AI Pulse · Frontier Labs Watch

Garry Tan is publicly urging US open-weight AI labs to use distillation to learn from frontier models. In simple terms, distillation is where a smaller or more accessible model is trained to reproduce useful behaviour from a stronger model. The goal is not necessarily to copy every capability. It is to make more of that capability available at lower cost and with fewer restrictions.

If this approach is adopted at scale, it could speed up the commoditisation of AI capability. Closed labs have pricing power when the best performance is available only through their own products and APIs. But if open-weight alternatives can narrow the gap quickly, buyers get more choice. That can mean lower model costs, less vendor dependence, and more ability to run models in your own environment.

For business owners, the practical point is to avoid building your AI plans around the assumption that today’s leading model will remain uniquely valuable for long. The advantage may shift from access to a model towards how well you connect AI to your data, processes, controls, and staff workflows. The original report is a useful signal of where that competition could head.

This is the kind of thing we build into an AI command centre, tracking which model options are becoming viable so your company can change suppliers or deployment choices without rebuilding everything.