However, I don't think that's the problem in the example of the marble bust not toppling. It would have toppled in a video game engine as well; they model gravity and inertia just fine.
I think what's happening is the model isn't being a world model at all, it's being a language model. It's trying to make the "scripted" thing happen, and it probably "feels" like the script should end in the bust on the pedestal.
I feeling I get from this is that this could be the mental model for a robot in the real world, continuously corrected by actual events.
We all have a world model in our heads of what we think is going to happen, we adjust quickly (or not) based on what does. This is pretty incredible stuff.
As an aside, why fall into the role of being an apologist for google-hype and Big Tech?
When you have a conversation, you have an idea in your mind of how the conversation is going to go. You predict what the other person is going to say before they say it.
When you are driving, you predict what the other drivers are going to do, you predict what signs say, how long you have until the green. If you see two red lights attached to the back of a car flash on, you may predict that the car is about to slow down and you start slowing down before their speed has changed.
Forget video games and movies. I believe these models will be that inner predictive model for robots.
They wouldn't be called "Laws of Physics" if obeying them was optional. If your center of gravity is outside your base, you _will_ topple over. That's just how gravity works.
And why not? Because the first model that can actually learn to handle the simulation not only can do the simulation internally too, but it can do it better than out simulation.
It's where (ex-)"AI luminaries" go to die. It's a way to make glacial, slow progress in the correct direction while never actually getting anywhere.
So we get to a point and people keep making complicated simulations of one kind or another ... and then someone (some model) who's never used the simulation, and doesn't need it for model training gets by far the highest score on every benchmark and everyone agrees the simulation was kind of beside the point.
If it's the latter, I don't see a 'hole'.
This has already happened to other ML fields. For example, there's not much point anymore to ASR models (just feed audio into LLMs) and the space that TTS models occupy is fast disappearing. Oh and it's not just that Qwen 3.6 is a good ASR in addition to being good at coding mario. Doing things this way, not using an ASR model is fundamentally more expressive. A whole bunch of companies started up on ASR + TTS "towers" (lots of models, Pipecat style) as they're called, and they're dead companies walking. Their tech has been surpassed by just bog-standard generic LLMs in a way that they will never get the upper hand again.
A chain of models approach, a separate ASR, will never do this: "you said you agree but I can hear you're sarcastic, so I'll explain again". Also getting a conversational tower to switch languages when the user does that (or when it makes sense in a language course) is not quite impossible but very, very hard. And yet feeding audio directly into Qwen will almost magically solve these problems better than I will ever do it (lightly finetuned). It's fundamentally better in every way (except perhaps speed, of course. Although arguably such a tower needs an LLM too, it'll just have way less tokens)
This point has long been passed for image recognition. Feeding a photo or screenshot into a full LLM along with a request on what to do beats the crap out of any OCR, image recognition, ... model every time these days.
And now there's also hints of world models just being bypassed. Just giving a finetuned model the last 50 video frames (surprisingly, that fits in context, especially at reduced resolution), movement statistics, motor power, status, ... of a robot and asking what to do actually works (with heavy finetuning, but I think it'll start working without). Well ... AND then distilling that knowledge into a much smaller model because B200 racks are more than a bit to heavy for a logistics robot to carry around. But we'll get over that.
The results so far from genie have been stunning, and its not obvious to me that another approach is guaranteed to end up better for all use cases.
Yes! Hence my comment about world models being where (ex-)"AI luminaries" go to die. They had a great idea and pushed it, advancing their career, but as you point out this does not provide even them with a better perspective on where things are going. When it comes to the next big thing they have no better chance than a fresh grad student has.
But simulation (in this round, world models) is how you get the current thing to perform 1% better. So they work on that, and because of their reputation loads of money and resources pours into it because they chose it ... and then the solution comes from a different place entirely.
Except things are changing too, and world models like Google and Meta and Mistral and ... are pursuing appear to be getting bypassed by the current tech. It's starting to look a lot like the "internal world model" of a transformer that you've shown 1 million youtube movies is better than a separate world model, probably because the resources they pour into their general chatbot models are still an order of magnitude above what they can pour into their world models.