Does this imply we don't need as much compute for models/agents? How can any other AI model compete against that?