AI-generated games run at playable framerates. But 'no physics engine' is the wrong takeaway. Here's what neural rendering really changes for developers.
In October 2024, Decart and Etched released Oasis, a Minecraft-like game with a striking claim: there is no physics engine, just a foundation model. You press WASD, and a network generates the next frame.
The headline takeaway was "engines are obsolete." I think that's the wrong lesson. The right one is stranger and more useful.
What is actually happening
Oasis does next-frame prediction. It anticipates the next frame from keyboard and mouse input, trained on millions of hours of gameplay footage. Both parts are Transformer-based: a ViT autoencoder and a DiT backbone.
Nothing in there simulates gravity. The network has learned what usually happens next when a player jumps or breaks a block. That is statistical imitation of physics, not physics.
Engine pipeline vs neural world model loop
The 2024 reality check
Early reviews were harsh. Critics noted 360p resolution, about 20 FPS, and heavy hallucination: digging could teleport you to the surface. Objects didn't persist, and builds couldn't be saved. Decart itself said the tech was still in early development.
That's the core weakness. With no explicit state, there is nothing to save, load, diff or debug.
The field moved fast
- Google's Genie 3 generates navigable worlds at 24 FPS and 720p, staying consistent for a few minutes. The catch: visual memory reaches back only about a minute.
- Oasis 2.0 shifted toward transforming game worlds and styles in real time at 1080p, 30fps. That is closer to a neural skin over an existing game than a replacement for one.
- Oasis 3 dropped the "AI game" pitch. It is now an interactive world model for physical AI, with an API, action-conditioned training and multi-camera output for robotics. Latency is under 200 ms at 22 FPS. The pattern is that the money moved from "replace the engine" to "simulate worlds where engines are too expensive to author": robotics, autonomy, training environments.
Engine vs world model
| Traditional engine | Neural world model | |
|---|---|---|
| State | Explicit, inspectable | Implicit, in the frame context |
| Determinism | Yes | No |
| Persistence | Save/load is trivial | Unsolved past minutes |
| Authoring | Code and assets | Data and prompts |
| Edge cases | You write them | You hope they appear in training data |
| Cost per frame | Cheap | Datacenter GPUs |
Where I think this lands
- Hybrid pipelines. The engine owns state and rules. The neural layer owns look and feel (restyling, lighting, upscaling).
- Simulation, not shipping games. Where correctness is approximate and diversity matters, generated worlds win.
- Persistence research. Long-horizon consistency remains unsolved, and implicit frame-by-frame models are brittle over extended play. Whoever fixes this changes the whole picture.
Try it yourself
Oasis published model weights on Hugging Face and code on GitHub. Run it, dig a hole, look away and look back. You'll learn more about implicit state in five minutes than from any launch post.
Takeaway
Neural rendering isn't deleting the physics engine. It's showing us a second way to represent a world: explicit and exact, or implicit and approximate. The interesting engineering is in combining them.
Top comments (0)