Statements
-
NVIDIA GTC talk: token efficiency is not about saving cost but about raising the ceiling of intelligence.
«Token efficiency is not just about efficiency. It is about raising the ceiling of intelligence itself.»
-
On the GPU gap with the US: says the gap has not narrowed.
«I don't think the gap has narrowed.»
-
On compute under export limits: trains on H800, admits a disadvantage in tier and quantity, but maxes out every card.
«We use H800 GPUs with Infiniband... but we made full use of every single card!»
-
Scaling-first principle: if scale solves it, don't invent a new algorithm; an algorithm's value is enabling better scaling.
«If scale can solve a problem, don't use a new algorithm; a new algorithm's value is enabling better scaling.»