For 68 years, mathematicians have stared at a one-sentence problem that refused to die. This month, an AI-generated proof killed it — and Terence Tao, after days of heavy AI-assisted work, turned the machine's roughly 90,000-line formalization into a clean argument anyone can follow. Along the way, he found the machine had actually proven something stronger.
The 68-year-old problem
Sendov's conjecture, posed by Bulgarian mathematician Blagovest Sendov in 1958, sounds almost too simple for an open problem: take any degree-n polynomial whose roots all lie in the unit disk. For every root a, there must exist a critical point (a root of the derivative) within distance 1 of a.
The Gauss–Lucas theorem guarantees that critical points live in the convex hull of the roots, so something close always exists — but "within 1 of every single root" is a much stronger demand. Decades of work settled the small-degree cases (Rubinstein; later Brown and Xiang pushed the range further) and, in 2020, Tao himself proved the conjecture for all sufficiently large degrees. The intermediate range sat untouched.
What the AI actually did
On August 5, Lech Mazur used ProofAtlas, an AI theorem-proving tool, to produce a proof covering all degrees at once — verified end-to-end in the Lean proof assistant. The catch: the machine proof was roughly 90,000 lines of Lean, effectively unreadable to humans, and not in a publishable form.
That is where the "digestion" began. Tao spent several days, with heavy AI assistance (including ChatGPT sessions), reworking the argument: connecting it to the existing literature, stripping out redundancy, and isolating the main ideas. What emerged is almost embarrassingly elementary — no complex analysis beyond the fundamental theorem of algebra, and the deepest inequality needed is a special case of Maclaurin's inequality, derivable from the AM–HM inequality plus induction.
The digestion produced two surprises. First, the argument actually proves the "interior" version of the conjecture, which implies both Sendov's conjecture and the stronger Phelps–Rodriguez strengthening — a stronger result nobody had noticed was hiding inside the machine's work. Second, Tao re-formalized the entire proof in Lean with an AI agent in about 15,000 lines, six times smaller than the machine's original.
Why the machine won where humans stalled
The four "communication identities" that power the proof (centroid, polar, and two origin identities) were all hidden in pieces of the AI's argument. The machine found them through sheer search; humans never combined those pieces because 68 years of elegant prior art creates a gravitational field around expected approaches. Machines carry none of that baggage — they do not need the proof to be beautiful, they just need it to exist.
What this means for AI and for math
This is the cleanest demonstration yet of the pipeline Tao has been describing: generation → verification → exposition → publication → canonicalization. AI now handles generation and verification (Lean checks the work); humans handle exposition and canonicalization — turning a verified-but-unreadable proof into something a community can teach, build on, and absorb into textbooks. The bottleneck has moved: finding proofs is becoming abundant; digesting them is the scarce skill.
For the AI industry, the signal is bigger than one theorem. An AI tool resolved a 68-year-old open problem in pure mathematics, and the strongest mathematician alive spent days interacting with machine output as a collaborator — not a calculator. The pattern mirrors what is happening in coding agents: generate broadly, verify rigorously, curate carefully. AI research agents like Polaris are already industrializing this generate-verify-digest workflow for science.
What you can do about it
- Invest in verification, not just generation. Lean and similar tools are what turn unreliable models into useful ones.
- Treat AI output as raw material. The 90,000-line proof was unusable; Tao's 15,000-line version is the real artifact.
- Digest in public. Tao's write-up and chat logs are a template for turning machine findings into community knowledge.
- Watch the pipeline. Teams that master generate–verify–digest will produce research outcomes, not just drafts.
