AI summary by TwinMind
You Are Your Own Existence Proof (Karl Friston) | AI Podcast Clips with Lex Fridman
Lex Fridman
45m 31s
Free Energy Principle Foundations
- Inherits from data analytic approaches for high-dimensional brain time series
- Borrows from machine learning, particularly variational inference
- Formal statement: existential imperatives for surviving systems can be cast as inference problem
- Probability of existing = evidence that you exist
- Statistical problem approach enables using inference mathematics for ensemble dynamics
- Anything existing separate from environment minimizes variational free energy
- Variational free energy = negative evidence lower bound (ELBO) from machine learning
- Maximizing marginal likelihood/evidence for own existence
Existence vs Living Systems
- Existence precedes living - fundamental property of separated systems
- Non-equilibrium steady state physics mathematics required
- If something exists, must exhibit optimization properties by mathematical necessity
- Optimizing quantity identical to evidence lower bound, Bayesian model evidence
- Bound on surprisal/self-information from information theory
- Log probability of being in particular state
System Examples and Boundaries
- Simplest example: single-celled organism identifiable from environment
- Oil drop in solvent as basic case study
- Interface persistence over time - why doesn't oil dissolve?
- Markov blanket defines boundary between internal/external states
- Four types of states: internal, external, blanket (surface), active states
- Blanket states carved into sensory and active states
- Active states not influenced by external states = autonomy
- Internal + active states = autonomous dynamics
Movement and Life Distinction
- Key difference: movement from internal states
- Oil drop with random internal states averages to no movement
- Sun example: lots of internal activity but no coordinated movement
- Living systems: structured internal dynamics enable coordinated movement
- Movement = only way to change universe (voice, walking, secreting)
- Non-random movement makes system alive
- Tadpole vs oil drop: active states changing external states
- Action-perception cycle with recurrent dynamics
- Active resampling of data through movement (chemotaxis)
Machine Learning Implications
- Current ML like oil drop with access to all data (big data approach)
- Missing active sampling/movement component
- Embodied intelligence importance not yet integrated
- Active vision with camera movement as exception
- Tadpole model: swimming to find right data vs passive exposure
Consciousness and Self-Awareness
- Planning capability as key distinction
- Virus has agency but no planning
- Ant has purposeful behavior beyond virus
- Consciousness may be vague concept (pile of sand analogy)
- Planning requires model of future consequences
- Planning as inference with generative models
- Movement essential for consciousness contenders
- Free will emerges from selecting among action policies
Self-Awareness Requirements
- Operates in world populated by similar creatures
- Social world necessitates self/other distinction
- No need for self-awareness if alone (Mars/feral child examples)
- Ambiguity from others requires "me vs you" inference
- Generative models must account for copies of self
- Turn-taking in conversation as highest planning manifestation
- Theory of mind: model of other's model
- Recursive modeling enables discourse
Practical Applications
- Write down generative model probabilistically
- Engineer gradient descent on objective function
- Heavy lifting in specifying generative model structure
- Similar to natural selection: explains existence but not specific features
- Tautological nature: things exist because they minimize free energy
- Practical value in creating self-evidencing artifacts
Meaning and Purpose
- Objective function: minimize expected free energy
- Personal level: fulfilling beliefs about self-type
- Childhood narratives/scripts provide framework
- Cultural stories define identity (princess/beast fairy tales)
- Active inference: creating own culture over time
- Personal example: Einstein/Sherlock Holmes mixture
- Journey of exploration narrative from childhood influences
