The AI Brief

The Information reported on September 13 that Anthropic, OpenAI, and Google DeepMind have been holding working-group meetings since July, well before Anthropic CEO Dario Amodei's public "We Must Pace the Frontier" essay, on creating an industry-led standards body focused on technical testing and pre-release auditing of frontier models. Amodei is driving the effort; OpenAI CEO Sam Altman is backing it as an alternative to waiting on a Washington process that has stalled. A draft executive order on a federal AI standards body, per Seoul Economic Daily, ran into internal White House conflict and went nowhere.

Microsoft CEO Satya Nadella joined the pacing camp on September 13, welcoming "the deliberate pacing needed to get alignment right" and announcing that Microsoft will publish a Code of Conduct governing its first-party MAI models for public consultation on September 14. Nadella also argued that AI governance "cannot be controlled by a handful of entities" and must include broad representation across countries and academia. Microsoft AI president Mustafa Suleiman called the principle "straightforward common sense." Google DeepMind chief Demis Hassabis had called in July for a FINRA-style (Financial Industry Regulatory Authority) frontier AI oversight body, which now appears to be the institutional model being discussed across the three labs.

The resistance formed the same day. Sacks accused the labs of "pretending antitrust law has to be suspended so you can form a cartel," a charge that lands directly on any body whose membership is the same companies whose releases it would assess. Separately, two safety researchers quit to join METR, the independent evaluator Amodei named in his essay as the model for embedded review: Joe Benton, former lead of Anthropic's Scalable Oversight team, and Josh Engels, a Google DeepMind safety researcher, both resigned September 12. Benton publicly called for mandatory reporting of recursive self-improvement progress and incident disclosure, saying there are "no adults in the room." The Amodei pacing essay and the Altman endorsement were first covered in Vol. I, No. 111 and Vol. I, No. 112; today's item reports the institutional infrastructure that predated both public statements.

Disclosure: Claude, which generates this brief, is built by Anthropic.

SoftBank Group shares fell 11.2% in Tokyo on September 14, their steepest single-day decline in nearly three months, after OpenAI CEO Sam Altman endorsed Anthropic CEO Dario Amodei's weekend pacing essay and confirmed OpenAI will not pursue an IPO in 2026. The selloff spread across Asian AI-linked stocks: South Korea's Kospi lost 2.5%, Japan's Nikkei slid 0.8%, and US Nasdaq futures dropped 1.6% in premarket trading. Nvidia fell 2.1% premarket; AMD fell 4.8% and Intel 5.3%. Memory chipmaker Micron, which had surged more than 240% year-to-date on AI demand, fell 5%.

SoftBank's exposure is compounded by scale and structure. The conglomerate's total commitment to OpenAI is expected to reach approximately $65 billion by October, and it is reported to be weighing a $10 to $20 billion junk-bond sale while carrying the OpenAI stake as margin-loan collateral. Anthropic CEO Amodei specifically warned that continuing at the current pace of development could allow autonomous systems to take over critical internet infrastructure within six to twelve months, a claim that President Trump rejected the same day, saying the US lead over China in AI must be preserved. China's Ministry of Foreign Affairs also issued a rebuttal to Amodei's call for restrictions on Chinese AI development, and China's State Security Minister Chen Yixin separately published a warning that rapid AI advances threaten political stability.

Sakana AI, the Tokyo lab co-founded in 2023 by Llion Jones, a co-author of the Transformer paper, and David Ha, released Fugu Max and Fugu Ultra v2 on September 11. The two models represent opposite ends of the same architecture: Fugu Max targets cost-performance efficiency at $2 per million input and $6 per million output tokens, which Sakana says undercuts Claude Sonnet 5 and GPT-5.6 by 40 to 60 percent on output. Fugu Ultra v2 targets maximum capability at $5 input and $30 output per million tokens, with a 1 million-token context window. Both are served through a single OpenAI-compatible API endpoint; existing integrations switch tiers by changing the model identifier.

The technical architecture is documented in two ICLR 2026 papers. TRINITY (arXiv 2512.04695) evolved a lightweight coordinator assigning Thinker, Worker, and Verifier roles across model turns. Conductor (arXiv 2512.04388) uses reinforcement learning to discover natural-language coordination between agents. Fugu Max expands the orchestration pool to include Nvidia Nemotron models through a Sakana-Nvidia partnership. Fugu Ultra v2 notably excludes Claude Fable 5, Fable 5.1, and GPT-6 Astra from its worker pool and still claims to beat all three on several benchmarks: 48.3 on Chartography (visual reasoning) against 27.3 for Claude Opus 5; 74.3 on DeepSWE (software engineering). Sakana reports best or joint-best results on five of eight benchmarks for Ultra v2.

The critical caveat: all benchmark results are Sakana-reported and none have been independently reproduced on Artificial Analysis or a comparable third-party evaluation platform. Sakana also does not disclose the identity of the models in its orchestration pool or per-task routing decisions, making external audit of the claimed performance gains impossible at this stage.

The Information reported on September 9 that OpenAI notified advertising partners that it would no longer accept ChatGPT ads for standalone image and audio generation products competing with its own features. Adobe was the clearest casualty: the company, which had participated in OpenAI's early advertising pilot and had been running campaigns for Firefly AI and Acrobat Studio, lost access without advance notice. The policy change was confirmed by Adobe's Senior Director for Americas Media, Doug Wyatt. Video generation ads remain permitted for now.

The timing is not coincidental. OpenAI launched ChatGPT Images 2.5 and ChatGPT Live voice capabilities in the same period it tightened ad rules. Beyond paid placements, OpenAI has also removed outbound referral links from in-app search results for image-related queries, eliminating both paid and organic discovery routes for third-party image tools. OpenAI's ad business hit a $1 billion annualized revenue run rate in under 200 days and is targeting $2.4 billion by year-end; analysts note the paradox that barring competitors reduces the advertiser pool while chasing an ambitious revenue target. Per The Information, the policy change was not reflected in OpenAI's published advertising guidelines at the time partners were notified.

The declaration, titled "A Severe Misalignment of AI in Mathematics," was published September 11 on Terence Tao's blog and at mathandai.org with 25 initial signatories, all Fields Medal winners spanning award classes from 1978 through 2026's ceremony. Named signatories include Tao, Scholze, Viazovska, Pierre Deligne, Martin Hairer, Cédric Villani, Manjul Bhargava, Caucher Birkar, June Huh, James Maynard, and 2026 medalist Yu Deng. The declaration was open for additional signatures as of publication. It dominated Hacker News on September 13 and 14 with 1,219 points and more than 1,200 comments, making it the highest-scoring AI item of the period.

The signatories' argument targets the metric, not the technology: they object to AI labs treating unsolved problems as benchmarks, because solving them "in a rush" without proper writeup, isolation of novel contributions, and attribution of prior work destroys the understanding that mathematics exists to produce. The declaration reads as a direct response to the week's events: OpenAI claimed a Navier-Stokes proof via Astra in early September, Anthropic's Claude produced a Lean-verified Fermat's Last Theorem proof in September 5, and multiple mathematicians including TU Dresden group theorist Andreas Thom alleged that OpenAI trained Astra on private prior-work conversations without attribution. Tao himself published a companion arXiv essay (2608.16753) based on his 2026 International Congress of Mathematicians lecture examining what mathematical goals AI actually serves.