The Turk Retires

The Expert Steps Out of the Box

America turned 250 on Saturday, and before the week's business, a word about the birthday — because it is the week's business. Everything this letter covers, every day, exists because the US still runs the freest market and the deepest capital markets on earth, and because a nation of immigrants keeps pulling in people who want to build. The man who has put more satellites into orbit than every government in history combined grew up in Pretoria. Microsoft and Google are run by men born in India. OpenAI is run by an openly gay man from St. Louis who spent the Fourth calling America "the most impressive social experiment in history." The press sells decline. The tape shows a country that keeps repricing itself and coming out ahead. Happy birthday.

And right on schedule, it repriced something else this week.

In 1770 — six years before the Declaration — a Hungarian engineer unveiled the Mechanical Turk, a chess automaton that toured Europe beating Napoleon and Benjamin Franklin. It was a hoax. A chess master was folded inside the cabinet. In 2005, Bezos rebuilt the trick as a business: Amazon Mechanical Turk, thousands of anonymous humans behind an API doing penny-a-click judgment. His own name for it was "artificial artificial intelligence." Last week AWS moved MTurk to its maintenance list; July 30, it closes to new customers. Twenty-one years in, the machine with people hidden inside is done. The crowd that labeled the first AI era just got let out of the box.

That would be an obituary, except the other end of the judgment market moved the same week. Bridgewater needed 80% accuracy on document filtering before its investors would trust a model. Naive prompts scored a coin flip. Expert prompts, 78%. Still short. So they fine-tuned an open Chinese model — Qwen, on Thinking Machines' Tinker — on decades of their own senior investors' decisions: 84.7%, roughly 30% fewer errors than the best frontier model, at 1/14th the cost. When training labels looked wrong, the fix was routing disagreements back to the senior investors. The alpha was in the filing cabinet the whole time. Anthropic's own 400,000-session study backs the pattern: people who know their field verify-succeed within five points of professional engineers, and management and sales are the fastest-growing users. Stanford's payroll data closes the loop — junior developers down 19%, the over-40 cohorts up double digits.

One repricing, two directions. Anonymous, crowd-scale judgment went to zero. Named, accountable, domain-expert judgment just became the scarcest input in AI.

Here's the call: smart open models plus true human experts win, and the frontier labs know it. Watch them enter the open-model game — and partner with domain experts to train frontier capability on human expertise. Nadella already tipped it, launching Frontier Co. with the line "human capital and token capital compound." He's describing Bridgewater's trade, sold as a product.

For 84 years there was always an expert in the box. There still is.

The expert was always in the box. Put yours to work.

At COAI today: the full Signal/Noise — the Turk's 250-year con, the Bridgewater math, and the expert-repricing playbook — is live at getcoai.com.

Outsider Labs Advisory →

250 years — America's age as of Saturday, and the number that explains the week. The same free market that has repriced everything from whale oil to seat licenses just repriced human judgment: Amazon retired Mechanical Turk, the anonymous penny-a-click crowd that trained the AI era, the same week Bridgewater's senior investors trained an open model that beat every frontier lab at their own job, at 1/14th the cost. The crowd went to zero. The expert became the scarcest asset in AI. The Turk retires; the expert steps out.

Inventory your filing cabinet. Bridgewater's edge wasn't a model — it was decades of documented senior-investor decisions, fine-tuned into 84.7% accuracy at 1/14th frontier cost. Today: name the three archives of expert judgment your firm owns and nobody else does. Credit memos, QC logs, pricing calls, claim decisions. That list is the newest asset on your balance sheet.

Put your domain experts in front of agents — stop routing everything through engineering. Anthropic's 235,000-user study says people who know the field succeed within five points of professional coders, and management and sales are the fastest-growing users. Today: hand your best non-technical expert one agent and one real task. The bottleneck was never the syntax.

Price one fine-tune pilot against your frontier bill. Naive frontier prompts scored 50% on Bridgewater's task — their tuned open model cleared the 80% trust bar. Pick a single narrow, high-volume judgment task, run an open-model fine-tune on your own decision archive, and put the per-token math next to your current invoice. One task. Real numbers. This week's renewal leverage — today's setup call.

Monday, July 6

Where the cycle caught up to us this week.

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: