Gary Marcus calls OpenAI's math result "amazing but vastly oversold."

His August 2 post concedes the result is real and attacks the inference as a fallacy of composition: success on one form of fancy cognition doesn't mean success on all forms is imminent. His sharpest technical objection is that math is uniquely favorable because it "allows for external tools to do verification and to create synthetic data," a property absent from open-ended real-world problems. He notes OpenAI published 249 pages on the results with not one page on how the model works, how proofs were verified, or what role humans played. Willison's complaint is narrower and more useful: he credits the Lean formalizations and reasoning-trace reconstruction as decent transparency, then names the gap that matters to practitioners. The prompts are withheld, and nothing says how many $2,000 failed attempts preceded the ten successes. A 10% hit rate turns a $2,000 proof into a $20,000 proof. That's the reusable check: published cost-per-success means nothing without the denominator.