Mira Murati’s Thinking Machines has successfully enhanced Bridgewater’s AI capabilities by making their private expert judgment trainable, resulting in a model that outperformed frontier models with 29.8% fewer errors. This advancement reflects a growing trend among investment institutions to leverage custom fine-tuning with internal expert labels for improved accuracy in nuanced tasks beyond what general models can achieve. The breakthrough arose from integrating high-quality labels from expert investors into the model, allowing it to effectively prioritize information in finance documents—an area where traditional models struggled.

Bridgewater: Bridgewater Associates is a prominent global hedge fund recognized for its rigorous, principles-based investment process and use of expert analyst judgment. The firm partnered with Thinking Machines to encode its private financial expertise into an AI model via high-quality labels from its investors. This collaboration allows the model to help analysts prioritize relevant signals from finance documents, reports, and market headlines.
Mira Murati: Mira Murati is an AI researcher and entrepreneur who founded Thinking Machines after her role as CTO at OpenAI. She publicly highlighted the Bridgewater partnership on X, noting how the collaboration uses the firm’s unique financial knowledge to fine-tune models that empower expert analysts. The project underscores her emphasis on practical applications that combine proprietary data with advanced training techniques.
Thinking Machines: Thinking Machines is an AI company founded by Mira Murati that develops advanced fine-tuning methods to incorporate specialized human expertise into large language models. It created a workflow that turns Bridgewater’s expert investor labels into trainable data, enabling the model to learn nuanced judgment patterns that general frontier models miss. The resulting system outperforms standard approaches on tasks requiring taste and prioritization in enterprise finance workflows.

`json
{
“Expert-Driven AI”: “Investment institutions are applying custom fine-tuning with internal expert labels to enhance AI accuracy on complex prioritization tasks that surpass the capabilities of general models.”,
“Training Innovations”: “Techniques like interleaved batching, stability-focused loss functions, and on-policy distillation are used to sustain aggressive learning while safeguarding specialized performance in domain-specific models.”,
“Enterprise AI Workflows”: “Substituting explicit written rules with high-quality expert judgments is an effective method for AI systems managing intricate, taste-dependent decisions in finance.”
}
`