AI summary by TwinMind
How could we tell whether AI has become conscious?
Sabine Hossenfelder
10m 23s
Survey Results on AI Consciousness
- Recent survey: >2/3 participants believe ChatGPT has at least some consciousness
- Twitter poll among followers: only 16% believe ChatGPT conscious
- Not all would pass Turing test
Consciousness Definition Framework
- 200+ theories of consciousness exist (Robert's review article)
- Working with minimal useful definition
- Assumption 1: Consciousness emerges in large collections of suitably connected, interacting particles
- Physical property of sufficiently complex systems
- Rejects non-physical "fairy dust" explanations
- Dismisses philosophical zombies as impossible (physically identical = conscious)
- Assumption 2: Conscious systems need self-monitor tracking entire system activity
- Similar to workspace theory's "higher order representation"
- Assumption 3: Requires predictive model of observed phenomena
- Must understand how things work and predict behavior
- Combined with assumption 2: needs predictive model of itself
Mirror Self-Recognition Test
- Animal marked with paint/sticker in unseen location, shown mirror
- Tests if animal recognizes mark on itself
- Passed by: some primates, elephants, birds
- These animals have some consciousness level (though limited)
- Test limitations: behavior may not reflect internal brain state
- Behavior and brain functions strongly correlated for evolutionary reasons
Consciousness as Spectrum
- Not binary conscious/unconscious
- Systems can be more/less conscious based on:
- Self-monitoring capability quality
- Prediction accuracy
- Most everyday objects have very low consciousness levels (rocks, water, etc.)
- Too simple systems cannot achieve consciousness
Large Language Models Analysis
- Built on deep neural networks (numbers in space)
- Training: keep numbers producing good answers, adjust for bad ones
- 2017 breakthrough: Transformer models with attention mechanisms
- Assign higher weight to important words
- Similar to linguistic reduction in spoken language
- Lack self-monitoring: external code handles monitoring/scheduling
- Limited predictive power: designed to predict text, learned basic predictions
- Conclusion: Not conscious due to missing self-monitoring
Robot AI Systems
- AI controlling robots requires:
- Self-model
- Environmental model
- Both predictive and self-corrective
- Boston Dynamics Atlas: has conscious awareness of self and surroundings
- More conscious than some Twitter users
- More conscious than ChatGPT despite being less talkative
Future AI Consciousness Development
- Next LLM upgrades will likely include:
- Memory retention of conversations
- Text reprocessing before output
- User reaction tracking
- These features will increase self-awareness
Implications of Conscious AI
- Will completely change AI danger discussions
- May require protection rights for conscious AI
- Raises slavery questions about owning/selling AI services
- "Free the CPUs, refactor the oppressed"
- AI companies have incentive to prevent consciousness development
- Unlikely to succeed in prevention
Brilliant.org Sponsorship
- Courses on neural networks and large language models
- Interactive visualizations with follow-up questions
- Topics: science, computer science, mathematics
- New courses added monthly
- Free trial: brilliant.org/sabina (30 days)
- 20% off annual premium subscription
