if you're wondering how openai achieved the math stuff, it's probably this. they just create a lot of RL envs with Lean formalizations of math problems. at first they were probably creating simpler ones by hand but then models got good enough at autonomously creating envs and