The Collection

Marcus opens with rare praise. Amodei’s 3,500-word essay may prove consequential. Altman and Musk endorsed pacing. Not one of the major labs, Marcus notes, seems to have good control of the technology. He wants the agreement signed in blood, with teeth, and with government willing to enforce something substantive. Transparency is the part he loves without caveat. It is also the part researchers such as Deb Raji, Meg Mitchell, and Timnit Gebru have been demanding for years.

Then the third cheer fails. Marcus calls out hyperbolic doom and cheer in the essay’s opening, the ladder-pulling suspicion shared by David Sacks and François Chollet, and Angela Rasmussen’s reading that “partner with government” can be a way to keep influencing the process while appearing pro-regulation. The proposal for independent evaluators points to METR, which Marcus and others treat as a Valley creature too close to the labs. Gillian Hadfield warns that any evaluator embedded too long can get swept up. Anthropic should not choose who grades Anthropic. China appears as both rival and excuse: scare enough about distillation and the slowdown dies the moment Beijing will not join. Marcus wants liability and even criminal charges for negligent agents, and product recalls when control fails. Agents are token-hungry. Liability would change the math. Trump, in a morning postscript Marcus flags, opposed a slowdown altogether and framed AI as a race China must not win.

Two cheers. Keep the third for a court.

Praise the essay if it forces daylight. Do not confuse a lab’s preferred auditor with an outside one.

Schwartz’s exclusive, visible in this inbox as a teaser, reports that Anthropic, OpenAI, and Google have been holding discussions about creating a standards body for the AI industry. The talks began, people familiar with the matter said, even before Amodei’s Saturday call for coordinated testing and auditing. Altman, at a companywide town hall earlier in the week, told staff he supported a testing and auditing organization, but believed the major labs would have to create that body on their own without U.S. government support.

The rest of the article sits behind The Information’s paywall in this mail. We label it unread rather than invent the membership list, timeline, or legal shape. What the free front door already says is enough for Monday’s claim: the slowdown essay did not invent the standards conversation. The labs were already shopping a private body. Altman’s reported stance is the sharp edge. Support testing. Assume Washington will not build the referee. Build one yourselves.

The body was meeting before the sermon.

Read the teaser as news. Treat the locked pages as locked. Do not fill them with guesses.

Azhar writes from a week that felt like March 2020 for AI assumptions. In Las Vegas he asked 250 IT executives whether they had serious results from AI initiatives. A year ago about a quarter of hands went up. This year he estimates roughly ninety-five percent. Almost all plan to spend more next year. Bloomberg, in his letter, has Microsoft planning AI capacity from about 2 GW today toward nearly 13 GW by 2032 inside a larger fleet move from 12 GW to 38 GW. Anthropic published economic scenarios; Azhar’s own models sit near the “substantial” case of about 8.3 percent added to U.S. GDP by 2030, with growth shifting from labor toward capital.

Sunday already held the fallow ground of Millennium compression. Azhar notes the same breakthrough and the same Tao warning, then turns to secrecy: OpenAI kept Astra dark for six months; the model behind the new proof work is newer still and almost unseen outside. Erik Hoel calls the research climate a dark forest. On safety, Azhar reminds readers that Clark and Amodei wrote about reinforcement learning risk in 2016. Now the same firms are oligopolistic powers proposing to collaborate on a slowdown and on independent evaluators with employee-level access. Walking past Adam Smith’s grave in Edinburgh, he quotes The Wealth of Nations: people of the same trade seldom meet together except in a conspiracy against the public, or in some contrivance to raise prices. The Collection line for Monday: demand can be real and the meeting still need an outsider in the room.

Spend more. Slow down. Pick your own auditor.

Believe the raised hands. Distrust a slowdown that never leaves the trade association.

CJ opens with Cousin Greg from Succession. Five million, Connor says, is a nightmare: not enough to retire, not worth working. Tom calls it the poorest rich person in America. CJ maps the joke onto private SaaS growing around twenty percent without being exceedingly profitable. You are growing. You have product-market fit. You may sit at hundreds of millions in revenue. In an AI-roiled market you are still the weakest strong man at the circus. Below about thirty percent growth, Gokul’s point in the letter, you lose the right to trade on a revenue multiple; profits, or the lack of them, start to set the terminal story.

Thursday’s news: Bending Spoons acquires Miro for $1.355 billion. Miro had raised at something like $17 billion. CJ puts the exit around 3× current ARR after more than $400 million of cash on the balance sheet, and assumes growth closer to twenty percent than fifty. He does not claim to know Miro’s exact rate. He insists on the larger lesson: what you raise at is not what you sell at. A minority round prices what can go right. A full sale prices what can go wrong for the buyer who must run the thing. Miro joins Airtable in the Bending Spoons aisle of big, loved products whose growth no longer supports the valuations they grew up with. Commenters nominate other late-stage names; CJ refuses to call them bad companies. Wrong window. Wrong moon. The Collection line beside the stopwatch: the frontier labs argue about pacing; the mature tools get repriced whether or not anyone slowed the models.

Sell early and still be right. Sell late and explain.

Keep the Succession joke. Apply it to growth rates, not just trust funds.