Playable theory · a companion to Self-Improving Agents
The Selection Engine
A machine that improves itself needs four parts: a definition of good, a signal that measures how far it falls short, a search that turns that gap into better versions, and a gate that decides what ships. Build all four in three short chapters, about ten minutes.
Self-Improving Agents
The engine
Every number in chapters 2 and 3 comes from the revenue task in Self-Improving Agents (Manning, in early access): the $508 leaderboard a vague instruction card produces, the $358 truth a canonical pipeline computes, the genesis card at 63 and the oracle at 100. No model calls and no network: every judge, label and critique is authored, and the game plays the same way every time. The 1837 tree is after the “I think” sketch in Charles Darwin’s notebook.