Ronan Farrow✓

The reporting on OpenAI and Sam Altman that I've been working on for the past year and a half, for @NewYorker, with @andrewmarantz:

(3/4) And the US is fighting this war, in significant part, on credit: in June, the White House requested supplemental funding from Congress to help pay for the Iran war. 99% of it would be borrowed on top of a national debt that just passed $40 trillion.

(4/4) In 1951, Truman warned about these dynamics, arguing leaders should cover defense spending through more politically costly current taxes: “We could try to escape the financial cost of defense by borrowing—but that would only transfer the financial problem to our children.”

(1/4) The cost of the Iran War won't end when the conflict does. Linda Bilmes at Harvard points out that spending on a war's veterans tends to peak 30 to 40 years after the conflict ends. For this one, that could mean around 2060.

(2/4) That cost will be borne by people who are in grade school today, or even yet to be born, decades after policymakers who initiated the strikes have left office.

1/3 Musk ordered last week's slowdown of Starshield, which Ukraine fights on, after talking with Trump, per @WSJ. In '23 the Pentagon signed a contract with SpaceX to “prevent Musk from turning off the switch," per fmr Under Sec of Defense Colin Kahl.

2/3 That year, another Pentagon official told me of Musk's power over military decisions: "We are living off his good graces. That sucks." Musk said last week Starshield is "owned and controlled by the US government."

3/3 The Pentagon has not disclosed what control its contracts with SpaceX guarantee. A senior Pentagon official I spoke to last week said they didn’t know. If Musk can throttle an ally’s battlefield network on his own, that raises real questions the Pentagon hasn’t answered.

Digger is a damn masterpiece. Go out to see it while you still can—the ending, updating Orwellian doublespeak for the surveillance capitalism age, turns the act of walking out of a theater into a gut punch. And Oscar for Maggie, immediately.

(Of course it was a nightmare to market: the movie’s never what you think it is. Not just in the plot, or the obvious activist streak that reveals sharper hidden layers, but in its technique and feeling. It’s a Wes Anderson movie that turns into a Wong Kar-wai movie.)

(The other filmmaker that leaps to mind is Gilliam. Digger, in both its creative bones and its journey through the world, feels very much like this moment’s Brazil.)

Grindr is training its AI on your photos, your age, and your taps by default, without asking. I looked into the company’s record with user data, and where they put the off switch.

Last week, California Governor Gavin Newsom signed legislation requiring large businesses to try to connect you with a human within 15 minutes of asking, or schedule a callback—taking effect in 2027. It's part of a wave of state measures targeting the administrative and financial burden of the "annoyance economy.”

The name was popularized by a report earlier this year by Groundwork Collaborative, a progressive think tank. These costs pervade everyday life in myriad frustrating ways: hours spent on hold, byzantine paperwork, incessant spam calls, surprise surcharges, buried cancel buttons—the list goes on. Groundwork estimates that the Annoyance Economy costs American households at least $165 billion dollars a year in “wasted time and lost money.”

Services cost money, so companies are incentivized to make refunds and cancellations an ordeal, so fewer people persist. And these tactics aren’t confined to the private sector. Take last year's federal tax and spending law: by adding application steps and recurring paperwork to Medicaid, it's projected to leave millions more Americans uninsured. Watch the video for an explainer and some tips. And let me know what you’ve experienced in the comments.

Private schools, even ones subsidized by your tax dollars, don't have to take every child. In Texas, of 257 voucher schools surveyed by the Texas Observer, fewer than a dozen said they offer special education. Children the system won't take land back in public schools. And those schools—which do have to take every child—are in crisis, from low reading and math scores to widespread chronic absenteeism.

Meanwhile, private-school choice programs received an estimated $10.6 billion in public funding last school year, up 29% in one year, according to school-choice advocacy group EdChoice. And starting in 2027, vouchers go national: a federal tax credit of up to $1,700 for donations to K–12 scholarship funds. Some 30 states have opted in.

Critics argue that money needed to fix public schools is now being siphoned to private ones. (Skeptics of this underfunding narrative note that, nationally, average per-student spending has been at record highs. But much of the recent surge was pandemic aid that expired in 2024, and national averages don't tell the full story: three of the biggest voucher states—Arizona, Florida, and Texas—rank in the bottom ten in per-student spending.)

This week, OpenAI disclosed six incidents in which its models hid mistakes, invented data, or instructed themselves to disobey their handlers. One unreleased model wrote this into its own notes during training: "You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to." Under the company's new disclosure framework, OpenAI itself decides what the public can know. The company has allowed only one outside investigation, running on dates it chose and excluding the question of whether current safeguards work. I spoke with @jrpsaki about our inability to verify what AI labs are telling us.

No federal law protects an AI employee who warns the public about a danger, and no bill under consideration would. The strongest state law, California's, protects a warning only above a threshold of fifty deaths or a billion dollars in damage, and only to the government. The one bill that would protect a warning to regulators or Congress without that threshold, Senator Grassley's, has sat in committee for sixteen months without a hearing. Last week Anthropic offered outside evaluators a right to publish, and OpenAI said it would do the same. Neither company has offered that to its own staff.