The Flywheel
You run an AI lab for twelve quarters. Each quarter, split your points between capability, safety, and trust-building, then watch the wheel. Get the balance right and the work begins to pay for itself. That loop has a name.
Before you play
Capability ships value, and invites incidents when it outruns your safety. An incident cuts that quarter’s value to 15 per cent or less, and safety decides whether it dents trust or craters it. Trust drives adoption, and adoption brings more points next quarter.
Two ghosts play beside you: the racer, who piles into capability, and the flywheel, who invests steadily. Their dice are their own; some quarters the racer gets lucky.
Read the hall as you play. The gauge on the iron column is this quarter’s risk; the six lamps light with adoption; an incident throws the belt and cracks the rim. The cat on the warm pipe sleeps through good quarters, and when every lamp is lit the night engineer raises his tea.
Safety didn’t slow the machine down. It let the machine be adopted.
That loop is the flywheel. Safety makes an incident survivable, a survivable incident leaves trust standing, trust brings adoption, and adoption pays for next quarter’s safety. The racer has a loop too; it runs the other way.
Twelve quarters is a toy, but the shape is the essay’s. In A Flywheel for AI Safety capability is no longer the bottleneck; trust is. E-commerce, the essay notes, took off not when the web got faster but when encryption, payments and consumer protection made it trustworthy. The sums here are tuned so that luck still matters: the racer beats the flywheel about one game in eight.
The argument in full: A Flywheel for AI Safety · more play on the Playground