The post argues that AI is not mysteriously “outthinking” mathematicians. It is winning in narrower ways that still matter a lot. Models can keep far more material in play at once, scan huge bodies of prior work, and keep grinding without fatigue. That makes them unusually strong at finding overlooked combinations, stitching together distant ideas, and pursuing many dead ends humans would abandon for time, ego, or career reasons. A lot of people found that framing basically right, especially for math and code, where outputs can be checked by compilers, theorem provers, tests, or other verifiers.
What people added is that this is not just “memory” in the everyday sense. It is breadth, persistence, and cheap iteration under validation. Several mathematicians said the bigger bottleneck in their field is social, not cognitive. Negative results, failed paths, and the messy route to a proof are useful but rarely published because academia rewards the polished rabbit-out-of-a-hat version. That is exactly the kind of search trace AI can generate and reuse at scale. Others pushed the same point from software. Models are great at chewing through boilerplate, porting patterns, and cross-referencing large codebases. They are much worse at taste. They miss when a helper function should exist, when five repeated chunks want one abstraction, or when a tiny naming or structural issue makes code harder to live with later. In practice they still need a human who understands the design and can manage context.
The strongest consensus was that current models are best where the problem can be scored. Code and
formalized math keep coming up because the loop is tight. Generate, test, refine, repeat. That same loop is weaker in law, medicine, and open-ended science, where correctness is contested, consequences are high, and there is no clean validator. People also drew a line between producing a correct proof and producing value. In math, a proof matters because humans can turn it into shared understanding, trust, and future work. A giant correct symbol dump that no one can interpret is less impressive than AI boosters suggest. Several commenters still expect AI to push into results that arrive before clean explanation, but even they mostly described that as a tool for centaurs, not a replacement for human judgment.
The mood around the article itself was skeptical. Many thought the title was a dressed-up restatement of something obvious, and some dismissed it as another round of redefining intelligence whenever AI gets good at a task. But the more grounded comments converged on a practical picture. Models are already very useful as superhuman recall plus relentless search. That is enough to move math, coding, and literature review. It is not enough to make abstraction, explanation, or scientific taste irrelevant.
davidepiffer.com
Not Even Wrong