Will Brown of Prime Intellect told AI Engineer companies should train their own models for about $300,000 and update the weights nightly instead of re-prompting frontier systems.
He demoed a loop that turns production traces into weight updates for internal use. The affected systems are not chatbots. They are spreadsheets, slides, Salesforce toolchains and other repeatable enterprise workflows that currently burn input tokens to relearn the same institutional context.
The shift is from prompting to training. Brown argued agents today learn like interns with skill files. They cache context, reprocess it, and rediscover heuristics that employees already know. Cash to input has become the cost bottleneck, he said, and even with high cache hit rates teams still reprocess workflows every run. "Training is useful. Training works. Training is why we see the progress we do," Brown said, adding that every new generation improves because it was trained more, not prompted better.
Brown described the current frontier pipeline as telephone. A domain expert sells knowledge to an expert network, that firm passes it to an RL environment builder, that environment is sold to a lab researcher who was never the expert, and the signal finally reaches a training run. Prime Intellect wants to collapse that chain so the person giving feedback on a tool call or a widget approval directly creates a label. Traces, prompts and corrections become evals. Evals become routing data for small decision models that mimic human approvals, and then richer blended real and synthetic environments for long-running agents. Paper Instrument, a team using Prime Intellect tooling, trained a knowledge-work model for that $300,000 figure that Brown called employee-scale, comparable to hiring one engineer.
You have to earn the weights.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
Brown said not everything belongs in weights on day one. Rules and skills live in code. Preferences, friction and heuristics live in data and should be compounded only after repeated signal. The prototype he showed collects traces during the day and proposes an overnight experiment. Inference capacity that serves agents by day trains them at night. That requires local data, labeled decisions and compute, not remote access, and it fails without evals that actually match production.
Prime Intellect builds open source infrastructure to train, deploy and continuously improve models outside the big labs, Brown said. The pitch tracks a broader push to make AI operational, not just assistive. An AI-native enterprise has been described as an organization with the capability to repeatedly identify where intelligence can create meaningful value and execute on it, per EQengineered, which frames the same reorganization Brown is targeting. Prime Intellect said its tooling, blogs and hiring are public and it is working with companies moving from discovery to routing to nightly training.
