A mathematician and an AI agent may have cracked a 167-year-old problem. Here is what it means for your work
An Anthropic mathematician says he and Claude have proved the Riemann Hypothesis. Whatever the final verdict, it shows what a domain expert can do with an AI agent, and why every professional will want one.
By Hivemind Labs
5 min read
- ai-agents
- research
- forward-deployed-engineers
Status as of 21 November 2026: the proof has been checked by computer and independently recompiled, and leading mathematicians say it is very likely correct. It is still under expert review, and no human-readable paper exists yet. We will update this post as the review concludes.
On Friday 20 November, mathematician Levent Alpoge posted a four-page note on his website. It opened with a sentence mathematicians have waited 167 years to read: “I believe I have, working with Claude, discovered a proof of the Riemann hypothesis.” (Phys.org)
Alpoge works at Anthropic, the company that makes Claude. We build with Claude, among other models, so read what follows knowing that. But the story matters well beyond any one company, and well beyond mathematics.
What the Riemann Hypothesis is, without the formulas
Prime numbers, the numbers divisible only by 1 and themselves, look scattered along the number line with no obvious pattern. In 1859 Bernhard Riemann suggested that their distribution actually follows a deep, hidden order, encoded in a function now called the Riemann zeta function (Al Jazeera).
No one could prove it. The Riemann Hypothesis became the most famous open problem in mathematics and one of the seven Millennium Prize Problems, each carrying a one million dollar prize from the Clay Mathematics Institute. Hundreds of other results have been proved assuming it is true. Some fast algorithms used in cryptography are only guaranteed to work if a version of it holds. As Imperial College’s Kevin Buzzard put it, if it is true, “we can sleep slightly safer in our beds at night.”
What actually happened
Alpoge did not publish a paper. He published a proof written in Lean, a programming language in which every step of a mathematical argument is checked by a computer. According to his note, the work ran from early October to mid-November, and Claude “far outpaced” him near the finish line (The Next Web). He says the same method also covers the broader generalised Riemann Hypothesis.
The proof is enormous: Buzzard describes the Lean code as about 17 million lines, which might correspond to tens of thousands of pages of human mathematics. That is why there is no paper to read yet.
Here is where things stand:
- The code checks. MIT’s Andrew Sutherland compiled it independently and reported that it was “rechecked by the Lean kernel with only the three standard axioms” (Scientific American via Yahoo).
- Experts are confident but careful. Oxford’s James Maynard is reported as “98 or 99 percent” certain. Buzzard’s open question is whether the theorem stated in the Lean code is exactly the Riemann Hypothesis and not something subtly easier.
- No one understands it yet. “Right now, we honestly can’t tell you, because nobody understands the proof of it,” Buzzard said. He expects that to change within weeks.
- It is not the only claim. On the same day, Stanford’s Ravi Vakil posted a separate, entirely human proof by another mathematician. Within hours, people using AI models had reported problems in it (The Next Web).
The Clay Institute requires a proof to be published and survive two years of scrutiny, so the prize is not decided. But even the cautious reading is extraordinary.
The real lesson: the agent did not work alone
It is tempting to read this as “AI solved maths”. That misses what made it work. Three things came together:
- A domain expert. Alpoge is a number theorist. He knew which questions were worth asking, which approaches were dead ends and what a real proof has to look like. He directed the work for six weeks.
- An AI agent. Claude brought speed and stamina no person has: exploring approaches, writing and rewriting millions of lines, and never getting tired near the finish line.
- A checker. Lean verified every step. That is why mathematicians can trust a proof no human has read yet. Without verification, a fluent but wrong argument looks exactly like a right one.
Remove any one of these and the result does not happen. An agent without an expert produces confident noise. An expert without an agent runs out of time. And neither can be trusted without a way to check the work.
The same pattern fits your job
Most professionals are not trying to prove theorems. But most have work with the same shape: deep judgment needed at a few points, and huge amounts of careful, checkable work in between.
- An accountant knows what a clean month-end looks like. An agent can match every invoice to its purchase order and delivery note, flag the twelve that do not match, and draft the credit requests. The accountant reviews twelve exceptions instead of twelve hundred lines.
- An engineer knows which design constraints matter. An agent can generate and test a hundred variants overnight against those constraints, so the engineer starts the morning choosing between the best three.
- A lawyer knows which clauses carry risk. An agent can read every contract in a portfolio, compare each clause with the firm’s standard positions and mark every deviation with its source.
- An agronomist knows what crop stress looks like. An agent can read thousands of soil and weather readings across a farm every day and recommend what each block needs. That is the idea behind Shaduf, our own hive of farm agents.
In each case the expert does not disappear. They move from doing the work to directing it and checking it, which is where their judgment is worth the most.
What makes it work in a business
The Riemann story also shows what separates useful agents from expensive demos. The same three parts apply:
- Start from the expert’s process, not from the technology. The agent should do a job your people already understand well enough to judge.
- Connect it to your real systems, so it works on your data in your ERP, CRM and tools rather than on copies.
- Build in the checker. Every step logged, every result traceable, and a person approves anything that moves money or reaches a customer.
This is how we build agents at Hivemind Labs. Our Forward Deployed Engineers sit with your experts, find the work that should go to an agent, and build it inside your systems with the checks your business needs.
You do not need a 167-year-old problem to get started. You need one expert, one process that eats their time, and a way to check the result.
Book a discovery call and we will find the work your experts should hand to an agent first.
Sources
- Phys.org (AFP): Anthropic researcher says he and Claude proved the Riemann hypothesis
- The Next Web: Anthropic mathematician says he and Claude proved the Riemann hypothesis
- Scientific American, via Yahoo: The AI Riemann hypothesis proof is almost certainly true
- Al Jazeera: Has the world’s most famous maths problem been cracked?