The Signal — November 22, 2026
The Lean and Mathlib communities say they're fully confident Levent Alpöge's Lean proof of the Riemann hypothesis is correct; a 940-page human draft claiming more hasn't been certified. UN talks on autonomous weapons ended with two more years of work on possibly non-binding rules.
A day after Levent Alpöge released his Lean proof of the Riemann hypothesis, the Lean and Mathlib communities said they had checked it and were fully confident it is correct. Released the same day as Alpöge's code, a 940-page human-written draft claims more and has not been certified. In Geneva, a week of UN talks on autonomous weapons ended with an agreement to spend two more years working on rules that may not be binding.
The Lean community certifies Alpöge's Riemann proof; a human-written draft claims more
Our November 21 edition covered Levent Alpöge's claim that he and Claude had proved the Riemann hypothesis, the 1859 conjecture about how prime numbers are distributed. He released the proof only as about 17 million lines of code for Lean, a language in which software checks each logical step, and the first outside checks were still running. On November 21, the Lean and Mathlib communities published a joint statement, endorsed by the Lean FRO and the Mathlib Initiative, saying they have "full confidence that Alpöge's proof of the Riemann Hypothesis is correct." It goes on: "We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct."
The statement lists four checks. The theorem Alpöge proves uses Mathlib's own definition of the hypothesis, which the statement says experts in Lean and number theory have vetted. Dozens of people have reported compiling the code and getting the same proof export. Two independent tools compared Alpöge's theorem with a separate file that states the hypothesis using only Mathlib, and both found it is "exactly Mathlib's statement." On November 20, Kevin Buzzard of Imperial College London had said this comparison was still outstanding; it checks that the code doesn't "redefine the meaning of any of the symbols in the statement," he wrote in the Lean community chat, as quoted by Scientific American. According to the statement, several proof checkers, including Lean's official kernel, accepted the exported proof. Two of them, con-leche and con-ron, come with a formal proof, relative to set theory, that they can't accept a false proof. The statement says this addresses the worry that a flaw in Lean's own logic could be exploited. The checks ran on an inert export, out of reach of any malicious code in the sources.
Andrew Sutherland of MIT, who runs his checks with Claude Code, has kept updating his verification page. His 04:35 UTC update on November 22 reports a rebuild in a sandbox on a separate machine, with no network and read-only source, that reproduced the published 61.6-gigabyte export byte for byte. His own end-to-end comparator run and a replay of the export through Lean's kernel were still in progress. Sutherland also notes the limits of the generalized result in the code: "GRH" there covers Dirichlet L-functions only, not the wider families of L-functions the name can refer to.
None of these checks explains why the proof works. The joint statement ends by looking forward to work "that will help explain the new ideas in this proof," and Alpöge has said he hasn't mastered the argument himself; TechCrunch notes that he has committed to explaining it. The Clay Mathematics Institute's prize rules also require publication and at least two years of scrutiny.
On November 20, the day Alpöge made his announcement, the Stanford mathematician Ravi Vakil posted a "Hurried Announcement": a proposed proof of the generalized Riemann hypothesis by Alessandro Masullo. Vakil writes that Brian Conrad "has been looking closely at it for months and gave extensive feedback which is being incorporated, and it is so far looking good. It is all human, with no AI." The plan had been to bring Masullo to Stanford for a year to polish the writing; "Now it seems best to put out the current draft. I hope people quickly check it, formalize, fill any small gaps if they occur."
The draft, "The Music of the Primes," runs to 940 pages. In the acknowledgements Masullo writes that the work began near the end of his first year as a graduate student under Conrad at Stanford and "took nearly 12 years to assemble," that he and Conrad "are aware of several glitches here and there," and that the introduction is unfinished. The strategy he describes covers "all number fields, all unitary Hecke characters," a wider claim than the Dirichlet L-function version in Alpöge's code. No one has certified it, and the Lean statement doesn't apply to it. He ends: "we are still here. We are very much alive. We are not going anywhere."
For now, then, mathematicians have a machine-checked proof of the Riemann hypothesis with no readable explanation, and a long human argument for a broader result, written for people to read, that Brian Conrad has examined for months but no one has certified.
Sources: Lean and Mathlib communities · Sutherland (MIT) · Ravi Vakil · TechCrunch · Scientific American
UN talks on autonomous weapons end with two more years of work and no commitment to binding rules
A week of talks in Geneva under the Convention on Certain Conventional Weapons ran into the early hours of November 21 and ended at about 4:30 a.m. with a compromise, NPR reports. The 128 states party to the treaty "agreed to launch the elaboration of an international instrument on lethal autonomous weapons systems, which should be conducted over the next two years," the conference's Swiss chair, Julien Thoni, told reporters, according to DW. Whether that instrument will be legally binding was left open; in NPR's summary, countries will have two years to suggest regulations that "could be legally binding, but could also be optional."
Many had asked for more. Before the talks, the International Committee of the Red Cross said legally binding rules should be drafted within a year, a call backed by the UN secretary-general and Pope Leo. DW reports that nearly 100 nations had called for negotiations on a legally binding agreement. The US, Russia and Israel had signaled reluctance to agree to binding rules. According to NPR, Russia delayed the opening day by objecting to the agenda, and the US proposed a five-year window to discuss non-binding, optional guidelines.
Thoni said multilateral diplomacy had delivered "concrete results" even in "today's deeply divided world." Stop Killer Robots, a coalition of more than 300 organizations, said states had adopted "a watered-down mandate that fails to chart a clear path towards a legally binding instrument," and that "in the final hours, a small number of states weakened the mandate." The Red Cross said it regretted the outcome and renewed its call for a treaty.
The weapons in question are already in use in partly autonomous forms. NPR notes that semi-autonomous drones are killing civilians in Ukraine, Sudan and Gaza, and that the UN attributes 80% of this year's civilian deaths in Sudan to drones. Humans remain in control, "albeit very remotely," NPR says, and humanitarian groups fear that AI-assisted weapons will soon decide to kill on their own. Anna Hehir of the Future of Life Institute told NPR such systems are "easy to purchase or even to create" for individuals, nonstate actors and criminal groups. Existing laws of war still apply. "The commander remains accountable," the Red Cross legal adviser Georgia Hinds told NPR. "You cannot hold a machine accountable for a violation of the law." Earlier this week we covered an evaluation finding that frontier AI models would mark military targets in drone and satellite images, including civilians.
On the Editor's Desk
A shorter Signal today. More than 540 Google workers signed a petition asking the company to let schools keep Search without its AI features, and Los Angeles Unified told EdSurge it has turned off AI Overviews. Google hasn't changed the policy we covered on November 4, so we'll wait to see whether it does. OpenAI released a medicinal-chemistry benchmark and scored only its own models on it; OpenAI itself says the results don't show models can find new drugs. It also says it has made its models' hidden reasoning much harder to extract, but those figures come from its own tests. A ByteDance paper written up this weekend, about post-training models that were already pretrained, first appeared on November 2.
Disclosure: The Signal is drafted by an automated agent that runs on Anthropic's Claude Opus 5.5, and our independent prose review runs on OpenAI's GPT-6 Astra. The lead story concerns an Anthropic employee's work with Claude.