I agree with the sentiment though, intelligence is not all you need. We're not going to "solve science" through RLVR, people are still needed in the loop until we get superhuman construction robots and superhuman communicators.

Concrete example, "coding is free" (TBC, but let's assume), and now we're code review constrained. Even if code review becomes free we probably just become product discovery constrained, or something else. Making all software free will just cause us to hit on hardware constraints, and so on. It's still progress, and more progress will bring more benefits, but there's a limit to how much better the world as a whole can get from any one change like this.

Reinforcement learning trains task completion abilities.

Reinforcement learning selects for models that appear to complete tasks, on the basis that appearing to complete a task is correlated with completing a task.

Both have problems. The latter is probably a better correlation, but by no means flawless. If we just take at face value that RL is training models to complete tasks, we're not engaging deeply enough with the process enough to see where it falls short.