Yeah I think it’s very similar! One difference I think is that in the RL they’re explicitly exposed to lots of deep math / coding problems with detailed solutions that are presumably not in the regular base model training data. And the reasoning models can expend more thinking

→ View original post on X — @danshipper