Your Company Should Train Its Own AI Models - Will Brown, Prime Intellect

In this AI Engineer talk, Will Brown of Prime Intellect argues that companies should train their own AI models on their own work. This explanation is based on the talk's published description and chapter list, not a transcript. According to the description, his starting point is that agents relearn the same things every session, because the knowledge that runs a company never gets written down. He proposes training models on a company's own traces (the recorded steps of past work), feedback and workflows, so that models learn on the job the way human coworkers do (talk on YouTube).

The chapter list shows how the argument is built. It moves from what belongs in a model's context, the material supplied with each request, versus what belongs in its weights, what the model learned in training, to evals used as test cases for a company's model, routing and decision models, and interfaces that turn people's decisions into training data, which the description calls widget-style. The talk ends with nightly training runs, "the self-improving organization," and an example model for knowledge work trained for about $300K. The description does not say what that model was trained to do, what the $300K covered, or how the model performed.

Brown made a related case in an earlier talk at LangChain, where he argued that continual learning, updating a model as it is used, should become as routine a part of building products as shipping software.