Daniel Kokotajlo comments on How much compute was used to train DeepMind’s generally capable agents?

[Daniel Kokotajlo]

Also for comparison, I think this means these models were about twice as big as AlphaStar. That’s interesting.