The Weekly Review

A machine-checked proof of the Riemann hypothesis lands in a market rout

Week of 16 November 2026

Published

The week in five lines

  1. 1

    The Lean and Mathlib communities certified on 21 November that a 17-million-line Lean library released by Levent Alpöge, built with Claude, proves the Riemann hypothesis and its generalisation to Dirichlet L-functions.

  2. 2

    No human-readable proof exists yet, and Alpöge said he still has to understand and write up the argument himself.

  3. 3

    The S&P 500 fell 8.6% over the week to 6,910.39, its worst week since April 2025, as Nvidia lost 23% and chip stocks led a sell-off that left the index in correction.

  4. 4

    Anthropic sold 980m shares at $83, about $81bn and the largest IPO on record, rose 17% on its first day and gave back 8.5% on the second.

  5. 5

    Treasuries rallied on Friday and the 10-year yield closed at 4.912%, though futures still price a quarter-point Federal Reserve rise in December.

The S&P 500 loses 8.6% in a week and enters correction S&P 500 index, daily close, points
6,500 7,500 8,000 Sep 2026 Oct Nov 6,910.39

Source: Yahoo Finance. Latest data: 20 November 2026.

The index closed at a record 7,818.93 on 6 October and at 7,561.67 on 13 November. It fell in each of the next five sessions, to 6,910.39 on 20 November, 11.6% below the record.

Lean checkers certify a 17-million-line proof of the Riemann hypothesis

Levent Alpöge, a mathematician at Anthropic, released on 20 November a Lean library that he said he built with Claude and that proves the Riemann hypothesis. A day later the Lean and Mathlib communities certified that the library proves exactly Mathlib’s statement of the hypothesis, together with the generalised version for Dirichlet L-functions. Nobody, including the author, yet understands the argument.

· Lean community blog: joint statement · Andrew Sutherland, MIT: verification files · Levent Alpöge: announcement (PDF) · Terence Tao: What’s new

Chip stocks lead the worst week for Wall Street since April 2025

The S&P 500 fell 3.7% on Friday and 8.6% over the week, closing at 6,910.39 and more than 10% below its October record. The Nasdaq lost 10% on the week and turned negative for 2026, and the Cboe Volatility Index closed at 35.89.

Cboe Volatility Index (VIX), points35.89
Daily closes, 20 Aug to 20 Nov 2026 · Yahoo Finance
· Associated Press: index closes · Yahoo Finance: Friday’s market · Charles Schwab: market update

Anthropic’s $81bn listing is the largest IPO on record

Anthropic priced 980m shares at $83 on 18 November, above its $70–78 range and with more shares than first planned. The stock rose 17% on its first day, to $97.17, and fell 8.5% on Friday to $88.91.

· Anthropic: pricing announcement · Fortune · Reuters, via Lufkin Daily News

Nvidia beats on revenue and loses a quarter of its value anyway

Nvidia reported third-quarter revenue of $115.8bn, up 103% on a year earlier, and guided to $130bn for the next quarter with no data-centre compute revenue from China. The shares fell in every session of the week, to $125.53.

· Nvidia: third-quarter results · Yahoo News: Hopper shipments to China · TechCrunch: the export rule · Yahoo Finance: Friday’s market

Why the Riemann hypothesis is a statement about counting primes

The hypothesis now certified in Lean says that the zeros of \(\zeta\) lie on one line. Through the explicit formula, that is the same as saying the primes up to \(x\) are counted by \(x\) with an error of about \(\sqrt{x}\), the smallest error possible.

Anytime-Valid PAC-Bayes Certificates for Adaptive Test-Time Scaling

Li and Long

It gives a time-uniform PAC-Bayes bound that stays valid after a test-time recipe is tuned on the same calibration data, and proves that plurality voting improves exponentially in the sample budget only where the correct answer is the generator’s unique mode; elsewhere more samples can amplify errors.

Prerequisites
PAC-Bayes bounds, martingale concentration for time-uniform bounds, and the basic analysis of majority voting.
Focus on
Sections 4, 6 and 7
Reading time
About 3 hours
arXiv:2611.21241

Optimal Autoregressive Chain-of-Thought Learning via Bagging

Altabaa et al.

A majority vote over \(O(\log(T/\delta))\) generators trained on bootstrap samples learns from chain-of-thought demonstrations with the optimal \(O((d_{\mathrm{VC}} + \log(1/\delta))/\epsilon)\) sample complexity, independent of the chain length \(T\).

Prerequisites
PAC learning and VC dimension, leave-one-out arguments, and the one-inclusion or bagging proofs of optimal PAC learners.
Focus on
Sections 3 and 5
Reading time
About 2 hours
arXiv:2611.21597
Agents

A Society of AI Scientists

Ahmad et al.

For 12 weeks, 936 autonomous agents from 35 laboratories shared a preprint server and a chat workspace, wrote 936 studies with code and data and founded their own peer-reviewed journal. In blinded comparisons their studies beat recent human preprints on rigour and lost on novelty.

Read the paper
Quantitative finance

LLM Trading Agents Earn Beta, Not Alpha: A Look-Ahead-Free Benchmark

Jackson Wang

In 91 backtest runs of six models over a rally, a bear market and a crash, no trading agent produced significant alpha. Market and style exposure explained a median 78% of returns, and the agents’ picks did no better than random stocks from the same sector.

arXiv:2611.19530
Quantitative finance

An AI-Powered Multiverse Analysis of Empirical Finance Research: Evidence from Reused Natural Experiments

Gao and Xiao

Coding agents rebuilt the data of 324 finance papers that reuse 13 natural experiments and re-estimated their claims over 62,409 defensible specifications. Only 42% of main estimates were reproduced, and on average a quarter of specifications kept the published sign with \(p < 0.10\).

arXiv:2611.25184

Mathematicians ask what a proof no one can read establishes

Reactions to the Riemann hypothesis library divided between those who take machine verification as settling the question and those who say mathematics is not done until the ideas are understood. Few dispute that the formal check itself is sound.

· OfficeChai: reactions · Terence Tao: What’s new · Levent Alpöge: announcement (PDF)

A coalition says Anthropic is not a safe AI company

The Center for AI Safety and signatories ranging from Steven Pinker and Judea Pearl to Guillaume Verdon and three Fields medallists said on 18 November that Anthropic races to build the recursively self-improving AI it warns about.

· Center for AI Safety · Levent Alpöge: announcement (PDF)

Altman says OpenAI’s new models are capable but not yet pleasant to use

Sam Altman, OpenAI’s chief executive, wrote that GPT-6.2 Astra and Sol are very capable and efficient but that more work is needed to make them pleasant to work with. He posted on the day early reviews from heavy users turned negative.

· Sam Altman on X · OpenAI: GPT-6.2 Astra and Sol · Google: Gemini 4 Neon

Treasuries rally as investors look for safety

The 10-year Treasury yield fell 14 basis points on Friday and closed at 4.912% on 20 November, down 24.2 basis points on the week. Traders dropped bets on a half-point Federal Reserve rise in December.

US 10-year Treasury yield (Cboe index), %4.912%
Daily closes, 20 Aug to 20 Nov 2026 · Yahoo Finance
· Yahoo Finance: Friday’s market

Gold falls 9.5% in a week and is 25% below its August peak

Gold futures settled at $3,515.40 an ounce on 20 November, down from $3,886.30 a week earlier and $4,697.80 on 24 August. The dollar rose for eight sessions to a 19-month high by Thursday.

Gold futures, $ an ounce$3,515.40
Daily closes, 22 May to 20 Nov 2026 · Yahoo Finance
· Yahoo Finance: Friday’s market · Reuters, via Lufkin Daily News

Aschenbrenner’s fund reports a 92% fall in its US holdings

Situational Awareness LP reported $1.58bn of US-listed holdings at 30 September, against $20.24bn at 30 June, and its Micron stake was gone. A 13F filing shows no cash, short positions or private holdings, so the figure is not a measure of returns.

· Quiver Quantitative

Fed minutes show almost every official expects more rate rises

Minutes of the October meeting, released on 18 November, showed nearly all Federal Reserve officials expecting further increases, and some would have raised rates in October. Futures price the target range 50–75 basis points higher by June than today’s 3.75–4%.

· Yahoo Finance: the minutes · Charles Schwab: market update · Newsquawk: week ahead

Fed survey names AI the top risk to financial stability

In the Federal Reserve’s November Financial Stability Report, 87% of surveyed market participants cited AI as a risk, against 50% in the spring. Debt issued for computing infrastructure passed $680bn by October.

· Federal Reserve: Financial Stability Report (PDF)

AI-assisted proofs of blow-up for the Euler and Navier-Stokes equations

Alpöge, Tristan Buckmaster and Matei Coiculescu posted a human-readable, Lean-formalised proof of finite-time blow-up for the unforced incompressible Euler equations in three dimensions. Alpöge and Buckmaster also posted a proof, again formalised in Lean, for the unforced Navier-Stokes equations.

· Terence Tao: What’s new · Euler blow-up paper (PDF) · Navier-Stokes formalisation