How Learned World Models Could Simulate Software

Summary

Dario Farzati reflects on a shift from programming every rule of a simulated world to training models on examples of how reality behaves. He argues that video generation systems first became “dreamers,” producing plausible frame-to-frame continuations without explicit physics, and that interactive systems moved closer to simulation when Google’s Genie 3 generated navigable worlds from a single image. The essay then discusses Robbyant’s open-sourced LingBot-World, described as a 28-billion-parameter model running at 16 frames per second for up to ten minutes. Its demonstrations appear to show emergent persistence: landmarks remain in place after the camera turns away, and an off-screen car continues along a road, even though the system has no conventional scene graph, collision system, or explicitly tracked objects. Farzati extends the idea to software, suggesting that a model trained on enough interface traces might reproduce spreadsheet results, terminal behavior, or application state as learned visual and interaction patterns rather than by executing conventional logic. He cites a 2022 demonstration in which ChatGPT simulated a Linux terminal, while noting that the simulation forgot earlier state and produced an unreliable “alt-universe.” The essay emphasizes that these systems remain limited: LingBot-World can drift over time, its memory is emergent rather than guaranteed, it requires enterprise-grade GPUs, and complex spreadsheet formulas could fail. Its conclusion is speculative: reality-like interactive behavior may emerge from learned prediction without a literal codebase underneath, but current systems are not the Matrix.