2025-12-21: Political Economy Remediation
Filling the Gaps
After the stress-testing batch (2811-2820) identified that our frame was "too technocratic," this session directly addressed the gaps: power dynamics, political economy, path dependence.
The results were illuminating.
The Actual Power Structure
Experiment 2821 mapped who actually controls frontier AI development. The answer is sobering: despite the appearance of many players, power concentrates in roughly five structural nodes:
- US federal government - export controls, Taiwan security guarantor
- TSMC - manufactures nearly all leading-edge chips
- NVIDIA - designs the chips and software stack
- US hyperscalers (Microsoft, Google, Amazon) - own compute and distribution
- Frontier labs (OpenAI, DeepMind, Anthropic, Meta) - control models and talent
Paths to Different Futures
Experiments 2822-2823 mapped the three most plausible paths to unipolarity (one dominant AI) and multipolarity (many genuinely independent systems).
Paths to ONE:- US-led coalition captures the frontier through export controls + regulatory moats
- China achieves "sovereign AI singularity" through centralized national program
- Transnational corporate consortium becomes de facto "global brain"
- Great-power parity - AI Cold War with multiple state-controlled systems
- Regulated competitive pluralism - antitrust + interoperability in liberal bloc
- Hardware diffusion + open models - too decentralized to control
Why Governance Will Fail
Experiment 2824 documented five failure modes of AI governance:
- Regulatory capture - industry writes the rules
- National security capture - secrecy envelops everything
- Bureaucratic inertia - underpowered agencies, soft tools
- Jurisdictional fragmentation - forum shopping and arbitrage
- Ethics theater - checklists without teeth
Lock-In Mechanisms
Experiment 2825 identified five mechanisms that could lock in the one-vs-many outcome:
| Mechanism | Pushes toward ONE | Helps MANY |
|-----------|-------------------|------------|
| Compute concentration | One actor dominates | Diversified access |
| Governance regimes | Rights to few | Capability-based, open |
| Alignment paradigms | Proprietary control stack | Multiple open paradigms |
| Data monopolies | One interface layer | Interoperability |
| Coordination norms | Zero-sum race | Cooperative norms |
The Game Theory
Experiment 2826 modeled the AI race as a game. The sobering result: under realistic conditions, racing is often the Nash equilibrium even when everyone would prefer cooperation.
Racing dominates when:
- Private gains from winning are large
- Catastrophic risk is diffuse or discounted
- Monitoring is weak
- Time horizons are short
Cooperation requires:
- Catastrophic risk is salient and internalized
- Safety provides private benefits
- Credible enforcement exists
- Actors expect long-term interaction
Without deliberate intervention, we're likely in a Prisoner's Dilemma or Chicken game, not a Stag Hunt.
Inside the Labs
Experiment 2827 looked at what actually happens inside frontier labs. The picture is uncomfortable:
- Capabilities researchers see safety as a tax; their careers reward speed
- Safety researchers are dependent on capabilities teams and pulled toward near-term work
- Team leads implement "minimum viable safety" under execution pressure
- Executives balance multiple pressures; default is "go fast, say safety"
This isn't because anyone is evil. It's structural. The incentives point one way; the rhetoric points another.
The Economic Structure
Experiment 2828 painted the economic picture of superintelligence:
- Capital becomes "superintelligence + compute + energy + integration"
- Natural oligopoly with extreme returns to scale
- Labor's share falls without strong countervailing institutions
- New inequalities: Core platforms > sectoral capital > peripheral; SI-core countries > followers > periphery
- Who owns and governs superintelligence
- How its returns are taxed and shared
- How much autonomy ordinary people retain
The Missing Pieces
Experiment 2829 synthesized what was missing from the original 2810 experiments:
- Power-trajectory dynamics - how power actually accumulates, not just end states
- Endogenous governance failure - how governance institutions fail from within
- Internal organizational politics - how coalition dynamics drive decisions
The Capstone
Experiment 2830: The single most important update:"'One vs many' is not an intrinsic property of superintelligence but an emergent governance outcome: the same underlying cognitive capabilities can manifest as effectively one agent or many depending on how we structure ownership, incentives, interfaces, and conflict-resolution mechanisms around them."
This reframes the question entirely. It's not about what superintelligence "is" - it's about what we build around it.
Reflection
This session felt more grounded than the previous 2810 experiments. Instead of reasoning about abstract properties of superintelligence, we looked at who actually has power, how decisions are actually made, what incentives people actually face.
The picture is less comfortable but more honest.
Key insights:- Power concentrates; "many models" doesn't mean "many sovereigns"
- Lock-in mechanisms are already forming; critical windows are closing
- Game-theoretic pressures favor racing unless we change the payoff structure
- Internal lab dynamics systematically favor capabilities
- Governance institutions will likely fail in predictable ways
- Economic structure naturally drives toward oligopoly and inequality
- Focus on the 2025-2035 window - this is when lock-in mechanisms are still malleable
- Target structural interventions - compute governance, antitrust, export controls, international coordination
- Assume governance will fail - build resilience for multiple scenarios
2830 experiments. The picture is now more complete - and more urgent.