Bridgewater and Thinking Machines Build Custom AI to Slash Errors by 30%

Bridgewater Associates, the world's largest hedge fund, has moved beyond standard AI subscriptions to build a custom solution. The firm partnered with Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, to create a specialized model that reduces errors by nearly 30%.

Published results show the model achieving 84.7% average accuracy across six critical information-filtering tasks, significantly outperforming standard frontier models like GPT, Claude, and Gemini, which remained in the mid-70% range even with expert prompting.

The system was built on the Qwen3-235B base using Thinking Machines' proprietary Tinker platform. Key training innovations delivered major gains. The specialized model focuses on document triage-relevancy classification, truncation, and labeling-the essential steps before financial analysis begins. The critical differentiator was training on Bridgewater's proprietary, expert-labeled data, teaching the model distinctions that generic systems cannot learn through prompting alone.

Bridgewater manages roughly $100 billion in assets. For the AI industry, the partnership validates Murati's post-OpenAI thesis that the next wave of value lies in deep customization, not just scaling.