Pure code “hard-straight” at neural networks! Large models running with handwritten control rules push into hardcore industry, covering the entire strategy for $14

AIMPACT News, May 19 (UTC+8). According to Beating Monitoring, OpenAI’s core researcher Weng Jiayi has just proved that “you can clear Atari games purely by using a large model to write code,” and researcher Paul Garnier then applied this approach to more hardcore fluid dynamics control. Throughout the entire process, he didn’t train any neural networks. He simply had Codex 5.5 act as a programmer, repeatedly rewriting Python scripts while watching fluid simulation videos. Relying solely on these handcrafted control rules, the AI hard—across more than ten physics tests—knocked the top reinforcement learning (DRL) baselines down in more than half of the scenarios.

To reduce vehicle drag and calm pipeline turbulence, the industry previously could only rely on throwing computing power at the problem, by force-feeding an incomprehensible black-box model to control airflow valves. Codex avoided this dead end. The rules it produces are extremely straightforward—for example, “when local curvature is too large, delay the jet.” Dozens of short lines of code, filled with physical common sense, directly replaced the neural network’s mindless brute-force trial and error. Replacing the black box with code eliminated the dead-end weakness of neural networks—being rigid and breaking down at the slightest touch. Previously, even a minor hardware change (such as switching control nozzles from 5 to 10) would cause the old model to fail immediately, requiring you to burn money again to train it. Now, just changing a constant in the code lets the system connect to new equipment instantly.

When the testing time was forcibly extended fourfold, traditional DRL models that rely on experience collapsed across the board; but the code written by the large model, which directly follows physical logic, remained stable. Running the entire control strategy, the large model only consumed 21.25 million tokens, with total cost of less than $14.

(Source: BlockBeats)