A Smaller, Custom AI Beat GPT and Claude at a Real Job — for a Fraction of the Cost
A giant hedge fund and a top AI startup showed that a free, downloadable model — trained on their own experts' judgment — outperformed every big-name AI on finance tasks, at roughly one-fourteenth the cost. The lesson applies to far more than Wall Street.
Bigger isn’t always better in AI — and here’s a concrete example. Bridgewater, one of the world’s largest hedge funds, teamed up with Thinking Machines Lab (a startup founded by former OpenAI CTO Mira Murati) and showed that a smaller, customized model beat every big-name AI they tested at a real financial job — while costing about 14 times less to run.
The task was the unglamorous, everyday work of an investment analyst: does this news article matter to us? Does this central-bank statement hint at future interest-rate moves? Simple calls for a human expert — and, it turns out, hard for general AI. Out of the box, versions of Gemini, Claude, and GPT scored around 50% accuracy. Careful expert-written instructions pushed them into the mid-70s, still short of the 80% the team wanted before trusting the output. Notably, newer and pricier models barely helped.
The fix was fine-tuning — the process of taking an existing open model and further training it on examples from your own experts, so it learns your specific judgment rather than generic knowledge. They built on Alibaba’s freely available Qwen3 model, trained it on how their analysts actually make these calls, and reached 84.7% accuracy — beating every frontier model. A clever cost-saver: instead of paying experts to label everything, they had a first model do a rough pass, then sent only the disagreements to humans. Think of it as teaching a bright new hire using your best employees’ decisions, rather than hoping a brilliant generalist guesses your house style.
One honest caveat: these results come from the two companies’ own evaluation, and both sell related products — no independent party has verified the numbers. But the direction lines up with what other teams keep finding: general-purpose AI hits a ceiling on tasks that hinge on specialized, in-house judgment.
What this means for you: Your business’s most valuable data is probably exactly the stuff you’d never paste into a public chatbot — and this is a sign that same data could power a smaller, private AI that beats the giants at your specific task. You don’t need a hedge fund’s budget: the pattern is a modest open model plus a good set of labeled examples from your own experts. Fine-tuning also means you keep the model, the data, and — if you want — run it on your own hardware, where nothing leaves the building.
Sources
Source: https://thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks/
A Vaccine Designed by AI Just Passed Its First Human Safety Test
For the first time, a vaccine ingredient designed entirely by AI has been tested in people — and cleared the first safety hurdle. It's an early step, not a finished product, but it's one of the clearest examples yet of AI doing real scientific work.
