The survey uses the same small set of ideas in every terrain. They are not predictions; they are the lenses the chapters keep reaching for. Read together, they explain why the frontier is jagged rather than straight.
Intelligence in these systems is a profile, not a single number. They are superhuman at some things and subhuman at others in the same moment, in a shape that resembles no human's. So capability does not advance as a line that crosses a wall on a date; it advances as an irregular coastline, gaining ground in more and more places while leaving odd blanks behind.
Where a cheap and reliable verifier exists — a game outcome, a passing test, a machine-checked proof — a system can be trained against it and discover things no human demonstrated. Where the verifier is expensive, slow, or inaccessible, more scale buys little. The price of the check, more than the size of the model, sets how far the factory reaches.
AI has collapsed the cost of producing plausible output — code, prose, molecules, hypotheses. It has barely touched the cost of deciding what is worth producing and confirming that it is correct. In most fields the binding constraint was always the second half, which is exactly the half that does not commoditize on the same curve.
Emergent, chaotic, computationally irreducible systems cannot be predicted faster than they can be run — a limit that binds any computer, not only this generation of them. The same shape appears in Hayek's knowledge problem: the relevant information is dispersed, tacit, and continuously regenerated, so no aggregator assembles it in time. A larger model does not dissolve either constraint.
A perfect forgery of a Vermeer loses almost all its worth the moment it is known to be a forgery, though the canvas is unchanged. Some of what people value lives in who made a thing, and in the fact that someone lived it — not in the artifact itself. That is a limit no quantity of parameters reaches, because it sits in the relational structure of human value.
When production approaches zero marginal cost, two things follow. Output divides into a synthetic floor optimized to a measurable proxy and a human premium that rises precisely because the floor exists. And the scarce input shifts — from doing the work to specifying it, judging it, and verifying it. The job does not disappear; its definition rotates toward the part the machine structurally cannot do.
See the threads at work, terrain by terrain, in the five chapters.