No neural networks were trained at all! A large model writing hand-crafted control rules has broken into industrial fluid dynamics—running the entire strategy for just $14. Weng Jiayi, a former core member of OpenAI’s post-training team, recently demonstrated that “purely using large models to write code can beat Atari games.” Now, researcher Paul Garnier has applied this same approach to the even more demanding domain of fluid dynamics control. He trained no neural networks whatsoever. He simply used Codex 5.5 as a programmer, having it repeatedly revise Python scripts by analyzing fluid simulation videos. With these hand-written control rules alone, the AI outperformed top-tier deep reinforcement learning (DRL) baselines in over half of the fourteen physical test scenarios. From reducing automotive drag to stabilizing pipe turbulence, industry previously relied on brute-force computation to train opaque black-box models that controlled airflow valves. Codex avoided this dead end entirely. The rules it generated were remarkably straightforward—for example: “Delay jetting when local curvature exceeds a threshold.” Dozens of concise lines of code, grounded in physical intuition, replaced neural networks’ mindless trial-and-error. By replacing the black box with executable code, it eliminated the rigidity and fragility inherent in neural networks. Previously, even minor hardware changes—like increasing nozzle count from five to ten—would render the old model useless, forcing costly retraining. Now, simply changing a constant in the code instantly adapts the system to new hardware. When test duration was forcibly extended fourfold, traditional DRL models collapsed outside their training distribution. But the large model’s code, directly grounded in physical laws, remained stable throughout. The entire control strategy was executed using just 21.25 million tokens, costing less than $14 total. #AI #AIAgent