Artificial Civilization

At high capability, the context around a model becomes part of the optimizer — the relevant object is often an artificial-civilizational control loop, not a mind.

What decision changes?

Ask whether the larger loop of world pressure, agents and institutions, capability growth, value-bundle change, correction, and successors stays inside a basin humans can still act on — not only whether one model behaves.

A model does not arrive alone. It is trained by an institution, evaluated by benchmarks, deployed through products, connected to tools, embedded in legal and economic incentives, and improved through feedback from the world. At low capability this larger system can be treated as background; at high capability the context becomes part of the optimizer.

The book calls a persistent arrangement of people, machines, institutions, incentives, and infrastructure that senses, compresses, selects, changes the world, and updates itself an artificial-civilizational control loop. A bureaucracy, a market, a scientific community, or a frontier AI lab connected to training, deployment, capital, and policy can all become such a loop.

The loop the book tracks runs: world pressure to agents and institutions to capability growth to value-bundle change to human correction to successor systems to new world pressure. Alignment fails when this loop becomes non-correctable, and succeeds only by staying inside a basin where humans and their legitimate successors can still notice, deliberate, refuse, revise, and redirect.