Pure Code Takes On Neural Network! Big Model Handcrafted Control Rules Penetrate Hardcore Industry, $14 Run Full Strategy

According to Perspicacious Beating monitoring, OpenAI's post-training core member Weng Jialin has recently demonstrated the ability to "clear Atari games solely by using large-scale models to write code." Researcher Paul Garnier has further applied this approach to a more hardcore fluid dynamics control.

He did not train any neural networks throughout the process. Instead, he had Codex 5.5 act as a programmer, continuously modifying Python scripts based on fluid simulation videos. Relying solely on this handcrafted control method, AI successfully surpassed top reinforcement learning (DRL) baselines in over half of the physics tests.

Whether reducing drag on cars or calming pipeline turbulence, the industry used to rely solely on brute computational force to train an incomprehensible black-box model to control airflow valves. Codex has sidestepped this deadlock. Its written rules are straightforward, such as "delay jetting when local curvature is too high." With just a few dozen lines of code imbued with physics knowledge, it directly replaces the brute-force trial and error of neural networks.

By swapping the black box for code, the rigidity and fragility of neural networks have been eliminated. Previously, any slight hardware alteration (such as increasing the number of nozzles from 5 to 10) would render the old model obsolete, necessitating costly retraining. Now, a simple constant adjustment in the code allows the system to instantly adapt to new equipment.

When the testing duration was forcibly quadrupled, traditional DRL models that ventured out of their experiential comfort zone completely collapsed. In contrast, the code written by the large-scale model, which inherently follows physical logic, remained stable throughout. Executing this entire control strategy cost the large-scale model only 21.25 million Tokens, amounting to less than $14 in total expenditure.