Clement Delangue asked to audit the frontier labs on Saturday, four hours after Dario Amodei said Anthropic would let third-party evaluators inside. His Open Alignment Initiative, run by co-founder Thomas Wolf, is Europe's loudest open-source voice volunteering to police the closed ones.
NVIDIA confirmed its $12.93bn purchase of Hugging Face on 3 September. Nine days later Reuters reported it was in talks to anchor Anthropic's IPO with up to $10bn.
The proposed auditor is being bought by a company negotiating to become one of the audited company's largest shareholders. Nobody has said who pays the embedded evaluators, or what happens when one of them reports something the lab disputes. (The Next Web)
Research & Labs
Twenty-five Fields medallists say the AI proof race is breaking mathematics
Terence Tao published a declaration on 11 September signed by 25 Fields medallists, among them Peter Scholze in Bonn, Maryna Viazovska and Martin Hairer at EPFL, and Maxim Kontsevich at the IHES. They call the goals of the AI labs and of the mathematical community severely misaligned.
The trigger was OpenAI's claim on 8 September that an unreleased model had settled the Navier-Stokes Millennium Prize problem in 88 hours, using roughly 10,000 parallel agents. Tristan Buckmaster at NYU had published related work hours earlier and asked whether OpenAI had drawn on it.
They are not saying the proofs are wrong. They are saying a solved problem is worth less than the methods the solving produces, and that unreferenced machine proofs break the chain by which those methods reach students. (What's new)
ElevenLabs ships a music model trained only on licensed catalogue
ElevenLabs released Music v2.5 on 13 September, trained under its Merlin and Kobalt deals covering 30,000 independent labels, with rightsholders opting in for royalties. Universal Music was not part of it, despite a separate agreement to build a remix platform together.
For a London company with Polish founders, licensed-only training is a commercial bet as much as a legal one. The AI Act's training-data summary duty has applied since August, and more than 30 providers have already been asked about copyright compliance. (Music Ally)
Deployment & Applications
The EU's own scaleup fund leads a $100m round in Stockholm
Tandem Health raised $100m on 14 September, led by the Scaleup Europe Fund that EQT manages on the Commission's behalf, taking the total raised to about $160m.
The company builds an AI clinical assistant that drafts notes during consultations; it runs in more than 10,000 care organisations and bought the Dutch scribe startup Juvoly in January.
Peer-reviewed results it cites report a 29% cut in admin time. That is a number European health systems can act on without waiting for a foundation model of their own. (Tech.eu)
Anthropic sells advisers a tool that promises not to advise
Claude for Financial Advisors launched on 14 September, wired into BlackRock, Charles Schwab, Vanguard, and iCapital. Anthropic says the model will not give investment advice and leaves that judgment to humans.
MiFID II does not work that way. A personal recommendation can be implied rather than stated, and ESMA has told firms that using AI changes none of their suitability obligations.
European advisers buy the tool and keep the liability. (The Next Web)
German AI use jumped to 57%, and almost nobody thinks they are using it well
Bitkom surveyed 603 German companies with 20 or more staff over July and August. Adoption reached 57%, up from 36% a year ago and 20% in 2024.
Of those using it, 59% say they are not tapping the potential at all. Three per cent say they are using it heavily, and none said fully.
The cost is not licences. Infrastructure was named by 51%, data preparation by nearly as many, and integration by 41%, against 21% for licences.
Two thirds of all respondents think the AI Act does German firms more harm than good. (heise online)
Governance & Safety
Hundreds of contractors are reading real ChatGPT conversations
404 Media reported that OpenAI pays hundreds of contractors to read and rate real ChatGPT exchanges under the internal name Project Lily. Reviewers do not see usernames, but they do see a memories summary that can reveal what the person uses the chatbot for and roughly where they live.
Asked where it had told users humans may read their chats, OpenAI did not answer, later pointing to a help page. The Court of Justice held in EDPS v SRB last September that the duty to inform lands on the controller at collection, whatever the recipient can work out.
Italy's Garante has already fined the company EUR 15m. The model-improvement setting is on by default for consumer accounts and off for the enterprise tiers. (The Decoder)
German researchers welcome the pause call, then pick it apart
Amodei proposed a coordinated slowdown of one to two years at the weekend. Altman and Musk backed it within hours, Asian AI stocks fell on Monday, and Trump waved it off from Ireland.
Jonas Geiping at the Max Planck Institute for Intelligent Systems told the Science Media Center the independent oversight bodies are the useful part. Thilo Hagendorff at Stuttgart noted that Amodei ties the plan to America extending its lead over China, which sits badly with a coordinated slowdown.
Philipp Hennig at Tubingen was blunter. Someone worried about the internet or human survival has options more concrete than a nebulous request for global coordination. (heise online)
Compute & Infrastructure
Sweden switches on a 60 petaflop machine for EUR 68.5m
EuroHPC and NAISS inaugurated Arrhenius at Linkoping University on 8 September. It runs above 60 petaflops, ranked 42nd on the June TOP500, and cost EUR 68.5m, with the Joint Undertaking covering up to 35%.
Europe is buying mid-range capacity for climate modelling, materials science, and life sciences while the AI Gigafactories tender launched in July is still running. Those are different machines for different jobs, and only one of them is being built at frontier scale. (EuroHPC JU)
AI & Society
Teachers in six countries say AI arrived faster than their schools could handle
Epson surveyed about 3,360 teachers and students in the UK, France, Italy, Germany, Spain, and Poland over July and August. More than 80% of teachers said AI is entering classrooms faster than schools can manage, and 68% said it is hurting learning.
The response is further along than the survey suggests. Italy started mandatory AI-literacy training for school staff on 3 August, and the Commission is drafting an AI literacy framework for schools with the OECD.
Whether either reaches the 82% of teachers asking for help supervising student use is the open question. (EdTech Innovation Hub)
One to watch
Who gets to certify that a frontier model is safe
Two labs have now offered to let outside evaluators in, and a European company has volunteered while being acquired by a chipmaker with a stake in one of them. None of that defines independence.
California's SB 813 would license a class of independent verification organisations. One METR evaluation ran up $400,000 in API tokens, paid by OpenAI.
The AI Office has held the power to demand evaluations since 2 August, with a Scientific Panel to advise it. Eliska Andrs at the Future of Life Institute argues the missing piece is funding for evaluators rather than more powers; watch whether the Commission budgets for that before the next incident forces the question. (EU AI Act Newsletter)