Research / Free energy, entropy, and life

Free energy, entropy, and life

Free energy, entropy, and life: why organisms predict — from Schrödinger to Friston to Seth

2026-03-05 · consciousness-cognition ai-architecture agentic-ai · medium · source → · wiki →
key claims
  1. Schrödinger's negentropy argument establishes that life requires thermodynamic openness, not a violation of the second law. Organisms maintain local order by continuously importing negative entropy (negentropy) from their environment and expelling disorder as heat and waste. Life is defined by sustained departure from thermodynamic equilibrium — death is equilibrium. This is the thermodynamic precondition for all subsequent arguments about prediction and consciousness
  2. Friston's variational free energy (VFE) is formally derived from Shannon entropy and shares mathematical structure with Boltzmann entropy; the connection to thermodynamics is structural, not merely analogical. VFE = KL − ln p(o), where minimising it pulls the organism's internal model toward an accurate representation of the world while bounding surprise (negative log model evidence). Both Shannon and Boltzmann entropy have the form −Σ p log p; the information-theoretic free energy concept derives from Jaynes's 1957 maximum entropy reformulation of statistical mechanics
  3. The Markov blanket is the formal structure of "maintaining a boundary with the environment," and any self-organising system with a stable Markov blanket necessarily minimises variational free energy — this is a mathematical result, not an empirical hypothesis. Friston is explicit: FEP is a principle like Hamilton's principle of stationary action; it cannot be falsified because it is true by mathematical construction. What are falsifiable are the process theories derived from FEP — predictive coding, active inference — which have empirical content
  4. Active inference extends passive predictive coding by incorporating action as a form of prediction fulfillment. In predictive coding, the brain minimises prediction error by updating beliefs (perception). In active inference, the brain also minimises prediction error by acting on the world to make sensory inputs match predictions. This distinction matters: active inference is the biologically realistic account of organisms that act, not just perceive, and it frames action and perception as two sides of a single inferential process
  5. Prediction is specifically required — not just metabolism — because allostatic regulation is more efficient than homeostatic correction. A purely reactive homeostatic system corrects deviations after they occur; an allostatic (predictive) system anticipates deviations and acts before they happen. Brains evolved as allostatic organs: they predict what the body will need before it needs it. This is why thermostats (homeostatic) are not sufficient for life at organism scale, and why the gap between Schrödinger's negentropy and Friston's FEP is bridged by the allostasis argument
  6. Seth explicitly distinguishes FEP (a functional description of what self-organising systems do) from his consciousness claim (a phenomenal claim about what embodied biological organisms feel). In Seth's own words: "The free-energy principle is not itself a theory about consciousness." Seth uses FEP to explain *why* brains predict (they need to maintain viable biological states) and *what* specifically they predict (interoceptive body states, not just external sensory inputs). The phenomenal claim — that there is something it is like to run these predictions — is not derivable from FEP and requires the additional condition of biological embodiment in a survival context
  7. The thermostat problem — that FEP applies to thermostats and candle flames as well as brains — is not dissolved by Seth's account; it is re-located. FEP does apply to any system with a Markov blanket. Seth's response is not to dispute this but to argue that consciousness requires a richer condition: biological, embodied, survival-oriented interoceptive prediction. A thermostat has no body to model for survival. This is Seth's extra condition — it narrows the domain of FEP that is relevant to consciousness without denying FEP's breadth
  8. The tautology objection to FEP is partially answered but not eliminated. The objection is that FEP describes any persistent system post hoc, making it vacuously true. Friston answers: FEP is a mathematical principle and was never claimed to be falsifiable at the general level; it functions like a modelling language. What can be tested — and often is — are specific implementations: predictive coding hierarchies, active inference agents, precision-weighting accounts of psychiatric disorders. The principle itself is not a theory; the process theories built from it are

Research Question

Why do living organisms need predictive brains? What is the precise relationship between the thermodynamic concept of entropy (disorder), the information-theoretic concept of free energy (surprise), Karl Friston's free energy principle, and Anil Seth's claim that consciousness is the subjective instrument of life's war against entropy?

Findings

Executive Summary

Living organisms need predictive brains because prediction is the efficient mechanism for maintaining the low-entropy states that life requires: any organism that cannot anticipate and correct deviations in its internal states will eventually reach thermodynamic equilibrium with its environment — death. The mathematical bridge from thermodynamic entropy to Friston's variational free energy is formal, not metaphorical, resting on the structural equivalence of Shannon and Boltzmann entropy demonstrated in information physics; Friston's free energy principle establishes that any system with a stable statistical boundary (Markov blanket) necessarily minimises an upper bound on its own surprise, and this is formally analogous to minimising thermodynamic free energy. Seth's beast machine thesis adds a phenomenal layer FEP does not provide: consciousness is what it is like to be a biological, embodied organism running survival-oriented predictive models — specifically interoceptive models that regulate the body's internal states — and this phenomenal dimension is not derivable from the thermodynamic or information-theoretic layers alone. The three-layer chain (Schrödinger → Friston → Seth) is logically coherent but the transitions between layers are motivated conjectures, not deductive entailments: thermodynamics does not entail prediction, and FEP does not entail phenomenal experience.

Key Findings

  1. Schrödinger's negentropy argument establishes that life requires thermodynamic openness, not a violation of the second law. Organisms maintain local order by continuously importing negative entropy (negentropy) from their environment and expelling disorder as heat and waste. Life is defined by sustained departure from thermodynamic equilibrium — death is equilibrium. This is the thermodynamic precondition for all subsequent arguments about prediction and consciousness. [confidence: high]

  2. Friston's variational free energy (VFE) is formally derived from Shannon entropy and shares mathematical structure with Boltzmann entropy; the connection to thermodynamics is structural, not merely analogical. VFE = KL[q(x) ∥ p(x|o)] − ln p(o), where minimising it pulls the organism's internal model toward an accurate representation of the world while bounding surprise (negative log model evidence). Both Shannon and Boltzmann entropy have the form −Σ p log p; the information-theoretic free energy concept derives from Jaynes's 1957 maximum entropy reformulation of statistical mechanics. [confidence: high]

  3. The Markov blanket is the formal structure of "maintaining a boundary with the environment," and any self-organising system with a stable Markov blanket necessarily minimises variational free energy — this is a mathematical result, not an empirical hypothesis. Friston is explicit: FEP is a principle like Hamilton's principle of stationary action; it cannot be falsified because it is true by mathematical construction. What are falsifiable are the process theories derived from FEP — predictive coding, active inference — which have empirical content. [confidence: high]

  4. Active inference extends passive predictive coding by incorporating action as a form of prediction fulfillment. In predictive coding, the brain minimises prediction error by updating beliefs (perception). In active inference, the brain also minimises prediction error by acting on the world to make sensory inputs match predictions. This distinction matters: active inference is the biologically realistic account of organisms that act, not just perceive, and it frames action and perception as two sides of a single inferential process. [confidence: medium]

  5. Prediction is specifically required — not just metabolism — because allostatic regulation is more efficient than homeostatic correction. A purely reactive homeostatic system corrects deviations after they occur; an allostatic (predictive) system anticipates deviations and acts before they happen. Brains evolved as allostatic organs: they predict what the body will need before it needs it. This is why thermostats (homeostatic) are not sufficient for life at organism scale, and why the gap between Schrödinger's negentropy and Friston's FEP is bridged by the allostasis argument. [confidence: medium]

  6. Seth explicitly distinguishes FEP (a functional description of what self-organising systems do) from his consciousness claim (a phenomenal claim about what embodied biological organisms feel). In Seth's own words: "The free-energy principle is not itself a theory about consciousness." Seth uses FEP to explain why brains predict (they need to maintain viable biological states) and what specifically they predict (interoceptive body states, not just external sensory inputs). The phenomenal claim — that there is something it is like to run these predictions — is not derivable from FEP and requires the additional condition of biological embodiment in a survival context. [confidence: high]

  7. The thermostat problem — that FEP applies to thermostats and candle flames as well as brains — is not dissolved by Seth's account; it is re-located. FEP does apply to any system with a Markov blanket. Seth's response is not to dispute this but to argue that consciousness requires a richer condition: biological, embodied, survival-oriented interoceptive prediction. A thermostat has no body to model for survival. This is Seth's extra condition — it narrows the domain of FEP that is relevant to consciousness without denying FEP's breadth. [confidence: high]

  8. The tautology objection to FEP is partially answered but not eliminated. The objection is that FEP describes any persistent system post hoc, making it vacuously true. Friston answers: FEP is a mathematical principle and was never claimed to be falsifiable at the general level; it functions like a modelling language. What can be tested — and often is — are specific implementations: predictive coding hierarchies, active inference agents, precision-weighting accounts of psychiatric disorders. The principle itself is not a theory; the process theories built from it are. [confidence: high]

  9. Seth's beast machine thesis generates indirect but tractable empirical predictions despite the hard problem remaining unsolved. Predicted test domains include: anaesthesia should preferentially disrupt interoceptive predictive hierarchies; disorders of consciousness (vegetative state, dissociative states) should exhibit aberrant interoceptive precision-weighting; creatures with richer interoceptive machinery should show behavioural signatures of richer subjective experience. None of these are direct tests of phenomenal experience, but they are tests of the interoceptive-prediction account. [confidence: medium]

  10. The three-layer argument (Schrödinger → Friston → Seth) is logically coherent but transitions between layers are motivated conjectures, not deductive entailments. Thermodynamics does not entail prediction — Schrödinger establishes the precondition; Friston formalises one efficient mechanism for meeting it. FEP does not entail phenomenal experience — Seth locates where consciousness fits in the biological economy without deriving it from physical principles. The chain is the strongest available account connecting physics to life to consciousness; it is not a proof.

Assumptions

Analysis

The three-layer argument is the strongest available account connecting physics to life to consciousness, but each transition between layers requires additional premises. Schrödinger establishes why life needs to resist entropy; he does not establish how prediction achieves this more efficiently than, say, robust physical shielding. Friston provides the how — via the formal machinery of variational inference — but at the cost of generality: FEP is true of every self-organising system, not just biological ones. Seth rescues the specificity by anchoring the account to interoceptive prediction in biological organisms with survival stakes — but this move is a motivated narrowing, not a mathematical derivation.

The key intellectual achievement of this three-layer picture is that it relocates the mystery of consciousness. Instead of asking "why does brain activity produce experience?" (Chalmers' hard problem, which has made no empirical progress), Seth asks "what is the specific kind of predictive process that produces this specific structure of conscious experience?" The answer — interoceptive, survival-oriented, allostatic prediction — generates testable predictions and a research programme. Whether it ultimately dissolves the hard problem or merely defers it remains genuinely open.

Risks, Gaps, and Uncertainties

Open Questions

  1. Does any published derivation go formally from the second law of thermodynamics to the claim that organisms with brains minimise VFE, in a single mathematical chain? Or is this always an assemblage of three separately-motivated claims? (This would be a meta-scientific question about the FEP literature.)
  2. The allostasis/prediction argument — that prediction is specifically required because reactive homeostasis is insufficient — deserves a standalone deep-dive with empirical support.
  3. How does IIT's information-theoretic account of consciousness compare to Seth's FEP-grounded account? Both use entropy/information concepts; their relationship is unstated in existing research items.
  4. Seth's indirect empirical predictions (anaesthesia, disorders of consciousness, interoceptive precision) — is there a review paper that tests these specifically? This could validate or constrain the beast machine thesis.

Output

Open Questions

  1. Is there a crisp mathematical derivation that goes from the second law of thermodynamics all the way to the claim that organisms minimise prediction error? Or is it an analogy?
  2. How does active inference (acting to reduce surprise) differ from passive predictive coding (updating beliefs)? Which is more relevant to consciousness?
  3. Does the FEP entail that any self-maintaining system (a thermostat, a candle flame) is proto-conscious? Seth says no — what is the extra condition?
  4. What are the failure modes? What would falsify the claim that consciousness is entropy-resistance?
  5. How does this connect to Integrated Information Theory (IIT) — another thermodynamic/information-theoretic approach?

sources

← newerThe hard problem vs. the real problem of consciousness
older →Exploit versus explore Artificial Intelligence (AI) investment classification

Connected items

Loading…

View full knowledge graph →