x.com/willdepue/status/2099746385486635469

OpenAI directly contradicted this characterization in its September 8, 2026 article about the Navier–Stokes result. In that post, OpenAI wrote that the internal model used for the result "was developed through large-scale reinforcement learning on top of a previously pretrained model."

The same post also says "Since August 28 we have been training a new internal model", but that does not mean the system was a brand-new model being pretrained from scratch. OpenAI's explicit clarification is that the result came from RL applied to an already pretrained base model.

So the claim that it was "a new pretrain and not some RL run" is contradicted by the company's own description of how the model was developed. It also means any further inference based on the idea that this was only a couple weeks of from-scratch pretraining is not supported by the source OpenAI published.

This claim treats the model as if it were a fresh system only ~11 days into pretraining when OpenAI used it for the Navier–Stokes effort announced on September 8, 2026. But OpenAI's own description says the opposite about the training stage involved.

In the same announcement, OpenAI wrote that "Since August 28 we have been training a new internal model" and that its training was still ongoing. However, OpenAI then explicitly clarified that "The OpenAI internal model used for this result was developed through large-scale reinforcement learning on top of a previously pretrained model."

That means the relevant system was not merely a model less than two weeks into from-scratch pretraining. It was a previously pretrained model that had undergone additional large-scale RL. So the post's phrasing is inaccurate about what stage of training the model was in.