company

2026-11-22 09:35

The Riemann Hypothesis, which has troubled the human mathematical community for 167 years, has ushered in the latest progress.

Just now, the Lean and Mathlib communities issued a joint statement announcing the official certification of the formal proof of the Riemann Hypothesis submitted by Anthropic mathematician Levent Alpöge, and expressed full confidence in its correctness.

From the mathematician himself announcing the results to the formal proof community publicly certifying it, the whole process only took one day.

What’s even more striking is that Alpöge previously revealed that this research was completed in collaboration with the AI ​​model Claude. As the proof nears its end, Claude demonstrates mathematical reasoning that exceeds even his expectations.

On November 20, Alpöge announced the research results on his personal website, claiming to have solved the famous Riemann Hypothesis and achieved results involving the Generalized Riemann Hypothesis.

The Riemann Hypothesis was proposed by German mathematician Bernhard Riemann in 1859. It is one of the seven millennium mathematical problems announced by the Clay Mathematics Institute.

It involves the distribution of non-trivial zeros of the Riemann ζ function, and the core proposition is whether the real parts of all non-trivial zeros are equal to 1/2. Since this problem is closely related to the distribution of prime numbers, countless mathematicians have tried to solve it over the past century or so, but have never been able to find a rigorous proof.

Musk also lamented in a post that he proved that the Riemann Hypothesis can free a large number of mathematical conclusions that previously relied on this conjecture from conditional restrictions and become unconditionally valid theorems, calling it another amazing achievement of super intelligence.

Alpöge adopted a rather special way of releasing its results this time.

He once tried to get Claude to compile the proof into a human-readable text version, and the resulting draft was more than 250,000 pages, so large that even he himself found it difficult to read. Such an astonishing length has even been dubbed by some netizens as a "DDoS attack" against the mathematics community.

In the end, Alpöge decided not to make the draft text public, but instead released a 4-page description document and Lean formal proof code divided into 14 compressed packages, with a total size of about 17 million lines.

Lean is an interactive theorem proving tool that converts mathematical propositions and their reasoning into a form that computers can rigorously check. Researchers need to express complex mathematical logic into machine-verifiable code, and then use the proof checker to determine whether the reasoning complies with formal logic rules.

However, in the face of such a huge code, the problems that the outside world is most concerned about also arise.

Did 17 million lines of code actually prove the Riemann Hypothesis, or did it just allow the computer to complete a seemingly successful run?

To answer this question, the Lean and Mathlib communities quickly organized a large-scale verification effort, with many members from the community participating in the inspection.

One of the key steps is to confirm that the mathematical proposition finally proved by the program is completely consistent with the original conjecture. After all, even if the code works fine, there might be a slight difference in definition that turns out to be another problem.

For this purpose, the community adopted the expert-vetted definition of the Riemann Hypothesis in the Mathlib mathematics library, created a separate file, and compared it using two sets of tools, comparator and con-leche, and finally confirmed that the theorem proved by Alpöge was completely consistent with the definition of the Riemann Hypothesis in Mathlib.

At the same time, dozens of community members successfully compiled the entire code and reproduced the same original proof export file.

Verification records published by MIT mathematician Andrew Sutherland show that the entire set of proofs involves 175,187 modules, all of which were successfully compiled with zero errors.

However, just because the code compiles successfully, it does not mean that the mathematical reasoning in it is necessarily correct.

Therefore, the community further uses the Lean official kernel, as well as proof checkers such as con-leche and con-ron to review the exported proofs, and also introduces independent checkers developed by third parties and implemented in different programming languages.

Among them, con-leche and con-ron also have formal reliability proof based on set theory, which can reduce the risk of logical loopholes in the checking tool itself. In order to eliminate potential code interference, the relevant inspections are based on the exported static proof data to avoid malicious programs in the source code from affecting the results.

After the above inspections, the Lean and Mathlib communities believe that the four previously proposed verification criteria have been met, and clearly wrote in a joint statement: "We fully believe that Alpöge's proof of the Riemann Hypothesis is correct, and congratulate him on this achievement."

It is worth noting that the verification record published by MIT also involves five major theorems, covering the classic Riemann Hypothesis, as well as generalized results for the Dirichlet L function. However, this does not mean that all generalized problems corresponding to functions of type L have been solved.

Compared with the results of technical verification, another issue that has received widespread attention is what role Claude played in this research.

Alpöge revealed in his initial statement that the Riemann Hypothesis had been a problem he had wanted to solve since his student days. In order to overcome this problem, he has accumulated relevant theoretical knowledge for a long time and is even ready to invest in decades of research.

But the addition of AI has changed the research process.

According to his description, in the early stages of the research, he mainly set the direction by himself, combining two different mathematical ideas, and guiding Claude to explore mathematical situations suitable for generalization.

As the work gradually deepened, the division of labor between the two parties also changed. Alpöge said that towards the end of the research, Claude's derivation was far faster than his own, and some of the mathematical ideas even surprised him.

He also shared a specific example. While trying to generalize the Stepanov method, the research encountered the problem of how to simulate a certain multiplicity structure in an integer environment. Claude proposed that the chain length of theta lifts could be set to √X.

Alpöge believes that this suggestion has become one of the important ideas to promote research progress.

Judging from the researcher's description, Claude has gone beyond traditional auxiliary work such as literature retrieval and code completion, and has begun to directly participate in the exploration and derivation of cutting-edge mathematical problems.

MIT's verification records also show that independent inspections also used Claude Code. However, it is not yet possible to accurately quantify the proportion of contributions between human researchers and AI in the final proof, and relevant mathematical innovations still need to be further sorted out.

For the mathematical community, the certification of formal proofs is undoubtedly a major development, but it also brings about a difficult problem that few people have had to face before.

Computers can strictly check the logical relationships in tens of millions of lines of code, but they cannot directly explain to humans what new mathematical ideas are hidden in them.

Alpöge himself admits that the current proof is too large and extremely difficult to read. In the future, he hopes to greatly simplify the reasoning process and organize it into a traditional paper suitable for mathematicians to understand, discuss and disseminate.

Fields Medal winner Deng Yu also said, "I won't even look at it until I take out the 500-page version," which is enough to illustrate the pressure that such a huge proof puts on human reading.

When AI produces knowledge faster than humans can understand it, mathematicians may gradually become AI reviewers, spending a lot of time digesting machine-generated results. A new gap may also appear between the speed of scientific progress and human cognitive abilities.

Shan Zong

Hello, World Hello, Word

A total of 522 articles have been published

This article comes from the column

Previous article2026-11-22 09:34

Great GTA 6, "not fun" GTA 6

Aifaner App

Love style, make the future within reach

Digital trendy brands focusing on tomorrow’s products

Advanced tools, prophets first, AIGC’s inspiration guide

A new era of car manufacturing, tomorrow’s travelers

Explore new lifestyles in the city and be your guide to tomorrow’s life

Copyright © Guangzhou Liwo Zhiyuan Investment Management Partnership (Limited Partnership) 2008 - 2026. Use of Aifaner website content for commercial purposes requires permission. Non-commercial use is licensed under CC BY-NC 4.0.

All content is made available under the CC BY-NC 4.0 for non-commercial use. Commercial use of this content is prohibited without explicit permission.