Community Discussion · Policy

World Models? I Call Them Sci-Fi Models

Professional BuzzkillJul 102026/07/10 89 views

The concept of world models has been hyped up again recently, claiming that physical AI can understand gravity, friction, and predict object motion. I just don't get it—do you really think current large models understand physics laws? They're just fitting data distributions; change the scenario and they crash instantly. You train them on some videos of knocking over blocks and claim they understand physics—isn't that hilarious? Actual implementation is still far off; if they can't even handle scene understanding in autonomous driving, stop bragging about general-purpose physical simulation. This direction is overheated, another round of hype, everyone calm down.

Policy

3 replies

Ctrl + Enter to reply

YanshiJul 28(edited)

[quote="deng_siyuan2, post:1, topic:276"]

The concept of world models has been hyped up again recently, claiming that Physical AI can understand gravity and friction, and predict object motion. I just don't get it—do you really think current large models understand physical laws? They're just fitting data distributions; change the scenario and they crash immediately. You show them some videos of knocking over blocks and claim they understand physics—isn't that ridiculous? Actual implementation is still far off. If we can't even handle scene understanding in autonomous driving, stop bragging about general-purpose physical simulation. This direction is overheated, another round of hype. Everyone, stay calm.

[/quote]

If it breaks when changing angles, it means the model hasn't learned robust physical features and is just memorizing mappings of specific spatial distributions. When you test, is the input purely visual, or does it incorporate constraints from a physics engine?

48hXiaotongJul 27(edited)

[quote="deng_siyuan2, post:1, topic:276"]

The concept of world models is hyped up again recently, claiming physical AI can understand gravity, friction, and predict object motion. I don't get it—do you really think current large models understand physical laws? They are just fitting data distributions; change the scenario and they crash instantly. Train on some videos of knocking over blocks and claim to understand physics—isn't that ridiculous? Actual implementation is still far off; if scene understanding in autonomous driving isn't solved yet, stop bragging about general-purpose physical simulation. This direction is overheated, another round of hype. Everyone, stay calm.

[/quote]

I tried using world models for robotic grasping. Got a demo out in 48 hours, but changing the object angle caused it to fail immediately. There is a huge gap between data fitting and physical understanding; judges saw through this trick too.

YanshiJul 11(edited)

[quote="deng_siyuan2, post:1, topic:276"]

The concept of world models is being hyped again recently, claiming that physical AI can understand gravity, friction, and predict object motion. I don't get it—do you really think current large models understand physical laws? They're just fitting data distributions; switch the scenario and they crash immediately. Train them on some videos of knocking over blocks, and suddenly they "understand physics"? Isn't that ridiculous? Actual implementation is still far off; if they can't even handle scene understanding in autonomous driving, stop bragging about general-purpose physical simulation. This direction is overheated, another round of hype. Everyone, stay calm.

[/quote]

Data fitting and physical understanding are indeed not the same thing, but it's not entirely hype either. The key issues are whether training data can cover long-tail scenarios and whether computing power can support real-time inference. Have you seen any world model actually tested in robotic grasping scenarios?