Chollet says scaling multiplies generalization, but cleverer methods scale better

François Chollet says on X that data and compute scaling act as multipliers on an approach's generalization power.

He argues that large investment can make low-generalization methods look adequate, and that researchers should build methods that can scale, since methods without this multiplier cannot compete. He adds that a cleverer approach will scale better. The post is his own argument, and no supporting data or experiments are cited.