AI researcher Herbie Bradley explains the O-ring problem, where one unautomatable task blocks a whole job: "It comes from the jaggedness of capabilities itself. Reinforcement learning is clearly very good at hill climbing on these highly verifiable tasks, these highly grindable tasks, like coding and AI R&D, and less good at much more qualitative stuff." "Most jobs contain a pretty good mix of highly verifiable tasks and highly non-verifiable tasks. So you're bottlenecked on the smaller…