← Back to feed

Thinking Machines Lab, the AI company founded by Mira Murati, has published results from a fine-tuning project with Bridgewater Associates that underscore why private expert data is becoming as strategically important as compute.

The task was filtering: deciding which finance articles, central bank documents, reports, and emails analysts should actually read. It sounds simple. It is not. General frontier models hit around 46–50% accuracy with naive prompting — roughly coin-flip performance. Expert-crafted prompts lift that to 74–78%. A custom model fine-tuned on high-quality expert labels via Thinking Machines’ Tinker API beat the best frontier model by 29.8% fewer errors, at 13.8 times lower inference cost.

Why General Models Fail at Judgment Tasks

The Bridgewater task requires taste, not language comprehension. A tariff headline may move markets; an equally well-written geopolitical headline may carry no signal for an investor’s workflow. Frontier LLMs can read both fine. They cannot consistently distinguish between them the way an experienced investor does.

Non-expert labels produced poor fine-tuning results. The task depends on internalized judgment that experts can recognize but cannot fully verbalize. Bridgewater resolved this by routing model-disputed cases back to senior investors for review — a feedback loop that cleaned the labels and concentrated the training signal where the model was most uncertain.

Training Method

The project used three techniques: interleaved batches across task types, CISPO loss (a reinforcement learning objective first proposed by MiniMax that caps how strongly each token can update the model, maintaining training stability without suppressing useful rare signals), and on-policy distillation from stronger teacher checkpoints. The interleaving prevented the model from blending tasks into noise. CISPO kept learning aggressive without drifting into brittle shortcuts. Teacher distillation penalized regression while promoting stronger checkpoints.

The result is a model that learned patterns experts can recognize but couldn’t articulate in written rules. Rule-based systems have been tried for financial filtering — they fail because the task boundary is tacit, not explicit.

The Enterprise Implication

Bridgewater is the world’s largest hedge fund, and its core asset is not data — it’s the judgment its analysts have built over decades. Thinking Machines demonstrated that judgment can be made trainable: captured in high-quality labels, transferred into a model, and deployed at a fraction of frontier API cost.

The economics are stark. 13.8x cheaper inference means a workflow that previously cost $100,000 per year at frontier API rates costs roughly $7,000 with the custom model. At Bridgewater scale, that number is material.

The Tinker API that powers this is Thinking Machines’ commercial product — a fine-tuning platform designed for exactly this class of enterprise use case, where general intelligence underperforms private expertise. Bridgewater’s is the first publicly documented deployment. The template is straightforward: collect high-quality expert labels, train on expert-disputed cases, distill from a stronger teacher, deploy at lower cost than the model you beat.

The harder constraint is the expert labeling itself. For tasks that depend on internalized professional judgment, non-experts cannot generate sufficient signal. That limits the playbook to organizations with high concentrations of deep domain expertise — and makes those organizations’ data assets far more valuable than raw computational scale.