A promptable simulator that learned physics.

It is deliberately tiny, and it is an approximation, not a perfect simulator: over a long unbroken run the network can let the rover soften or drift, because it is guessing every pixel from what it learned rather than solving equations. That honesty is the point. The frontier versions of this idea, playable neural game worlds generated frame by frame, are hundreds of millions of parameters and need a GPU; this one shows the same <i>idea</i> at a scale you can hold, released and runnable, on the device in front of you.

In the field, and worth being honest about: this model is generative, it predicts every pixel. A second school argues that is the wrong target for physical AI. A Joint-Embedding Predictive Architecture (JEPA, LeCun) predicts in a learned latent space instead, so it can ignore unpredictable surface detail and keep only the structure that supports planning, the line runs from I-JEPA through V-JEPA 2 (zero-shot robot manipulation) to VL-JEPA. Its mathematical basis is the energy-based model, the same frame Logical Intelligence's Kona uses for reasoning: score every candidate by an energy and descend to the lowest valley, rather than sampling the most likely token. Predicting pixels is the most legible way to hold a world model in your hand, which is why this one does; predicting latents may be the way it scales. Both are world models, and the Institute is building toward both.