The Flat Curve Society

The thesis: the AI capability curve is still climbing exponentially, but the most powerful models are about to be locked down like nuclear weapons — restricted by governments, gated by the compute supply chain, and priced out of reach — so for almost everyone the curve will only appear to flatten. Steve builds the case through two "horizons." The demand horizon is benign: you can't tell two models apart because your problems aren't hard enough to stretch either one. The discernment horizon is the dangerous one: you can't tell whether a model is right because checking its work is itself beyond you — "superhuman means unverifiable." From there he reads off the consequences: frontier intelligence sold like a vending machine, OSS models stuck a few months behind a hardening wall, and routing layers that send each task to the dumbest model that can handle it.

The back half turns practical. Leaning on Ezra Savard's Netflix training study, Steve defines AI literacy as measurable token-spend cohorts — no-agent, single-agent, multi-agent — and argues that the binding constraint of the coming plateau isn't model capability but teaching people to use, and then stop wasting, the models we already have. SaaS survives because rewriting it is too expensive; engineering survives because today's models still need grown-up supervision. The plateau, he insists, is good news: stable ground on which a real craft of building software can finally take hold. The keeper line — the curve is as real as the Earth is round, but just as flat from where you stand — gives the essay both its title and its argument in one breath.