Bridgewater fine-tuned Qwen3-235B to 84.7%, beating frontier models at 1/14 the cost
anacondainc · x · 2026-09-02
A case study from Bridgewater's AIA Labs (with Thinking Machines Lab) shows frontier models scored under 50% on six investor information-triage tasks, rising to 78.2% with expert-written instructions—still short of their 80% bar. Fine-tuning open-weight Qwen3-235B (44.8% out of the box) on their own expert-labeled data reached 84.7%, 30% fewer errors than the best frontier model at roughly one-fourteenth the inference cost. The moat: trustworthy labeling process, not the model.
More from Companies & People
- SGLang Summit sponsors grow from 11 to 36 in one week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-18
- OpenAI reportedly handled ~15T tokens daily in April, likely higher now — JoshPurtell · 2026-10-18
- Yacine MTB: Users will remember how Anthropic treated them once intelligence proliferates — yacineMTB · 2026-10-18
- Periodic's CoLM talk: when does training your own model beat frontier + harness? — heghbalz · 2026-10-18
- After GHC Cancellation, GitHub and Cloudflare Offer 175 Free Conference Passes — michellechen · 2026-10-18
- Ex-Apple Siri Engineer Builds 7 iOS Apps Solo, Hits $6.2K MRR — jdluk87 · 2026-10-18