What this covers

Sean Carroll speaks with Michael Levin, a biologist and information scientist at Tufts University, about how organisms build themselves and what counts as a self. The conversation challenges the standard picture of genetics: that DNA blueprints determine anatomy. Instead, Levin argues that cells form collectives governed by bioelectricity and gap junctions—electrical circuits that store an anatomical target and coordinate tissue behavior toward it, independently of what the genome encodes. The core claim is that form, identity, and goal-directedness emerge from reprogrammable multi-scale intelligences, not from genetic instruction. This has sweeping implications: it suggests that selfhood is computational, that target morphology can be rewritten without editing DNA, and that cancer amounts to a breakdown of collective communication rather than purely genetic corruption.

The conversation moves through several concrete demonstrations. Planarian regeneration—where cut fragments rebuild entire worms, and where bioelectric manipulation causes worms to grow heads of different species—reveals that anatomy is stored in electrical circuits, not read off DNA. Tadpole experiments where scrambled facial features still migrate to correct positions illustrate error minimization toward a remembered goal. Levin also connects these observations to broader frameworks: gap junctions as the substrate of scaling of selves, where merged cells internalize each other's payoffs and cannot cheat; the free energy principle as a unifying account of how organisms predict and minimize surprise across scales; and the bridge between biology and computation. He defends teleological language—goal-directed description—as empirically justified rather than philosophically taboo, arguing that choosing a level of description is pragmatic, not preferential. The discussion weaves developmental biology, physics, artificial intelligence, and philosophy of mind into a framework where organisms are not machines but persuadable systems.

Sharpest takeaway

Levin argues that organisms are not built from genetic blueprints but emerge from reprogrammable, multi-scale collective intelligences: cells use bioelectric circuits storing a 'target morphology' goal state that can be rewritten without altering the genome, implying form and selfhood are computational and persuadable rather than hardwired.

  • The genome encodes proteins (hardware), not anatomical shape; shape emerges from bioelectric circuits that store and pursue a target pattern
  • Tissue-level collectives perform error-reducing computations toward a remembered set point that individual cells cannot perceive
  • Selves form by gap-junction merging of cells into larger goal-pursuing agents, and shrink (as in cancer) when that communication breaks down

The argument · threads7 threads · 29 claims
0.74

Bioelectric circuits, not genetic sequences, store and regulate anatomical target patterns that cells build toward cybernetically.

5 pointscentrality 5/5
  • By interfering with the normal bioelectric circuit, planarian cells with an unchanged genome can be induced to build heads belonging to other planarian species up to 150 million years distant, demonstrating that the target morphology is not directly encoded in the genome.

    “you can get these cells to build heads that belong to other species of planaria 150 million years distant with no genetic change at all”

  • In planarian regeneration the early decision of what to build is made at a high-level bioelectric 'master regulator' circuit acting roughly three to six hours after cutting, with all transcriptional changes occurring downstream over the next 24 hours; this constitutes practical downward causation because intervening at the bioelectric level controls scaling and head-vs-tail identity far more efficiently than manipulating lower-level gene expression.

    “the decision, the early on decision of what this tissue level agent is going to build, is made at this higher level, which then inevitably filters down to the lower level”

  • Planaria store a representation of their correct anatomical pattern in a partly-electrical tissue circuit; by manipulating ion channels (without altering the genome) the stored pattern can be rewritten so that cut fragments build two-headed worms, and recutting in plain water reproduces the two-headed form, demonstrating separation of the pattern 'data' from the genetic machine.

    “a circuit, which is partly electrical, which stores the pattern of what a correct planarium is supposed to look like. And we can now see this pattern, we can rewrite it”

  • Gap junctions (voltage-gated proto-synapses acting as transistors) let connected cells share internal milieu so that signals propagate with their origin metadata stripped; a cell cannot tell whether a flux originated in itself or its neighbor, which makes cheating impossible (harms come back to you) and erases ownership of memories, merging cells into a compound self with larger goal horizons.

    “both cells get access to each other's internal milieu... all the metadata as far as what the originator of that information is, is stripped”

  • Cryptic worms, which on each cut make a biased-coin-flip decision about head number despite normal molecular features, have an abnormal stored bioelectric pattern creating destabilized anatomy; this is analogous to bistable perception (Necker cube, rabbit-duck) in the nervous system, where identical input yields flipping interpretations driven top-down by Fristonian expectations.

    “what's abnormal is the bioelectric pattern that is stored by their real-time electric circuit”

0.70

Organisms pursue hierarchically nested goals from homeostasis through anatomy to behavior across expanding spatiotemporal scales using the same algorithmic principles.

4 pointscentrality 5/5
  • Because we only observe organisms that survived competition, all sophisticated life is excellent at pursuing goals starting from simple homeostasis; goals scale from tiny homeostatic set points (chemical/metabolic ranges) up to anatomical goals (a properly shaped hand) and then pivot to behavioral goals in 3D space, running the same algorithms over progressively larger spatio-temporal horizons.

    “life has to start out with homeostatic loops... And those loops scale into progressively more impressive goals”

  • A larger tissue- and organ-level organization performs computations that deform the action space of subunits, so that cells following local mechanical rules (like free-energy minimization) end up serving a global body plan; the collective makes decisions (e.g., correct number of fingers, limb length) by reducing error from a stored anatomical set point in a cybernetic, not designer-driven, sense.

    “there is a larger level of organization... performing some interesting computations that deforms the possible action space for the sub-units”

  • Per the free energy principle, living things must not only predict their environment but also predict themselves; organisms are patchworks at different scales of systems that cooperate and compete, each guessing what will happen next and having preferences, all working to minimize the delta between expectation and outcome, which is critical for evolvability.

    “not only do living things have to predict their environment, but they actually have to predict themselves”

  • When tadpole faces are experimentally scrambled ('Picasso tadpoles' with eyes, jaws, and nostrils in wrong positions), the organs nonetheless move in novel, unnatural paths until they reach correct frog positions, showing the developmental system continuously reduces the error between current configuration and a remembered target configuration, stopping movement and proliferation when error is small enough.

    “all of these little pieces that make up the head and the face will move in novel, unnatural configurations until they land in the correct positions to make a frog face”

0.57

Future agents will be continua blending human, nonhuman, and artificial substrates, requiring rethinking of what we owe to agents and whether hybrid systems genuinely have preferences.

8 pointscentrality 3/5
  • In a prisoner's dilemma simulation where agents can merge (so the number of agents is not held constant), cooperation does not merely emerge but becomes inevitable, because once agents merge you cannot cheat against yourself—the larger self internalizes the payoffs.

    “if you don't fix the number to be a constant and you let agents merge, you find out that cooperation doesn't just emerge, it's inevitable, because you can't cheat against yourself”

  • Since all cognitive agents (including humans, who are 'a bag of cells') are made of parts that already have memory and goals, the real question is not whether cells can have goals but how a collection of competent agents scales up into a unified larger self—a problem Levin models via gap junctions.

    “all cognitive agents are made of parts. There's no such thing as a sort of monadic, kind of diamond-like cognitive thing”

  • If the bioelectric target-pattern code were understood, regenerative medicine could build an 'anatomical compiler'—like computer-aided design—where one draws the desired organism and the system decomposes it into stimuli that drive cells to build that structure, rewriting the goal state of the cell collective rather than micromanaging or 3D-printing cells.

    “I think we would be looking at something like an anatomical compiler... you basically just draw a picture of, a schematic of the animal or plant that you want”

  • Because future agents will be continua—e.g., 80% human brain cells, 20% drosophila cells, 10% electronics—the binary distinctions between human and non-human, real preferences and as-if algorithm-following, are not robust; we must rethink what we owe to agents with different cognitive capacities, an ethics problem at least as large as the biomedical and robotics ones.

    “we could now make a system that's 80% human brain cells, 20% drosophila cells and 10% electronics with machine learning. Is that a human? Is it not?”

  • With any living system we can quickly determine how to motivate it via positive and negative reinforcement because it has things it likes and dislikes (infotaxis: networks constantly search for information to update self-models), unlike current robots whose parts don't have intrinsic motivation—raising the open question of whether artificial or hybrid agents 'really care.'

    “with any living system, we can tell very quickly how to motivate it, we know what it cares about... because we can train it with positive and negative reinforcement”

  • Whether cells can have goals and memory is already settled, because you yourself have memory and goals and you are a collection of cells; the genuinely open problem is only scaling.

    “the question isn't whether cells can have goals and hopes and dreams, that question has been settled, at least if for no-one else but for yourself”

  • Planaria are so regenerative that they do not age—there is no such thing as an old planarian—meaning current lab worms are in direct physical continuity with worms from half a billion years ago, which implies that complex brained organisms can in principle avoid death and that supposed thermodynamic limits on aging are not correct.

    “they are so regenerative that they're immortal, they don't age, there's no such thing as an old planarian”

  • Evolution favors developmental machines with a separation of hardware from software: cells reliably build toward a set point, but the set point itself can be represented and changed without rewiring the cells, making the system reprogrammable—if the set point changes, the cells will build something different without genetic modification.

    “it favors machines with a particular capability, which is as a separation of hardware from software”

0.56

Genomes encode proteins and timing, not size, shape, symmetry, or anatomical outcomes; development exploits physics and information processing to reliably generate forms.

4 pointscentrality 4/5
  • Like a Galton board producing a bell curve whose shape is not encoded in the marbles or board, embryos and living structures harness the laws of physics and computation; the genome describes a machine that generates and processes information by exploiting these laws to reliably produce outcomes.

    “what embryos do and all living structures do, is they harness the laws of physics and the laws of computation”

  • The genome does not encode the size, shape, or symmetry of an organism; it encodes a recipe for proteins (the low-level hardware that cells build structures from) and some timing of their appearance, so you cannot in practice read a genome and predict what anatomy will result.

    “if you actually look at what's in the genome, you're not going to find anything in there about the size, the shape, the type of symmetry of the organism that's going to come out”

  • Because planaria reproduce by fission with each fragment rebuilding from its own cells, they inherit somatic mutations (unlike most animals), accumulating mutations over ~400 million years into extremely messy, mixaploid genomes with varying chromosome numbers per cell—yet their anatomy regenerates perfectly 100% of the time, revealing a large gap between messy genetics and rock-solid anatomy.

    “they do inherit every mutation that doesn't kill the neoblast, the stem cell that got it”

  • Classical epigenetics shows information can be stored outside the genome in subcellular architecture: rotating a patch of a single cell's surface 180 degrees produces a lineage that all inherit the misoriented structure (templated off the mother cell like crystal templating), even though the genome is wild-type and 'powerless to help'—and similar subcellular polarity scaling contributes to planarian pattern inheritance.

    “their normal genome is powerless to help them”

0.53

Goal-directed teleological language is empirically justified and necessary for prediction and control; whether to use it depends on which intentional level best explains behavior.

3 pointscentrality 4/5
  • Biology has been trapped in a counterproductive 'teleophobia'; goal-directed, teleological language is not only acceptable but necessary, and whether to use it is an empirical/practical question (like Dennett's intentional stance) of what level of intentionality best yields prediction and control, not a philosophical taboo.

    “I think biology been trapped in this teleophobia, which is doing us a great disservice”

  • You cannot pick a level of description simply because you prefer it (e.g., insisting everything be done at the lowest biochemical level); the best level is determined by how much effort is needed to control the outcome, with systems ranging from non-persuadable (cuckoo clocks requiring rewiring) to highly persuadable (humans best influenced by stimuli and training rather than molecular micromanagement).

    “you can't simply pick a level because you like it, you have to pick the best level, and how do you know what the best level is? And the best level is how much effort do I need to put in to control the outcome”

  • Critiques of panpsychism err by scaling down the physical system (to a rock) without proportionately scaling down the intentionality; if you scale cognition down to the simplest case, freedom-as-indeterminacy would look like quantum indeterminacy and minimal goal-directedness would look like the action principles of physics.

    “they scale down the physical system... and then they fail to scale down the intentionality”

0.35

Cancer is multicellular breakdown caused by loss of electrical cell-to-cell communication, reverting cells to unicellular goals and can be reversed by restoring collective connection.

2 pointscentrality 3/5
  • Cancer is a breakdown of multicellularity: closing gap junctions (an early step in carcinogenesis since the late 1970s) shrinks a cell's computational boundary back to a single cell, reverting it to its ancient unicellular goals (migrate, proliferate, treat the body as environment), causing metastasis; this can be triggered by purely physiological disruption of electrical communication without a genetic component, and can in principle be reversed by re-connecting cells to the collective.

    “in cancer, they've known since the late '70s that one of the first steps of carcinogenic transformation is a closure of the gap junctions”

  • Robots don't get cancer because their parts have no sub-goals—so no part defects—but that same lack of multi-scale agency makes them inflexible; the design challenge is to create systems that reap the benefits of higher computational capacity and cooperation from merging while retaining the agency and adaptability of the pieces.

    “we make devices where the pieces don't have sub-goals... yes, no, you're not going to have a robots where part of it decides to defect... but on the other hand, the robots aren't very good, they're not very flexible”

0.25

Computer science and biology both study computation in different substrates; concepts flow bidirectionally and illuminate how organisms and machines work.

3 pointscentrality 2/5
  • Computer science and biology are not really different fields—both study computation, just in different media—so concepts can flow both ways: robustness and brainless intelligent problem-solving from biology can inform machine learning and robotics, while cognitive/computer-science concepts can illuminate how organisms work.

    “I fundamentally think that computer science and biology are not really different fields, I think we are all studying computation just in different media”

  • Brains are an elaboration of an electrical signaling system evolution found around the time of bacterial biofilms; bacteria are electrically active and biofilms drive brain-like potassium-wave electrical dynamics in which individual bacteria coordinate to form a proto-body, a larger agent able to do what single cells cannot.

    “bacterial biofilms drive a lot of very brain-like electrical dynamics... what you get are waves, potassium waves, that in many ways are similar to what happens in brains”

  • Cellular plasticity (e.g., eyes placed on tadpole tails still see) and modern hybrid bioengineering blur the continua between robot, program, machine, evolved, and designed, so these inherited categories no longer carve nature at its joints and must be redefined, with major implications for biomedicine, robotics, and ethics.

    “all of these words like robot, program, machine, all of these things I think are operating on categories that are no longer good categories”