Inside OpenAI’s Breakthroughs in Mathematical Reasoning

播客中，来自OpenAI的Mark Selkey和Meitab Swani，两位前数学家，讨论了AI对他们领域带来的变革性影响。他们将“可达结果的复兴”（renaissance of reachable results）归因于AI的独特能力，这些能力与人类的数学实践有所不同。 Meitab分享了他的“转变”时刻，当时“GPT-5”（如转录稿所述）迅速为一个未解决的埃尔德什问题找到了参考文献，而他和同事们曾为此花费数小时却一无所获。这凸显了AI在快速搜索大量文献和建立联系方面的优势，这对于人类而言通常是“难以做到”（humanely hard）的任务。 Mark和Meitab指出了AI在数学领域的几个关键优势： 1. **高效执行：** AI擅长执行复杂的想法，在人类经常“迷失于琐碎细节”（finicky details）的地方，它“总是能精准地完成这类论证”（always nail[ing] these kinds of arguments）（例如，epsilon-delta证明）。 2. **毅力和冒险：** 与人类数学家可能在数小时或数周后因感知到的低“风险回报”而“放弃”一个想法不同，AI在得到指示后会“就这么做”（just do this），以“极其顽强”（extremely dogged）的毅力追求探究方向。这种能力有助于将“可达结果”（reachable results）呈现在眼前。 3. **干净的回溯：** AI可以从失败的路径中回溯，而其“上下文窗口”（context window）不会被先前的尝试“污染”，这对于人类认知来说是一个挑战。它能够有效地“像上周的你一样，再造一个克隆体，然后说，你知道，别这么做，试试别的”。这使得解决方案空间可以并行探索。 4. **涌现推理：** AI在解决问题时产生的“总结性思维链”（summarized chain of thought）“惊人地像一位人类专家”，类似于“阅读同事的笔记”。这表明AI正在做出“正确的决策，并最终能够使用‘良好的数学判断力’来修剪搜索树”，而不是暴力穷举所有可能性。 讨论随后转向了OpenAI的AI“Astra”所解决的具体问题。一个值得注意的例子是**球体堆积问题**，它询问在“d”维度中球体可以堆积得多密。尽管对于 d=1、2、3、8 和 24 的情况存在解决方案，但高维情况在数值猜想之外基本上是未知的。Astra 为高维线性规划（LP）边界提供了一个渐近界（大约是 2^(-0.604d)），这一结果之前仅通过“极其巧妙的数值计算”（extremely clever numerics）被猜测。Mark指出这个证明“相当短”，“令人难以置信”，并且解释了“其来源”（where this is coming from），展示了一种优雅的复分析方法。 Astra 还改进了**球面码和二元码**的界限，这些问题分别与弯曲表面和超立方体上的球体堆积有关，并应用于纠错。这涉及利用“复杂得多的表示论”（much more sophisticated representation theory）。独特的是，这个问题涉及了一些人类交互；最初被要求改进码的界限，Astra在被提示后进一步深入，揭示了与球体堆积问题的深层联系。这凸显了AI以任务为导向的特性，以及在指导下提升复杂度的能力。 另一个重要成就是Astra反驳了每个群都是**sofic**（意味着它可以被有限群近似）的猜想。尽管一个相关的、更强的猜想（Aldous-Lyons）此前已被一个涉及量子复杂性理论的250页证明所推翻，但Astra对非sofic群的证明“短得多且更容易”，大约15页，并且停留在群论范畴内。它识别出了一种以前的数学家无法排除的“具体组合结构”（concrete combinatorial structure），展示了“非常精妙的组合学”（very delicate combinatorics）和一个被AI推翻的“奇怪的阴谋”（weird conspiracy）。Mark指出，这些AI生成的证明通常“非常短且优雅”，与冗长的人类证明形成了令人惊讶的对比。 关于数学界的适应，发言人认为AI将改变数学进步中的瓶颈。尽管证明结果将变得不那么具有挑战性，但理解、吸收、内化和解释新数学的任务将变得至关重要。AI可以加速理解，使得“使用模型理解一段数学知识比不使用模型快得多”（much faster to understand a piece of mathematics with a model than without it）。这种提高的可及性可以允许更多人参与数学。数学家的角色可能会演变为侧重于提供“共同理解”（communal understanding），并与像P vs. NP这样可能超出AI能力范围的“大谜团”作斗争。最终，AI有望加速应用数学的发展，并为长期存在的问题带来答案。

The podcast features Mark Selkey and Meitab Swani from OpenAI, both former mathematicians, discussing the transformative impact of AI on their field. They attribute a "renaissance of reachable results" to AI's unique capabilities, which differ from human mathematical practice. Meitab shared his "conversion" moment when "GPT-5" (as stated in the transcript) quickly found a reference for an open Erdős problem that he and colleagues had spent hours unsuccessfully researching. This highlighted AI's strength in rapidly searching vast literature and making connections, a task often "humanely hard." Mark and Meitab identified several key strengths of AI in mathematics: 1. **Efficient Execution:** AI excels at executing complex ideas, "always nail[ing] these kinds of arguments" (e.g., epsilon-delta proofs) where humans often get lost in "finicky details." 2. **Persistence and Risk-Taking:** Unlike human mathematicians who might "give up" on an idea after hours or weeks due to perceived low "risk-reward," AI will "just do this" if directed, pursuing lines of inquiry with "extremely dogged" persistence. This capability helps bring "reachable results" into view. 3. **Clean Backtracking:** AI can backtrack from failed paths without its "context window" being "polluted" by the prior attempts, a challenge for human cognition. It can effectively "make another clone of yourself from, like, last week and say, you know, don't do this, try something else." This allows for parallel exploration of solution spaces. 4. **Emergent Reasoning:** The "summarized chain of thought" generated by AI during problem-solving is "shockingly like an expert human would," resembling "reading a colleague's notes." This demonstrates that AI is making "the right decisions and is eventually able to prune the search tree" using "good mathematical judgment," rather than brute-forcing every possibility. The discussion then moved to specific problems tackled by OpenAI's AI, "Astra." One notable example was the **sphere packing problem**, which asks how densely spheres can be packed in 'd' dimensions. While solutions exist for d=1, 2, 3, 8, and 24, high-dimensional cases were largely unknown beyond numerical conjectures. Astra provided an asymptotic bound for the linear programming (LP) bound in high dimensions (roughly 2^(-0.604d)), a result previously only conjectured through "extremely clever numerics." Mark noted this proof was "quite short," "unbelievable," and explained "where this is coming from," demonstrating an elegant complex analysis approach. Astra also improved bounds for **spherical and binary codes**, problems related to sphere packing on curved surfaces and hypercubes, respectively, with applications in error correction. This involved leveraging "much more sophisticated representation theory." Uniquely, this problem involved some human interactivity; initially asked to improve code bounds, Astra then pushed further when prompted, revealing a deep connection to the sphere packing problem. This highlighted AI's task-oriented nature but also its ability to ascend in complexity when guided. Another significant achievement was Astra's disproof of the conjecture that every group is **sofic** (meaning it can be approximated by finite groups). While a related, stronger conjecture (Aldous-Lyons) was disproved earlier with a 250-page proof involving quantum complexity theory, Astra's proof of a non-sofic group was "much shorter and easier," around 15 pages, staying within group theory. It identified a "concrete combinatorial structure" that previous mathematicians couldn't rule out, showcasing "very delicate combinatorics" and a "weird conspiracy" that the AI disproved. Mark noted that these AI-generated proofs are often "very short and elegant," a surprising contrast to lengthy human proofs. Regarding the math community's adaptation, the speakers believe AI will shift the bottlenecks in mathematical progress. While proving results will become less challenging, the tasks of understanding, absorbing, internalizing, and explaining new mathematics will become paramount. AI can accelerate understanding, making it "much faster to understand a piece of mathematics with a model than without it." This increased accessibility could allow more people to participate in mathematics. Mathematicians' roles might evolve to focus on providing "communal understanding" and grappling with "big mysteries" like P vs. NP, which might remain beyond AI's reach. Ultimately, AI offers the prospect of accelerating applied mathematics and bringing answers to long-standing problems.