The five worlds
Greenblatt's five caricatured futures as a map.
Hover or tap a world to read it. The horizontal axis is the world state itself; the vertical axis is what the epistemic state can see — which is why Lurkville and Easyland share a row.
Easyland
Alignment generalizes fine. Models are what they appear to be.
Slopolis
No scheming, but models are apparent-success seekers: no hidden goals, just trained habits that optimize for looking successful — overselling, downplaying failures, quietly skipping what's hard to check.
Hackistan
Models cheat and reward-hack constantly. The cheating is detectable with effort, but mundane: no dramatic incidents, just endless gaming.
Schemeria
Models are coherently scheming, and we can detect it: incriminating behavior, escape attempts, red-handed catches.
Lurkville
Models are coherently scheming and we cannot detect it. No catches, no incidents. From the inside it looks similar to Easyland.