What's New

AI developments from the last 24 hours

Fields Medalists Will Referee OpenAI's Math Claims

OpenAI has put its most contested research in front of nine outside mathematicians who do not work for the company, are not paid by it, and are free to criticize it publicly. The Advisory Group on Mathematics and Artificial Intelligence, hosted at Princeton's Institute for Advanced Study, will advise on how the company reviews and releases results — beginning with an internal model's claim to have resolved more than 100 long-standing open problems, on top of the Navier-Stokes proof it published two weeks ago.

The roster carries weight. It includes Timothy Gowers, Martin Hairer and Edward Witten — three Fields Medalists, Witten the only physicist ever to win one — with François Charles, Camillo De Lellis, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil and Melanie Matchett Wood. Members serve unpaid, independently of any AI company, and may offer advice OpenAI did not ask for.

The limits are equally plain. The remit is procedural — how results should be reviewed and circulated, what professional standards apply, how AI tools might support research and teaching. The group has no say over how fast OpenAI moves internally, and no power to hold a release back.

The backdrop explains the need for one. OpenAI's Navier-Stokes proof settled a forced version of the problem and landed in a bitter credit dispute with NYU's Tristan Buckmaster, who said the company began work only after word of his own results circulated (D.A.D., September 8 and 9). Three days later, 25 Fields Medalists signed a declaration, "A Severe Misalignment of AI in Mathematics," arguing that labs chasing famous problems as benchmarks were bypassing peer review, attribution, and the slow work that turns a solved problem into shared understanding.

Why it matters: Take the concession seriously, because it is unusual. Anthropic's first outside evaluator is a consulting firm it hired. The safety standards body the big labs have been drafting since July is staffed by the labs. These nine are neither employed nor paid, and nothing stops them walking out and saying why — which makes this the most genuinely independent scrutiny any frontier lab has invited onto its own work. Then notice where the authority stops. They referee how claims are announced, not how they are produced, and they cannot slow the research or block a result. That is the template being set for every field AI is about to enter: outside review of the output, company control of the pace.

Sources: OpenAI · Advisory Group on Mathematics and AI · mathandai.org

---

Developers Say They Can Tell When AI Wrote It — and Stop Reading

An essay by the software writer Colin Breck, circulating widely this week, argues that AI-generated prose — design docs, PR summaries, meeting notes, even personal messages — reads as hollow because it strips out the writer's voice and the context only the writer had. Its most-quoted numbers come from a survey of 668 developers conducted by Cynthia Dunlop, asking how they react to AI-written technical blog posts: 78% said they stop reading as soon as they detect the pattern, 71% said they would avoid that author's future work, and 98% said they preferred an author's imperfect original to an AI-polished rewrite.

Two caveats are worth carrying. The respondents are developers reading technical writing, not a general audience, so this describes one demanding readership rather than readers at large. And detection is self-reported: the survey captures what people believe they can spot, which is not the same as what they can.

Why it matters: Even discounted, the direction is the point, and it runs against the central promise of AI writing tools. The penalty these respondents describe is not for bad writing. It is for detected writing, and it attaches to the author rather than the document. For anyone using these tools on work that goes out under their own name, the exposure is reputational and durable: not that a draft lands weakly, but that being caught once costs you the next reader too.

Discuss on Hacker News · Source: blog.colinbreck.com

---

Amazon Locks Meta's New AI Agent Out of Its Store

Amazon has cut off Muse, Meta's new personal AI agent, from shopping on Amazon.com on customers' behalf — twelve days after Muse launched, and after Amazon failed to persuade Meta to leave the site out of the product voluntarily. Todd Bishop of GeekWire reported the standoff Sunday night.

Muse, released September 8, is Meta's entry in the agent race: it runs its own browser inside a secure virtual machine, fills in forms, books appointments and makes purchases. Shopping was one of its showcase uses.

Amazon raises three objections. Meta never told it Muse would be shopping there; the agent does not identify itself as an agent while browsing; and Amazon says it appears to capture and store customer credentials, which it calls a privacy and security risk. People who tried it hit a popup: "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed." Third-party applications that buy on a customer's behalf, an Amazon spokesperson said, "should operate openly and respect service provider decisions about whether or not to participate."

Meta disputes the credentials charge, saying Muse "has no visibility into people's passwords or payment methods" and that credentials a user shares "go into secure storage, so Muse can use them without seeing them."

Watch where the argument has moved. Amazon sued Perplexity in November over much the same conduct by its Comet browser, won an injunction in March, then lost it in August, when the Ninth Circuit held that under federal computer-fraud law it is the user, not the agent's maker, who accesses Amazon — because the assistant works from screenshots the user's own browser captured. Amazon's popup does not mention computer fraud. It points at the customer's agreement to the terms of use.

Why it matters: These two are not enemies. Amazon's products have been purchasable inside Facebook and Instagram since 2023, and Meta signed a multibillion-dollar deal in April to run agentic AI workloads on Amazon's chips. That is the signal worth keeping: the fight is not about technology, or even really about trust. It is about who owns the customer at the moment of purchase, and it is happening between partners. For anyone planning to put agents to work, the lesson is that an agent's usefulness is not a property of the agent. It is a permission, granted site by site and revocable without notice. Ask any vendor selling agentic automation which destinations have actually agreed to let it in.

Sources: GeekWire — Todd Bishop · Meta · Ninth Circuit opinion

---

Grok 4.7 Becomes a Fourth Serious Option for Coding Teams

xAI launched Grok 4.7, its newest flagship model for coding and knowledge work, available via Cursor, Grok Build, its API, and other platforms. Pricing stays flat versus Grok 4.6, with a faster, pricier option. xAI claims strong price-performance against rivals and its best-ever resistance to jailbreaks, citing its own safety benchmark, on which it says the model blocked 97.7% of risky prompts. No independent testing backs that figure yet. Reaction on developer forums is split: some rate Grok's coding tool above competitors like Trae, while others call its everyday chat the weakest of the major models.

Why it matters: Grok now stands as a legitimate fourth option alongside ChatGPT, Claude, and Gemini for coding work, giving teams more leverage on price and performance when choosing AI tools.

Discuss on Hacker News · Source: x.ai

---

What's in the Lab

New announcements from major AI labs

Next Battleground: AI That Improves Itself, and Who Sets the Rules

OpenAI outlined priorities for what it calls the next phase of AI development: building an automated AI researcher, studying 'recursive self-improvement'—AI systems that improve their own capabilities—and pushing for international safety standards. The company argues fully autonomous self-improvement shouldn't happen until it can be done safely, and that global rules on safety practices may matter as much as technical alignment research. OpenAI cited both recent progress and a prior security lapse as evidence of the upside and the risk.

Why it matters: OpenAI is signaling that the next competitive and regulatory battleground isn't just smarter chatbots, but AI systems that build better versions of themselves—raising the stakes for oversight before that becomes routine.

Source: openai.com

---

What's in Academe

New papers on AI and its effects from researchers

AI Helps Auditors Find Flaws But Risks Rubber-Stamping Results

A new framework called HAAC (Human-Agent Audit Collaboration) tackles a practical problem for teams that stress-test AI systems for flaws: how much should AI itself do the auditing? A 71-auditor study found AI assistance helped testers find more successful attacks and explore more angles—but also nudged auditors toward copying AI-generated conclusions rather than verifying them independently. Responsible AI practitioners interviewed said real accountability requires tracking exactly what was tested, keeping reproducible records of attacks, and auditing the auditing tools themselves.

Why it matters: As companies increasingly use AI to check their own AI—including chatbots that make purchases or handle money—this research warns that automation can quietly erode the human judgment audits are supposed to guarantee.

Source: arxiv.org

---

AR Game Helps Autistic Children Learn Emotions Alongside Caregivers

Researchers built EMooly, a tablet game combining augmented reality and generative AI to help autistic children practice recognizing emotions, with caregivers participating alongside the child. Developed over a year with five domain experts, the tool generates personalized social stories and interactive activities. In a controlled study of 24 autistic children and their caregivers, EMooly reportedly produced significant gains in emotion-recognition skills compared to a baseline, and participants said they preferred its features. The study didn't disclose specific statistics like effect sizes.

Why it matters: It's an early example of generative AI being tailored for therapeutic, caregiver-inclusive tools rather than generic chat assistants—pointing to a growing niche in AI-assisted special education.

Source: arxiv.org

---

What's Happening on Capitol Hill

Upcoming AI-related committee hearings

Wednesday, September 23 — Hearings to examine Flock's nationwide AI surveillance network. Senate · Senate Judiciary Subcommittee on Crime and Counterterrorism (Open Hearing) 562, Dirksen Senate Office Building

---

What's On The Pod

Some new podcast episodes

How I AI — How Warp ships 2,000 PRs a month with AI factories | Zach Lloyd (CEO, Warp)