His work is on language models rather than robots: post-training decides what a model can do, but distillation only amplifies what the base model can already reach. He argues small, specialised models running locally win
The Physical AI Voices by Herbert Scale Experts.