[Sharing] Wei Zhewei: Three Questions on Mathematics in the AI Era

Original Media · 11m · 中文 · YouTube · Best Partners TV

TLDR

A mathematician and AI researcher raises three profound questions about AI's impact on mathematics: whether mathematics is truly elegant, whether human mathematical intuition is reliable, and whether the feedback loop between mathematical discovery and application must remain long. Using examples from Go (AlphaGo), theoretical computer science (3SUM conjecture), and a discrepancy theory problem solved with AI assistance, the speaker argues that AI may shatter human aesthetic expectations, overturn deep-seated intuitions, and dramatically compress the time from problem formulation to solution.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

The optimal strategy for the discrepancy problem (Tusnády's problem) is to place points completely randomly — a result that 'completely violates all human intuition' of experts in the field.

•

AlphaGo's 'move 37' against Lee Sedol had a 1 in 10,000 probability in human play and was initially judged a mistake by commentators, yet turned out to be a brilliant, game-winning move.

•

The 3SUM conjecture, a cornerstone of fine-grained complexity theory used as a hardness assumption in hundreds of papers, was broken not by a theoretical breakthrough but by an algorithmic improvement found by researchers (likely aided by computational search).

Key takeaways

•

AI systems like AlphaGo have already destroyed human aesthetic concepts in Go (e.g., the 'six stones on the second line' move), suggesting a similar fate may await mathematical beauty.

•

The 3SUM conjecture in theoretical computer science — long believed to have an elegant n² lower bound — was recently broken with an n¹·⁹⁹⁹⁹² algorithm, challenging the notion that fundamental problems have 'beautiful' complexity.

•

Human intuition in mathematics can be profoundly wrong: AI discovered that the optimal point configuration for a classic discrepancy problem is simply random points, contradicting all expert intuition.

•

The feedback loop between mathematical discovery and real-world application (historically decades or centuries) could shrink to hours with AI, fundamentally changing the role of mathematicians.

•

Mathematics may return to an era of tight coupling with physics and engineering, where humans formulate problems and AI solves them.

Notable quotes

“AI从某种意证来说,因为他获得了更强的起力,它是摧毁了人类对于围棋的很多的美感”

▸ Notable quotes · 2:13

“这样重要的一个问题它都不是N平方的,因为N平方我觉得是很自然的,很很很具备对称美的一个东西”

▸ Notable quotes · 4:36

“特别蓝染色的点长什么样子呢?特别蓝染色的点就是随机在平面上撒一把点,这个就完全违背了人类的所有的直觉”

▸ Notable quotes · 8:19

“以前觉得可能是不可能解的啊也没有任何的数据工具解的,但是以后呢用AI可能就数小时就把它解决了”

▸ Notable quotes · 9:49

“数学将来如果AI发现了它不是一个或者说它在很多结构上并不如我们想象的优雅那我们应该怎么办”

▸ Notable quotes · 9:49

Chapters

•

0:00 — Introduction: AI's Deep Impact on Mathematics: The speaker, a Peking University mathematics graduate now working in AI and theoretical computer science, introduces three open questions about how AI may fundamentally change mathematics, not just accelerate it.

•

0:37 — Question 1: Is Mathematics Truly Beautiful?: Examines the belief that mathematics is inherently elegant. Uses AlphaGo's destruction of Go aesthetics (e.g., six stones on the second line) and the recent breaking of the 3SUM conjecture (n² → n¹·⁹⁹⁹⁹²) as evidence that AI may reveal mathematical structures as less beautiful than humans assume.

•

5:10 — Question 2: Is Human Mathematical Intuition Reliable?: Discusses how AlphaGo's 'move 37' (1/10,000 human probability, initially deemed a mistake) overturned Go intuition. Parallels in mathematics: AI solved a discrepancy problem (Tusnády's problem) showing the optimal point set is random — contradicting all expert intuition that it should be structured.

•

8:48 — Question 3: Must Mathematical Feedback Loops Be Long?: Highlights historical gaps of 50–100 years between mathematical discovery (e.g., Radon transform, 1917) and application (CT scans, 1970s). Argues AI could compress this to hours: domain experts formulate problems, AI solves them, revolutionizing the math-science-engineering pipeline.

•

10:33 — Conclusion: Three Questions for the Community: Poses the three core questions: (1) What if math isn't elegant? (2) How do we learn from AI when our intuition is wrong? (3) What happens when feedback loops shrink from centuries to hours? Calls for collective reflection.

Speakers

•

Speaker 1 — Presenter/Researcher

⋮