Bridgewater and Thinking Machines Fine Tune AI That Beats Frontier Models on Financial Documents
Bridgewater Associates and AI startup Thinking Machines have fine-tuned a custom language model that outperforms frontier systems in classifying financial documents, according to a case study released Tuesday. Bridgewater, one of the world’s largest hedge funds, partnered with Thinking Machines to develop the model on Tinker, an open platform co-founded by OpenAI and NVIDIA veterans Mira Murati and Soumith Chintala that lets developers distill and run AI models efficiently.
The hedge fund supplied its own institutional knowledge and a dataset labeled by financial experts to train the model using on-policy distillation, a technique that aligns AI outputs directly with human decision-making patterns. Thinking Machines claims the specialized model filters financial data and identifies high-signal news at 13.8 times lower inference costs than top-tier general-purpose AI systems. The write-up highlights a shift among hedge funds toward building narrow, cost-efficient AI tools tailored to specific research workflows rather than relying on expensive off-the-shelf frontier models.
From the sources (6 posts)
@soumithchintalaBridgewater, one of the worlds largest hedge funds, a Tinker customer talks through how they've carefully fine-tuned a model focused on what makes interesting financial news. Their fine-tuned model is more effective and cheaper than any fro
@miramuratiBridgewater used their unique financial knowledge and partnered with us on @tinkerapi to fine-tune a model that helps their analysts focus on what's important. Experts improving AI that empowers experts.
@soumithchintalaBridgewater, the worlds largest hedge fund, a Tinker customer talks through how they've carefully fine-tuned a model focused on what makes interesting financial news. Their fine-tuned model is more effective and cheaper than any frontier mo
@tinkerapiSorting which financial docs are worth an analyst's time is surprisingly hard for frontier LLMs. With an expert-labeled dataset and on-policy distillation, Bridgewater fine-tuned a model to do it reliably and cheaply.
@mtsliveSITUATION DETECTED: Bridgewater partnered with Thinking Machines to fine-tune a custom model on expert investor judgment that outperforms frontier models on financial information filtering tasks at 13.8x lower inference cost.
@luke_wood_mlBridgewater write up of how they used both expert labeled data and on-policy distillation to beat current frontier models in financial document classification. Excited to see more similar uses of Tinker in the future.