The AI Frontier: from Gemini 3 Deep Think distilling to Flash — Jeff Dean

Jeff Dean, Google's chief AI scientist, says in a Latent Space interview that Google's cheaper Flash models and its frontier models are linked, because distillation, training a smaller model to imitate a larger one, needs the larger model to exist first. Dean calls distillation a key technique, set out in a paper co-authored with Geoffrey Hinton and Oriol Vinyals, and says that for several Gemini generations the next Flash has matched or beaten the previous Pro.

Asked whether Flash will eventually cover everything people need, Dean says that holds only if demand stays fixed. As models improve, people ask for harder things, such as a whole software package instead of a loop, and those requests reveal where the frontier still fails. Dean adds that Flash's lower latency matters as tasks grow longer. These are Dean's views as a Google executive, not benchmark results. Intel also covers Dean's Sequoia conversation.