Cut a salamander's leg off and it grows back. Cut yours and you get a stump. The two genomes are not that different. So why can one regenerate and the other can't?

For a century, biology answered "we don't fully know — it's complicated." Michael Levin's lab has spent twenty years arguing we have been looking in the wrong place. The information that tells cells what to build does not live only in DNA. Much of it lives in the electrical patterns cells maintain across their tissues — patterns we can read with fluorescent dyes and rewrite with off-the-shelf ion channel drugs.

If he is right, three things you probably assumed were permanent become editable:

Birth defects. His lab has rescued severe brain, eye, heart, and gut defects in frog embryos caused by nicotine, alcohol, and aberrant genetic signalling — without editing a single gene — by forcing the correct voltage pattern into the developing tissue. The genetic damage is still there. The intervention works around it.

Cancer. Frog cells expressing the same human oncogenes that drive lethal tumours (Ras, p53) can be talked back into normal behaviour by restoring electrical communication with their neighbours. The mutation stays. The cancer stops. The framing shifts from "kill the rogue cells" to "remind them they are part of a body."

Regeneration. A 24-hour drug cocktail delivered through a wearable patch triggered 18 months of progressive limb regrowth in adult Xenopus frogs, which normally cannot regenerate limbs at all. Mammalian work is underway. Even a partial mouse limb result would be one of the most significant findings in biology.

None of this is FDA-approved. None is in human trials. The frog and planaria results have to translate to mammals, and that takes years. Levin is careful: this is what becomes possible if the framework holds up in human biology. It might not. But the experimental record is now strong enough that the question has shifted from "is any of this real?" to "how far does it go, and how fast?"

The rest of this wiki is the evidence — 332 peer-reviewed papers — for why a growing number of researchers are taking that question seriously. If even half of what Levin's framework predicts holds up in mammals, the second half of the 21st century in medicine looks very different from the first.

Every cell in your body maintains a voltage across its membrane. For decades this was considered a purely local housekeeping function, relevant to neurons firing and muscles contracting. Michael Levin's lab at Tufts has spent twenty years showing that these voltages are also doing something far stranger: they form spatial patterns across tissues that function as a computational layer directing large-scale anatomy.

In this framework, morphogenesis is a form of basal cognition. Cell collectives solve the problem of "what shape should we build?" using bioelectric networks as their substrate. The code is instructive (not merely permissive), rewritable (without genetic modification), persistent (across cell divisions), and ancient (predating nervous systems by hundreds of millions of years).

If DNA is the hardware specification, bioelectric patterns are the software. You can change the programme without changing the hardware.

The framework has produced a series of results that would have been considered impossible a decade ago. Levin's group has grown functional eyes on the tails and guts of tadpoles by altering local membrane voltage, without touching a single gene. The ectopic eyes are anatomically correct, connect to the nervous system, and respond to light.

They have induced tails to regrow in place of heads and heads to regrow in place of tails in planaria by temporarily disrupting gap junction coupling. More striking, these animals remember their altered anatomy through subsequent regenerations even after the drug is removed. The information is stored somewhere other than the genome.

In frog embryos, Levin's team has rescued severe brain defects caused by teratogens (nicotine, alcohol) by forcing the correct bioelectric prepattern with an ion channel. The genetic damage remains. The voltage pattern is sufficient to produce a normal brain anyway.

They have built xenobots, reconfigured collections of frog skin cells that self-organise into novel living machines capable of locomotion, cooperation, and a primitive form of self-replication. And anthrobots, similar constructs made from adult human tracheal cells. Both exhibit behaviour not specified anywhere in the donor genome, suggesting that cellular collectives carry a latent repertoire that the standard genetic-programme view does not explain.

Cancer, in this framework, is not primarily a genetic disease. It is a disconnection problem. When cells lose their bioelectric coupling to the surrounding tissue, they revert from participants in a larger anatomical project to autonomous proto-organisms. They stop solving the collective problem (build a liver, maintain a skin) and start solving a simpler one (maximise local survival). The mutations come second.

The practical consequence is that tumours driven by aggressive oncogenes, including ras and tp53 mutations that are typically considered deterministic, can be normalised by restoring bioelectric and gap junction connectivity. The oncogenic mutations remain expressed. The cells stop behaving like cancer.

If this generalises, the implications are enormous. It suggests cancer treatment could shift from killing cells that cannot be reformed to reconnecting cells to the anatomical context they have forgotten. Not a replacement for chemotherapy in aggressive disease, but potentially a new axis of intervention that is orthogonal to the genetic one. The early results in frogs are striking enough that translation is the question, not the premise.

If anatomy is maintained by an active bioelectric process rather than being a static endpoint of development, then aging may be partly a failure of morphostatic information. The pattern that tells tissues "stay this shape, replace cells in this configuration" degrades over time. Gap junction coupling declines. Voltage gradients flatten. Cells stop hearing the signal that keeps them coordinated.

The testable prediction is that restoring youthful bioelectric patterns should restore some fraction of tissue coherence. This has been shown in Xenopus tail regeneration: adult frogs, which normally cannot regrow limbs, can be induced to do so through targeted bioelectric intervention at the wound site.

Mammalian extension is the frontier. Mouse studies are underway. A mammalian limb regeneration result, even partial, would be one of the most significant findings in biology.

The deepest open problem is whether bioelectric patterns constitute a proper language, with something like grammar, semantics, and compositionality, or whether they are a set of loosely coupled signals that happen to correlate with developmental outcomes.

If it is a language, cracking it would let us write anatomical instructions directly, rather than waiting for evolution to encode them genetically. The implications cascade quickly. Regenerative medicine becomes a compiler problem rather than a stem cell problem. Birth defect correction becomes a matter of writing the correct prepattern. Replacement organs could be grown in situ rather than transplanted. Cancer treatment becomes bioelectric normalisation with targeted mutations addressed separately.

Beyond medicine, the cognitive implications are stranger. If cell collectives solve anatomical problems using bioelectric computation, then cognition is much older and more distributed than we thought. The same mathematics that describes a planarian regenerating its head may describe a brain forming a thought. Levin's TAME framework (Technological Approach to Mind Everywhere) makes this claim explicit: intelligence is a property of any system that can pursue goals across a problem space, and biology is full of such systems at every scale.

This wiki catalogues what is known, what has been demonstrated, and what remains contested. The experiment the wiki supports, generating novel hypotheses from the framework's grammar rules, is a test of whether that grammar is structured enough to reason over symbolically. If it is, the Rosetta Stone question becomes tractable. If it is not, the framework may be a collection of striking phenomena in search of a theory.

Every active system pursues homeodynamic goal states, and the size of those goals in space, time, and abstraction defines its cognitive light cone. An amoeba's is tiny, measured in micrometres and seconds. A liver cell's is larger, coordinating with neighbours across years. A brain's is much larger still, operating over abstract categories across decades.

The key move: evolution scales cognitive light cones. Morphogenesis is the process by which cells assemble into collectives whose light cones are larger than any individual cell's. You build a body by combining small agents into a larger agent with bigger goals.

This framing gives Levin a way to talk about "cells having goals" without anthropomorphism. A cell's goal is whatever homeostatic state it is working to maintain. Measurable. Testable. Not a metaphor.

Levin's framing: "When cells lose their bioelectric connection to surrounding tissue, the cognitive light cone shrinks from the large goal of maintaining a nice organ to a tiny goal of I'm a single cell, and the rest of this body is just external environment. My goals are to become two amoebas."

Standard oncology targets the cell: kill it, damage its DNA, block its proliferation. Bioelectric medicine targets the relationship: reconnect the cell to the anatomical context it has forgotten. The mutations remain. The cell stops behaving like cancer.

Levin's team has demonstrated this in frog models, normalising tumours driven by ras and tp53 mutations through bioelectric and gap junction restoration alone. His consistent framing: "We're not trying to kill the cells, we're trying to talk them back into the collective."

From the 2020 TED talk, unusually clear for a scientist: "Most of the problems of biomedicine, birth defects, degenerative disease, aging, traumatic injury, even cancer, all boil down to one thing: cells are not building what you would like them to build."

If this is right, medicine has been solving the wrong problem. Rather than trying to directly engineer tissue (hardware-level intervention), we should be trying to rewrite the goals that cell collectives are pursuing (software-level intervention).

"The critical opportunity is to learn to communicate with cell groups, not to micromanage them, not to force the hardware, to communicate and rewrite the goals that these cells are trying to accomplish."

The experiment Levin returns to most often in interviews, because it contains the entire theoretical challenge in one picture.

Cut a planaria in half. Normally both halves regenerate into normal worms. Expose them briefly to octanol (a gap junction blocker), and the middle fragments regenerate into two-headed worms. So far, interesting but not shocking.

The shocking part: these two-headed worms, without continued octanol exposure, regenerate as two-headed in all subsequent amputation cycles. The pattern is maintained indefinitely. When Levin's group checked the genome, there were no fixed mutations. The information is being stored somewhere other than DNA.

His conclusion: "The simplest explanation is that somatic tissues are doing exactly what the brain does." Bioelectric circuits hold long-term, re-writable memory, and that memory determines anatomy.

Another experiment Levin cites often. Tadpole faces develop from a specific arrangement of precursor tissues: eyes in one place, mouth in another, nostrils in a third. Levin's lab rearranged these precursors at the embryonic stage, putting them in the wrong positions.

The tadpoles developed normally anyway. The eye primordia navigated, sometimes along strange paths, to their correct adult positions. The mouth formed where a mouth should be.

His conclusion: development is not a sequence of fixed chemical steps. It is problem-solving in morphospace. The cells know the target. They will find a path to it from whatever starting configuration you hand them. This is the property Levin exploits in every bioelectric intervention: he is not drawing the anatomy, he is changing the target.

From the 2024 Lex Fridman conversation, Levin's most speculative framing. Biological systems get "free lunches" from a structured latent space of possibilities. Xenobots exhibit behaviours no frog ancestor selected for. Anthrobots, made from adult human cells, cooperate in ways their evolutionary history cannot explain.

The implication: the space of possible minds and bodies is pre-structured, and evolution is not creating it from scratch but discovering regions of it. When you rearrange cells into a xenobot, you are not forcing them to invent new behaviour. You are giving them access to regions of possibility space their usual context hides.

Compatible with a Platonic reading of mathematics applied to biology. Levin marks this as speculation, but it is the speculation that drives his most ambitious claims about what bioelectric engineering could unlock.

Across both Lex Fridman conversations and the David Eagleman episode: the interesting question about alien intelligence is not whether it exists elsewhere in the universe. It is whether we can recognise the forms already present on Earth.

Xenobots are genuinely alien. Their behavioural repertoire was not selected by evolution. Planaria retain memories through decapitation and regeneration. Ant colonies and slime moulds solve optimisation problems without brains. Individual cells navigate morphospace with goal-directed behaviour.

"If we can't recognise intelligence in a two-day-old embryo, we're not going to recognise it when it shows up from somewhere else."

Across multiple interviews, Levin is consistent about his vision of 2050-era medicine. Regeneration as compilation: give the body the target morphology as a bioelectric instruction, cells execute it. No stem cell transplants, no scaffolds, no 3D-printed organs. Cancer as bioelectric normalisation: first-line treatment shifts from chemotherapy to reintegration; genetic mutations addressed separately, if at all.

Birth defects corrected in utero: teratogen exposure detected, correct prepattern imposed, genetic damage irrelevant to outcome. Aging as morphostatic information loss: restoring youthful bioelectric patterns restores tissue coherence. Not immortality, but longer morphological integrity. Synthetic biology as medicine: anthrobots as targeted therapeutic agents.

He is careful to mark this as "if the framework holds up." He is not promising any of it. He is describing what becomes possible if bioelectric engineering works the way the frog and planaria results suggest it should.

The analogy that recurs in almost every interview. The genome is hardware: the specification of proteins, receptors, channels. Bioelectric patterns are software: the programme running on that hardware, specifying what anatomy to build.

The analogy is useful because it makes a specific prediction: you should be able to change the programme without changing the hardware. And that is what Levin has done in every signature experiment. Two-headed planaria, ectopic eyes, teratogen rescue, xenobots. All demonstrations that the software layer can be edited independently.

His most provocative framing, from Tim Ferriss: "We've been trying to do medicine by editing the hardware. What if we just rewrote the software instead?"

From the GoodAI interview: "Evolution doesn't produce specific solutions to specific environmental problems. It produces problem-solving machines." A cell is not a lookup table of stimulus-response pairs. It is a competent agent navigating a high-dimensional space, exploiting whatever affordances physics and chemistry provide.

Important framing for thinking about medicine and engineering. When you intervene in a biological system, you are not pushing a button on a machine. You are communicating with an agent that has its own goals and its own repertoire of problem-solving moves. Good interventions work with that agency. Bad ones fight it.

Levin gets emails from people unsettled by his work: they read his papers, accept that they are a collective intelligence of cells, and write back asking, in effect, "now what?" His response is consistent across interviews: discovering the substrate doesn't dissolve the self, it relocates it. "The hardware does not define you," he tells Curt Jaimungal. "You are still the amazing integrated being with potential and a responsibility to do things." Knowing you are built from cogs and gears no more diminishes you than knowing a symphony is built from air pressure waves diminishes the music.

He pushes the point further on Lex Fridman's second appearance: even within a single human life, "you are in no way tied to the story that was told by your past self." The self is an ongoing self-referential process, not a fixed thing being preserved. Identity is what the system continues to do, not what it is made of. This is why the standard reductionist move ("you are just neurons firing") feels deflating but is empirically empty. The "just" is doing all the work, and it shouldn't be.

One of Levin's stranger and more consequential extensions. Drawing on William James's line that "thoughts are thinkers," he argues that patterns in excitable media (including ideas and memories) sit on a continuum of agency the same way cells and organisms do. A memory has a survival problem: what does it need to do to persist into the future? Some memories find ways to recruit attention, replicate themselves through retelling, and shape the brain that hosts them. That is niche construction, performed by the pattern, on the substrate.

He pushes this into the metamorphosis case. The caterpillar's brain dissolves almost completely, and yet trained associations survive into the butterfly. But the butterfly's body, sensors, and behaviours are utterly different, so the memory has to be reinterpreted, remapped onto a substrate it was never made for. Levin frames the paradox cleanly: "If you don't change, you will die. If you do change, you're no longer yourself." The resolution is that the pattern is the agent doing the migrating, not the meat. His standing methodological refrain applies here too: patterns can be agents too, but you don't know until you do the experiment.

Across interviews Levin returns to a specific epistemological complaint: the damage done by binary categories. Living vs non-living. Cognitive vs mechanical. Real intelligence vs "just physics." His objection is not philosophical squeamishness. It is that these categories prevent scientists from deploying tools. "What those categories do is they prevent you from hoarding tools," he tells Lex. Once you have decided a sorting algorithm or a gene regulatory network is categorically not the kind of thing that could have goals, you stop trying behavioural assays on it. You never run the experiment, so you never find what's there.

He replaces the categorical question ("is it intelligent?") with operational ones: how much intelligence, of what kind, in what problem space, with what cognitive light cone? The point is not to make grand pronouncements about consciousness in thermostats. It is to keep the toolkit open, because the universe does not owe us crisp lines where our concepts wanted them.

Most discussions of cancer use a moral vocabulary: selfish cells, defectors, betrayal of the collective. Levin reframes the geometry. Cancer cells haven't become more selfish; they've become smaller. Their cognitive light cone has collapsed from "I am part of a tissue with goals at the scale of an organ" down to something close to "I am a single-celled organism with goals at the scale of my own membrane." They aren't doing something new and aggressive. They are doing what unicellular life always did, but inside a body that depended on them not doing it.

The therapeutic implication is sharp. If the cells are genetically irrevocably damaged, you cut, burn, or poison them. But if the problem is a shrunken light cone (a coordination failure in the bioelectric network) you can try to enlarge it. Levin's lab has expressed strong human oncogenes in frog embryo cells and then, instead of removing the oncoprotein, forced the cells back into tight electrical coupling with their neighbours. The cells normalise. They make nice skin, nice muscle. The oncogene is still there. The selfishness vocabulary couldn't have suggested the experiment; the geometry one did.

Levin is candid about how unfashionable his work was when he started. Membrane voltage was treated as a "housekeeping parameter," something cells needed to keep the lights on, but not a place where information lived. When his postdoctoral mentor mentioned Levin's plans to manipulate ion channels to control development, a senior scientist told him, in earnest, that Levin was "having a psychiatric break." The expected outcome of perturbing cellular bioelectrics was assumed to be "uninterpretable death."

He started the lab in September 2000 with himself and one technician named Adam, having spent his postdoc years quietly building tooling for a research program he wasn't sure would survive its first grant cycle. He kept his programming skills sharp from childhood as a backup plan, assuming he would "eventually get kicked out" and have to return to coding. The thing he keeps saying in interviews is that the bet was not philosophical conviction but the wager that if you could show the experiments worked, the philosophy would take care of itself: "If I can show I got to new discoveries that you didn't get to, there you go."

In conversation with Anna Ciaunica, Levin floats a contrarian frame for senescence. Standard accounts root aging in molecular damage: telomere attrition, mitochondrial decay, wear in the substrate. Levin proposes a complementary story at the cognitive layer. A goal-directed system that has achieved its established goals, with no new goal and no fresh reinforcement, may simply start to degrade. The teleology is the scaffold. Pull it out and the structure relaxes.

He is careful that this is not a replacement for the molecular story; it is an additional axis. But it has a startling clinical flavour. Humans, with their long timescales and abstract goals, may need genuinely new challenges (not entertainment, but real cognitive engagement with novel problem spaces) to keep the integration coherent over decades. The framing connects to his observation that planaria, which are biologically immortal, never run out of regenerative work to do. They are never finished. There is a research program lurking in here about whether persistent novel challenge has measurable effects on biological aging, separate from the usual lifestyle variables.

When pressed on whether his ideas about cognitive light cones, platonic space, and agential matter are "really science" or "really philosophy," Levin gives a consistent answer that is worth taking on its own terms. He is not arguing for the metaphysics on metaphysical grounds. He is arguing that frameworks earn their keep by enabling experiments that the previous framework couldn't even formulate. The cancer-as-shrunken-light-cone view earned its keep by suggesting bioelectric normalisation as a strategy, which worked. The "patterns are agents" view earns its keep, or doesn't, based on whether it generates experimental programs that hardware-only frameworks miss.

He told Curt Jaimungal it is important to him that all these ideas "don't just remain as kind of philosophical musings — they have to make contact with the real world. And specifically, not just explaining stuff that was done before, but facilitating new advances." This is unusually clear epistemic hygiene for a field where big-picture talk often floats free. It also explains why he is harder to argue with than expected: he doesn't really care if you think his metaphysics is overreach, as long as the experimental program keeps producing results the orthodox frame couldn't predict.

In the AI House Davos conversation, Levin draws a contrast he keeps returning to: the deep architectural difference between biological systems and current AI. Cells, tissues, and organisms operate under conditions of constant perturbation, with novel damage and configurations they were never trained on, and they improvise solutions. Cut a planarian into pieces it has never been cut into, and each piece works out how to become a whole worm. Move the pieces of a tadpole's face into a scrambled arrangement, and the cells rearrange themselves toward a normal frog, solving a problem evolution never previewed for them.

Current AI systems, by contrast, are fundamentally error-correction architectures. They are optimised to converge on training-distribution behaviour and tend to fail in brittle ways outside it. Levin is not making the tired "AI is fake intelligence" argument. He is pointing at a research program. If you want machines that handle genuinely novel situations the way biology does, you probably need to build with agential, partially-autonomous components that have their own small goals, rather than fully top-down controlled systems. This connects to his prediction that future engineering will increasingly mean collaborating with materials that have preferences, not commanding inert ones.

Levin is careful. He distinguishes what his lab has demonstrated (eyes, planaria, tumour normalisation in frogs) from what he thinks is possible (mammalian regeneration, bioelectric cancer therapy in humans) from what is speculative (Platonic space, cognitive light cones at scale). Interviewers sometimes blur these lines. He does not.

He is generous with analogies. Hardware/software, compilers, Rosetta Stones, alien minds, target morphology, cognitive light cones. He chooses metaphors his interviewers can follow without sacrificing scientific content.

He is consistent. The framing in 2020 (TED) is structurally the same as the framing in 2026 (Tim Ferriss), just more confident as experimental results accumulate. Evidence of a coherent research programme rather than a moving target.

He walks into the philosophical questions. Most developmental biologists stay away from questions about cognition, agency, or consciousness. Levin walks directly into them. This is either his greatest strength (enabling conceptual breakthroughs) or his greatest weakness (inviting the overextension critique), depending on whether the experiments keep validating the theory.

Long-form podcasts: Lex Fridman Podcast #486 (2024) and #325 (2022) · Tim Ferriss Show #849 (2026) · Sean Carroll's Mindscape #132 (2021) · David Eagleman's Inner Cosmos #111 (2024) · Theories of Everything with Curt Jaimungal (in-person at Tufts, June 2024; "Breakthrough Research in Platonic Space" Dec 2024; with Anna Ciaunica, Jan 2025) · Machine Learning Street Talk (Oct 2024) · Cognitive Revolution (Feb 2024).

Talks & written interviews: TED 2020 · AI House Davos / "Beyond the Brain" (Dec 2025) · GoodAI conversation · Psychology Today interview (June 2025).

Why used: Transparent embryos, well-characterised development, amenable to microinjection and pharmacology, established bioelectrical recording techniques.

Key findings: Ectopic eye induction in gut tissue via voltage manipulation. Left-right asymmetry controlled by early ion fluxes. Brain defect rescue via HCN2 channel modulation. Bioelectric tumour detection and normalisation. Collective teratogen resistance (CEMA).

Why used: Extreme regenerative capacity, simple body plan, can be sectioned repeatedly for testing.

Key findings: Two-headed regeneration via gap junction blockade (permanent, heritable). Long-term bioelectric memory of anatomical polarity. Gap junction RNAi phenocopies pharmacological effects.

0.5-2mm living robots from Xenopus embryonic stem + cardiac cells. Self-organise into novel anatomies outside normal developmental context. Exhibit autonomous locomotion, self-healing, and kinematic self-replication. Cells retain morphogenetic computation when dissociated and recombined.

30-500μm robots from adult human tracheal cells, cilia-propelled. Demonstrate that bioelectric self-organisation occurs in human cells. Promote neural tissue regeneration without genetic modification.

Levin has confirmed ongoing mouse studies. Major challenges include immune complexity, limited regenerative capacity, and ethical constraints. Translation to humans requires solving mammalian bioelectric control first.