The Hodge Conjecture Rumor: OpenAI, a Viral Tweet, and the Math World’s Quiet Fury

Around September 9, 2026, a post on X began to be shared repeatedly. The user synthwavedd wrote that the Hodge conjecture “is very close to being verified by OpenAI,” and said that either OpenAI or Anthropic was also close to solving the Birch and Swinnerton-Dyer conjecture. He added that the intensity of the behind-the-scenes race for “the next one” was unlike anything he had ever heard described.

If true, this would mean three Millennium Prize Problems had been conquered within a single month. Chinese internet discussion quickly followed, and a related question on Zhihu received more than 40,000 views.

The Clay Mathematics Institute issued no statement. OpenAI issued no statement. At present, the Hodge conjecture has only internet rumors. There is no corresponding paper, no code, and no confirmation from any authoritative institution.

This is not the first such rumor in 2026, and it will not be the last.

The Hodge conjecture was proposed by the Scottish mathematician William Vallance Douglas Hodge in 1950. Its core claim is: on a nonsingular complex projective algebraic variety, every Hodge class can be expressed as a rational linear combination of algebraic cycle classes.

For the vast majority of people, this formulation is gibberish. If one uses an analogy to understand it, topology studies the “soft” features of shapes, such as “how many holes does this shape have,” where paths can be bent arbitrarily. Algebraic geometry studies the “rigid” structures defined by polynomial equations, which cannot be deformed at will. The Hodge conjecture asserts: for a certain class of well-behaved spaces, every abstract “hole” discovered by topologists is underlain by a hard skeleton defined by algebraic equations.

In other words, any topological feature can ultimately be described by precise algebraic equations.

This conjecture brings together the most abstract concepts in modern mathematics: cohomology, Hodge structures, algebraic cycles, and the theory of motives. Only a very small number of professional mathematicians can understand its deeper meaning. A mathematics researcher wrote on his blog: “I know some things about the Hodge conjecture, but I have absolutely no reason to believe it is close to being solved, with or without large language models.”

The Hodge conjecture rumor appeared the day after OpenAI announced that its internal model had solved the existence and smoothness problem for the Navier-Stokes equations.

OpenAI claimed that an undisclosed model stronger than GPT-6 Astra organized about ten thousand AI agents, produced a proof in 88 hours, and released a 166-page paper and Lean theorem prover verification code at the same time. OpenAI researcher Sébastien Bubeck estimated the computational cost at several million dollars.

New York University mathematician Tristan Buckmaster then publicly accused OpenAI of attempting to “scoop” nearly a year of research by him and Anthropic researcher Levent Alpöge. Buckmaster said that he and Alpöge used work by Diego Córdoba and Luis Martínez Zoroa as a starting point, and after about a year of effort, made progress on August 15 and completed verification on August 22. After learning of their progress, OpenAI began training a new model that used a very similar proof path.

Buckmaster described the negotiations with OpenAI. He said OpenAI proposed two options: either he publish the Euler equation result first and OpenAI publish the Navier-Stokes result the next day, or he alone write a paper presenting the Navier-Stokes result while acknowledging that OpenAI’s internal model had solved the problem. Bubeck also twice asked to remove Alpöge from the author list, on the grounds that Alpöge worked at competitor Anthropic.

After Buckmaster refused and said he would make the process public, he reportedly received the reply: “Why are you trying to destroy your career?” He answered: “I am a scholar. Why do you think making this public would destroy my career?” The response was: “If you don’t want me to be nice, then I don’t have to be nice.”

Several mathematicians who received the relevant emails called OpenAI’s behavior “research misconduct.” OpenAI denied the accusation, calling the claims “completely false” and saying that the company’s researchers “in no case saw their work until they publicly released it.”

The Hodge conjecture rumor appeared while the Navier-Stokes controversy had not yet subsided. It looks more like a spillover effect of a public relations strategy.

On mathematical difficulty, the two problems are fundamentally different. Although the Navier-Stokes equations are difficult, proof paths can be advanced by constructing counterexamples or establishing blow-up conditions. This is closer to the “search-and-verify” mode at which AI excels. The Hodge conjecture requires insight into deep structural connections between algebraic geometry and topology, and such insight cannot be obtained through brute-force search. A mathematics blogger wrote: “You cannot brute-force the Hodge conjecture, but you can formalize the relevant machinery and ensure your reasoning is airtight.”

On existing progress, the Hodge conjecture has been verified only in a very small number of special cases, such as divisors on abelian varieties and certain K3 surfaces. The general case remains completely open. An August 2026 arXiv paper explicitly stated that “the Hodge conjecture itself remains open beyond this computational range, and three independent obstructions indicate that no currently available method can advance it.” As of February 21, 2026, the status of the Hodge conjecture was “open; no known accepted complete proof or disproof.”

The rumor itself comes from an anonymous social media post. There is no paper, no code, and no confirmation from the Clay Mathematics Institute. With the Navier-Stokes controversy having already severely damaged OpenAI’s credibility, the Hodge conjecture rumor is more likely an emotional projection based on the narrative that “AI is about to conquer everything.”

2026 has seen multiple breakthroughs by AI in mathematics. In May, an internal OpenAI model disproved the Erdős unit distance conjecture, a problem that had troubled the mathematical community for 80 years. The model constructed new point sets, provided an infinite family of examples, and achieved a polynomial-level improvement. Fields Medalist Tim Gowers verified the result.

In August, OpenAI disclosed its next-generation model Astra, publishing 10 new results on research problems that had seen no progress for at least a decade, covering high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. It also released a 249-page research paper and computer-verifiable Lean code. Astra scored 97.6% on FrontierMath Tier 4, while GPT-5.6 Sol scored 83%.

DeepMind’s AlphaProof Nexus solved nine long-standing open Erdős problems, the earliest of which had troubled the field for 56 years, and also proved 44 OEIS mathematical conjectures.

These are verifiable advances. But they also reveal the boundaries of AI’s mathematical ability.

In June 2026, the results of a rigorous test called “First Proof” were released. The test selected 10 frontier research-level mathematics problems, none of which had ever appeared in model training data, and had them reviewed by professional mathematicians. Four AI systems participated; the best solved only 6 problems, while human mathematicians solved all of them correctly. The results of the Riemann-Bench benchmark were even starker: all frontier models scored below 10%.

Terence Tao offered a framework for judgment: AI’s mathematical ability is “uneven.” It can quickly solve certain types of problems, especially those that can be advanced through search, combination, and formal verification, but it remains far behind on problems requiring deep structural understanding and the creation of entirely new concepts.

In his 2026 International Congress of Mathematicians lecture, Terence Tao described a transition now underway. He said mathematics is moving from an era of “proof scarcity” into an era of “proof surplus.”

In the past, the core value of mathematical research was linked to “discovering and proving new theorems.” When an important problem was solved, it often meant that mathematicians had found new methods, built new theories, and opened new paths for successors. Proofs were scarce because generating them required months or even years of effort by human mathematicians.

But when AI can generate large numbers of proofs at acceptable cost, what becomes scarce is no longer the proof itself, but the eye for posing questions, the ability to understand proofs, and the skill of converting machine output into human knowledge. Tao warned that “using powerful problem-solving tools indiscriminately can achieve the short-term goal of solving the problem at hand, but at the cost of the system needed to sustain the next wave of progress, or of understanding the progress already made.”

This judgment connects to the “convex hull” metaphor. If AI excels at combining, extending, and searching within an existing body of knowledge, that is, generating a “convex hull,” then the place where mathematical breakthroughs occur is often outside the convex hull. Wiles’s proof of Fermat’s Last Theorem was not a simple patch on existing knowledge; it introduced an entirely new mathematical framework. AI does not yet possess this ability to “create a framework from zero.”

But this does not mean the “convex hull” is unimportant. A convex hull can be much larger than the original set. If the future pattern is that AI handles the convex hull, humans organize and summarize, leading mathematicians extend the boundary a little further on that basis, and then AI generates an even larger convex hull, then the territory of mathematics will be greatly expanded. The problem is that the starting point of this cycle, extending the boundary outward, is precisely the part that most requires human judgment.

In June 2026, 16 prominent mathematicians issued the Leiden Declaration at Leiden University in the Netherlands, explicitly opposing AI misuse. It has now received support from more than 600 mathematicians worldwide as well as endorsements from institutions such as the International Mathematical Union.

They oppose several specific things. AI itself is not the target. They oppose the flood of low-quality AI-generated papers; technology companies using mathematical results to train AI and then monopolizing them in reverse, raising the threshold for research; and the possibility that if AI problem-solving becomes the sole benchmark, the understanding, insight, and judgment that humanity has long cherished may gradually be squeezed out.

On Mastodon, Terence Tao issued a more specific warning: AI is consuming the supply of high-quality open problems in mathematics. Difficult problems are relatively scarce, and researchers usually choose carefully which problems are worth months or even years of effort. AI’s indiscriminate “solving” of problems may produce a large number of results in the short term, but it damages the problem-selection mechanism needed to sustain the long-term health of the mathematical field.

The deeper worry concerns talent cultivation. Leading mathematicians do not grow out of thin air. Mathematical judgment, knowing which problems are worth doing, which directions are promising, and which proofs are elegant, needs to be cultivated through extensive training in technical details. If more and more of this technical work is handed over to AI, how will young mathematicians develop that judgment? If the middle layer of “tweak the conditions and publish a paper” is replaced by AI, where will the career path of ordinary mathematics researchers be?

Daniel Litt, a mathematics professor at the University of Toronto, expressed a more pessimistic view in a talk titled “The End of Mathematics”: “I think we may eventually enter a world without high-quality mathematical research, in which human mathematical expertise is completely lost.”

The Hodge conjecture rumor will most likely be forgotten within a few days. On the X timeline, the next sensational AI news is already on its way.

But the problems it reflects will not disappear because of that. The Navier-Stokes controversy revealed a more difficult question than “Can AI do mathematics?”: when AI output and human mathematicians’ results become intertwined, how should the rules of academic priority, knowledge attribution, and authorship ethics be updated? When commercial companies have the ability to “race” to publish a team’s year-long results within days, can the incentive mechanisms of academic research still function normally?

In the era of “proof surplus” that Terence Tao described, the core challenge is not the number of proofs, but how to maintain quality amid surplus, preserve judgment amid speed, and protect the value of human understanding amid automation. Mathematics is more than a collection of proofs. AI can generate proofs, but understanding what proofs mean, weaving them into the picture of human knowledge, and deciding where to go next remain human work.

At least for now.

Late at night in an office, a mathematician closes Astra’s output window, picks up the stomach medicine on the desk, unscrews the cap, and shakes out two pills. He stares for a long time at the proof step highlighted green by the AI, then opens a new blank document and begins to write the next line.