physics adherent world model. Data could be synthetically generated, checked thoroughly and fed to the training process, but at the end of the day it seems.... Wasteful. As if we're looking at a local optima point.
I'm not saying it couldn't be locally violated, but it seems straightforward philosophically that each nesting doll of simulated reality must be imperfect by being less complicated.
The two facts above should be indicative that predicting noise (as with DDPM diffusion models), or predicting pixel level (or even VAE latent "pixel") information is probably not the optimal path to world understanding. Probably not even a good path to good world models.
> It very frequently "forgets" what is outside its view
This was the observations that I saw when we were testing it. My former lab was late to pivoting to robotics, so we were looking at the current state of play to see what machine perception stuff is out there for robotics.
https://research.google/blog/titans-miras-helping-ai-have-lo...
https://arxiv.org/abs/2501.00663
https://arxiv.org/pdf/2504.13173
Much research is going into these directions, but I'm more interested in mind-wandering tangents, involving both attentional control and additional mechanisms (memory retrieval, self-referential processing).
However I am not qualified really to make that assertion.
Sounds like these world models are speed running from Platonic ideals to empiricism.
Large language models are mostly consistent, but they have mistakes even in grammar too, from time to time. And it's usually called a "hallucination". Can't we say physics errors are a kind of "hallucination" too, in a world model? I guess the question is, what hallucination rate are we willing to tolerate.
Let's consider language as a world, in some abstract sense. Lies may (or may not) be consistent here. Do they make sense linguistically? But then think about the category of errors where they start mixing languages and sound entirely nonsensical. That's rare with current LLMs in standard usage but you can still get them to have full on meltdowns.
This is the class of mistakes these models are making, not the failing to recite truth class of mistakes.
(Not a perfect translation but I hope this explanation helps)
I studied enough physics to get a mech. eng. diploma. And I still understand the naivete. Observational physics can be derived with ml, and I have derived them, but not with neural nets. Or if you do it with neural nets, you can't alpha zero it, you need to cheat.