I'm a big fan of Two-Minute Papers. It's one of the primary ways I keep up-to-date with the data science world. Today the channel posted this video covering the contents of the math proofs OpenAI published last week:
Károly Zsolnai-Fehér’s breakdown exposes the central tension of modern AI: the growing divergence between computational verification and human comprehension. He notes that frontier reasoning models have decisively moved past basic arithmetic errors, Yet the resulting outputs frequently resemble what he frankly terms a “big ball of barf:” dense, unreadable computational artifacts that confirm an assertion without conveying why it holds. However, is this really a problem? Is math simply an optimization function to output binary truth values, or is it an epistemic discipline designed to expand human understanding?
In mathematical philosophy, this distinction marks the collapse of the assumption that formal verifiability equates to scientific progress. Weatherall and Wolfson (2026) argue “that a mathematical proof be correct is neither sufficient nor necessary for it to have epistemic value.” They caution that “the mere generation of correct proofs is not, in itself, a contribution to mathematics,” because it bypasses conceptual synthesis, definitions, and reusable abstractions that mathematicians rely on to build theory. If a machine takes three hours of test-time compute to brute-force a lemma across thousands of un-annotated verification steps, it creates an answer that requires another neural model just to audit.
This output deluge threatens to choke academic peer review. Tao (2026, p. 10) believes that as we “transition from an era of proof scarcity to an era of proof abundance,” our institutions (”journals, priority conventions, hiring and promotion criteria, prizes, the very notion of a research program”) will likely “behave poorly” because they “were designed under the assumption of scarcity.” When proof generation is cheap but human digestion remains bounded by working memory, the scarce resource becomes expert attention. Tao (2026, p. 11) establishes a necessary boundary: “A proof that no human can properly explain should be viewed as incomplete, even if it has been formally verified.”
….But again, I have to ask myself, is this really a problem? We’re talking about human understandability as the method to verify AI-generated proofs. We are convinced that humans have to be in the loop, kind of like a micromanaging boss. But what if we could ensure results were accurate without a human understanding the output? There’s a tinge of sadness for me as I write that out, as if I’m abdicating a core aspect of human experience to a machine, but for the sake of progress I have to consider that accuracy without understanding could be viable. Humans in organized society already specialize into skill-sets and outsource other work and knowledge to trusted professionals; if we had a way to establish trustworthiness of an AI, how is abdicating our understanding of mathematics to AI any different?
I admit this is pie-in-the-sky thinking, and does not truly address the reality we live in right now. AI is still prone to hallucinations and mistakes, as made evident by the forced retraction of several OpenAI papers in the last few days. If mathematical discovery becomes concentrated in proprietary server farms executing “drive-by provings” (Avigad, 2026, p. 2), independent validation becomes nearly impossible. For the time being, democratizing open-weights models and formal verifiers is essential to prevent frontier labs from turning scientific research into a black box.
References
Avigad, J. (2026). Mathematicians in the age of AI. arXiv. https://doi.org/10.48550/arXiv.2603.03684
Tao, T. (2026). Mathematics in the age of AI. arXiv. https://doi.org/10.48550/arXiv.2608.16753
Weatherall, J. O., & Wolfson, J. (2026). Correctness, artificial intelligence, and the epistemic value of mathematical proof. arXiv. https://doi.org/10.48550/arXiv.2602.12463
Zsolnai-Fehér, K. [Two Minute Papers]. (2026, October 13). OpenAI just changed what it means to do math [Video]. YouTube. https://www.youtube.com/watch?v=z_sXLlLVUls