A lot of algorithmic improvement in AI is ultimately bottlenecked by compute. It is very easy to come up with ideas that could improve models! But to prove that they do, especially at scale, is expensive and takes a long time.