The canonical public record of elliptic curve ranks, maintained by mathematician Andrej Dujella, now lists a 2026 entry at rank 30 or higher, publishing one explicit curve together with 30 explicit independent rational points that any mathematician can verify. A separate leaderboard's commentary attributes the find to Claude working with two named researchers. The artifact is real and checkable; almost everything being said about how the AI contributed is not documented in any primary source.
Key facts
- The record page publishes an explicit curve plus 30 independent rational points, establishing a verified lower bound of rank at least 30.
- Dujella's rank history table records rank at least 28 in 2006, at least 29 in 2024, and at least 30 in 2026.
- The named human collaborators are Levent Alpoge, a mathematician at Anthropic and a junior fellow in Harvard's Society of Fellows, and Ava Howell, who describes herself as a systems engineer, cryptographer and AI researcher.
- No Anthropic announcement covering this specific result exists.
An elliptic curve is an equation whose solutions in whole-number ratios have a beautiful and stubborn structure. Some curves have only finitely many such solutions. Others have infinitely many, generated from a small set of starting points, and the size of that generating set is the curve's rank. Nobody knows whether ranks are bounded -- whether there is some ceiling beyond which no curve can go -- and that open question is why the record matters. Every few years someone finds a curve with one more independent point than the previous best, and the ceiling recedes a little further.
What makes this a good subject for AI-assisted work is that the search space is astronomically large and the verification is cheap. Finding a candidate curve requires enormous computational hunting. Checking one is arithmetic: the points are published, you plug them in, and either they satisfy the equation and are independent or they do not. This is the ideal shape for machine assistance -- expensive to find, trivial to confirm -- and it is the same shape as AlphaEvolve tightening the matrix multiplication exponent and Claude designing protein binders for 14 of 15 targets.
An analogy: finding a very long unbroken sequence in a lottery's history takes months of searching, and confirming someone else found one takes an afternoon.
So the verified core is solid. There is an explicit curve on a respected mathematician's canonical record page, with 30 explicit independent points, and any competent number theorist can check it this week.
Three things are not verified, and the distinction is the story.
First, the exact rank. The published artifact establishes a lower bound: the rank is at least 30. Establishing that it is exactly 30 requires assuming two major unproven conjectures. A lower bound backed by explicit points is a strong, checkable claim. Exactness is conditional, and the two should not be blurred.
Second, and most importantly, what Claude did. The record page attributes the submission to a handle. A separate leaderboard's commentary history carries the line crediting Claude alongside Alpoge and Howell. That is attribution, not a description of work. Did the model propose the search strategy? Generate candidate curves? Write the search code? Execute the computation? Review results a human generated? Every one of those is a wildly different claim about AI capability, and no published source distinguishes between them. There is no Anthropic blog post about this result.
Third, a specific claim circulating about the history is wrong, and the primary source is right there. The idea that the previous step -- rank 28 to rank 29 -- took ten years does not survive contact with Dujella's own history table, which records rank at least 28 in 2006 and rank at least 29 in 2024. That is eighteen years. The 29-to-30 step then took roughly two.
That eighteen-to-two compression is genuinely striking and does not need embellishment. It is also worth noting that the 2024 record predates the current generation of models, so a two-year gap is not by itself evidence of anything about AI.
The honest caveat, and the reason this story is worth telling carefully rather than loudly, is that mathematicians will notice the difference between "a rank 30 curve was found with AI assistance" and "Claude found a rank 30 curve." The first is defensible and interesting. The second is a stronger sentence than any primary source supports, and overclaiming on a result whose verifiable core is unusually solid is a self-inflicted wound. The artifact deserves better than the framing it is getting.
Related reading on this site: what is a proof assistant, recursive self-improvement, and reinforcement learning with verifiable rewards.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)