Bridgewater fine-tunes Qwen3-235B with Tinker, beats GPT-5.5 at 1/14th the cost
simonguozirui · x · 2026-10-16
frank shared Thinking Machines' Tinker case studies with a thesis: "AI labs are the new startups"—twenty years ago every company became a software company; twenty years from now every company will train its own models in particular ways.
Highlights from Tinker's enterprise fine-tuning cases:
- Bridgewater: fine-tuned Qwen3-235B for financial information filtering, averaging 84.66% accuracy across six held-out tasks vs GPT-5.5's 78.2%—roughly 30% fewer errors at one-fourteenth the inference cost.
- Lightning Rod Labs: fine-tuned gpt-oss-120b to predict clinical events (dialysis, transfusion, ventilation) from patient notes, cutting prediction error by 27% and calibration error by 69%, approaching GPT-5 performance.
The broader argument: as general intelligence spreads, software becomes the shared foundation and what gets built on top is "judgment"—companies will own models trained on their own taste and domain expertise, and future tools will look less like tools and more like us.
More from Companies & People
- ai& × AMD Build Day hackathon in Tokyo: 8 teams, 4.5 hours, Chinese models allowed — DavidBennett__ · 2026-10-16
- Sakana AI opens first dedicated ISV, SI and CS roles in Japan GTM push — SakanaAILabs · 2026-10-16
- Kaggle teams with Google DeepMind and Zapier for inaugural Bench Summit — jocarrasqueira · 2026-10-16
- Ethan Mollick jokes people take his shipping-timeline tweet too seriously — emollick · 2026-10-16
- Mantic AI cofounders on building AI that forecasts world events better than top humans — tshevl · 2026-10-16
- Anthropic readies work-email discount path to annual Claude Pro, checkout spotted — ryanmerket · 2026-10-16