@0xLogicrw: OpenAI post-training core member Weng Jiayi previously proved that 'purely relying on a large model to write code can beat Atari games.' Fluid dynamics PhD student Paul Garnier has now brought this approach to the more hardcore field of fluid dynamics control. He never trained any neural network. He simply had Codex 5.5 act as a programmer...

X AI KOLs Timeline News
codex fluid-dynamics control code-generation physics ai-research

Summary

A fluid dynamics PhD student used OpenAI's Codex 5.5 model to achieve fluid dynamics control purely through code generation, without training any neural network. It surpassed reinforcement learning baselines in multiple tests, with low cost and interpretable results.

OpenAI post-training core member Weng Jiayi previously proved that 'purely relying on a large model to write code can clear Atari games.' Now, fluid dynamics PhD student Paul Garnier has applied this method to the more hardcore field of fluid dynamics control. He didn't train any neural network at all. He simply let Codex 5.5 act as a programmer, reading simulation test scores and diagnostic logs, debugging itself, and repeatedly rewriting Python scripts. In this industrial script test, Codex 5.5 outperformed Claude Opus 4.7 and Gemini 3 Pro, managing to defeat top deep reinforcement learning (DRL) baselines in over half of the physical scenarios. To reduce drag on cars or calm pipeline turbulence, the industry previously had to pour computational power into training an uninterpretable black-box model to control air valves. Codex avoided this dead end. The rules it writes are extremely straightforward, such as 'when local curvature is too high, delay jet injection.' Dozens of lines of short code with physical common sense directly replaced the neural network's brute-force trial and error. Replacing black boxes with code eliminates the neural network's vulnerability to rigidity and brittleness. Previously, any slight hardware change (e.g., switching from 5 to 10 control nozzles) would render the old model useless, requiring expensive retraining. Now, just change a constant in the code, and the system can instantly adapt to new equipment. When the test duration was forcibly extended fourfold, the traditional DRL models venturing outside their training experience completely collapsed; but the code written by the large model, directly following physical logic, remained stable throughout. Running this entire control strategy consumed only 21.25 million tokens at a total cost of less than 14 USD.
Original Article
View Cached Full Text

Cached at: 05/20/26, 06:27 AM

OpenAI post-training core member Weng Jiayu previously proved that “pure large language model code can beat Atari games.” Now fluid dynamics PhD student Paul Garnier has applied this approach to the far more demanding field of fluid mechanics control.

He never trained a single neural network. He simply let Codex 5.5 act as a programmer, reading simulation test scores and diagnostic logs to self-debug and iteratively rewrite Python scripts. In this industrial script test, Codex 5.5 outperformed Claude Opus 4.7 and Gemini 3 Pro, beating the top deep reinforcement learning (DRL) baselines in over half of the physical scenarios.

Reducing drag on cars, calming turbulence in pipes — industry previously could only throw compute at the problem, brute-forcing an opaque black-box model to control airflow valves. Codex sidesteps this dead end. The rules it writes are extremely straightforward, e.g., “when local curvature is too large, delay jet injection.” A few dozen lines of short code grounded in physical intuition directly replace the neural network’s mindless brute-force trial and error.

Replacing the black box with code eliminates the brittleness of neural networks. Previously, if the hardware changed even slightly (e.g., the number of control nozzles increased from 5 to 10), the old model would immediately break and require expensive retraining. Now, just change a constant in the code, and the system instantly adapts to new hardware.

When test time was forcibly extended fourfold, the traditional DRL models left their comfort zone and completely collapsed; but the LLM-generated code, which directly follows physical logic, ran steadily. The entire control strategy cost only 21.25 million tokens, with a total expense under $14.

pg (@pg_dons): 1/5 TLDR; We used Codex to discover and maintain heuristic learning for hard fluid dynamics control cases.

I’ve been applying DRL and GNN to physics since 2019, and over the past 3 months I’ve been toying with the idea of using agents in our processes. Inspired by the blog post

Similar Articles

@0xLogicrw: Former OpenAI post-training core member Jiayi Weng proposed a new reinforcement learning paradigm called "Heuristic Learning" in his personal capacity and open-sourced all experimental code. He used Codex (GPT-5.4) to repeatedly play the Atari game Breakout, but GPT-5.4 was never retrained...

X AI KOLs Timeline

Former OpenAI researcher Jiayi Weng proposed a new paradigm called "Heuristic Learning", which uses large language models to generate and iteratively modify Python code to solve reinforcement learning tasks. Knowledge is stored in interpretable code rather than neural network parameters, effectively avoiding catastrophic forgetting. It has achieved excellent results on Atari and MuJoCo benchmarks and the code has been open-sourced.

@php_martin: OpenAI Codex is now free. But what really shocked me isn't the free part — it's that local open-source models can deliver AI Agent performance close to the cloud experience. The video demonstrates 4 real-world scenarios: fixing a crashed space game, building a Whac-A-Mole web game in minutes, generating an Apple-style product homepage, and even launching a browser to search, download, and save files on its own.

X AI KOLs Timeline

OpenAI Codex is now free, but even more surprising is that local open-source models can achieve AI Agent performance close to the cloud, demonstrating scenarios like fixing games and developing web games without requiring API keys or internet.

@berryxia: Small model, big wisdom? It's now real! A 7B small model now acts as the boss of top large models like GPT-5, Claude Sonnet 4, Gemini 2.5 Pro. A new paper shows an RL-trained 7B model learned to write natural language subtasks, assign them to different models, precisely...

X AI KOLs Timeline

A new paper proposes training a 7B small model via reinforcement learning as a task scheduler, automatically decomposing subtasks and assigning them to top models like GPT-5 and Claude. It surpasses individual frontier models on several hard benchmarks, demonstrating that end-to-end reward learning can effectively replace manual prompt engineering and multi-agent pipeline design.