Intelligence Paradigms
Overview
Section titled “Overview”This section explores different paradigms for building intelligent systems. Understanding architectural diversity helps anticipate which safety approaches may be needed and which risks are most relevant.
Current Paradigms
Section titled “Current Paradigms”The dominant current architecture:
- Massive parameter counts, uniform activation
- Strong general capabilities
- Well-studied safety properties
- Examples: GPT-4, Claude, Gemini
Mixture of Experts architectures:
- Conditional computation for efficiency
- Routing mechanisms select experts
- Potentially different failure modes
- Examples: Mixtral, Switch Transformer
State Space Models:
- Alternative to attention mechanism
- Linear scaling with sequence length
- Emerging research area
Hybrid neural-symbolic systems:
- Combine learning with logical reasoning
- Potentially more interpretable
- Research stage
Scaffolding Approaches
Section titled “Scaffolding Approaches”Foundation models with basic prompting
Chain-of-thought, few-shot learning
Extensive tool use, agentic frameworks
Alternative Pathways
Section titled “Alternative Pathways”Learning predictive world representations
Architectures with formal safety guarantees
Biological Approaches
Section titled “Biological Approaches”Architectures not yet conceived
Why Paradigms Matter for Safety
Section titled “Why Paradigms Matter for Safety”Different paradigms may have:
- Different failure modes and risk profiles
- Different amenability to interpretability
- Different alignment properties
- Different governance implications