This sorta framing might—to a surprising degree—overlap with how I got quaked reading Thomas Nail’s Being and Motion several years back (2021, I believe), and started looking into Kinocomputing (alternative paradigms for computing and intelligence). It was the first time I’d really thought about intelligence conceptually in a non-trivial way and I recall realising that intelligence, as currently conceived, wasn’t enough. By far.
If you focus on intelligence you’ll get the future wrong. That’s already gotten to boring. Understood correctly, we hit the intelligence sigmoid plateau a few years ago. Gains since then have been a mix of horizontal scale-out and liveness gains. RLVR solves for liveness, not intelligence because it locks on to highly playable domains. These are valuable because you can continue to play indefinitely in a way that’s impossible in RLHF domains, which unravel as they go longer because that’s what stochastic phenomena do.