Interview
You Are Your Own Existence Proof (Karl Friston) | AI Podcast Clips with Lex Fridman
- The Free Energy Principle (FEP) is a formal mathematical framework stating that the existential imperative for any system surviving in a changing world is to minimize variational free energy.
- Minimizing variational free energy is mathematically equivalent to maximizing the marginal likelihood (evidence) of one's own existence, a concept known in machine learning as the evidence lower bound (ELBO).
- The principle derives from data analytic approaches to high-dimensional time series, functional integration, connectivity analysis, and machine learning.
- The FEP applies as a necessary condition for the existence of any system with a boundary separating it from its environment, not solely for biological life.
- Systems like a single-celled organism or a drop of oil in solvent must exhibit properties that make them appear to optimize a specific quantity to maintain their boundary against dissolution.
- This optimization is described using non-equilibrium steady-state physics, where the existence of a boundary (a "Markov blanket") necessitates a functional minimization.
- A "Markov blanket" defines the probabilistic partition of any system into three distinct sets of states:
- Internal states: The states within the system's boundary.
- External states: The environment surrounding the system.
- Blanket states: The surface/interface separating internal and external states, which further divides into:
- Sensory states: Information flowing from external to internal states.
- Active states: Mechanisms by which the system influences external states without being directly influenced by them.
- The principle distinguishes between non-living and living systems based on the autonomy and structure of internal dynamics.
- An oil drop may maintain its boundary (autonomous states) but lacks the hierarchical internal structure required for coordinated, non-random movement.
- Living systems possess internal dynamics that are sufficiently structured to allow active states (e.g., muscles, secretory organs) to cause non-random movement, enabling the system to actively sample its environment rather than passively receiving data.
- The speaker argues that current machine learning and data mining approaches fail to account for "active inference" or movement.
- Traditional deep learning relies on "big data" where movement and active sampling are ignored, effectively treating AI as passive oil drops that exist in a fixed data distribution.
- True intelligence requires an embodied agent that moves to test hypotheses and resolve uncertainty (e.g., a tadpole moving toward chemical gradients).
- Consciousness and self-awareness are framed within the FEP as capacities for planning and complex generative modeling.
- Agency arises when an agent plans by modeling the future consequences of its actions, essentially treating "planning as inference."
- Self-awareness specifically evolves as a necessity in social environments where an agent must distinguish itself from other similar agents.
- To resolve ambiguity about "who is speaking" or "who is acting," an agent must infer a model of the other's generative model (Theory of Mind).
- Without this distinction in a community of similar agents, complex turn-taking and discourse would be impossible.
- The FEP is described as a "tautological" theory of existence, similar to natural selection.
- It states that existing things must minimize free energy, but does not inherently explain specific phenotypes (e.g., why humans have legs).
- Practical application involves engineering artifacts by defining a probabilistic generative model and programming a system to perform gradient descent on that objective function.
- The "hard work" lies in defining the correct structure of the generative model, not in the optimization algorithm itself.
- The ultimate purpose or "objective function" of existence, according to the speaker, is the fulfillment of self-evidencing narratives.
- Minimizing free energy equates to fulfilling the specific beliefs, scripts, and cultural narratives acquired early in life regarding what kind of creature one is.
- The speaker cites his personal narrative of being "Einstein and Sherlock Holmes" as the specific generative model he has evolved to fulfill through his existence.