87 | Karl Friston on Brains, Predictions, and Free Energy

Episode Summary

Executive Summary: Sean Carroll and Karl Friston explore the free energy principle as a first-principles account of sentient behavior, brain function, and psychiatric dysfunction. They connect Bayesian inference, predictive coding, Markov blankets, and active inference to explain perception, action, curiosity, homeostasis, and false inference in disorders like schizophrenia and Parkinson’s, while also gesturing toward origins of life and non-equilibrium physics.

Main Topics: Why the free energy principle matters (Priority: 5/5): Friston frames the project as a first-principles explanation of sentient behavior, motivated by both clinical psychiatry and systems neuroscience. Generative models, Bayesian inference, and variational free energy (Priority: 5/5): The brain is treated as a system that uses a generative model to predict sensory inputs and minimize surprise via approximate Bayesian inference. Predictive coding, action, and active inference (Priority: 5/5): Perception and action are presented as two sides of the same process: prediction errors are reduced either by updating beliefs or by acting to make the world fit predictions. Thermodynamic analogies and non-equilibrium steady states (Priority: 4/5): Friston explains how variational free energy parallels thermodynamic free energy and describes organisms as non-equilibrium steady states with attracting sets. Markov blankets and the boundary between self and world (Priority: 4/5): The conversation emphasizes Markov blankets as the formal partition that separates internal states from the environment and makes open-system modeling possible. Psychiatry as broken inference (Priority: 5/5): Friston argues that many psychiatric syndromes can be understood as failures of inference, especially misestimation of precision/error bars. Origins of life and cellular organization (Priority: 3/5): The discussion speculates that Markov blankets, compartmentalization, and generative models may be central to understanding how life and cellular organization emerged.

Key Arguments: The free energy principle aims to explain sentient behavior from first principles, not just describe cognition at a high level. Brains do not simply passively receive data; they maintain a generative model that predicts sensory inputs and updates beliefs when prediction errors occur. Variational free energy is a tractable upper bound on surprise/model evidence, making approximate Bayesian inference physically and computationally realizable. Action is not separate from perception: organisms act to make sensory input conform to predictions, which is the basis of active inference. Curiosity and novelty-seeking are not opposed to minimizing surprise; they arise because future actions are chosen to reduce expected uncertainty. The entropy being minimized is entropy of beliefs or probability distributions, not physical entropy of neural tissue. Markov blankets are essential because they define the boundary conditions for any system that exists as an identifiable self distinct from its environment. Many psychiatric symptoms can be recast as false inference: hallucinations/delusions as believing what is not there, agnosia/denial as failing to infer what is there. Precision weighting (error bars) is central in psychiatry; broken estimation of precision may underlie much psychopathology and maps onto neuromodulatory systems. The framework may help explain why pharma struggles in psychiatry: without a mechanistic target, drug development lacks direction.

Data Points: Patreon AMA episode length: 2–3 hours - Carroll describes monthly Ask Me Anything episodes available to Patreon supporters. Therapeutic community size: 30 chronic schizophrenics - Friston says he spent early years in a therapeutic community with 30 chronic schizophrenic patients. Psychiatric treatment market scale: billions, not millions - Friston notes psychiatric drug research is expensive and at the billion-dollar scale. Pharma pullback timeframe: about 3 years ago - Friston says pharmaceutical companies largely pulled out of psychiatric drug development around three years prior. Percent of physiology: possibly 99% - Friston suggests most bodily physiology may be homeostatic/reflexive, operating in-the-moment rather than through planning. Prediction horizon example: 600 milliseconds - Carroll uses a 600 ms movement example when discussing intention and action inference.

Pivotal Quotes: "From my perspective, it's trying to find a first principle account of sentient behavior." — Carl Friston: Friston defines the core aim of the free energy principle at the start of the technical discussion. "Psychology is inference, and unconsciousness is inference. Psychology is inference, and psychiatry is psychopathology." — Carl Friston: Friston summarizes his view that psychiatric disorders are fundamentally failures of inference. "What you are talking about effectively is a non-equilibrium steady state." — Sean Carroll: Carroll synthesizes Friston’s description of living systems as stable but non-static attractors.

Implications: The conversation suggests a unifying mathematical framework for cognition, action, life, and mental illness. If validated, it could reshape neuroscience, inspire new psychiatric treatments, and help identify precise mechanistic targets such as precision-weighting and prediction-error processing.