OpenAI can't rule out that it stole its most recent breakthrough

A mathematician, Tristan Buckmaster, and his collaborator Levent Alpöge spent a year working on a difficult problem in fluid dynamics, the finite-time blowup for fluid equations. They used various large language models (LLMs) to assist in their work, including OpenAI's Codex. After making significant progress, Buckmaster inquired about the potential for his work to be used in training the model.

OpenAI responded that while it was "unlikely," they could not rule out that de-identified user data from their product usage had helped improve their models. This means that the researchers and agents involved in the project were not aware of the work that had been included in the training data. OpenAI clarified that no specific user data was accessed or visible to them, and that the proofs and results differed significantly from their own.

This situation highlights the fact that consumer Codex sessions are automatically considered training data, and thus, everything users input may potentially be used in model training.