Wouldn’t the Navier-Stokes result indicate that it is possible to get crystallized competence for long-term strategy via the current paradigm? In particular, “keeping yourself on track”/coordinating this across thousands of parallel agents was a very important subskill to generate this result (and maybe this is what you mean when you say parts of this are RLVR-able). Are there particular subskills that you have in mind that seem less RLVR-able?
Wouldn’t the Navier-Stokes result indicate that it is possible to get crystallized competence for long-term strategy via the current paradigm? In particular, “keeping yourself on track”/coordinating this across thousands of parallel agents was a very important subskill to generate this result (and maybe this is what you mean when you say parts of this are RLVR-able). Are there particular subskills that you have in mind that seem less RLVR-able?