AI Access Inequality
The growing gap between who can access frontier AI capabilities and who cannot. As models become more capable, restricted access creates a two-tier system: those with frontier AI and those without.
The Argument
From The Closing of the Frontier:
- The internet was the last permissionless frontier — same tools for everyone
- Frontier models are now restricted to enterprise partners (Glasswing model)
- “Intelligence is the greatest creative force in the world. If a small group has a monopoly on it, you are the permanent underclass” (George Hotz)
- Capital now trivially converts to superhuman labor (Rudolf Laine)
Tension: Safety vs Access
- Safety argument: frontier capabilities (zero-day discovery) are too dangerous for unrestricted access
- Access argument: restricted access stifles defensive innovation, forces safety researchers to use weaker models
- MATS researchers (serious safety program) use Chinese open-source models because they can’t get frontier access
Historical Parallels
- Frederick Jackson Turner (1893): American frontier closing
- Nuclear non-proliferation: wrong analogy because intelligence is economically valuable, not just destructive
- “Mainframe era”: open-source models 3-12 months behind frontier, like Apple II vs mainframes
Proposed Solutions
- FOIA-style transparency obligations
- Clear access criteria with appeals mechanisms
- API access as minimum viable surface (programmable, doesn’t foreclose possibility)
- Government-level accountability for government-level capabilities
Connections
- The Closing of the Frontier — primary analysis
- Glasswing — the triggering announcement
- Export Controls — related policy debate
- AI Alignment — safety vs access tradeoff