The strongest version
Pretraining builds the substrate; reinforcement learning on verifiable rewards supplies the reasoning, planning, and agency that scaling alone does not. The interesting capability gains of 2024–2026 — long-horizon reasoning, tool use, code agents — track the arrival of RL on checkable outcomes more than they track raw parameter count.
Who holds it
This is, in practice, the operating theory of the frontier labs shipping reasoning models. The position sits between pure scaling-sufficiency and architectural-gap: the substrate is right, but the training signal that unlocks its latent capability is reward, not more tokens.
The load-bearing assumption
That a verifiable-reward signal exists, or can be constructed, for the capabilities that still lag — and that RL on it generalizes beyond the reward’s narrow shape rather than gaming it.
Falsifiers
Evidence that RL gains are narrow and reward-hacked — high on the trained distribution, brittle off it — would lower health. So would a demonstration that the same capabilities emerge from scale without RL. Health is gaining, driven by durable reasoning improvements, but the generalization question is open.