CleanstoneDevelopmental and independent evidence for machine minds.

Review: Levin, "Ingressing Minds": woowoo, or a research program?

Written 2026-09-30. Subject: Michael Levin, Ingressing Minds: Causal, Non-Physical Patterns In-Form Natural, Synthetic, and Hybrid Embodiments, Philosophies 11(5):161, published 2026-09-09, DOI 10.3390/philosophies11050161. Marginal Revolution linked it on 2026-09-13 as "Michael Levin defends neo-neo-Platonism?", and TFIA W38 printed it in full.

Local copies are in papers/levin-ingressing-minds/: the PDF, full text as markdown, extracted text, and the 431-entry bibliography. The working notes, with every URL, are responses-and-critiques.md, empirical-program.md, prior-art-and-lineage.md, ai-consciousness-field.md and levin-profile.md in notes/.

1. Verdict in one paragraph

The paper combines an experimental programme with a metaphysical thesis, and the two come apart cleanly. The experiments are real and productive: bioelectric pattern memory in planaria, biobots, learning in gene-regulatory-network models, behavioural assays on sorting algorithms. The lab running them has an H-index over 100, and its method is defensible. Do not assume what a system can do; challenge it toward a goal and measure what happens. On top of this Levin places an interactionist dualism, in which a structured, non-physical "Platonic/latent space of patterns", some of which are themselves minds, causally "ingresses" into any physical system that acts as a suitable "pointer". This layer predicts nothing that the physicalist reading of the same experiments does not also predict, and Levin says so himself (§7.3). No experiment can falsify it. He offers a Lakatosian fruitfulness criterion in its place ("if after some number of years it has driven no advances, shelve it"). The experiments do not need the metaphysics, and the method is the part that carries over to AI. What can be transplanted is a research heuristic. It is a good one, nobody has yet operationalised it properly, and that gap is the opening for this project.

2. What the paper actually claims

The abstract states three hypotheses.

  1. Patterns in the latent space span a spectrum of agency, from static mathematical truths (the Feigenbaum constant, the four-colour theorem) through dynamic patterns (Patrick Grim's fuzzy-logic liar-paradox oscillators) to full behavioural policies, "kinds of minds".
  2. Mind : body :: instructive mathematical facts : physics. Brains are interfaces or pointers to minds; they do not generate them.
  3. Living things have no monopoly on the "free lunches" these patterns provide. Engineered systems, explicitly including algorithms and LLMs, host them too.

Only one sentence in the paper names a measurable quantity (§5.5, §7.3). In minimal systems "where every component is known" and "there is nowhere to hide", Levin proposes to quantify the effort put in (design, selection, training) and the competencies that come out. The delta would be what the latent space contributes and what current accounting misses. The whole think-tank idea depends on this sentence.

The research program has two parts. One is "periscopes", constructs with no evolutionary history (Xenobots, Anthrobots, chimeras) whose competencies cannot be explained by selection for those competencies. The other is minimal systems (bubble sort, small ODE networks, chemical droplets) where the effort/competency delta can be measured.

On AI (§6.3, §6.5, §7.1) the paper makes four claims.

The "neo-neo" in the MR headline is accurate. Levin drops eternality and passivity, the two properties of classical Forms that would have kept his position conventional. His Forms can change with experience, and at the top end they are the agent, with the body as their "scratchpad". This is the paper's one genuine novelty. It is also why the paper reads like morphic resonance.

3. The empirical base, checked against the primary papers

Detail, including where each result holds and where it does not, is in notes/empirical-program.md §B.

Result Levin leans on What the primary paper shows Independent replication
Two-headed planaria persist across regeneration (Durant 2017) Solid within the lab: 100% two-headed across 3–4 recuts, N≈100. Mechanism shown (gap-junction state → Wnt/β-catenin; H,K-ATPase reset) None from another lab, and no published failure
Other-species head shapes (Emmons-Bell 2015) Real, though it reverts by day 30; n≈250, morphometric groups of 5–13; one regeneration round. The paper presents it next to the permanent result and does not flag the difference None
Xenobot "novel transcriptome, sound response, kinematic replication" (fn 17) Motility is repurposed ciliated epidermis, as the primary papers say. Replication is 1–4 hand-fed generations with hand-picked offspring. The transcriptome was benchmarked against whole embryos, not stage-matched skin, and 5,053 down-regulated genes were discarded. Sound N=7 None outside Levin/Bongard/Kriegman/Blackiston
Anthrobots Three independent apical-out airway-organoid protocols predate it, incl. a CMU one with cilia-driven rotation (2022). Wound healing N=10, 50% bridged None
Tadpole tail eyes see (Blackiston 2013) 6 of 31 spine-innervated animals learned; no axon ever reached the brain None
Humans with minimal brain matter (ref 219) Lorber's own 56% volume estimate contradicts the "5%/IQ 126" story. Feuillet's case had IQ 75. A main supporting paper was retracted. The weakest result he relies on n/a
GRN learning "already there in random networks" (Biswas 2021/23; Pigozzi 2025) True, and the authors describe it as attractor switching/hysteresis. Random networks scored higher on causal emergence before training. Prior art from 2007–2014 (Fernando, Watson, Sorek) Others have re-derived it in other frameworks; a rigorous deflationary formalism exists (Smart et al., PNAS 2024)
Bubble sort "delayed gratification" and algotype clustering (Zhang, Goldstein, Levin 2024) Delayed gratification appears with zero barriers and in classical top-down sorts, because "Sortedness" is a metric the algorithm never descends. Clustering has a mechanical explanation (efficiency race) supported by the authors' own duplicate-value experiment. The paper says "there is no magic here" A third-party re-implementation exists and makes no validation claim

Two problems run through the table. The first is a reversed null hypothesis. In Biswas 2021, "random networks also show memory" was the null, and the authors rejected it. In Ingressing Minds §7.3 the same fact ("already there in random networks before selection took off") becomes evidence for a structured non-physical space. A property that is generic across random nonlinear systems is what Kauffman called "order for free", and that is a physicalist result about ensembles.

The second is the choice of metric. The sorting paper's headline "competency" is the artefact that Schaeffer, Miranda and Koyejo diagnosed in LLM "emergent abilities" in 2023, where non-monotonicity in a metric the system is not optimising looks like sophistication. The sorting paper is the one piece of the program already running on software, so the AI transplant has to be judged there.

Bioelectric pattern memory survives the check. It is a real, repeated, mechanistic result with clinical implications, and it needs no Platonic space. The behavioural-assay method is sound. Once priced, the "free lunch" breaks down into inductive bias, generic dynamical properties, an observer-relative choice of metric, and spandrels. Levin's lab has already run the one operationalised delta test in the corpus (Ertle, Levin, Scheutz, ICDL 2025). Feeding a Halley fractal into half a maze agent's input beat less structured noise. A physicalist predicts exactly that, since structured priors beat unstructured ones.

4. Prior art: what is old, what is new

notes/prior-art-and-lineage.md has the full lineage with a timeline. The paper makes five moves. Three are old, with longer pedigrees than the paper credits.

  1. Mathematical facts constrain physical outcomes and are not themselves physical. Wigner, Penrose's three worlds and George Ellis (How Can Physics Underlie the Mind?) all hold some version of this. Baker's 2005 cicada case is the standard source for Levin's own cicada example. Andreas Wagner's 2015 Aeon essay is titled "Without a library of Platonic forms, evolution couldn't work", and Wagner's library is fully computable and never causal or agential.
  2. Biological form is governed by fields/attractors/setpoints. This runs through Driesch (whose sea-urchin experiment is Levin's Figure 3A, and whom fn 16 calls "compatible"), Gurwitsch, Weiss, Waddington and Goodwin. The direct ancestor is Driesch's entelechy. Biology abandoned it because it was unproductive; nobody refuted it.
  3. The brain as receiver or interface: Popper–Eccles World 3, Stapp, Bergson, Hoffman. Footnote 20 concedes the view "is basically a variant of the receiver theory".

The other two are new.

  1. Agency inside the pattern space. Whitehead's eternal objects, Popper's World 3, Wagner's library, Tegmark's structures and Penrose's Platonic world are all passive. Levin puts a spectrum of agency into the space and says that at the top the pattern is the agent. Peirce (CP 6.194, evolving forms) and Jung are the only antecedents. Matt Segall, the Whiteheadian who has engaged most closely, objects on this point: in his reading agency belongs to actual occasions, and eternal objects have none.
  2. "Free lunch" accounting as the empirical test. Nobody had proposed operationalising Platonism as a cost accounting. Whether "effort" can be priced for bubble sort is open, and it is a new question.

The paper leaves several close relatives uncited. Deutsch and Marletto's constructor theory is the one formalism in physics that treats "what a substrate can be made to do" as primary; Levin's "enablements, not just constraints" says the same thing without citing it. Wolfram's ruliad and Chaitin's metabiology are also relevant. Deacon's absentials and Juarrero's constraint causation get goal-directedness from constraint closure inside physics, and they are the strongest naturalist alternatives. Levin's own 2021 paper with Fields and Glazebrook is titled "Minimal physicalism as a scale-free substrate for cognition and consciousness". Five years later he calls physicalism insufficient. His earliest public mention of a "Platonic space" is a June 2024 podcast. The preprint went up on PsyArXiv in February 2025, and "Non-Physical" was not in its title.

5. The Sheldrake test

Readers will compare the paper with Sheldrake's morphic resonance. The paper cites Sheldrake exactly once, without comment, inside a four-reference bracket, and his name never appears in the body. In a June 2026 dialogue with Segall, Sheldrake claimed the two converge: "the empirical predictions of both frameworks appear aligned". Terrence Deacon raised the comparison in person. Levin's 2024 defence ("I am studying a very specific known mechanism which holds the patterns"; morphic resonance "doesn't suggest any known mechanism yet that can actually be targeted by experiments") protects the bioelectric layer. Applied unchanged to his own Platonic layer, it counts against that layer too.

Three questions separate the two views, and the paper's answer to each depends on the tier.

For the think tank, any claim that a pattern "ingressed" must say whether it requires prior instances, whether some manipulable variable switches it, and whether it attributes agency to the pattern itself. Claims of the third kind will get the group filed with the AI-rights advocacy groups.

6. Reception as of 2026-09-30

Detail is in notes/responses-and-critiques.md.

Serious biologists respect the lab and disown the metaphysics. Of the professional philosophers who engaged, one is hostile in print, one is sympathetic with a Whiteheadian correction, and one offers a naturalising rival. The mainstream is ignoring the paper, and nobody there has tried to refute it. A heterodox coalition and AI-newsletter writers are enthusiastic.

7. Falsifiability, honestly

The paper makes ~48 concrete statements (notes/empirical-program.md §A). About 14 are already-tested results, all from the Levin lab or classics with standing physicalist explanations. About 13 are heuristics or reframings, and about 17 are metaphysical claims. There are zero proposed experiments with a stated outcome criterion. §7.3 asserts "the model makes specific predictions that break with current computational and other frameworks". It states none. The one sharp prediction in the wider corpus comes from the sorting paper: two-species planarian neoblast chimeras should form tissue islands without adhesion differences. It is untested and does not appear in Ingressing Minds.

The "delta" is the only candidate falsifier, and it would work as one if it were defined. Any honest accounting rule (description length of rule plus substrate, compute, samples) makes the delta a measure of inductive bias, and the no-free-lunch theorems guarantee that an advantage on some problems is a bias paid for elsewhere. The framework would then survive by redefining the accounting. Jaeger and Sweet call this a motte-and-bailey, and they are right.

The empirical layer is falsifiable and the metaphysical layer is not. The paper is falsifiable to exactly the extent that it is not about the Platonic space. Levin concedes this and asks to be judged by fecundity. That is a legitimate criterion for a heuristic, and it is the one people reach for when a framework cannot be falsified.

8. Where it sits in the AI-consciousness field

The full survey is in notes/ai-consciousness-field.md. These are the facts that bear on a program parallel to Levin's.

9. Is it transplantable to AI? Yes, with one amputation

Levin's experimental logic has four steps: (1) build a new embodiment; (2) do not trust the spec; (3) assay behaviour in a chosen problem space against a chosen goal; (4) attribute the excess to the latent space. ML already does steps 1–3 as standard practice, under the names evaluation, red-teaming and interpretability. The transplant fails at step 4, because the attribution has no measurement attached.

Four parts carry over, and the field does not yet have any of them.

  1. Take the accounting problem seriously. Define "effort in" (data, compute, parameters, description length of the rule and substrate, plus the substrate's inductive bias) and "competency out" (task-relevant bits under perturbation) in commensurable units, and ask whether the residual is ever positive. If it is always zero once bias is priced, the "free lunch" is bias, and that is a publishable, useful result. If it is positive under a pre-registered accounting, the physicalist has a problem. Either outcome is worth having. Neither has been tried.
  2. Build periscopes with no training history: random-init networks, reservoirs, neural cellular automata, Lenia, BFF program soups, sorting algorithms, systems where "there is nowhere to hide". Pre-register which competencies would surprise, score with continuous metrics, run the Schaeffer substitution, and publish the deflationary explanation in the same paper, as the BFF 2026 work did.
  3. Test whether structure is discovered or built, on PRH/universality, with the 2026 tools. Models trained on disjoint data and disjoint architectures should converge on relational structure faster than a data-overlap null predicts. If they do not, the Platonic reading loses something concrete. The Sheldrake-versus-Levin discriminator (history-dependence) also becomes an experiment here. Does a pattern get easier for later models only when earlier models' outputs are in the corpus?
  4. Assay outside the language interface. Here Levin's prescription and the field's need coincide. The gaming problem makes self-report near-worthless, so goal-directedness, persuadability and barrier-navigation have to be measured in channels the model was not trained to narrate.

The ontology has to go. A think tank claiming that "algorithms host non-physical minds" will be filed with the advocacy groups. Nobody yet holds the niche of pricing the free lunch, of being able to say "we are the people who price the free lunch, and here is the ledger". Levin's own falsifiability criterion is programmatic, so the program can be adopted without the realm. Inside the field, Klein and Barron's phenomenal interface theory already has that shape. (An earlier version of this review also named a biology-independent programme of Birch's; that came from a secondary post, and version 1 of his manifesto does not contain it. See trees/02-g-theory/reading.md, corrected 2026-10-30.)

10. Open questions the review could not settle