# AI News — November 21, 2026: Alpöge Posts Lean-Verified GRH Proof, Tao Adds Navier-Stokes Blowup

> Source: <https://ai0.news/posts/2026-11-21-daily-digest/>
> Published: 2026-11-21 07:00:08+00:00

**Recommendation:** Verify AI-generated formal proofs with two independent kernels before accepting them; the GRH artifact passed both, closing off the single-verifier-bug failure mode.
More in the playbook "Verifying model output": <https://wpnews.pro/playbooks/output-verification>

Good morning. If yesterday was about Anthropic going public, today is about what one of its mathematicians did while waiting for the lockup to lift. Levent Alpöge is claiming a Lean-verified proof of the Generalized Riemann Hypothesis, Terence Tao is claiming heavy-AI-assist proofs of Navier-Stokes and Euler blowup, and the market is responding to all of this by erasing the Nasdaq’s entire 2026 gain. It’s a strange juxtaposition.

**A possible proof of the Riemann Hypothesis.** Anthropic mathematician Levent Alpöge has [posted a claimed proof](https://alpo.ge/remark.pdf) of the Generalized Riemann Hypothesis, formally verified in Lean — a 1.3GB, roughly 15-million-line artifact that MIT’s Andrew Sutherland reportedly checked using Claude Code. [The Verge notes](https://www.theverge.com/ai-artificial-intelligence/1010051/anthropic-riemann-hypothesis-millennium-prize-problem) Alpöge delayed publication until after Anthropic’s November 19 IPO, which is either prudent or very bad optics depending on your priors. [TechCrunch reports](https://techcrunch.com/2026/11/20/ai-claimed-a-solution-to-the-167-year-old-riemann-hypothesis-and-a-human-did-too/) multiple mathematicians have now compiled the proof, and a separate, fully human proof from a different author appears to exist too.

**Tao piles on with Navier-Stokes.** Not to be outdone, Terence Tao [announced three results](https://terrytao.wordpress.com/2026/11/20/unforced-euler-singularity-unforced-navier-stokes-singularity-and-the-generalized-riemann-hypothesis/) in a single post: a human-readable Lean-formalized proof of finite-time blowup for the unforced Euler equations, a rougher proof of the same for Navier-Stokes (another Millennium Prize problem), and his own Lean-only GRH proof via Alpöge. Two independent Lean kernels have validated the GRH artifact, which rules out the usual “maybe the verifier has a bug” escape hatch. One HN commenter captured the mood: “I thought we’d have more time.”

**Claude solves two black hole problems.** Anthropic also published [results in theoretical physics](https://www.anthropic.com/research/claude-black-hole-physics): Claude completed a 20-year-old calculation showing a black-hole-formation phase transition persists in shapeless spacetime, and derived an exact upper bound on a quantum-gravity testing technique. Neither applies to our universe, and HN commenters noted the papers list no human authors alongside Claude — a framing choice that’s drawing more pushback than the math itself.

**METR updates the time-horizon chart.** [MIT Tech Review covers](https://www.technologyreview.com/2026/11/20/1149399/the-most-misunderstood-graph-in-ai-just-got-a-major-update/) METR’s refresh of its influential capability-doubling plot, which previously showed frontier models doubling their task horizon every four months on verifiable work. The update adds harder, judgment-heavy tasks to test whether that curve generalizes beyond checkable coding and math — a question that lands differently in a week where AI is co-authoring Lean proofs of century-old conjectures.

**Language models rewriting their own weights.** A [new paper](https://arxiv.org/abs/2611.21885) claims a 27B model modified its own weights through reasoning alone, gaining 6.6 points on unseen problems and restoring a damaged copy of itself to ~80% accuracy over eight iterations. HN was unimpressed on inspection: the actual “weight edit” amounted to shortening reasoning traces to dodge token limits, and commenters pointed to Google’s earlier Sleep framework as prior art. The safety angle is the one worth watching — the same mechanism could, in principle, ablate alignment-relevant activations.

**AI-designed CPU claims 4.59x perf-per-watt.** Architect Labs says its AI stack autonomously designed [Frostbite](https://architectlabs.com/blog/frostbite), a general-purpose CPU based on interaction nets, from ISA through GDS layout. HN picked it apart quickly: the baseline is an in-order superscalar rather than a modern out-of-order design, and interaction-net and dataflow architectures have been explored since the 1980s (WaveScalar got multiple mentions). The more honest framing, as one commenter put it, is that AI is synthesizing decades of buried research literature into working artifacts — which is still something.

**The market is not celebrating.** Anthropic stock dropped 9% on its second trading day, nearly erasing the IPO-day pop, and the broader AI complex took about $4 trillion in market cap with it — enough to wipe out the [Nasdaq’s entire 2026 gain](https://www.theguardian.com/business/2026/nov/20/global-stock-markets-slump-anthropic-ipo-ai-sell-off). Nvidia and TSMC both posted double-digit losses. HN commenters flagged the $2 billion in insider sales on day one and are openly wondering whether Anthropic was “the last through the door before the house goes up in flames” for AI IPOs. Meanwhile [the Equity podcast](https://techcrunch.com/podcast/the-fallout-from-nvidias-jensen-huang-scandal-continues/) is still working through the Jensen Huang fallout alongside Waymo’s freeway suspension and Tesla’s Cybercab orders.

Mathematicians are having the week of their lives and shareholders are having a very different one. We’ll see tomorrow which of these proofs survives peer review — and whether the market finds a floor before Anthropic’s lockup even matters.
