This is the interesting bit. As we move to the next phase of scaling reasoning models w RL, data and compute converge. Next breakthroughs require moving from 100’000’s of GPUs to millions of GPUs.

→ View original post on X — @ninadschick,