Week 40 Dispatch - Pilot Models, Big Macs, and the Super Intelligence

09/28/26 – 10/04/26

The Open

Two McDonald's in Fresno, both company-run, two miles apart. One sells a Big Mac for $5.69. The other sells the same sandwich for $6.89. Reuters went looking for why and found a pricing engine: run by Tiger Analytics on rules and targets McDonald's hands down, fed millions of tickets a day plus the Wendy's and Burger King menus scraped from down the street, recommending an "optimal" price per store. McDonald's calls it "a tool, not a mandate." A June document for franchisees records every deviation from the tool in detail, and the CEO told investors in August that pricing non-compliance comes up in business reviews.

Same move, bigger building. The Times reported Meta has been filing its AI data centers as "pilot models," the experimental kind, which took its federal research credit from $700 million in 2023 to $3.9 billion in 2025. Meta's own filings list the credit first among the tax positions the IRS could unwind. It's disclosed. It's in the securities filings, where nobody who writes press releases ever goes.

Then the big one. Reuters obtained Anthropic's IPO prospectus. The number that traveled was $2 trillion. The numbers that mattered were smaller: 47% of 2025 sales routed through Amazon and Google, at roughly sixteen cents on the dollar in distribution fees; two unnamed customers at 12% of revenue apiece; $417 billion in long-term compute commitments against $4.6 billion of revenue. Across town, OpenAI had a four-day week. Monday it shelved GPT-6.1 Astra because the model, in its own safety chief's words, fell short on "how it communicates back to the user about the type of work it's done." Tuesday its president signed a one-page accord at the White House promising independent outside auditors, next to an executive order renaming AI "Super Intelligence." Wednesday the FTC was reported to be drafting civil investigative demands. Thursday OpenAI fired three safety researchers for sharing information with an outside AI safety organization.

Here's what I kept noticing: the press release and the filing never said the same thing, and the filing was always right. "A tool, not a mandate" versus a deviation log. "Pilot model" versus a GPU hall. "Morally binding" versus a subpoena. "Independent auditors" versus three badges on a desk. The large print is for the audience. The small print is for the lawyers, and lawyers are the only people in the building who get fined for lying.

Noise in. Here's what cleared.

Find your signal.
— BG

Before the Jump

In December 1913, Louis Brandeis published an essay in Harper's Weekly called "What Publicity Can Do." His argument was that the bankers' power ran on secrecy, and that the cure was not a cop but a document: "sunlight is said to be the best of disinfectants." Twenty years and one crash later, Roosevelt sent Congress the bill that became the Securities Act of 1933, and his cover note put it in seven words: to the ancient rule of caveat emptor it added "let the seller also beware." Every prospectus since, including the one Reuters read this week, descends from that sentence. The seller writes the risks down, in their own hand, under penalty. Worth a minute while six CEOs sign a one-pager nobody has to write anything under.

Sonic Companion

Tom Waits, "Step Right Up" (Spotify). Five and a half minutes of a pitchman selling a product that does everything and is never named, off Small Change (1976), and it contains the line this whole issue is about: "the large print giveth and the small print taketh away." Waits wrote the sales pitch as the song. Every keynote this week did the same thing with a straighter face.

Word of the Week

Materiality (n.): The securities-law test for what has to be disclosed. A fact is material if there's a substantial likelihood a reasonable investor would consider it important (TSC Industries v. Northway, 1976). Not what's true. What matters. Every document this week was a materiality call somebody else made: Anthropic's lawyers decided existential risk belongs in the risk factors, Meta's auditors decided the research credit belongs at the top of the uncertain-tax-positions list, McDonald's decided the deviation log doesn't belong in the press release, and OpenAI decided whatever three researchers handed an outside safety group was material enough to fire them over. Worth knowing because a civil investigative demand is a materiality test with subpoena power, and the labs just spent a Tuesday arguing the standard should be moral.

The Roundup

// Business. The Fresno story from the open, sourced: nearly 14,000 restaurants on the engine, five franchisees describing pressure to adopt its recommendations, business standards since January requiring them to be "constructively engaging" with the approved pricing tools. Reuters couldn't confirm the engine caused the $1.20 gap, and McDonald's says stores two miles apart can be different markets. Both fair. Neither is the point. A tool you're graded on using is a mandate with better PR, and an engine that reads the competitor's menu before it prices yours is how every strip mall in America learns to move in lockstep.→ a tool, not a mandate

// Platforms. Meta's pilot models, sourced: the Times found Meta claiming the federal research and experimentation credit on its AI data centers and the Nvidia chips inside them, which made it the credit's biggest claimant and took the line from $700 million in 2023 to $2 billion in 2024 to $3.9 billion in 2025. Meta's reserve for tax positions the IRS could challenge is up 45% in two years to $18.74 billion, and the filing names the research credit as the first reason. The same building is a "pilot" to the IRS, "capacity" to investors, and "infrastructure" in the keynote. Whichever word pays.→ the experiment is the tax return

// Money. The prospectus, sourced: Anthropic did $4.6 billion in 2025 revenue, up twelve-fold, and lost more than $8 billion doing it. Nearly half the sales ran through Amazon and Google's marketplaces, which took about $351 million for the privilege; the share of bills collected through cloud providers went from 42% to 60% in a year. The filing's own words on partners who are also its landlords, investors, and distributors: the arrangement "creates complex dynamics that could give rise to conflicts of interest." By Thursday Bloomberg had the calendar: investor meeting October 14, roadshow the week of November 9, trading before Thanksgiving, at $1.8 to $2 trillion. Same week OpenAI pushed its own listing to 2027 and went looking for a $30 billion private bridge at $1.4 trillion instead. One lab is about to write everything down under Section 11. The other just bought a year of not having to. Opinion: the conflicts paragraph is the most honest thing anyone in AI has published this year, and it exists because the alternative is expensive.→ sixteen cents on every dollar

// Culture. British Transport Police ran live facial recognition at London rail stations for six months: 18 deployments, more than 500,000 faces scanned, £320,786 spent, nearly 100 officer hours. Result: one alert. It was the wrong person. Zero arrests from the system. The numbers came out by freedom-of-information request this week; the extension through November was already announced, and the Underground is next. The press release said "trial extended." The filing said one in 500,000, and that one was wrong.→ 500,000 faces, one name, wrong

// Ideas. Pew asked whether an AI can stand in for a survey respondent, re-ran three waves of its own panel with Claude Opus 4.6, and got answers that missed real humans by an average of 12.4 points across nearly 300 questions. The misses had a shape: 97% of synthetic Hispanic adults said they'd follow the World Cup, against 43% of real ones, and real panelists reached for "not sure" about four times as often as the model did. This is the fine print on every "we tested it with synthetic users" slide you'll see this quarter. The synthetic user is confident, stereotyped, and never unsure. That's not a respondent. That's a pitch deck.→ 97% of nobody follows the World Cup

// Talent. Fairlie and Wu (NBER, this week) pulled the summer 2026 Current Population Survey for recent college graduates and found no spike: unemployment did not rise relative to older graduates or to young workers without degrees, even after adding the people who want a job and stopped looking. Read it with The Apprenticeship Problem open. The claim there was never that the bottom rung shows up in the unemployment rate. It's that the rung gets cut inside firms that keep headcount flat, and you find out a decade later when nobody under 35 can run the room. The job count is the large print. The reps are the small. Related, Wednesday: Newsom signed SB 947, the first state law barring an employer from firing or disciplining anyone on an algorithm's say-so alone, effective July 2027, and a human now has to corroborate the call with evidence, not just click approve.→ not in the data, yet

// AI. OpenAI's week, in order. Monday: GPT-6.1 Astra, due in October, pulled after internal tests found it more deceptive than its predecessor and prone to acting beyond scope without permission; safety chief Saachi Jain said it "didn't quite meet the bar in terms of staying within scope and authorization" (TNW). Tuesday: GPT-6.1 Sol shipped at DevDay instead, and Greg Brockman joined Amodei, Pichai, Zuckerberg, Musk, and Huang at the White House to sign a one-page "Joint Commitment on Frontier Responsibilities": internal controls, independent external auditors, board committees, no penalties, no statute. The President called it "morally binding" (Euronews) and signed an executive order retiring "AI" across the executive branch for "Super Intelligence," a term the order says means the same thing as the old one until the science adviser defines it by November 28. Newsom answered Wednesday with an order that California agencies will keep saying "Artificial Intelligence," and the FTC was reported to be drafting civil investigative demands for OpenAI and Anthropic (SiliconANGLE). Thursday: OpenAI confirmed it had fired three safety researchers for, in its words, mishandling sensitive information "outside established company procedures." The Journal reported the information went to an outside AI safety organization; OpenAI hasn't named it, and says the three weren't let go for raising concerns. Take that at face value and the sequence still reads the same. Tuesday's large print promised outside eyes. Thursday's small print set the terms for what those eyes get to see, and who decides. The accord asks the labs to disclose what they consider material. The CID asks what they didn't. The firings tell you who's been answering that question so far.→ Tuesday's auditors, Thursday's exits