Complementary remarks from Gary Marcus and Terence Tao on OpenAI’s giant math drop

Gary Marcus published a critique of OpenAI's announcement of a new math milestone in an essay he described as written in extreme haste. The piece, titled with complementary remarks from Marcus and Terence Tao, is split into two parts; only Marcus's take is included, while Part Two is labeled for Tao.

Marcus argues the significance is not the result itself but the lack of disclosed information. He points to OpenAI's report using phrases such as "Same procedure" and "Using an unreleased model," and says such a report would never pass peer review. He lists unknowns: the procedure, the architecture, the failure rate, and details on training, post-training, and data augmentation. He also asks whether proofs were generated in one shot and then verified by the symbolic system Lean, or whether an iterative process was used.

Because those details are missing, Marcus says there is zero idea of how generalizable the result is outside math. He criticizes social media discussion as an ignorant cheering section that applauds without asking basic scientific questions. The new system, he writes, could be a legitimate step toward AGI, or it could be a clever leveraging of Lean and synthetic data in a verifiable domain with no generality whatsoever. From the initial report, he concludes, almost nothing can be told.

Why it matters: For developers and product teams, the critique highlights that without published details on procedure, architecture, failure rates, and training, the new math result cannot be evaluated or assumed to transfer to other domains.