A 100,000-line Lean proof of a 28-year-old conjecture. Written by an AI. In one week.
The paper dropped on arXiv. "Spectral Gap for the Binary Fixed-Margin Swap Chain" resolves the KTV conjecture, posed in 1997. The authors—affiliated with UC Berkeley, Brown, Yonsei, Concordia, and the Korea Institute for Advanced Study—used a large language model to generate the proof and OpenAI’s Codex to formalize it in the Lean proof assistant. The proof generation required "only a handful of prompting rounds." The formalization ate "about 100 hours" of Codex time, consumed "an entire week of my Pro quota," and spat out roughly 100,000 lines of Lean code.
This is not a toy problem. The binary fixed-margin swap chain is a real, open problem in probability and combinatorics. It sat unsolved for nearly three decades. An AI didn't assist a human breakthrough. It did the heavy lifting.
What took 28 years?
The KTV conjecture concerns the spectral gap of a Markov chain on binary matrices with fixed row and column sums. The spectral gap determines how fast a random process mixes—how quickly it reaches equilibrium. The problem sits at the intersection of probability, combinatorics, and statistical physics. A solution was sought for decades because the combinatorial structure is deep. Previous attempts demanded flashes of human insight that never fully materialized.
Lean is a proof assistant. It checks the logical validity of every step in a mathematical argument. Formalizing a proof in Lean is normally the bottleneck: a human mathematician can sketch an intuition in hours, but translating it into machine-checkable code can take years. This is the part the AI handled.
The machine did not think. It searched.
The mechanism is the story. The authors describe an iterative loop. They prompted the LLM, which generated candidate lemmas. A human checked whether the lemmas were plausible. Codex translated plausible lemmas into Lean. The verifier accepted or rejected the code. Rejections fed back into the prompts.
This is not a single flash of silicon genius. It is a brute-force search guided by a formal framework and a human with taste. The key insight: the LLM did not need to "understand" the problem. It needed to generate valid steps in a language, Lean, whose rules are absolute and machine-readable. That capability—generating plausible paths through a combinatorial explosion of proof states—is not the same as human intuition. It is orthogonal. It is also real.
The authors' own words land hardest. "Our theorem resolves the binary-matrix case after nearly three decades: it gives a worst-case optimal spectral-gap bound for every feasible set of margins, a result stronger than the original rapid-mixing conjecture itself." Then the admission: "Yet what stunned me even more was how the result was obtained and checked. In both proof generation and formalization, LLMs performed far beyond what I had expected."
A scientist surprised by their own tool’s capability. That is the signpost.
The human becomes the editor
The KTV proof is a proof of concept. The mechanism—LLM search plus formal verification—now works on an open problem of genuine difficulty. The consequences cascade.
The hardware race starts now. Within 24 months, every top-50 mathematics department will need a dedicated LLM cluster for automated theorem proving. Compute budgets will shift from human salaries to GPU time. The precursor is already visible: three AI systems just solved open math problems, and departments are beginning to fund compute, not postdocs.
The human role inverts. The scarce skill becomes prompt engineering and verification, not solitary proof construction. A graduate student in 2028 who can craft a search strategy that guides an LLM through a formal library will be more valuable than one who can solve a problem from scratch on a chalkboard. Mathematics curricula will teach Lean alongside, or instead of, LaTeX. The aesthetic of the field flips: the most admired minds will be those who can ask the machine the sharpest questions, not those who hold the longest chains of deduction in their own heads.
The career ladder breaks. Publishing a proof ceases to be a career-defining event. Verifying an AI-generated proof, curating the library of lemmas the machine generates, becomes the new unit of prestige. Tenure committees will scramble for new metrics. An assistant professor who maintains a widely used formal-math codebase and produces a stream of verified results may matter more than a Fields Medalist who works in isolation.
Here is the mechanism of displacement. The human in the KTV loop did two things: judged lemma plausibility and steered the prompt. Both are judgment tasks. The LLM did the construction. As formal libraries grow denser and verifiers faster, the judgment tasks shrink. Plausibility checking becomes automated when the verifier can reject false lemmas instantly. Prompt steering becomes automated when the system learns which search strategies succeed. The human is not replaced by fiat. The human is squeezed out as each subtask yields to automation. The endpoint is a closed loop: LLM proposes, verifier disposes, LLM adapts. No human in the loop.
A specific, falsifiable prediction follows. Within 24 months, a peer-reviewed journal will accept the first fully AI-generated proof with no human co-authors. The paper will likely land in combinatorics or number theory, where formal libraries are mature. The KTV paper had human co-authors. The next one will not need them.
The contrarian take is not that human mathematics ends. It’s that mathematics as a purely human endeavor ends. The best mathematicians will become interpreters of machine-generated structures, the way experimental physicists interpret data from a particle collider they did not design and cannot fully model. The field bifurcates: those who build the machines, and those who read their output. The middle ground—the lone human constructing a proof—shrinks fast.
If you run a math department
Start budgeting for a GPU cluster now. The window to adapt is 18 to 24 months. Shift grants from "proofs" as a deliverable to "verification infrastructure" as a public good. The relevant resource is no longer a single brilliant mind with a light teaching load. It is a compute cluster, a library of formalized math, and a small team of engineers who know how to talk to both.
If you are a graduate student in mathematics, learn Lean and learn prompt engineering. The job market you enter will not be the one your advisors entered. Pure problem-solving, unassisted, becomes a hobby—beautiful, but not bankable.
The last human proof?
The 100,000-line proof took a week of compute. The next one will take a day. The one after that, an hour. The question is no longer whether AI can do mathematics. The question is whether human institutions can process machine-generated truth at the speed it arrives.
The KTV conjecture was one of the last great problems to yield to a human-guided effort. The next one will be solved by a machine, and a human will click to read the answer.
