Claude Helps Harvard Mathematician Crack 78-Year-Old Six-Dimensional Sphere Conjecture
Nashnova编辑部
Harvard mathematician Levent Alpöge and Anthropic's Claude jointly proved that a complex structure exists on the six-dimensional sphere S⁶, closing a 78-year open problem; this marks the first time AI has led the construction of a geometric object that did not previously exist, redrawing the boundary of what machines can do in pure mathematics.
What was this 78-year-old question actually asking?
Can the six-sphere S⁶ — the "shell" of a ball in six-dimensional space — carry a complex structure (a set of rules giving it complex-number coordinates, unlocking the tools of complex analysis)?
The question was posed in 1948. No one had produced a definitive answer — until August 24, 2026, when Levent Alpöge and Claude submitted a 108-page proof answering "yes."
This means → The AI did not search a known solution space or verify an existing candidate. It participated in constructing from scratch an object that had never existed — creation, not filtration.
How does the proof work? Three steps to build a complex structure on S⁶
Step 1: Fold the upper half-plane using a (3,4,∞) triangle group, producing a base sphere with three special points (t = 0, 1, ∞).
Step 2: Over every non-special point, attach a complex 2-torus (a four-dimensional "doughnut"), forming a fiber bundle. The three special points are left empty.
Step 3: Fill the three gaps with classical methods — Mumford's torus degeneration at t = ∞; Kodaira's logarithmic transforms (multiplicity 3 and 4) at t = 0 and t = 1. Once sealed, a compact complex threefold X is born.
In plain terms = build the skeleton, tile the surface, plug three holes — and once plugged, the whole thing is topologically S⁶.
How do we know what was built is actually S⁶?
Section 7 computes the fundamental group of X: substituting parameters (ℓ₀, ℓ₁, ℓ₂) = (0, 1, −1) yields the trivial group, matching a sphere.
Combined with the Hurewicz theorem, the Whitehead theorem, and Smale's 1961 generalized Poincaré conjecture, X is confirmed homeomorphic to S⁶.
This means → Because Kervaire and Milnor proved in 1963 that no exotic spheres exist in dimension six, homeomorphism upgrades directly to diffeomorphism — X is S⁶ itself, with no alternative.
Three breakthroughs in 35 days — how is this one different?
July 20: Claude Fable 5 produced a counterexample to the Jacobian conjecture (open for ~87 years).
August 10: A research-grade Claude raised the proven proportion of Riemann zeta zeros on the critical line from 41.6% to 67.2%, deploying ~60 sub-agents, executing over 2,400 shell commands, and consuming ~31 million output tokens.
This reflects a qualitative shift: the first two results can be framed as efficient search within known solution spaces. The S⁶ complex structure did not previously exist — this time the model participated in construction from nothing.
What does the math community say? What's still missing?
Mathematician Qiaochu Yuan audited the proof with GPT-5.6 Sol in ~21 minutes, finding no gaps. Sol's verdict: if the 108 pages hold, "arguably the most important AI math result to date."
Justin Curry, associate professor at SUNY Albany, said: if the proof is correct, "this is absolutely the most remarkable recent AI achievement."
The proof has not yet undergone peer review; the final verdict awaits formal verification. This means → it is a very strong candidate answer, not yet a settled theorem.
What does this mean for the boundary of AI capability?
Alpöge's own research is in number theory and arithmetic geometry — not complex geometry. He simultaneously holds an Anthropic postdoctoral position — a dual role that itself signals AI played far more than an assistant's part.
In plain terms = previously, AI doing math looked like a "super-calculator": you told it what to search, and it searched fast. This time it participated in designing the blueprint — constructing a new object at the conceptual level.
This reflects a deeper signal: if the proof ultimately stands, the default assumption that "AI can assist but cannot create" will be formally broken.
Content is for reference only, not financial advice.