Tri Dao proposes diffusion LLMs as fast worker models paired with an orchestrator

Tri Dao, on X, says he has wanted to pair a smart orchestrator or architect model with a very fast worker model, and suggests a diffusion LLM could fill the worker role.

In his post he cites Mercury 3, a release from Stefano Ermon, who says it runs at more than 1,000 tokens per second.

The post is a suggestion from Dao, not a reported test of this pairing, and the report contains no benchmark results or evaluation of how such a system would perform on tasks.