The Agency-Consciousness Matrix

This is a quad chart containing some random notes about where AI might be headed, (dis)organized along two axes: Agency and Consciousness.

Not Conscious

Conscious

Not Agentic

Agentic

Oracle systems

Language Models without agent scaffolding

This is pretty satisfying for my personal aesthetics, seems economically valuable, will take time for humans to adapt and fully integrate.

However attractive this option is, it doesn't seem stable. It's too easy to make an MVP agentic system.

Seems less likely than the agentic options. My impression is that we are further along in operationalizing agency than we are in operationalizing consciousness.

• Plenty of commercial interest in this. E.g.
  • LangChain,
  • BabyAGI
  • McKinsey Blog Post
• Seems likely that we will be somewhere in this quadrant in the next couple of years. Less certain whether we will be close to human-level agency, for example whether AI systems will be able to operate on multi-year planning horizons autonomously.

• Seems to be Jürgen Schmidhuber's view
• From an impersonal perspective, seems more aesthetically satisfying than the agentic-but-not-conscious scenario.
• Seems unpredictable.
• Whether "adding consciousness" tends to improve predictability seems like an interesting question.

Increasing Consciousness →

Arguments for Increasing Consciousness

Keywords:

• Consciousness Studies
• Consciousness Prior (Bengio)
• Integrated Information Theory
• Global Workspace Theory

Increasing Agency ↓

Arguments for why agency might increase

• Why Tool AIs Want to Be Agent AIs (Gwern)
• What Failure Looks Like (Christiano)

Keywords:

• AI Agents