Hardware Overhang

The price performance of compute has actually been super exponential for almost that whole timeframe, meaning it gets better and better every year. And so now we're sitting at this place where we have this incredible embarrassment of riches in terms of available cheap compute. It's like an environment where there's a lot of dead trees and kindling on the ground, and you're like haphazardly striking sparks off of a flint. You can quickly and easily turn an enormous amount of compute toward any given new ML task that you want to try to figure out.