AI Is Solving 'Impossible' Math Problems. Can It Best the World's Top Mathematicians?

Why It Matters

AI’s growing capability to generate and test conjectures could dramatically speed mathematical discovery, influencing fields from physics to finance. However, its current unreliability means human expertise remains essential for validation.

Key Takeaways

• Meta AI solved 10.1% of Lyapunov problems, up from 2.1%.
• DeepMind's AI earned IMO silver, Gemini gold in 2024.
• Experts see AI aiding conjecture generation, not replacing proofs.
• Human oversight still required for AI‑generated mathematical insights.
• AI could accelerate research by tackling thousands of problems annually.

Pulse Analysis

Artificial intelligence has moved from mastering games to tackling abstract mathematics, a transition highlighted by Meta’s recent model that identified Lyapunov functions for more than ten percent of randomly generated dynamical‑system equations—a tenfold jump over legacy methods. Earlier this year DeepMind’s AlphaProof and AlphaGeometry 2 earned an IMO silver‑medal score, while Google’s Gemini Deep Think secured a gold‑medal equivalent by solving five of six Olympiad problems within competition time limits. These milestones demonstrate that large‑language and specialized reasoning models can now handle multistep symbolic reasoning, a prerequisite for genuine mathematical problem‑solving.

Despite the headline‑grabbing results, leading mathematicians remain cautious. Kevin Buzzard notes that no AI has yet produced a proof beyond human reach, and Neil Saunders warns that probabilistic outputs can mislead when absolute certainty is required. The recent FrontierMath workshop revealed that even the most advanced models need extensive human‑crafted prompts and post‑processing to be useful. Failures such as the incorrect conjecture generated in Marc Lackenby’s topology project underscore the necessity of expert validation. Consequently, AI is currently best viewed as a hypothesis‑generation tool rather than a substitute for rigorous proof.

The consensus among scholars like Fields Medalist Terence Tao is that AI will reshape the workflow of mathematics rather than eliminate the discipline. By automating the initial sweep of thousands of low‑hanging conjectures, researchers could allocate more time to deep, high‑impact problems that resist brute‑force methods. This acceleration has ripple effects across industries that depend on mathematical innovation, from cryptography to climate modeling. As models become more transparent and trustworthy, the partnership between human intuition and machine speed is likely to become a standard component of future mathematical research.