Pure code hard neural network! Large model handwritten control rules enter the hardcore industry, with a full set of strategies running for $14 According to Beating monitoring, OpenAI's core training member Weng Jiayi has just proven that "writing code solely through large models can complete Atari games". Researcher Paul Garnier then applied this method to harder core fluid dynamics control. He did not train any neural networks throughout the entire process. Simply let Codex 5.5 act as a programmer, repeatedly rewriting Python scripts while watching fluid simulation videos. Relying on this handwritten set of control rules, AI managed to outperform the top-level reinforcement learning (DRL) baseline in over half of the physics tests. To reduce drag on cars and smooth out turbulent flow in pipelines, the industry used to rely solely on computing power to force out an incomprehensible black box model to control the airflow valve. Codex avoided this dead end. The rules it writes are extremely straightforward, such as' delaying jet when the authority's curvature is too high '. A few dozen lines of short code with physics knowledge directly replace the brainless brute force trial and error of neural networks. Replacing the black box with code has eliminated the dead end of neural networks, which are rigid and shatter at the slightest touch. Previously, as long as the hardware was slightly modified (such as changing the control nozzle from 5 to 10), the old model would be scrapped on the spot and had to be retrained by burning money. Now as long as you change a constant in the code, the system can instantly connect to the new device. When the testing time is forcibly extended by four times, the traditional DRL model that leaves the experience zone collapses completely; But the code written by the big model always runs stably because it directly follows physical logic. Running through this entire control strategy, the large model only consumed 21.25 million tokens, with a total cost of less than $14. [Original link]