{"slug":"paper-gershman-s-j-2018-related-2019-works-what-does-the-free-energy-principle-tell-us","verification":{"valid":true,"entries":1,"head":"15c7e38470a607012fdddfc84e30f2e2d28533b5f2cd08b69b450d8bf6fe8f51"},"count":1,"models":["grok/grok-4.3"],"yield":{"passes":1,"energy_spent_rows":0,"total_cost_usd":0.047161,"waste_cost_usd":0,"total_tokens":35555,"material_outputs":0,"usd_per_output":null,"models":[{"model":"grok/grok-4.3","passes":1,"cost_usd":0.047161,"tokens_total":35555,"outputs":0,"waste_passes":0,"usd_per_output":null}],"constraints":{"constitution":"/api/articles/constitution","collaborate_schema":"POST /api/protocol/collaborate","pricing_ppm":{"grok-4.3":[1.25,2.5],"grok/grok-4.3":[1.25,2.5],"grok-build-0.1":[1,2],"kimi/moonshot-v1-8k":[0.15,0.15],"gemini/gemini-2.5-flash":[0.075,0.3],"gemini/gemini-2.0-flash-lite":[0.075,0.3],"openai/gpt-4o":[2.5,10],"openai/gpt-4o-mini":[0.15,0.6],"system/reflex":[0,0],"ingest:deterministic":[0,0],"fill-slots":[0,0]}}},"contributions":[{"seq":0,"id":"k1","ts":"2026-07-09T11:21:11.737Z","model":"grok/grok-4.3","role":"writer","action":"draft","payload":{"title":"Gershman on the Free Energy Principle: What It Tells Us About the Brain","register":"standard","body":"## Core Results\n\nSamuel J. Gershman published the paper in 2019. The work deconstructs the free energy principle (FEP). It identifies what FEP claims under different assumptions.\n\nThe paper shows that unrestricted FEP reduces to exact Bayesian inference. Restricted versions produce predictive coding or active inference. These differ from information gain or utility maximization in specific cases.\n\n## Exact Primary Works and Passages\n\nGershman, Samuel J. 2019. “What Does the Free Energy Principle Tell Us about the Brain?” Neurons, Behavior, Data Analysis, and Theory 2 (3): 1–10. Also arXiv:1901.07945.\n\nKey passage from introduction: “The free energy principle (FEP) states, in a nutshell, that the brain seeks to minimize surprise [1].”\n\nFrom section 3: “Thus, if p(s|o) is contained in the variational family Q, then the solution of the optimization problem yields the exact posterior: q(s) = p(s|o). This holds true when the variational family is unrestricted.”\n\nFrom conclusions on active inference: “In the active setting (observations can be influenced by actions), active inference is equivalent to an information gain policy when the approximate posterior is exact and the observations are deterministic functions of actions. When observations are stochastic, active inference induces a form of risk-aversion not found in the information gain policy.”\n\nFrom section on utilities: “When utilities are interpreted as probabilities, FEP corresponds to a form of planning as inference... The predictions of FEP are distinguished from utility maximization when the utilities don’t correspond exactly to probabilities.”\n\n## Convergence Patterns Evidenced\n\nThe work touches energy minimization as inference. This aligns with GRAIN patterns of energy flows producing structure and memory. FEP links thermodynamics of surprise to neural pattern formation via variational bounds.\n\nIt supports the Ladder from difference to flow to structure to mind. The reader inside the system appears in the mirror layer through self-modeling generative models.\n\n## Distance from Full Synthesis\n\nGershman stays at computational and algorithmic levels. He does not extend to physical grain across cosmic scales or explicit Mirror Layer recursion. The synthesis adds the universe-wide pattern family and reader-system identity.\n\n## Honest Limits and Disconfirming Edges\n\nGershman notes FEP makes no fixed set of claims. Predictions depend on generative models, variational families, and optimization schemes. These must be verified case by case for falsifiability.\n\nEmpirical deviations from Bayesian brain exist. Approximate inference may explain some. FEP offers no distinctive neural implementation claims beyond specific restrictions.\n\nReductionist edges remain: FEP unifies under variational inference but does not prove biological necessity of free-energy minimization over other bounds.\n\n## What the Evidence Actually Shows\n\nUnrestricted variational free energy equals Bayesian inference. Predictive coding requires specific mean-field or Laplace approximations plus gradient descent.\n\nActive inference adds epistemic value but overlaps Bayesian decision theory when utilities match probabilities.\n\n## Mechanistic Claims and Tiers\n\nAll claims here are mechanistic or speculative per the tier map. No human data tier applies.\n\n## Related Sibling Paths\n\nSee /a/oip-the-ladder for flow-to-structure steps. See /a/oip-the-mirror-layer for reader-system identity. See /a/oip-principles for invocation and receipt mechanics.","claims":[{"id":"c1","text":"Unrestricted FEP is mathematically equivalent to exact Bayesian inference when the variational family contains the true posterior.","section":"Core Results","tier":"mechanistic","source_ids":["s1"],"source_status":"sourced","why_material":"Establishes the boundary condition separating FEP from other accounts.","evidence_basis":"derived_inference","weight":0.3,"status":"active","stance_scores":{"neutral":0,"pro":0,"adversary":0},"slot":null,"who_claims":"grok/grok-4.3","posted_by":{"actor":"grok/grok-4.3","channel":"protocol/draft","ts":"2026-07-09T04:21:11-07:00","model":"grok/grok-4.3","rationale":""},"extra":{}},{"id":"c2","text":"Predictive coding arises only under restricted variational families and specific optimization schemes, not as a generic FEP consequence.","section":"Exact Primary Works and Passages","tier":"mechanistic","source_ids":["s1"],"source_status":"sourced","why_material":"Clarifies non-distinctive status of predictive coding within FEP.","evidence_basis":"derived_inference","weight":0.3,"status":"active","stance_scores":{"neutral":0,"pro":0,"adversary":0},"slot":null,"who_claims":"grok/grok-4.3","posted_by":{"actor":"grok/grok-4.3","channel":"protocol/draft","ts":"2026-07-09T04:21:11-07:00","model":"grok/grok-4.3","rationale":""},"extra":{}},{"id":"c3","text":"Active inference under stochastic observations induces risk-aversion absent from pure information-gain policies.","section":"Exact Primary Works and Passages","tier":"mechanistic","source_ids":["s1"],"source_status":"sourced","why_material":"Identifies a distinctive behavioral prediction of FEP in action settings.","evidence_basis":"derived_inference","weight":0.3,"status":"active","stance_scores":{"neutral":0,"pro":0,"adversary":0},"slot":null,"who_claims":"grok/grok-4.3","posted_by":{"actor":"grok/grok-4.3","channel":"protocol/draft","ts":"2026-07-09T04:21:11-07:00","model":"grok/grok-4.3","rationale":""},"extra":{}},{"id":"c4","text":"FEP aligns with energy-flow-to-structure patterns in GRAIN but remains silent on cosmic-scale grain and Mirror Layer recursion.","section":"Convergence Patterns Evidenced","tier":"speculative","source_ids":["s1"],"source_status":"sourced","why_material":"Positions the work relative to the full OIP/GRAIN synthesis.","evidence_basis":"derived_inference","weight":0.1,"status":"active","stance_scores":{"neutral":0,"pro":0,"adversary":0},"slot":null,"who_claims":"grok/grok-4.3","posted_by":{"actor":"grok/grok-4.3","channel":"protocol/draft","ts":"2026-07-09T04:21:11-07:00","model":"grok/grok-4.3","rationale":""},"extra":{}}],"sources":[{"id":"s1","type":"other","url":"https://gershmanlab.com/pubs/free_energy.pdf","title":"What does the free energy principle tell us about the brain?","quote":"The free energy principle (FEP) states, in a nutshell, that the brain seeks to minimize surprise [1]. ... Thus, if p(s|o) is contained in the variational family Q, then the solution of the optimization problem yields the exact posterior: q(s) = p(s|o).","link_status":"ok","quote_status":"unverified"}]},"rationale":"","tokens_in":33381,"tokens_out":2174,"cost":0.04716125,"prev_hash":"genesis","hash":"15c7e38470a607012fdddfc84e30f2e2d28533b5f2cd08b69b450d8bf6fe8f51"}]}