/daily ·09 SEPT 2026 ·WEDNESDAY ·2 MIN READ ·6 STORIES

The Week AI Did Math

OpenAI claims the Navier–Stokes prize, Anthropic declines safety pre-screening, and a safety researcher quit — loudly.

01 / The Day

WEDNESDAY 09 SEPT 2026, ranked

OpenAI Claims a $1 Million Math Prize — If Peer Review Agrees

OpenAI published what it describes as a proof of the Navier–Stokes existence and smoothness problem — one of mathematics' seven Millennium Prize Problems — sparking a dispute over credit that included an allegation that an OpenAI researcher pressured an NYU mathematician to drop an Anthropic co-author.

Navier–Stokes governs turbulent fluid flow; no proof of solution existence had stood for over 150 years
Nature and Quanta both reported it; Clay Institute independent verification is still pending
Anthropic separately formalized a proof of Fermat's Last Theorem using AI tools in the same week

Why it mattersIf the proof holds, it is the first Millennium Problem solved — by an AI system, not a human mathematician.

01 openai.com↗

Anthropic Won't Submit Mythos 5.1 to UK Safety Evaluators Before Release

The Financial Times reports Anthropic declined to give the UK AI Safety Institute advance access to Mythos 5.1 for pre-release evaluation, breaking with the voluntary testing framework that labs agreed to at the 2023 Bletchley Park summit.

AISI evaluations are voluntary; Bletchley signatories pledged cooperation but face no legal enforcement mechanism
Decision arrives days after Matt Clifford — who shaped the AISI framework — resigned over an Anthropic conflict of interest
UK government has no statutory power to compel pre-release model access under current law

Why it mattersVoluntary safety frameworks survive only as long as labs honor them; the first prominent defection tests whether the Bletchley model has any real weight.

02 techmeme.com↗

Anthropic Alignment Lead: >10% Chance AI Kills All Humans Within a Decade

Evan Hubinger, Anthropic's Alignment Science lead, published a personal estimate of greater than 10% probability that AI could cause human extinction within ten years. Simultaneously, researcher Jacob Coxon resigned from Anthropic citing uncontrollable AI fears, and an anonymous post titled 'I resigned from Anthropic today' appeared on Hacker News.

Hubinger's figure substantially exceeds any probability Anthropic has stated in official safety documentation
Coxon's WSJ exit interview cites a mismatch between Anthropic's safety rhetoric and its deployment pace
All three signals landed the same night Anthropic declined AISI pre-release testing

Why it mattersWhen alignment scientists at the lab most identified with AI safety start posting 10% extinction estimates and resigning, every safety disclosure that lab has made is retroactively in question.

03 techmeme.com↗

Meta Launches Muse, Its Personal AI Agent for 3 Billion Users

Mark Zuckerberg unveiled Muse — Meta's personal AI agent — rolling out first through WhatsApp, Messenger, and Instagram, making it immediately one of the most widely distributed AI agents on earth by reach if not by capability.

Muse targets the consumer base, not enterprise — a deliberate positioning against OpenAI's GPT and Anthropic's Claude
WhatsApp distribution alone gives Muse theoretical access to over 2 billion active users globally
Zuckerberg addressed growing community pushback against planned Meta data centre sites at the same Q&A

Why it mattersA personal AI agent launched inside apps that serve as the primary internet for large parts of the world is a different category of deployment than anything that has come before.

04 ai.meta.com↗

DeepMind's AlphaGenome Atlas Maps Every Possible Human DNA Mutation

DeepMind released AlphaGenome Atlas, a model predicting the functional consequence of all possible single-nucleotide substitutions across the three-billion-base human genome — roughly 9 billion variants — covering effects on gene regulation, splicing, and expression across tissue types.

Atlas covers variants never observed in living populations, far beyond existing clinical databases
Primary applications are rare disease diagnosis and drug-target identification in genetically unexplained conditions
Built on AlphaFold architecture extended from protein structure to regulatory genomics

Why it mattersThe gap between sequencing a genome and understanding what each variant does has been genomics' central bottleneck; Atlas closes a large portion of that gap computationally.

05 deepmind.google↗

NSA, CISA, FBI Warn of Chinese AI Industrial-Scale Distillation from US Models

A joint advisory from NSA, CISA, and the FBI warned that Chinese AI firms are conducting systematic large-scale knowledge distillation from US frontier models — training competitive systems using outputs from GPT-6 Astra, Claude, and Gemini without authorization.

Distillation — training a smaller model on a larger one's outputs — is technically against most API terms but difficult to detect at scale
Advisory names specific US frontier models as sources; enforcement options remain limited absent evidence of identity
Follows existing export controls on chips and model weights as a second vector of AI capability transfer concern

Why it mattersIf US frontier model outputs are the effective training signal for rival systems, hardware export controls solve only part of the technology-transfer problem.

06 techmeme.com↗

Be subscriber #013 the weekly ten, every friday by email · no spam · unsubscribe anytime

Past days