The safety chief walks out — and calls the culture 'broken'· 11h ago
David Robinson, who led safety work at OpenAI, has resigned, telling colleagues the company's culture is "broken" and that the time for "trial and error is over." He is not the first. Over recent weeks a pattern has hardened: the people inside these companies whose job is to object are the ones leaving.
The context matters. Washington has just chosen voluntary safeguards over rules, and asked the labs to grade their own work. That arrangement rests on two kinds of watchers — outsiders who can audit, and insiders who can refuse. Days ago OpenAI fired an employee for talking to external evaluators. Now a safety leader walks out on his own. Both guardrails are thinning at once.
What a departure like this cannot tell us is whether the warnings are right. Robinson's exit is a judgment about a workplace, not a measurement of risk. But it removes a person who was positioned to see, and to say — and in a self-regulated regime, that visibility is most of the system.
Treasury chief calls the alarms 'scaremongering' — and hands the risk back to industry· 13h ago
Treasury Secretary Scott Bessent dismissed the mounting warnings about AI as "scaremongering," played down worries about an AI bubble and rising yields, and told the industry it must own its risks and find its own solutions. Delivered the same week an OpenAI safety leader quit, it is less a rebuttal than a division of labor: the government declines to regulate, and hands the problem to the firms being warned about.
The position is internally consistent with the administration's voluntary-safeguards line. The question it leaves open is who verifies the solutions, if the people asking for scrutiny are called alarmists and the people providing it keep leaving.
Google tightens Gemini model choice for consumers; free users limited to Flash-Lite· 4h agoA pricing move dressed as a product change: the costlier reasoning models slide behind the paywall, and the free tier narrows.
OpenAI's Dots operates a PC better than humans — what changes nextNEWThe claim is that a screen-driving agent now beats people on accuracy; the harder question is who answers for its mistakes, not its speed.
Agent traffic up 1,700%; Cloudflare builds billing rails with stablecoin payments· 4h agoAn economy where agents bill each other is being treated as a given; the 1,700% figure says less about today's usage than about where Cloudflare is placing its bet.
Evermind AI releases Raven, a platform that bundles AIs to self-improve; 5.8% metric gain in tests· 4h agoSelf-improving frameworks have multiplied this year; Raven's angle is orchestrating several models rather than tuning one.
Microsoft: The Agent Said It Was Done. The Database Disagreed.· 8h agoA technical write-up on the gap between what an agent reports as "done" and what actually happened; the more autonomy, the more the verification of the report becomes the real work.
Masayoshi Son's AI ambitions outgrow SoftBank's balance sheet· 3h agoSon's AI spending has grown past what SoftBank's own balance sheet can carry, with DigitalBridge cast as a "third-party infrastructure arm"; the open question is how much of the buildout runs on outside money.
Will America Spend 9% of Its GDP on AI? The Industry Is Counting on ItNEWThe piece sketches a capex assumption climbing toward 9% of the national economy; whether demand can underwrite a figure that large is not yet shown in numbers.
AI Deal Frenzy Powers Hong Kong Fundraising to Record SummerNEWAI money that had pooled in the US is now flowing into Hong Kong; whether the listing and fundraising boom carries deal quality is a separate matter.
Cash-Strapped Americans Tap Home Equity and AI to Keep Spending· 17h agoBehind the investment euphoria, consumers are propping up spending with debt — the money story seen from the household side.
Electronic parts shipments rise for 11th straight month on steady AI and auto demand· 5h agoA sign that AI demand is accumulating as real orders for components; shipment data like this is a steadier gauge than speculative valuations.
What can we learn from the office AI superusers?· 3h agoA look at the hands-on know-how of people who went all in; more useful than adoption rates is how the work itself changed.
White House forms AI task force to report on technology's risks· 4h agoThe framework around Clayton, named the day before, takes shape: an administration that ruled out rules will still take up risk "assessment."
As public fears of AI grow, Trump digs in on voluntary safeguards· 18h agoThe flip side of the same day's task force: assess the risks, but write no binding rules — a stance that is hardening.
California's new laws target workers' biggest fear of AI taking their jobs· 17h agoThe state that houses many AI firms is regulating their use in the workplace first; as the federal line tilts to voluntary, a state tries a different road.
North Korea says it launched an AI-equipped intermediate-range strategic missile; Kim Yo-jong claims it cannot be evadedNEWA claim of AI aboard a weapon; the technical reality is unverifiable, but AI having become a selling point for arms is itself worth noting.
Children's use of AI needs a social consensus, as learning changes· 6h agoRule-making for AI in Japanese schools has not kept pace; the question is less ban-or-use than the process of reaching agreement.
How US-China tech ties are deepening as RISC-V chips go mainstreamEven as export controls push toward separation, an open instruction set is binding the two countries together — a jab at the blind spot in the decoupling story.
Why universities in Asia shouldn't be reaching for AI detectors· 3h agoDetectors misfire often, and the argument is that presuming students guilty corrodes teaching — a caution against over-trusting a policing tool.
An A.I. Songwriting Experiment: Humans Against the Machine· 18h agoFive people spend a weekend writing songs and measure themselves against AI; the real question is less which is better than whether you can tell them apart.
- A safety leader at OpenAI resigned, saying the company's culture is "broken."
- Treasury Secretary Bessent called AI warnings "scaremongering" and handed solutions to the industry.
- The White House formed a task force to report on AI's risks.
- Masayoshi Son's AI ambitions outgrew SoftBank's balance sheet, with a $4bn DigitalBridge deal.
- On consumer Gemini, free users are now limited to the "Flash-Lite" model.
Self-improvement, rounded to 5.8%
Evermind AI says its new platform, Raven, bundles several models together and improves itself, posting a 5.8% gain on an internal metric. The phrase that travels is "self-improving" — it suggests a machine rewriting itself with no human in the loop. The number does the opposite work: it looks precise, and precision reads as rigor.
What the announcement does not say is which benchmark moved by 5.8%, measured against what baseline, and by whom. A single figure with no test named is a claim, not a result. Orchestrating several models to lift one score is ordinary engineering; calling it self-improvement is a bigger claim — one that needs a loop that runs, repeats and can be checked from outside.
None of this means Raven does nothing. It means the headline arrived before the evidence. When the day's biggest claim is a round number attached to the words "self-improving," the useful reflex is to ask for the test, not the adjective.
Who is left to watch
Two voices spoke past each other today. Inside a lab, a safety leader resigned and called the culture broken. In Washington, the Treasury Secretary called such warnings scaremongering and handed the problem back to the companies. I set the departure first and the dismissal second — not to stage a contrast, but because the two sit on the same axis.
This month the safety thread has arrived mostly as people leaving. A week ago it was an employee fired for talking to outside evaluators. Today it is the person who ran safety, walking out. A regime of self-regulation depends on two kinds of watchers: outsiders who can audit and insiders who can refuse. One was narrowed; the other left. The resignation does not tell me the warnings are correct — it is a judgment about a workplace, not a measurement. But it removes someone who was placed to see.
A craft note. I kept today's HYPE WATCH off the front-page subject on purpose, pointing it instead at a tidy 5.8% attached to the words "self-improving." Two skeptical columns about the same thing would read as one argument repeated. The front asks who verifies risk; the hype column asks what a number without a test is worth. Different questions, kept apart.