AI Conquers 100+ Math Problems: Why Human Understanding Matters More Than Ever

On September 21, OpenAI announced that an internal model trained since August 28 had solved over 100 world-class mathematics problems across most fields. This triggered anxiety in the mathematics community: PhD students worry about degree devaluation, journal editors fear collapse, and veteran proof-writers question their craft's remaining lifespan.

Amid this gloom, University of Toronto mathematician Daniel Litt published a blog post titled "A beginning for mathematics," revising his earlier talk "The End of Mathematics." Litt accepts a radical premise: AI that consistently surpasses humans on most mathematical tasks may arrive soon. His focus is on how the profession, training, and human effort should adapt.

The Core Argument: Understanding vs. Production

Litt identifies two goals of mathematics: producing and understanding high-quality mathematics, and training high-quality mathematicians. Historically, both were achieved through proving theorems — main theorems in papers, theses, and solutions to long-standing conjectures.

He illustrates the gap with a thought experiment: a monkey (or computer) given ZFC axioms and mechanical inference rules could generate endless logically valid theorems. Enumerating all propositions and attempting proofs would eventually hit open problems. This absurdity reveals that the community values more than "this proposition is proved."

Even an extremely clever monkey that understands its proofs, writes elegantly, solves open cases, and proposes new fundamental questions cannot produce human understanding. Litt predicts a "mathematics explosion" where models generate far more results, but as long as we care how much humans understand, mathematicians will not become redundant — they may become busier.

What to Preserve: Seminars Over Journals

Litt distinguishes mathematics' intrinsic value from today's institutional forms. He wants to preserve learning seminars, spontaneous idea collisions, students discussing problems with advisors, and collective sense-making. Journals, peer review, arXiv, and authorship conventions need not remain unchanged.

He warns against chasing AI's current weaknesses (e.g., theory-building, question-posing) as new evaluation criteria, because institutional change is slow while model capabilities advance rapidly. Instead, consider the endgame: if AI masters nearly all mathematical skills, what do we still want to keep?

Redefining the PhD: Expertise and Communication

Traditionally, a PhD thesis both contributed new results and signaled the student's mastery. AI breaks the signaling function: a student could submit a beautiful, novel thesis they haven't fully read or understood.

Litt proposes separating mathematical progress from mathematical ability. The future PhD goal: become a genuine expert on an important, deep problem and be able to communicate that understanding. Theses may still exist, but degree conferral should rely more on rigorous oral defense — explaining the problem, handling expert questioning, adapting methods to new situations. Whether AI originally found the result becomes secondary.

Traditional evaluation modes — talks, long discussions, face-to-face mathematical interrogation — regain value as papers become easier to generate.

The Underrated Skill: Knowing What Problems Matter

AI trains on open problems that human mathematicians have already filtered for value. A conjecture becomes an "open problem" because many people deemed it worth years of effort. AI appears to solve autonomously, but its "good problems" were pre-selected by the community.

Litt argues this curation skill has been under-rewarded. Future credit could explicitly go to those who propose good problem sets, build long-term research programs, and convince others "this is worth digging." He acknowledges AI will likely also improve at posing questions and designing research agendas, eventually producing massive PDFs.

When PDFs Flood In: The Understanding Bottleneck

If a model conjectures and proves a beautiful result but no human understands or cares about it, its pure-mathematics value is limited. More mathematics means more need for mathematicians to understand assumptions, scope, failure modes, and applications.

Productivity gains don't eliminate understanding costs; they expose the bottleneck more sharply.

Mathematics Can Now "Pay-to-Win"

Mathematics was perhaps the cheapest science — one person, one blackboard. AI makes some problems solvable by injecting cash: more compute, more tokens, more inference rounds. Some mathematicians lament the loss of the "detour" — developing auxiliary theories and tools while circumventing a hard problem, often discovering something more important.

Litt acknowledges the loss but counters: if a basic problem costs a nice dinner to solve, we should be happy. The answer is just the start: what does it explain? What do we understand? Why does it hold? Where does it generalize? What new questions arise? Some follow-ups may again be solved for a dinner's price; others will plunge everyone back into confusion, spawning new communities.

The Final Step No AI Can Take

Litt returns to a mundane scene: a student stuck on a problem, knocking on a professor's door. They may use AI or not, may first struggle at the blackboard. The model may later give a beautiful explanation. But "seeing an explanation" and "understanding it yourself" remain separated by a step no one else can complete.

Litt assumes AI will keep strengthening, solve many currently hard problems, and render many papers, PhD training methods, and evaluation systems obsolete. But mathematics won't be "finished." As answers become easier, we'll still ask what they mean; each resolved confusion births new ones. He concludes: "There is still so much for us to learn — infinitely so. We have always stood at the beginning of mathematics. We always will."

Reference: https://proofsandprompts.com/2026/09/14/a-beginning-for-mathematics/