Game theory is everywhere in the extended total AI stack. Much of it achieving sub optimal local equilibria (or Prisoners’ Dilemma (PD) outcomes). This is very preliminary and I have missed a lot…
Level 1: Geopolitical Competition
Compute
US: CHIPS Act protectionism vs global supply chain efficiency
China: Indigenous semiconductor development vs immediate capability gaps
Europe: Digital sovereignty vs technological dependence
PD: Export controls reduce global efficiency, create supply vulnerabilities
People
Brain drain competition: Immigration policies vs domestic talent development
Academic capture: Funding foreign students vs security concerns
Defection incentives: Individual researchers vs national loyalty
PD: Talent hoarding reduces global research productivity
Data
Data localization mandates vs cross-border AI training
Strategic data denial: Restricting adversary access vs commercial openness
Surveillance capitalism: State data collection vs citizen privacy
PD: Fragmented data reduces model quality globally
Architecture/Algorithm IP
Open source vs national champions: basically China v ROW…
Standards warfare: Technical leadership vs interoperability
Export control regimes: Denying capabilities vs innovation diffusion
PD: Closed development reduces collective progress
Level 2: National Competition (Firms Within Countries)
Compute
Cloud oligopolies: AWS/Azure/GCP vs frontier lab expenditure
Preference deals: Government cloud contracts vs competitive procurement
Infrastructure sharing: Consortium models vs proprietary advantage
PD: Duplicate infrastructure investment vs collective efficiency
People
Salary arms races: Individual talent retention vs industry sustainability
Non-compete enforcement: Labour mobility vs trade secret protection
Academic partnerships: University capture vs open research
PD: Talent costs escalate, reducing aggregate R&D spending
Data
Platform competition: User data moats vs data portability
API access pricing: Ecosystem development vs direct monetization
Synthetic data generation: Reducing external dependencies vs shared datasets
PD: Data silos reduce model training effectiveness
Architecture/Algorithm IP
Patent thickets: Defensive patenting vs innovation freedom (eg where would we be without the publication of “Attention is all you need”?)
Open source contributions: Ecosystem building vs competitive advantage
Standardization: Industry compatibility vs differentiation
PD: Fragmented standards increase development costs
Level 3: Intra-Firm Competition (Teams/Ideologies)
Compute
Resource allocation: Research vs product deployment compute
Architecture choices: Transformer scaling vs novel approaches
Infrastructure decisions: Cloud vs on-premise vs hybrid
PD: Competing teams hoard resources, reducing firm-wide efficiency
People
Internal talent competition: Cross-team poaching vs collaboration
Research vs engineering: Long-term research vs immediate product needs
Safety vs capability teams: Risk mitigation vs performance optimization
PD: Internal competition reduces knowledge sharing
Data
Dataset access: Shared resources vs team-specific advantages
Labeling priorities: Quality vs quantity vs speed trade-offs
Synthetic vs real data: Control vs authenticity debates
PD: Teams duplicate data work instead of sharing
Architecture/Algorithm IP
Technical approach competition: Transformer vs CNN vs hybrid architectures
Safety integration: Constitutional AI vs reward modeling vs other approaches
Deployment strategies: Edge vs cloud vs hybrid decisions
PD: Teams reinvent solutions instead of building on each other
Level 4: Intra-Model Competition (Internal Components)
Constitutional Layer vs Core LLM
Compute allocation: Safety checking vs generation efficiency
Training priority: Helpfulness vs harmlessness objectives
Response selection: Multiple generations vs single-pass optimization
PD: Safety measures reduce capability metrics, creating internal tension
Attention Head Competition
Information routing: Which tokens get attended to
Feature extraction: Competing representations of same input
Layer depth allocation: Early vs late processing trade-offs
PD: Heads optimize locally rather than globally
Training Objective Conflicts
Loss function components: Next-token prediction vs human preference vs safety
Multi-task learning: Task-specific vs general capability optimization
Fine-tuning vs pre-training: Specialization vs generalization
PD: Objectives pull model in competing directions