Skip to content

Intelligence Paradigms

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.

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

Foundation models with basic prompting

Chain-of-thought, few-shot learning

Extensive tool use, agentic frameworks

Learning predictive world representations

Architectures with formal safety guarantees

Architectures not yet conceived

Different paradigms may have:

  • Different failure modes and risk profiles
  • Different amenability to interpretability
  • Different alignment properties
  • Different governance implications