Post-Training

Enterprise AI adoption 1.0 was getting everyone access to Claude, or using Haiku to automate repetitive work across the company. The next phase is custom models built for concrete business tasks. They can be cheaper to run, faster at inference, and, on a well-defined task, more capable than a generalized frontier model.

We are starting to see what this looks like in practice. Thinking Machines and Bridgewater used post-training to teach a model the repeated judgments inside an analyst’s investment workflow. Applied Compute and Mercor trained a model for long-horizon corporate-law work. Trajectory’s work on Harvey LAB showed that an open model could move toward closed-model performance on legal tasks in less than a day. The lesson is simple: once a business can define and measure the behavior it wants, the largest general-purpose model may no longer be the best tool for the job.

I post-train models for clients using Prime Intellect’s infrastructure, from evaluation and reward design through SFT and RL, to make their agents more capable at the work that matters to the business.