Week 37 Dispatch — Fraudulent Accounts, Samuel Slater, and the Humanisers

09/07/26 – 09/13/26

The Open

Somewhere this year, a developer paid a Chinese AI lab for access to its model, typed a prompt, and got back an answer from Claude. The lab kept the transcript and trained on it. Anthropic put that in writing this week, in a threat report that names seven labs and thousands of fraudulent accounts.

Nobody waited for the ethics to resolve. AT&T measured a 2% quality drop against a 56% cost cut and took the deal. Uber has held total AI spend flat since March. Cognition shipped a coding model that claims frontier parity at 64% less, built on a base model from one of the labs in the report.

I kept reading these as theft stories but I'm not 100% sure that is the case.

Related: Why You're All Suddenly Talking About Human Judgment — what gets scarce when the output stops being the hard part.

Over and over this week, somebody looked at a copy, measured the gap, and decided the gap didn't matter. Two percent worse. Sixty-four percent cheaper. Three years of experience in fourteen months.

Noise in. Here's what cleared.

Find your signal.
— BG

Before the Jump

On September 13, 1789, 237 years ago today, a 21-year-old named Samuel Slater boarded a ship in London and told the authorities he was a farm laborer. He wasn't. He had spent his apprenticeship inside Jedediah Strutt's mill learning the Arkwright water frame, and Britain had banned the export of both the machinery and the men who understood it. Slater carried no drawings and no models. He carried the weights in his head. Within a year he was running a mill in Pawtucket, and Derbyshire was calling him Slater the Traitor, while Andrew Jackson called him the father of American industry.

Britain successfully embargoed the hardware and lost anyway. Funny innit?

Sonic Companion

Talking Heads, "Once in a Lifetime" (Spotify). And you may ask yourself, well, how did I get here is the correct reaction to a $517B compute commitment financed against customers actively cutting their bills. This is not my beautiful house is the correct reaction to the LG story below. Same as it ever was.

Word of the Week

Appropriability (n.): how much of the value of your own innovation you actually get to keep. David Teece named the problem in 1986 and his point was brutal. The inventor often loses to the imitator, and whether you win has almost nothing to do with whether your thing was good. It depends on whether the knowledge can walk out the door, and timing. And timing is hell of a drug. Worth knowing because a model whose entire product surface is its output has close to the weakest appropriability regime ever built. Every answer it sells is a free training sample for the competition.

The Roundup

// AI. The story from the open. Anthropic's threat report names Alibaba, DeepSeek, Moonshot, and Xiaomi among seven Chinese labs running distillation through thousands of fraudulent accounts, with Moonshot and DeepSeek in some cases serving Claude to their own paying customers as their own model and training on what came back. I am shocked. Shocked I tell you → 150 pages of own goal.

// Money. Two public companies and one startup just repriced the frontier. AT&T: 56% cheaper through model routing, 2% quality drop, 100,000 employees. Uber: cost per session down 52%, total spend flat since March despite rising usage. Cognition's new SWE-2 claims parity with Fable 5.1 and Astra on some coding benchmarks at 64% less, and it's built on Kimi K3, which belongs to Moonshot, which is on Anthropic's list above. Follow that chain slowly. The cheap model eating the frontier's enterprise business may be descended from the frontier→ 2% worse, 56% cheaper.

// Talent. Anthropic's economists modeled the 2030 economy: in the extreme scenario GDP hits $44.4T, labor's share of every dollar falls from 60¢ to 45¢, knowledge-worker wages drop more than 10%, and coders get advised to retrain as electricians. In the same week, Goldman, Morgan Stanley, and Citi told their law firms to cut fees because AI now does the document review, Citi's global head of legal wants costs down per transaction, and nearly half of large firms say AI has already moved their pricing → the hour stops being the unit.

This is When Time and Materials Go to Zero, happening on the record.

// Ideas. Thompson's argument against AI watermarking is the sharpest thing written this week. Mandating disclosure assumes the model is an author who must identify itself, which quietly concedes that the human who directed it is not. A ballpoint pen does not sign your letter. If your defense of your own work is a label certifying a machine didn't touch it, you have already accepted the machine's claim→ the pen doesn't get a byline.

// Platforms. A Gamers Nexus teardown alleges LG smart TVs capture microphone audio while the screen is off, transcribe it to plain text on-device, scan the home network for other devices, and ship it to LG's ad division, across roughly 216 million units in the US. LG says audio is only recorded on the wake word or the remote button. Both things can't be true, and the interesting part is that nobody needs the scandal resolved to know how this ends. The TV was never the product→ the standby light is on.

// Talent. The best AI deployment story of the week is a 23-year-old junior PM who built a Claude project trained on ~100 transcripts of his manager's feedback, with a Coach Mode that refuses to produce anything until he can defend the thinking. Ask it to just write the PRD and it says no. His manager clocked roughly three years of development in fourteen months. Everyone is asking whether AI eliminates the junior role. Wrong question. The constraint was never junior labor, it was senior attention, and that is the thing we just learned to copy→ the coach that says no. See also: Corporate Children of the Magenta Line.

// Culture. And here's the same trade without the happy ending. Roughly 40,000 people in Nairobi made a living writing essays for cheating students overseas. ChatGPT took that industry down in about two years. The work that's left is "humanising", getting paid to rewrite AI output until it reads human enough to beat the detectors. Every incentive in the watermarking fight above is visible here in its final form: the machine writes, the human launders, and the human is the cheaper input→ the last job in Nairobi.