OpenAI post-training core member Weng Jiayi just proved that a large language model can beat Atari games by writing code. Now researcher Paul Garnier has taken that same approach into the harder world of fluid dynamics control.
He never trained a single neural network. He simply let Codex 5.5 act as a programmer, watching fluid simulation footage and rewriting Python scripts over and over. Using those hand-crafted control rules, the AI took on a dozen physical tests and beat state-of-the-art deep reinforcement learning (DRL) baselines in more than half of them.
Reducing drag on cars, calming turbulence in pipes — industry used to just throw compute at the problem, training an opaque black-box model to control air valves. Codex avoids that dead end. The rules it writes are dead simple: "When local curvature is too high, delay the jet." Dozens of lines of short code, baked with physics intuition, replace the mindless brute-force trial and error of neural networks.
Replacing the black box with code solves the brittleness problem. Before, if you changed the hardware — say, swapping from 5 nozzles to 10 — the old model was instantly useless and you had to burn cash retraining. Now you just change a constant in the code, and the system adapts on the fly.
When test duration was stretched to four times longer, conventional DRL models that strayed from their training regime collapsed entirely. But the code written by the large model, which directly follows physical logic, stayed rock solid. To run the entire control strategy, the large model consumed just 21.25 million tokens — costing less than $14.