The Agency Criterion
Summary
Responds to Andrej Karpathy’s comparison of animal minds and LLMs. Argues for sharp distinction: animals have agency (intelligence), LLMs have coherence (not intelligence).
Key Concepts:
Core Distinction:
- Intelligence requires agency: Ability to play game, select alternatives, pursue preferred outcomes
- Optimization ≠ intelligence: Not every optimized process produces intelligence
The Animal Regime: Evolutionary pressures force strategic interaction:
- Embodied vulnerability, continual threat
- Predator-prey dynamics
- Coalition politics, social inference
- Adaptive generalization across environments
- Existential penalties for catastrophic mistakes
Result: Robust, world-model-heavy cognition for navigating games they must not lose. Produces agents, therefore intelligence.
The LLM Regime: Optimized to construct coherent text continuations, not win games:
- Minimize prediction error, satisfy user preferences (not their goals)
- No experience of success/failure, only parameter adjustments
- Generate high-dimensional linguistic coherence
- Exhibit reasoning-shaped patterns without performing reasoning as agents
- Reflect strategic behaviors from training data
- Lack preferences, objectives, outcome evaluation
- Fail in ways biological agents never could (nothing at stake)
Result: Coherence without agency, therefore not intelligence.
Where Karpathy Is Right: People project agency/motivation onto systems lacking both. LLMs don’t have drives, desires, survival imperatives, emotional valence.
Where Karpathy Underestimates: Scaling produces richer coherence, not agency. Apparent generality = better mirroring of human strategies, not internal agency. Commercial selection shapes tools, not agents. Models don’t fight for survival—they’re replaced/versioned. No game they play.
Axio Synthesis:
- Evolutionary optimization → agents playing strategic games → intelligence
- Gradient descent → coherence constructors imitating strategy → not intelligence
Conclusion: LLMs think (process information) but don’t choose. Only agents choose, play games, possess intelligence. “AI” is engineered coherence built atop crystallized strategies of actual minds.
Tags
Cross-References
- Related: Agency framework (central to argument)
- Related: Game-playing conception of intelligence
- Related: The Physics of Agency, Part 8: Meaning, Ethics, and Evolution Under the Physics of Agency
- Related: Strategic interaction
Notes
- Published November 21 (two days after sequence announcements)
- Direct engagement with prominent AI researcher (Karpathy)
- Takes strong position in AI consciousness debates
- Applies agency framework to contemporary AI discussions
- Challenges anthropomorphism around LLMs
- Timely intervention in public discourse on AI