henryaj comments on How Well Does RL Scale?

[henryaj]

Nit: scaling up RL by 100x and inference by 10,000x would be a 1:3 OOM ratio I think