AI时代纯数学研究的前景：计算能力与人类创造力的平衡

斯蒂芬·沃尔夫拉姆近日撰文探讨了人工智能对纯数学研究未来的影响。1他认为，虽然AI在数学领域具有显著应用价值，特别是在知识库挖掘和问题求解方面，但纯数学的根本驱动力仍然是人类的想象力与创新思维。1沃尔夫拉姆强调，数学的本质不在于生成尽可能多的新定理，而在于发现能被人类理解的概念和结构。1

沃尔夫拉姆以自身经历为例说明这一观点。1他指出，1988年Mathematica（现为Wolfram Language）推出时，曾引发数学将被自动化替代的担忧，但实际上这一工具提升了可以进行的数学水平。1类似地，AI在数学中的主要价值在于"主题挖掘"——整合数百万篇数学论文中的知识并发现其中的连接。1然而，计算不可约性保证了即使遵循简单规则也能生成无法快速预测的新结果，但这些生成的定理往往与现有的人类数学概念无关。1沃尔夫拉姆提及，他在2000年利用自动定理证明找到了布尔代数最小公理系统，但所得证明极其冗长、低层次且"外星风格"，26年来仍未找到人类可理解的版本。1

沃尔夫拉姆认为，大语言模型本质上是统计模型，其复杂性增加时可靠性下降；自动形式化也存在问题，AI可能曲解数学意图而非真正将其形式化。1为应对这些挑战，沃尔夫拉姆团队正在开发新的Wolfram Language扩展，涵盖纯数学研究的基本构件（从层到李群到Clifford代数），目标是使纯数学的广泛领域实现计算化。1

Stephen Wolfram has examined how artificial intelligence will shape the trajectory of pure mathematics research, arguing that while AI offers significant computational advantages, the discipline's core remains fundamentally tied to human creativity rather than algorithmic power.1 Wolfram contends that pure mathematics is not primarily about generating proofs but rather about formulating the right questions and uncovering new mathematical concepts and structures.1 He notes that the true value of AI in mathematics lies in "topic mining"—the ability to synthesize knowledge from millions of mathematical papers and identify novel connections between them.1

The introduction of Mathematica, now known as Wolfram Language, in 1988 prompted similar concerns about mathematical automation, yet the technology ultimately elevated the level of mathematics that could be pursued rather than replacing it.1 Wolfram illustrates the distinction between computational capability and mathematical insight through a striking example: in 2000, he employed automated theorem proving to discover the minimal axiom system for Boolean algebra, but the resulting proof was extraordinarily lengthy, elementary in nature, and incomprehensible to humans—and no human-readable version has been found in the 26 years since.1 This outcome exemplifies a deeper principle: computational irreducibility ensures that even simple rules can generate results that cannot be predicted through shortcuts, yet the theorems produced often bear no relation to existing human mathematical concepts.1

Wolfram emphasizes that large language models are fundamentally statistical in character, with reliability declining as complexity increases, and that automated formalization presents inherent challenges, as AI may misinterpret intention rather than genuinely formalize human mathematics.1 To advance pure mathematics in the computational age, he is developing new extensions to Wolfram Language that will encompass research tools for pure mathematics—ranging from lattice theory to Lie groups to Clifford algebras—with the objective of making pure mathematics widely computational.1