Mentatcurated
Artificial Intelligence medium · first-party

The answers were never public

Bridgewater tuned a downloadable Chinese model to sort its financial documents better than Claude, GPT and Gemini, and it won because the right answers only ever existed inside the firm.

Six unglamorous jobs: decide whether a news article is worth an analyst's time, whether a central bank document is relevant, where the boilerplate ends in a forwarded email. On those, a model Bridgewater fine-tuned for itself was right 84.7% of the time. Claude Opus managed 78.2%, GPT-5.5 78.0%, Gemini 74.3%. Running the tuned model cost about $10 per thousand documents against $80 to $100 for the frontier services. Bridgewater's own AI lab did the training, on infrastructure rented from Thinking Machines Lab, which published the results.

Nothing here picks a trade. The win is in deciding what an analyst should bother reading.

The frontier models did not lose for want of intelligence. They lost because nobody has ever written down what Bridgewater considers relevant. Whether a given paragraph of Fed minutes matters to a macro investor is a taste the firm spent decades forming and never published, so it sits in no dataset any lab has scraped. Fine-tuning is how that taste gets transcribed. The firm put its own experts on every case where the model and the human labellers disagreed, and their rulings became the training signal.

That inverts the usual scarcity story. The base model was an Alibaba release anyone can download for nothing; the scarce asset was a written record of the firm's own judgment, and that record is what converts into a system running at a fourteenth of the price. The numbers are Bridgewater's own, measured on a private test set it built and graded, so read the margin as a claim rather than a settled result. The recipe is open to anyone sitting on expertise that has never been published and willing to spend the months writing it down, which describes a good deal of finance, law and medicine.

The lenses

Novelty 3
Impact · breadth 3
Impact · depth 4
Actionable 3
Delight 3
Substance 4
Hype 4

The facts

What it doesDecides which articles, filings and emails an analyst should read, and where the boilerplate ends
Accuracy84.7%, against 78.2% for the best frontier model tested
Cost to runAbout $10 per 1,000 documents, versus $80 to $100 for the commercial services
Checked byNobody outside Bridgewater: private test set, in-house experts, no model released
Open thinkingmachines.ai →

How this connects

Tap a node to open it

How this connects