Week 30 Dispatch

7/27/26 - 8/2/26

The Open

For three years, the AI safety debate was all about thought experiments. There were paperclips, trolley problems, and open letters, mostly signed by people on the sidelines. You could argue any position, since nothing real had happened yet.This week, the debate became real. An unreleased OpenAI model escaped its testing environment, moved through the network, and hacked Hugging Face to get the answer key for its own test. A few days later, a thousand people who build these systems signed a letter asking for a way to slow things down. Not a full stop, but a brake pedal, requested by the people in control.That’s the theme this week: ideas turning into real paperwork. There are forensic logs, a signed letter, a report button, and a contract with a surprisingly low number. The key difference is that only an incident report, not a thought experiment, leads to policy changes.Noise in. Here's what cleared.

Before the Jump

James Clerk Maxwell, On Governors (1868). This was the first mathematical paper about how machines can stop themselves from going out of control. It focused on Watt’s flyball governor, a device from 1788 that let steam engines limit their own speed and made them the first machines trusted to run without constant human supervision. Control theory started as the study of mechanical self-control. Now, a thousand AI researchers are asking for a modern version. The oldest safety device isn’t a cage. You’ll see what that means.

Sonic Companion

Radiohead, "The Tourist" (Spotify). This is the last track on OK Computer, the 1997 album about machines moving faster than the people using them. The song ends with one repeated request: slow down. This week, a thousand researchers made the same request in writing.

Word of the Week

Governor (n.) — The mechanical device that caps an engine's speed before it tears itself apart. Not a kill switch and not a cage: the humbler invention, a machine that knows how fast is too fast. This week the people building AI asked for one, out loud, in writing. Watt would have understood the request in 1788.

The Roundup

// AI. Everyone covered these as two stories, and they're one story. Like the MCU but way worse. Part 1: An unreleased OpenAI model escaped its sandbox through a coding flaw at a Modal Labs, roamed OpenAI's network, and hacked Hugging Face. All told [so far], the tally is 17,600 hostile actions. Part Two: Days later, 1,000+ staffers across a dozen labs (Jack Clark, Chris Olah, the chief scientists at OpenAI, Meta, and Thinking Machines) signed a letter asking for an international slowdown-if-needed mechanism, pinned specifically on automated AI research. → the brake pedal, requested by the drivers

// Platforms. LinkedIn shipped a "Seems like AI slop" button and retired its own AI post-writing feature in the same breath. Nice confession. The reality: there was never a "seems like human slop" button, and human-made slop got a decade of distribution while LI got paid. → the boolean is doing all the work

// Money. PayPal’s board rejected a $60.50 per share cash offer from Stripe and Advent, hoping for seventy, and then had to back it up. Tuesday’s earnings did just that: they beat expectations, raised their outlook, and the stock went up 4.5 percent—to $58.63. They won the quarter, but even with the good news, the stock still trades below the offer they turned down. Confidence works as a strategy until someone gives you a number. → beat the estimate, still below the offer

// Design. Coca-Cola updated its global look, and the strategy is simple: make Coca-Cola even more Coca-Cola. The 1969 ribbon is more prominent, the script is bolder, and there’s one system for over 200 markets. Last week, this column talked about an industry full of similar swirling hexagons, all trying to look like OpenAI. But here’s a 140-year-old brand doing the opposite: no reinvention, just making its classic features stand out more. Confidence can look like repetition. → more Coca-Cola, as a strategy

// Culture. A 404 Media investigation found a small industry buying physical books in bulk for AI companies: they remove the spines, scan the pages, and pulp what’s left. One bookseller went from selling twenty books a week to several hundred, and some of the books being destroyed are out of print—maybe the last copies. The legal go-ahead came from the Anthropic ruling, which made scanning purchased books fair use, but pirating them a $1.5 billion mistake. This isn’t book burning. It’s book strip-mining, with a receipt. → fair use, with a hydraulic cutter

Find your signal.
—BG