What researchers are saying

Chollet says scaling multiplies generalisation and cannot replace it

Francois Chollet, creator of Keras

Chollet argued that data and compute scale up the generalisation power of a method, and that inefficient methods only look adequate while investment is enormous.

Francois Chollet responded to a statement by Richard Sutton that brute-force computation is not enough for intelligence. Chollet framed data and compute as multipliers on the generalisation power of an approach, in place of the usual picture of them as substitutes for better methods.

On that view, inefficient approaches can appear adequate under massive investment, because a large multiplier hides a small base. Methods with stronger inherent generalisation, he argued, will in the end scale more effectively.