Fences, Not Sandboxes

The author describes operating a team of 50-60 AI agents, primarily using Anthropic's Claude Fable model, spending approximately $120k monthly in API tokens to develop a video game. He argues that current AI governance approaches focused on sandboxes and safety controls are misguided, predicting instead that future AIs will be regulated through laws rather than technical constraints, though he acknowledges that current models still make unreliable decisions requiring constant human oversight.