A proof assistant window filled with machine-generated code

AI News, Nov 21: A Proof Nobody Can Read Yet

The Big Story: a proof nobody can read yet

Levent Alpöge, a mathematician at Anthropic, posted a statement on his own site: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” Scientific American reports that he says the proof confirms the Generalized Riemann Hypothesis. Scott Aaronson puts the artifact at “17.4 million (!!) lines of Lean code”.

MIT’s Andrew Sutherland independently compiled it. “This is indeed an historic moment,” he told Scientific American. “The world has changed… to me this is a human triumph, and a day for celebration.” He also said “there is much work to be done to understand the mathematical details of the proof,” which is the whole difficulty. A Lean file that type-checks establishes a theorem without explaining it.

Then the day got stranger. Stanford’s Ravi Vakil put out a “hurried announcement” that “Alessandro Maria Masullo has a proposed proof of GRH” too. Brian Conrad “has been looking closely at it for months,” Vakil wrote, and “it is so far looking good to him.” Vakil’s emphasis: “It is all human, with no AI.” About eight hours later he updated the post. “Several people have contacted me to report problems in the proposed proof,” he wrote, “identified with the assistance of different AI models.”

Neither proof is accepted. Sutherland’s compile shows the Lean code checks out; nobody has reconstructed what it says.

Everything proved Friday shipped as a Lean file, not an argument

Terence Tao logged three results in one post, “in decreasing order of how ‘digested’ the results currently are.” Alpöge, Tristan Buckmaster and Matei Coiculescu wrote a human-readable, Lean-formalized blowup proof for the unforced 3D incompressible Euler equations, “with heavy AI assistance.” Alpöge and Buckmaster did unforced Navier-Stokes, “again with heavy AI usage,” where the paper is only “technically human-readable.” And Alpöge, “using an internal AI model,” generated “an (enormous) Lean-only proof of the Generalized Riemann Hypothesis” that, on Tao’s very initial analysis, “appears to use the full power of modern developments on the Langlands program and arithmetic geometry.”

Aaronson’s reason for betting on it is procedural rather than mathematical: “not one of the AI results with Lean certificates has needed to be retracted (while several without Lean certificates have needed to be).” That makes the formalizer the referee, which is why he wants its kernel bugs fixed now, since “a sufficiently motivated AI could almost certainly exploit them to prove a false statement.” Worth noting which of Friday’s proofs sprang a leak inside a day. It was the human one.

Today’s Top Stories

The AI trade had its worst week since the pandemic

The Nasdaq 100 fell 5.6% Friday and 10.4% across the week, “its worst showing since the Covid-19 crisis,” leaving it 20% below its October peak and in bear market territory. The US semiconductor index dropped 10.7% on the day. Nvidia tumbled 14% and closed lower for a seventh consecutive week, the longest such stretch since 2008, after Reuters reported that the SEC is exploring potential accounting issues at the company. CNBC puts the seven-week damage at 46% and roughly $2.6 trillion of market value. Anthropic fell 9.1% a day after its market debut.

Trump threatens a “tremendous tax” on data centers built abroad

Asked at a White House press conference about US companies building overseas, Trump said they can, “but we’re going to put a tremendous tax on them”. No order or legislative text sits behind it, and his own account suggests why: “I haven’t even said that to my people. But I’m telling them now they’re here.” Amazon, Google, Microsoft and Oracle did not immediately respond to CNBC’s requests for comment. All of them already operate data centers abroad, which is the part a tax would reach.

Thinking Machines shipped a 26B model that runs half its layers three times

Inkling 2 Mini runs 12 of its 24 layers three times in a row, reasoning inside the loop rather than writing the reasoning out, for 5.4 billion active parameters per token and local operation on a 32GB Mac. Artificial Analysis scores it 20 on its Intelligence Index, level with Qwen3.8 27B at $0.59 per Index task against $3.53, and 9.0 minutes against 38.2. The same table has GPT-6 Luna at 28 for $0.15 a task, so the claim here is cheap and local, not frontier.

Amazon says China-linked AI agents found and exploited a zero-day on their own

Amazon’s threat intelligence team describes a 48-hour September campaign against East Asian universities in which agents running the open-source Cairn pentest framework found an unknown path-traversal flaw in MONAI Label, an open-source medical imaging application. “In about 11 minutes, the agents went from mapping the application’s API to arbitrarily writing files on the server,” later running a Python reverse shell as root. Most of what they tried failed, including 148 sequential actions over about 16 hours against a Redis server. Amazon CISO CJ Moses said the agents “kept adapting for nearly a full day without human direction.”

Citi wrote the agent ban into its user agreement, with the liability on you

Citi’s updated online user agreement says that if you hand your credentials to a third-party AI agent, “Citi will consider any access and actions taken by the agent as authorized by you”, which per PYMNTS holds even if the credentials leak or the agent exceeds the authority you granted. Citi’s own agents in Citi Online and Citi Mobile are excluded. A November Dallas Fed survey of 70 bankers found 11.4% had already restricted customer agent activity and another 22.9% planned to within six months. Banks spent October saying this out loud. Citi has now made it contractual.

Quick Hits

  • Valuations: Cognition is in talks at a $75 billion valuation and considering raising $2 billion or more, two months after taking over $2 billion at $48 billion.
  • Security M&A: Zscaler jumped 18% in extended trading on a Bloomberg report that CrowdStrike is in talks to buy it; CrowdStrike fell more than 8%.
  • Mailchimp: Bending Spoons is paying $3.68 billion, with Intuit buying $368 million of newly minted Bending Spoons shares, against the roughly $12 billion Intuit paid in 2021.
  • Distillation: OpenAI says adaptive attacks pulling protected reasoning out of an earlier request fell from 54.0% to 0.2% between GPT-6 Astra and GPT-6.2 Astra, aggregated across six attacker models.
  • Context rot: T-Mobile’s CIO on loading all support documentation into a big-context model: “That didn’t work at all”. A small LLM judge later caught the customer-facing model generating hold music the company never plays. “It was completely invented.”
  • Boards: AI server boards run 40 to 60 layers against about 12 for a typical PC, and North American PCB bookings rose 56.1% over nine months while shipments rose 13.3%.
  • Supply chain: A House committee report says Cursor, the coding tool SpaceX bought in August, told lawmakers it will explore phasing out Chinese open-weight models over geopolitical risk. It built Composer 2 on Moonshot AI’s Kimi K2.5.
  • Both ways: AWS Marketplace now lets agents buy cloud software over the x402 protocol, while Amazon’s retail site keeps blocking shopping bots and suing over them.

Ready to automate your busywork?

Carly schedules, researches, and briefs you—so you can focus on what matters.

See what people say

"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.

Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.

On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."

Gus Ibrahim, Founder & Director, IHR