预训练。工程师搭建一个系统,让它根据前文预测下一个 token [8]。这个阶段它学到的几乎一切,都来自人类写下的海量非结构化文字。有一篇 AI 稿件里写着这样一步:“by the splitting theorem for vector bundles on the projective line, E is a direct sum of line bundles”。能写出这一步的模型,学到的是后半句极有可能紧跟在前半句之后。这个事实不是它发现的。1957 年格罗滕迪克证明了它 [9],此后数学家在论文、教材和讲义里一遍遍写下这一步。概率就是从这里来的。数学家说的直觉,是知道下一步该用哪个引理;模型的直觉,是从人类写下的千千万万个“下一步”中估计出来的条件概率。从信息论看,最小化预测误差就是压缩 [10, 11];参数能装下什么,取决于输入了什么。
他们定义什么算正确。训练时,验证器决定奖励;推理时,它决定许多次尝试里留下哪一个。标准写得松,模型就会钻空子 [58, 59]。数学之所以能成为试验场,是因为数学家早已把“证明”定义得极为精确。OpenAI 的 Navier–Stokes 论文写道,它确立的是情形 (C),“in the Millennium problem statement for Navier–Stokes as stated by Fefferman” [60],还专门核对了这份问题陈述的勘误中对压力的要求。它的 Lean 形式化建立在 Mathlib 之上 [61],这是一个由数学家和程序员多年共同建设的开源数学库。验证器本身,也是人类的工作。
OpenAI. OpenAI Math Release: research manuscripts and Lean formalizations. 2026. Apache 2.0; 722 manuscripts released 6 October 2026, 719 current after 7 October. github.com/openai/math
A. S. Besicovitch. On Kakeya’s problem and a similar one. Mathematische Zeitschrift 27, 312–320, 1928. doi.org/10.1007/BF01171101
Roy O. Davies. Some remarks on the Kakeya problem. Proceedings of the Cambridge Philosophical Society 69, 417–421, 1971. doi.org/10.1017/S0305004100046867
Thomas Wolff. An improved bound for Kakeya type maximal functions. Revista Matemática Iberoamericana 11, 651–674, 1995. doi.org/10.4171/RMI/188
Nets Hawk Katz, Izabella Łaba and Terence Tao. An improved bound on the Minkowski dimension of Besicovitch sets in R³. Annals of Mathematics 152, 383, 2000. doi.org/10.2307/2661389
Larry Guth. The endpoint case of the Bennett–Carbery–Tao multilinear Kakeya conjecture. Acta Mathematica 205, 263–286, 2010. doi.org/10.1007/s11511-010-0055-6
Hong Wang and Joshua Zahl. Sticky Kakeya sets and the sticky Kakeya conjecture. Journal of the American Mathematical Society 39, 515–585, 2026. doi.org/10.1090/jams/1067
Alexander Grothendieck. Sur la classification des fibrés holomorphes sur la sphère de Riemann. American Journal of Mathematics 79, 121, 1957. doi.org/10.2307/2372388
Larry Guth, Hong Wang and Ruixiang Zhang. A sharp square function estimate for the cone in R³. Annals of Mathematics 192, 551–581, 2020. doi.org/10.4007/annals.2020.192.2.6
Laurent Fargues and Peter Scholze. Geometrization of the local Langlands correspondence. 2024. arxiv.org/abs/2102.13459
Caucher Birkar. Anti-pluricanonical systems on Fano varieties. Annals of Mathematics 190, 345–463, 2019. doi.org/10.4007/annals.2019.190.2.1
Omprokash Das, Christopher D. Hacon and Mihai Păun. On the 4-dimensional minimal model program for Kähler varieties. 2024. arxiv.org/abs/2205.12205
Hugo Duminil-Copin, Clément Hongler and Pierre Nolin. Connection probabilities and RSW-type bounds for the two-dimensional FK Ising model. Communications on Pure and Applied Mathematics 64, 1165–1198, 2011. doi.org/10.1002/cpa.20370
Bertrand Duplantier, Jason Miller and Scott Sheffield. Liouville quantum gravity as a mating of trees. Astérisque 427, 2021. doi.org/10.24033/ast.1149
László Lovász and Balázs Szegedy. Limits of dense graph sequences. Journal of Combinatorial Theory, Series B 96, 933–957, 2006. doi.org/10.1016/j.jctb.2006.05.002
Jean Bourgain, Nets Hawk Katz and Terence Tao. A sum-product estimate in finite fields, and applications. Geometric and Functional Analysis 14, 27–57, 2004. doi.org/10.1007/s00039-004-0451-1
Grigori Perelman. The entropy formula for the Ricci flow and its geometric applications. 2002. arxiv.org/abs/math/0211159
Pierre Deligne. Hodge cycles on abelian varieties (notes by J. S. Milne). Hodge Cycles, Motives, and Shimura Varieties, Lecture Notes in Mathematics 900, 1982. www.jmilne.org/math/Documents/Deligne82_2003.pdf
Hong Wang and Joshua Zahl. Volume estimates for unions of convex sets, and the Kakeya set conjecture in three dimensions. 2025. arxiv.org/abs/2502.17655
Jonathan Bennett, Anthony Carbery and Terence Tao. On the multilinear restriction and Kakeya conjectures. Acta Mathematica 196, 261–302, 2006. doi.org/10.1007/s11511-006-0006-4
Shing-Tung Yau. On the Ricci curvature of a compact Kähler manifold and the complex Monge–Ampère equation. I. Communications on Pure and Applied Mathematics 31, 339–411, 1978. doi.org/10.1002/cpa.3160310304
Shiing-Shen Chern. A simple intrinsic proof of the Gauss–Bonnet formula for closed Riemannian manifolds. Annals of Mathematics 45, 747–752, 1944. doi.org/10.2307/1969302
Shiing-Shen Chern and James Simons. Characteristic forms and geometric invariants. Annals of Mathematics 99, 48–69, 1974. doi.org/10.2307/1971013
Shaoming Guo, Hong Wang and Ruixiang Zhang. A dichotomy for Hörmander-type oscillatory integral operators. Inventiones Mathematicae 238, 503–584, 2024. doi.org/10.1007/s00222-024-01288-8
Shaoming Guo, Changkeun Oh, Hong Wang, Shukun Wu and Ruixiang Zhang. The Bochner–Riesz problem: an old approach revisited. Peking Mathematical Journal 8, 201–270, 2025. doi.org/10.1007/s42543-023-00082-4
Terence Tao. A sharp bilinear restriction estimate for paraboloids. Geometric and Functional Analysis 13, 1359–1384, 2003. doi.org/10.1007/s00039-003-0449-0
Chenyang Xu and Lei Zhang. Nonvanishing for 3-folds in characteristic p>5. Duke Mathematical Journal 168, 1269–1301, 2019. doi.org/10.1215/00127094-2018-0066
Yu Deng, Zaher Hani and Xiao Ma. Long time derivation of the Boltzmann equation from hard sphere dynamics. 2025. arxiv.org/abs/2408.07818
John W. Morgan and Gang Tian. Ricci Flow and the Poincaré Conjecture. American Mathematical Society and Clay Mathematics Institute, 2007. www.claymath.org/library/monographs/cmim03.pdf
Jian Ding, Allan Sly and Nike Sun. Proof of the satisfiability conjecture for large k. Annals of Mathematics 196, 1–388, 2022. doi.org/10.4007/annals.2022.196.1.1
T. D. Lee, Kerson Huang and C. N. Yang. Eigenvalues and eigenfunctions of a Bose system of hard spheres and its low-temperature properties. Physical Review 106, 1135–1145, 1957. doi.org/10.1103/PhysRev.106.1135
Jean Leray. Sur le mouvement d’un liquide visqueux emplissant l’espace. Acta Mathematica 63, 193–248, 1934. doi.org/10.1007/BF02547354
Luis Caffarelli, Robert Kohn and Louis Nirenberg. Partial regularity of suitable weak solutions of the Navier–Stokes equations. Communications on Pure and Applied Mathematics 35, 771–831, 1982. doi.org/10.1002/cpa.3160350604
John Jumper et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589, 2021. doi.org/10.1038/s41586-021-03819-2
The mathlib Community. The Lean mathematical library. Proceedings of the 9th ACM SIGPLAN International Conference on Certified Programs and Proofs (CPP 2020), 367–381, 2020. doi.org/10.1145/3372885.3373824
Ilia Shumailov et al. AI models collapse when trained on recursively generated data. Nature 631, 755–759, 2024. doi.org/10.1038/s41586-024-07566-y
引用本文
@misc{li2026shoulders,
author = {Qiongxiu Li and {Sea-Fill Team}},
title = {What AI’s science is built on},
howpublished = {Sea-Fill, On Whose Shoulders},
year = {2026},
url = {https://seafill.info/Shoulders/blog/}
}