Saturday, November 21, 2026

TL;DR

Two mathematicians posted proposed proofs of the Riemann Hypothesis family on November 20, one written with Claude and one with no AI, and neither is verified. An AI-assisted attack on the European Commission has run more than 10 days. Meta employees say its Muse agent shows them other people's bank balances. Paul Christiano gave about a 1-in-3 chance of superintelligence within a year.

Signals

Claims of AI-assisted proofs are now being checked in public within hours, and AI models are among the checkers.  #

Levent Alpöge posted a note on November 20 saying he and Claude found a proof of the Riemann Hypothesis and formalized it. Separately, Ravi Vakil announced Alessandro Masullo's human-only draft proof of the Generalized Riemann Hypothesis. About eight hours later Vakil added that readers had found problems, some with the help of AI models. Both are unverified claims. Simon Willison quoted both. Alpöge note · Vakil announcement · Willison, Nov 20

Insiders are shortening their timelines for loss of control, and they are predicting it rather than showing it.  #

Zvi Mowshowitz summarized Paul Christiano's Dwarkesh Patel interview: about a 10% risk of losing control within a year, about a 1-in-3 chance of superintelligence within a year if development is not slowed, and a call for labs to pause algorithmic progress until security is adequate. Christiano put interpretability at 5 to 10% odds of working. This is a prediction, read secondhand through Zvi's post. It follows Ryan Greenblatt's November 9 update toward superintelligence in one to two years. Zvi on Christiano · Zvi, AI R&D acceleration

Beyond Willison and Mowshowitz, it was a quiet week: no new Karpathy, Cherny or Ronacher output turned up.

News

ACTARTEX attack on the European Commission passes 10 days  #

The Register says attackers using the ARTEX pentest framework took Sysper payroll and MyPMO reimbursement systems offline. Salaries were paid on time. Whether staff data was taken is unconfirmed. CERT-EU warned on November 5. The Register

WATCHMeta employees say Muse mixes users' conversations #

Internal posts reported by Reuters describe bank balances and reminders from other accounts appearing in chats. Meta says staff test versions the public does not have. Muse has over 10 million downloads. TechCrunch

WATCHCalifornia attorney general subpoenas Nvidia over China chip transfers #

Rob Bonta opened the investigation November 19, citing the Huang spy allegations and investor disclosures. Nvidia separately confirmed it stopped licensed H200 shipments to China, which brought $2.4 billion last quarter. AI Weekly · The Neuron, H200

SHIPGemini for Mac and a reported Apple agent put computer-control agents on the desktop  #

Google's Gemini can operate Mac apps for US paid subscribers. Bloomberg reports Apple is building its own agent and adding extra permission warnings for rival agents. TechCrunch, Gemini · Bloomberg, Apple

SHIPTencent releases Hy4 Flash weights; Moonshot loses Nvidia access deal #

Hy4 Flash takes text and images with 1 million tokens of context and runs locally. Moonshot scrapped a $150 million Indonesian chip-access deal after new US limits on remote use. TechNode · The Information

SHIPThree used V100s hit about 100 tokens per second on Qwen 3.6 35B  #

Tom's Hardware reports a $1,042 build beating a $5,399 M5 Max Mac Studio at long context. The comparison omits some components. Tom's Hardware

WATCHLean and Mathlib groups certify Alpöge's Claude-assisted Riemann Hypothesis proof Updated this evening  #

The Lean FRO and the Mathlib Initiative said on November 21 they have full confidence the proof is correct. Two checking tools, comparator and con-leche, confirmed the formal theorem matches Mathlib's statement of the Riemann Hypothesis. The certification covers the formal proof, not the mathematics: no paper has been posted, and the human write-up Alpöge promised has not appeared. Lean FRO post on X · OfficeChai

Also noted

  • OpenAI says Codex agents ran 1,700 ChatGPT slowdown probes and found 600 root causes; 15 of about 400 fixes shipped. AI Weekly #
  • Google built a simulated internet to train and attack Chrome browsing agents safely. Wired #
  • Anthropic's IPO raised $81.3 billion, the largest on record. Semafor  #
  • Nvidia reportedly in talks to buy Inception; its Mercury API shuts down December 28. The Neuron #