AI Pulse · AI Trends Pulse

Garry Tan is arguing that US open-weight AI labs should use model distillation more aggressively, following a playbook he associates with Chinese competitors. His point is simple. If a smaller model can learn to reproduce much of what a frontier model does, it may deliver useful performance at a far lower cost to run.

That matters because the cost gap is becoming a practical business issue, not just a research debate. Many companies do not need the largest available model for every task. They need a model that handles support queries, document extraction, internal search, forecasting assistance, or workflow automation reliably and at a cost that works at volume. Distillation is one route to that. It involves training a smaller model using outputs from a stronger one, with the aim of retaining enough capability for a defined job.

Tan’s argument has drawn attention, with the discussion sitting at 204 points and 95 comments on Hacker News. You can read the details in the original report.

For operators, the takeaway is not to wait for one perfect model. Start separating work that truly requires a frontier model from work that could run on a smaller, cheaper system. This is the kind of thing we build into an AI command centre, matching the right model and cost profile to each business process.