I think it’s impressive that they’re doing it on a single (large) node. Costs matter. Efficiency improvements like this will probably increase capabilities even

I think it’s impressive that they’re doing it on a single (large) node. Costs matter. Efficiency improvements like this will probably increase capabilities eventually. I’m also optimistic about building better (rather than bigger) datasets to train on.