Meta Reportedly Testing a Personal AI Agent That Browses the Web for You

Meta is reportedly testing "Muse," a personal AI agent with its own built-in browser, said to handle tasks across a user's daily life. Details are thin—no official Meta announcement or documentation has surfaced, and the news arrives via developer chatter rather than a product page. Commenters on Hacker News voiced distrust given Meta's history with user data, worrying Muse could push shopping suggestions and ads or scrape personal information through its browser access, though one tester called it more polished than rival agents.

Why it matters: If Meta builds an agent that browses and acts on your behalf, it inherits both the promise of AI agents and the company's long-running trust deficit on data privacy—making adoption a harder sell than the technology alone would suggest.

Discuss on Hacker News · Source: ai.meta.com

'Gambling With Our Lives': An Anthropic Researcher Quits, and His Safety Chief Agrees

Jacob Coxon, a 27-year-old pretraining researcher who spent three years inside OpenAI and then Anthropic, quit this week with a blunt public warning that rocketed past 20 million views: the labs are "racing straight to self-improving superintelligence and gambling with our lives." Neither company is acting responsibly, he told the Wall Street Journal, and the race has entered what he called the "endgame." Safety-motivated exits aren't new—but this one didn't stay a lone voice.

Anthropic's own alignment science lead, Evan Hubinger, publicly agreed. Endorsing Coxon's thread, he said what most executives only imply in careful press quotes: he personally puts the odds that AI kills all humans within the next decade at more than 10%. Anthropic is trying its best, he added, but has no plan yet to solve alignment for superintelligence and is not clearly on track to find one. He was careful about the timeline—today's models he considers low-risk; the danger is a superintelligence emerging from the recursive self-improvement loop that OpenAI and others now say is accelerating faster than expected (D.A.D., September 8).

The episode forces the obvious question, and Coxon answered it himself. Why keep building something you think might kill everyone? At OpenAI, he wrote, many haven't internalized the civilizational stakes; at Anthropic they have, but the company is locked in a race, convinced that if it doesn't reach superintelligence first, someone less careful will. Entering that race, he argued, is "a hubristic gamble that should not be launched from a private company's Slack." It is the same logic Helen Toner, the former OpenAI board member now at Georgetown, captured in a widely shared parable—monsters in the forest, rivals charging in for the treasure, so the safety-minded go faster to tame them first—while granting it is exactly the posture that makes outsiders wonder whether you are "crazy or lying or both." It maps onto the fault line we covered last week (D.A.D., September 5), where critics like Timnit Gebru dismissed extinction talk as fear ginned up to lift valuations. The uncomfortable possibility is that both hold at once: the fear can be sincere and still serve the company that professes it.

Sources: Business Standard · The Wall Street Journal, via TradingView · posts by Jacob Coxon, Evan Hubinger, and Helen Toner on X

Why it matters: The useful signal isn't the exact percentage—no one can truly calibrate a number like that—but who is saying it and what they concede. A researcher walked away from one of the best-paid jobs in tech rather than keep building, and the person whose job is to stress-test that company's models for catastrophe says he is not reassured and has no plan for the hardest part. You don't have to share their estimates to take the tell seriously: when the safety staff are this alarmed out loud—some heading for the exits—"move fast and trust the vendor" is not the posture the moment calls for.

Source: business-standard.com

'We Heard Rumors': OpenAI Publishes Its Navier-Stokes Proof, and Deepens the Credit Fight

Yesterday we covered the bitter dispute over an AI-assisted assault on the Navier-Stokes equations (D.A.D., September 8). Overnight, OpenAI escalated it by publishing the work at the center of the fight: a formal writeup and a machine-checked proof, in the verification language Lean, that a smoothly moving fluid can—under a smooth external force—accelerate to infinite speed in a finite time. It credits an unreleased internal model, more capable than GPT-6 Astra, running some 10,000 coordinating agents that reached the result in about 88 hours. If it holds up, it would be the first time an AI system has cracked a problem of this stature.

Two caveats come with it. OpenAI settled the forced version of the problem—the breakdown scenario mathematicians generally treat as the easier cousin of the celebrated question of whether ordinary, unforced fluids always stay smooth—which may be why it says it won't claim the $1 million prize. And the proof is only hours old: the Lean formalization means the logic is machine-verified, but human mathematicians haven't worked through it, and their early reaction has been sharply skeptical.

The document's real revelation is how the effort began. OpenAI's Sébastien Bubeck said the team started on the Millennium problems "due to viral twitter rumors that Anthropic had resolved 2 Millennium problems." By the company's own account, this was less a research program than a sprint to see whether its system could match a rival's rumored feat—launched, as Buckmaster alleged, only after word of his and Alpöge's work had spread. The admission drove the backlash, as critics fixed on the implication that a lab could watch a researcher near a breakthrough and point its own agents at the same target to get there first. OpenAI rejects the darkest reading: Bubeck says his researchers never saw the mathematicians' work before it was public, that the two proofs differ, that OpenAI assumed it was racing Anthropic only because Alpöge's posts ran on Anthropic's models—and that the mathematicians made contact first, a sequence they in turn dispute. (Terence Tao, whose earlier praise for the pair's work was widely cited, has since said his remarks were being misread as confirming a breakthrough that hasn't been verified.)

The quietest thread is the most consequential. OpenAI denies accessing any user data to solve the problem, yet concedes it cannot rule out that reworked, de-identified data from the mathematicians' use of its products helped improve its models. An unlikely source explained why that hedge matters: Cohere chief executive Aidan Gomez wrote that training on reworked user data is routine at the major labs, that novel problems like advanced math are the likeliest to be harvested, and that even a "we won't train on you" pledge usually protects only the exact text you enter—rewritten versions are fair game. Anthropic is contesting the framing too, with Alpöge posting a year-old email to show his collaboration with Buckmaster long predated OpenAI's interest.

Sources: OpenAI — "On the Navier–Stokes Millennium Prize Problem" · Scientific American · Hacker News · posts by Sébastien Bubeck (OpenAI), Aidan Gomez (Cohere), Tristan Buckmaster, and Levent Alpöge on X

Why it matters: Hold both halves at once. If the proof survives scrutiny, it is a genuine landmark—a machine-verified answer, in under four days, to a question that resisted humans for 90 years. That capability story is enormous. But the surrounding one is a warning. A company announced a historic result it won't take the prize for, in a document conceding it began only after hearing a rival's rumor and can't be sure a competitor's data didn't seep into its model. Whatever the mathematical verdict, this is what a prestige race looks like when the players are labs rather than scholars—and it leaves every professional a blunt lesson: whatever you feed an AI tool, a rewritten version of it may train the machine, even where you were promised otherwise.

Source: openai.com

An MIT Student Handed Routine Quantum Lab Tests to ChatGPT

An MIT graduate student in the Engineering Quantum Systems Group connected GPT-5.6 Sol, via OpenAI's Codex coding tool, directly to lab software to run routine calibration tests on superconducting qubit chips. On a standard six-qubit test chip, the AI reportedly handled a full measurement sequence largely on its own—finding transition frequencies, tuning control pulses, and clocking how long the chip retained quantum information—when signals were clean. It struggled and needed hand-holding when readings were weak or noisy. No success-rate or time-savings figures were disclosed.

Why it matters: It's an early sign that AI agents can take over the repetitive instrument-running grunt work in physical science labs, letting researchers spend more time designing experiments and interpreting results instead of babysitting equipment.

Source: openai.com

ChatGPT Hits 1 Billion Weekly Users as OpenAI Touts Compounding Gains

OpenAI used a blog post tied to GPT-6 Astra's launch to detail how it converts research gains into business scale: ChatGPT and its products now reach 1 billion weekly users and 2.5 million businesses, and users who stick around six months send 50% more messages and take on twice as many tasks as newcomers. This follows Astra's rollout (D.A.D., September 4).

Why it matters: The usage numbers are the point—evidence, in OpenAI's telling, that AI is compounding into genuine product stickiness rather than novelty traffic. It's the metric rivals and investors will track as the company pitches its next act.

Source: openai.com

Clinical AI Beats Physicians on Diagnosis in Simulated Primary-Care Test

A new study pitted a clinical AI system called Doctorina against eight physicians and four leading language models on 150 simulated Polish primary-care visits. Doctorina correctly identified the top diagnosis 82% of the time versus 57% for physicians, and scored higher on treatment quality too. It also beat general-purpose models like Kimi and Claude Opus, though those trailed closely on management decisions. The results held up when researchers reran the test.

Why it matters: It's an early signal that AI tuned specifically for clinical workflows—not just a chatbot with medical knowledge—can outperform doctors on diagnostic accuracy, raising the stakes for how health systems evaluate and deploy these tools.

Source: arxiv.org

How You Build a Bias Audit May Skew Its Results More Than the AI Does

A closely watched finding—that AI models flip from favoring minority applicants in one-by-one evaluations to penalizing them in side-by-side rankings—doesn't hold up outside its original context of charity aid decisions, according to a new audit spanning hiring, lending, and medical triage. Testing 40,726 requests across five models, researchers found the earlier effect shrank by half and, more strikingly, that models showed as much bias toward whichever candidate was listed first as toward any demographic trait. They also spotted every planted test case, raising questions about whether audits capture real-world behavior.

Why it matters: Companies and regulators leaning on bias audits to certify AI hiring or lending tools should note that how a test is built may skew results more than the model's actual discrimination does.

Source: arxiv.org