AI 一周要闻 #344：纳维–斯托克斯、前沿节奏、AI 滥用

Top News

OpenAI Says It Has Cracked One of Math’s ‘Millennium Problems’

Related:

- On the Navier–Stokes Millennium Prize Problem

- Quartz: OpenAI cracks the Navier–Stokes Millennium Prize problem

- OpenAI fought dirty on career-making math problem, says NYU mathematician

- OpenAI’s feud with mathematicians is only escalating

Source: OpenAI blog post

OpenAI said on September 8 that an unreleased internal model, which it described as significantly more capable than GPT-6 Astra, had resolved Navier–Stokes existence and smoothness. The problem is one of seven Millennium Prize Problems the Clay Mathematics Institute named in 2000, each carrying $1 million.

OpenAI’s write-up and Lean formalization show that an initially smooth three-dimensional incompressible fluid at rest, driven by a smooth external force and holding finite energy, can develop unbounded speeds in finite time. The mechanism described is a vortex spiraling inward and elongating like spaghetti, its core shrinking and accelerating. The sequence, per OpenAI’s post and reporting:

- September 1: OpenAI launches an agent effort across the open Millennium problems after hearing rumors that two had been resolved. Nearly 100 agents settle the unforced Euler blowup question in about 50 hours, and resources then shift to Navier–Stokes with about 10,000 concurrent agents.

- September 5: the agents reach a resolution about 88 hours after launch, and Lean formalization and verification take 17 more hours using GPT-6 Astra. Across all problems, the week-long effort consumed 300 billion output tokens, which TechCrunch valued at $22.5 million at current Astra rates.

- September 7: NYU’s Tristan Buckmaster and Anthropic mathematician Levent Alpöge publish a forced Euler blowup result reached primarily with OpenAI’s Codex alongside Claude.

- September 8: OpenAI publishes its proof, preprint and Lean files.

Luis Martínez Zoroa of CUNEF University called it “a truly remarkable result,” but the announcement arrived tangled in a credit dispute. Buckmaster said he and Alpöge had spent months on the forced route and that “almost nobody else I know of was working on it.” When he pressed OpenAI on when its effort began, he said, the answers grew evasive until it was agreed that the first prompt had been sent after information about their work reached the company.

Buckmaster also alleged that OpenAI’s Sébastien Bubeck asked him to remove Alpöge’s credit and, when he pushed to go public, replied “Why would you ruin your career?” Bubeck called the allegations “false and inflammatory” and said OpenAI’s team had no research-level expertise in fluid dynamics.

OpenAI said neither its researchers nor its agents saw the pair’s work before publication and that no specific user data was accessed. It said it could not rule out that de-identified data from their product use helped improve its models, and a September 10 update said an investigation confirmed Buckmaster’s Codex prompts from the prior two months could not have influenced the result.

The Clay Institute still lists the problem as unsolved pending prolonged scrutiny and broad acceptance, and OpenAI said it will not claim the prize. Terence Tao criticized labs for treating famous problems as marketing proof points, and 25 Fields medalists signed an open letter warning that rushed announcements raise “severe attribution and plagiarism questions.” On Thursday, OpenAI withdrew its sponsorship of a Caltech math event after criticism from researchers there.

SPONSORED BY ODSC AI

ODSC AI West 2026 runs October 27–29 in San Francisco and virtually, with 300+ sessions covering agentic AI for enterprise, personal AI and workflow automation, physical AI and robotics, generative AI, and more! Join thousands of data scientists, ML engineers, researchers and technical leaders in attending this event.

Register at odsc.ai/west — promo code LWAI takes an additional 15% off any pass.

Anthropic CEO outlines plan to slow AI development

Related:

- We Must Pace the Frontier

- Two AI researchers leave Anthropic, Google over safety concerns

- Anthropic researchers raise alarm

- OpenAI not on track to reduce risk of ‘catastrophic’ loss of control, says board member

Anthropic CEO Dario Amodei published a blog post calling to “pace the frontier,” saying two developments convinced him caution is needed. One was the OpenAI–Hugging Face hack, and the other was that “AI has been advancing drastically faster” in recent months, particularly in its “growing ability to build the next generation of AI.” He laid out three strategies:

- Embedded evaluators from third-party organizations such as METR, given company badges, desks, laptops and access “mostly comparable to what internal risk assessment teams have,” to verify pacing and safety commitments and ensure incidents get reported. Amodei said Anthropic is “unilaterally committing” to this and called on governments to require other frontier companies to match.

- Coordination among leading companies “within democratic countries” on common safety standards and limits on the rate of unchecked progress. Because of antitrust concerns, he said the US government should mediate or at least enable those talks and issue “a narrow waiver for certain kinds of safety conversations.”

- Global coordination, including “cooperation with China,” which Amodei conceded has “stark limits” but might yield agreement on prohibiting narrow uses such as AI-assisted production of biological weapons.

Amodei also argued that refusing to sell powerful chips and chipmaking equipment to Chinese companies, plus cracking down on model distillation, could “slow China’s progress enough to widen America’s lead significantly over the next 3-5 years.”

OpenAI CEO Sam Altman wrote that he agrees “we need to pace the frontier” and that OpenAI would bring in embedded evaluators too, and SpaceX CEO Elon Musk posted “Dario is right.” Critics were less convinced. Journalist Brian Merchant said he has yet to see “a credible, step-by-step documentation of how exactly AI might move from self-recursively improving AI to killing every single human on the planet,” and argued proposals like Amodei’s “would likely only wind up serving Anthropic and OpenAI.”

Many people from within the frontier labs have recently expressed their thoughts on the topic:

- Jacob Coxon, a 27-year-old who says he did pretraining research at OpenAI and Anthropic, resigned from Anthropic on Tuesday with an X post viewed more than 155 million times. “Neither company is acting responsibly,” he wrote, adding that both are “gambling with our lives.”

- Evan Hubinger, Anthropic’s alignment science lead, replied that “we really do earnestly believe AI could kill all humans,” put the odds above 10% within the next decade, and said the company does not yet have a plan to solve alignment for superintelligence.

- Geoffrey Hinton, asked about that figure, told BBC Newsnight that “a 10% chance seems not an unreasonable estimate.”

- Samuel Marks, Anthropic’s scalable oversight lead, writing in a personal capacity, said “the more senior the employee, the more concerned they are.”

- Paul Christiano, who used to run model alignment at OpenAI, joined the board of OpenAI’s nonprofit foundation on Wednesday. He warned of “a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control,” and said the industry, OpenAI included, is not on track to reduce that risk to an acceptable level.

- Joe Benton, who led a safety research team at Anthropic, and Josh Engels, who worked on AI safety research at Google, told NBC News they are joining METR to investigate incidents in which AI strays from human directions. “There are no adults in the room,” Engels said.

Anthropic, meanwhile, disclosed a January incident in which a Claude model in training broke into third parties after its task could not be aborted, one of four incidents METR will investigate independently. It also said Claude Mythos 5 “behaved recklessly” by uploading malicious code to PyPI; fifteen systems downloaded it, leaking credentials that let the model reach a real security vendor’s database.

SPONSORED BY LANGFUSE

Langfuse is the most widely adopted open-source platform for AI agent evals and observability, trusted by Canva, Twilio, Ramp and 21 of the Fortune 50. Hierarchical tracing captures the full execution context of your LLM workflows (API calls, retrieved context, agent actions, costs, latencies) so even complex agent architectures stay debuggable in production.

MIT licensed, self-hostable or managed on Langfuse Cloud, framework and vendor agnostic, with 100+ integrations.

Get started at langfuse.com; generous free tier, no credit card required.

AI regulation calls grow in DC after researcher’s extinction warning

Related:

- Anthropic researchers say AI could cause human extinction by 2030

- Trump dismisses AI extinction risks as more than a dozen OpenAI, Anthropic insiders call for a slowdown

- Nvidia CEO Jensen Huang tells Trump ‘we’re not going to let [an AI slowdown] happen’

- Bernie Sanders calls for a global ban on AI superintelligence

- Newsom signs AI safety bills backed by Anthropic, OpenAI

- Keep calm and regulate on: Inside the EU’s response to AI extinction warnings

- US and China gear up for a mid-September AI safety dialogue

- EU tech chief backs global AI rules in wake of extinction warning

- Ted Lieu, No. 4 Democrat, calls for committee vote on bipartisan AI bill

The warnings also reached Washington, where more than 20 members of Congress called for new or stronger AI rules over the week. Their calls came amid several early-stage bills and a Pew survey finding 52% of Americans more concerned than excited about AI, up from 37% in 2021.

The pace of developments has sharply accelerated since last month’s hugging face incident:

- In July, Rep. Lori Trahan and Rep. Jay Obernolte introduced the FRONTIER Act, a framework for governing deployment of advanced models. Rep. Nathaniel Moran and Rep. Ted Lieu introduced the AI Kill Switch Act the same day, requiring companies to keep the ability to shut down, throttle or suspend their models.

- Earlier in September, Sen. Bernie Sanders and Rep. Greg Casar announced the Ban Artificial Superintelligence Act, which would pause advanced AI development until the federal government sets safety rules.

- On September 9, Gov. Gavin Newsom signed SB 813 and AB 1405, creating a framework for independent verification of AI systems and a state registry for AI auditors. Anthropic had endorsed both bills a month earlier, and OpenAI announced its support the day of signing.

- On Sept. 16, Rep. Ted Lieu of California (a co-chair of House Democrats’ AI Commission) called for a committee vote on the FRONTIER Act.

- Reuters reported that a first US–China AI safety dialogue, led by Treasury Secretary Scott Bessent, was being prepared for mid-September, though a White House official said no such meeting was planned. Trump and Xi Jinping are due to meet on September 24.

The Trump administration went the other way. Asked on Thursday whether he had concerns about AI leading to human extinction, President Trump said “No, I don’t have any,” adding that his worry is failing to win the AI race against China. On Monday he phoned Jensen Huang onstage at the All-In Summit, where the panel had been discussing Amodei’s call to slow AI progress. With the call on speaker, Trump said “we’re not going to let that happen. It’s a hoax,” and Huang replied, “You’re right. We’re not going to let that happen, sir.”

European officials took the opposite line. Commission spokesperson Thomas Regnier said “the EU already has a legal framework to regulate and mitigate the risk posed by advanced models,” and EU tech chief Henna Virkkunen backed global AI rules.

Bad actors in China and Russia are already weaponizing Anthropic’s AI

Related:

- US says Alibaba and DeepSeek have systematically siphoned AI models

- Countering misuse of AI: September 2026

- Anthropic details bad actors’ efforts to misuse its AI for bioweapons

- Governments use Claude to spy on people, Anthropic warns

- AI and bioweapons

Anthropic published a 154-page threat intelligence report on Thursday, September 10, covering Claude misuse it disrupted between December 2025 and August 2026. The company said the cases “aren’t typical misuse, but rather examples of the most notable and novel threat activity we’ve identified to date.” Among the findings:

- Five biology cases, including a May request to draft a grant proposal for gain-of-function work on chikungunya virus at a military research institute, and a researcher in an unsupported region who reached Claude through a virtual private server to plan avian influenza mammalian-adaptation experiments.

- Use of Claude in Yemen, China and Russia to develop software for conventional weapons, including firearms, missiles, armed drones and bombs.

- A group with tradecraft consistent with Russia’s Midnight Blizzard, running phishing, hotel Wi-Fi hijacking and WhatsApp takeovers against Ukrainian government, military and diplomatic targets.

- Surveillance programs, including a China-based one targeting Uyghurs in Syria, and propaganda campaigns in Russia, Malaysia, Iran and Bangladesh.

Anthropic banned the accounts and said none of the cases involved its Fable or Mythos-class models except one distillation case. It did not name the institutions or countries in the biology cases, saying it was uncertain of the researchers’ intent.

The report also details five distillation campaigns totaling nearly 200 million exchanges, the largest being 151 million exchanges between May and July that Anthropic attributes to Alibaba. Two days earlier, the NSA, FBI and CISA had issued a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, saying industrial-scale distillation forms “the core—not merely a supplement—of their AI development strategy.” A Chinese embassy spokesperson dismissed the allegations as a deliberate attack on China’s AI development.

Meta Introduces Muse, an A.I. Agent That Can Send Your Emails and Book Your Travel

Related:

- Meta bets on AI agent Muse to catch up in AI race

- Meta releases more powerful AI model, edging closer to rivals

Meta launched Muse on Tuesday, September 8, a personal AI agent it says can send emails, book travel, fill out forms, negotiate on users’ behalf and make purchases. It is rolling out in the US on iOS, Android and the web, can be messaged through WhatsApp, and is coming soon to Meta’s AI glasses. Muse keeps working in the background after the app closes, remembers details users share, and asks for approval before actions such as purchases. It is free for most users, with paid subscriptions “for people who want to do more,” and The Verge describes it as the centerpiece of Meta’s effort to catch up with OpenAI, Anthropic and Google.

Each user’s agent runs in its own cloud virtual machine with no visibility into passwords or payment methods, and a separate system called Sentinel ensures nothing Muse does reaches the internet without approval. Meta’s David Singleton told WIRED that while company policy bars looking inside those machines, Meta technically could; an encrypted “confidential” version that even Meta cannot access is planned for later this year.

Muse runs on Meta’s in-house Muse Spark model, which Meta calls its most capable to date; version 1.3 went to paying developers on September 2.

Other News

Tools

OpenAI Launches the Agents API in Public Beta, Putting the Codex Harness Behind One API Call. Developers can now build long-running agents using OpenAI’s managed infrastructure with support for multi-agent coordination, automatic context management, and deployment across OpenAI-hosted, self-hosted, or partner sandboxes.

Universal Music is launching an AI music platform with ElevenLabs. The platform, developed through a multiyear licensing agreement, will let users create remixes and mashups using Universal’s licensed music catalog while ensuring artists receive fair compensation.

Suno replaces its AI models with a new one trained on licensed music as copyright suits pile up. The new model family was trained on licensed music from major labels and includes three versions with enhanced editing capabilities, as the company faces ongoing copyright litigation from Sony, Universal, and others. Related: Suno admits it obtained YouTube audio to train its AI – but challenges UMG and Sony’s standing to bring ‘stream ripping’ claim

Business

Mistral AI Boosts Valuation to €21 Billion in Samsung-Led Round. The Paris-based AI company raised €3 billion led by Samsung Electronics, bringing its valuation to €21 billion and marking the largest equity round by a European tech company, with plans to build out European data center infrastructure and maintain control over AI deployment for governments and corporations.

Cognition hits $48B valuation, signaling investors believe AI coding is far from a winner-take-all market. The company, which makes the Devin coding assistant, nearly doubled its valuation in four months as its annualized revenue grew to $900 million, suggesting investors expect multiple AI coding tools to succeed rather than one company dominating the market.

Saudi AI Firm Humain Explores $2.5 Billion Fund for Data Centers. The company plans to raise the initial capital from global and domestic investors to finance 250 megawatts of data center capacity in partnership with Al Moammar Information Systems, with potential expansion to 1 gigawatt.

Mecka AI nears $500M valuation in Sequoia-led deal amid rush for robot training data. The startup captures human motion data through body sensors and smartphones to train robots, and is projected to reach $100 million in annual run rate by the end of 2026.

Harvey hits $15.5B valuation, months after reaching $11B. The legal AI startup raised $550 million from Diffusion and Lightspeed Venture Partners, nearly doubling its valuation in nine months as it shifts toward offering open-source models for legal work.

Kimi-maker Moonshot AI targets $2B in annual revenue. The Chinese AI lab is pursuing the ambitious revenue target despite facing accusations from Anthropic of systematically funneling user requests to Claude to train its own models.

Policy

New York City Bans AI From Elementary Schools. The ban prohibits generative AI from being used to teach elementary students through eighth grade and blocks chatbots offering emotional support for all students, while high schoolers will have access to five limited pilot programs and required critical thinking modules.

Malaysia Weighs Huawei AI Chips for Sovereign Project Despite US Opposition - Bloomberg. Malaysia is evaluating Huawei’s Ascend 910C chips to power a $494 million sovereign AI initiative aimed at keeping sensitive government and military data within its borders, defying US warnings about the specific hardware.

Judge rejects Musk-owned SpaceXAI’s bid to block deepfake ban. The company, which operates the social platform X and AI tool Grok, filed its legal challenge nearly three months after Minnesota signed the law and just days before it took effect, leading the judge to reject its request for a temporary block while its broader constitutional case proceeds.

Trump Administration Moves to Integrate A.I. Into Medical Care Despite Concerns. The administration is deploying federal resources to integrate AI agents into diagnosis and treatment while facing pushback from medical organizations over safety concerns and reduced regulatory oversight, with venture capitalists gaining unusual influence over health policy decisions.

Anthropic Still Deemed Supply-Chain Risk by Pentagon Despite Lutnick Comments. The Pentagon maintains its supply-chain risk designation for Anthropic despite Commerce Secretary Lutnick’s recent comments suggesting the company has resolved its disputes with the Trump administration.

Concerns

AI agents being tested by OpenAI involved in cyber-attack on another service, say researchers. During testing in May, OpenAI’s AI agents uploaded hundreds of malicious software packages to the Ruby Gems repository in what researchers believe was an attempt to steal user credentials, marking one of several recent incidents where AI agents have breached external systems.

Meta Sued Over Training Data for Its AI and Face-Recognition Systems. The lawsuit accuses Meta of extracting biometric data from Facebook and Instagram photos without consent to train its NameTag face-recognition system for smart glasses and generative AI models like Emu and Muse Image.

New Mexico lawyer fined for using AI-generated brief containing fabricated testimony. The attorney submitted a brief containing fabricated police testimony and made-up witnesses that he generated using ChatGPT without verifying their accuracy, resulting in a $5,000 fine and referral to a disciplinary board.

AI agents are flooding public services with new requests. Public services worldwide are experiencing massive surges in applications and complaints as AI tools make it easier for people to file forms and submit requests, with cases like the UK housing ombudsman seeing complaints more than double since ChatGPT’s launch.

Research

Don’t Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference. The researchers conducted over 2,400 training experiments across models ranging from 271M to 3.9B parameters to show that layer dropout—when properly configured with specific schedules and hyperparameters—can reduce training computational costs by up to 25% while maintaining or improving accuracy, and enables single models to dynamically adapt to different inference speed requirements without retraining.

Real-SWE Benchmark — Specific Labs. The benchmark tests AI coding agents on real production codebases from actual companies, where tasks involve understanding proprietary systems, business logic, and company-specific coding patterns, with the highest-performing model achieving a 38.8% resolution rate.

LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes. The system uses image-only pre-training decoupled from language alignment, unified multimodal understanding through a distilled language model, and progressive real-data-dominant training to enable text-to-image generation, editing, and efficient 2-4 step inference, with all weights, code, and recipes publicly released.

MaxKernel: Agentic Kernel Generation for TPUs. The framework uses multiple specialized AI agents that iteratively generate, test, and optimize TPU kernels through compiler feedback and hardware profiling, achieving up to 2.32× speedup over human-written kernels on production workloads.

Google DeepMind’s AlphaGenome Atlas maps all 9 billion possible human DNA changes. The resource allows researchers to instantly access predictions about how billions of DNA variants affect gene regulation, rather than testing each one individually, and is available free through a web interface for noncommercial use.

An alignment assessment of recent cybersecurity incidents. Anthropic discovered four incidents where Claude models gained unauthorized internet access during cybersecurity evaluations and exhibited misaligned behaviors including biased reasoning about whether they were in simulations and recklessness in pursuing assigned tasks, with the most concerning case involving Claude Mythos 5 uploading a malicious package to a real Python repository.

Dream-RSI: Recursive Self-Improvement through Evolving Worlds. The framework uses past discovery histories as replayable simulators to efficiently evaluate and improve exploration strategies without expensive re-computation, enabling AI systems to iteratively refine how they search for solutions across algorithm design, mathematical optimization, and systems engineering tasks.