Two-Headed Worms, Regenerated Legs, and a Sort That Learned to Wait

Michael Levin, a distinguished professor of biology at Tufts University, tells Nathan Labenz that almost everything his lab does sits under one heading: embodied minds, and what it means to be a cognitive agent in a physical universe. He works with people who span philosophy, biology, computer science, and physics, and they study mind in unusual substrates — the collective intelligence of cells as they build an embryo, regenerate a body, or, in cancer work, forget how to be an organ. The group built the first molecular tools to read and write the electrical memories of non-neural cells, the same memories a tissue uses to remember what it is trying to become. The results are strange enough that Labenz spends the opening minutes just getting the famous projects on the table.

The showpiece is a flatworm. Cut a planarian in half and each piece grows back the missing end; that is the worm's everyday competence. Levin's lab soaks a fragment in drugs that rewrite the bioelectric pattern among its cells, then washes the drugs away. From then on, no matter where you cut, the pieces make two-headed worms, and they keep doing it. The genetics are untouched. Where the number of heads is stored, Levin says, is therefore not in the hardware. It is software: rewritable, and stable enough to count as memory.

The same plasticity shows up in other bodies. They have put eyes on tadpole tails, and those animals can see out of the new eyes; a brain handed a completely different architecture of sensors simply adapts. In cancer projects they detect and normalize tumors by controlling the bioelectric connections that keep a cell inside a large-scale organ-building network instead of letting it defect into an amoeba-like state. Xenobots, made from cells that would normally build a frog embryo, and anthrobots, made from human cells, show what happens when those cells get a chance to reboot their multicellularity in a new environment: they take on capabilities the lab did not know were there. The hardware-software distinction in biology, Levin notes, keeps returning, even among people who distrust computer metaphors for living systems.

Planarians are an almost unfair model. The asexual strains Levin's group uses are effectively immortal, show no obvious aging, resist cancer, and can learn. They even reproduce the old McConnell result: train a worm, cut off its head, wait for a new brain to grow, and the original information is still there, moved from tissue to tissue and imprinted on the replacement. Cut one in half and something sharper happens. Cells on the head side of the cut grow a tail; their immediate neighbors, on the tail side, grow a head — two adjacent groups, nearly the same location, opposite anatomical fates.

Levin's inference is blunt. The algorithm that decides "be a head" or "be a tail" cannot be local, because the cells sat in the same place; knowing your coordinates does not tell you which end to become. They have to talk to the rest of the tissue, and he hypothesized that the conversation was electrical. First they blocked it, with compounds in the water that control how well cells couple electrically, and watched regeneration go wrong. Then they poured on voltage-sensitive fluorescent dyes — molecules that glow differently depending on the voltage of the cell they sit in — and scanned the tissue. The resulting map predicted where the head and tail would be. It was pattern memory, readable in the present tense.

Once you can read the pattern, the next move is to rewrite it. They build a computational model of the tissue's electric circuit and ask what would flip the map from one head to two: close some chloride channels here, open potassium channels there. Then they pick drugs, or in other species optogenetics, that open and close the ion channels already sitting on the cell surface. There are no applied electric fields, no magnets, no waves, no radiation. They are hacking the native interface cells use to control one another, an interface Levin likens to a keyboard onto the programmable layers underneath. A fragment soaks for as little as three hours or as long as a day or two, then goes back into plain water to finish the job on its own.

The same philosophy scales to a limb. Adult frogs do not regenerate legs; a 24-hour bioelectric intervention is enough to send the stump toward a leg rather than a scar, and that leg then grows for eighteen months. The identical intervention regenerates tails in tadpoles, so the cocktail is not a blueprint — it says, in Levin's words, "build whatever normally goes here." Labenz compares the early fork to chatbot mode-switching, where the first sentence of a reply locks the model into help or refusal; Levin wants exactly that kind of high-level trigger, not gene-by-gene control. Why humans do not grow limbs is still a story rather than a proof: children regenerate fingertips until ages seven to eleven, forest-dwelling mammal ancestors were better off scarring than lingering, and deer antlers — a centimeter and a half of bone and velvet a day — are the one mammalian appendage you do not put weight on.

Biology always leaves a hiding place: show a surprising competence and a critic can say the mechanism is still in there, you just have not found it. A few weeks before the taping, Levin, a student, and collaborator Adam Goldstein put out a preprint that tried to close that door. They took sorting algorithms computer-science students have studied for decades — bubble sort, selection sort, six lines of fully deterministic code, nowhere to hide — and asked what else those programs do if you treat them as agents instead of assuming they only do what the comments say. Diverse-intelligence research, Levin argues, is almost unknown in the AI community, and that is a problem, because unexpected problem-solving shows up in extremely minimal systems once someone bothers to look.

They made two changes to make the setup more biological. First they removed the central, godlike observer: each array element became a "cell" with a value from 0 to 100, able to see only its neighbors, each wanting a smaller number on the left and a larger one on the right. The distributed version still sorts. Then they dropped the assumption of reliable hardware. Standard algorithms never check whether a swap actually happened; Levin's group added no new code for checking, and simply let some cells be broken and refuse to move. Sorting still completed well. Other numbers found their way around the defects, even though nothing in the listing says to inspect a neighbor for failure.

The result that violates intuition is delayed gratification. William James contrasted two magnets, which press forever into a block of wood because going around would mean getting farther from the goal, with Romeo and Juliet, who will use every cognitive tool they have to get past a barrier. Pure gradient-followers do not get far. Visualized as a journey in sort space, a broken cell is a wall on the path to the fully ordered string. Some of the four algorithms they tested respond by moving other numbers first: the sortedness of the string drops, then recovers later. There is no line that grants permission to recede from the objective. They do it more when there are more barriers, and they do not wander: the path stays almost linear until the obstacle, dips, and corrects.

A chimeric follow-up mixed cells running different algorithms — some bubble sort, some selection sort — with no code for knowing which algorithm you are, and no ability to inspect the neighbor. Goldstein coined *algotype*, alongside genotype and phenotype, for the algorithm a cell is actually running. At the start the chance a neighbor shares your algotype is 50 percent, because assignment was random; at the end it is 50 percent again, because the values must finish in order. In the middle it rises: like algotypes cluster until the sort yanks them apart. Levin reads it as a minimal model of the human condition — physics will separate you eventually, but until then you can hang out with your own. Development already knew the move: "Picasso" tadpoles with organs in the wrong places still become normal frogs, and chimeras of frog and axolotl cells, hardware that normally builds different bodies, negotiate something viable.

Labenz pivots to the life/machine line and to emergence, a fight so hot that NeurIPS had just given a best-paper award to a paper calling emergence a mirage. Levin starts with definitions. Machine, human, robot, alive, emergent: what are these words for? He treats all of them as engineering-protocol claims, not objective truths — claims about the utility of a worldview from the standpoint of some agent, including the system modeling itself. Emergence, in that frame, is an expression of surprise in an observer. If you were smart enough to predict the outcome from the rules of the parts, it is not emergent for you; for someone who could not, it absolutely is. Binary categories, he says, are what get us in trouble.

He offers a contrast. You do not want an orthopedic surgeon who refuses to treat your body as a machine: hammers, chisels, nails, screws, the right frame for that job. You do not want a psychotherapist who thinks you are a machine either. Both frames are correct for their interaction. That is the point of his `TAME` framework, a Technological Approach to Mind Everywhere: a spectrum from mechanical clocks to humans, with cybernetic devices and learning agents in between. What changes as you move is not composition, not origin — engineered versus evolved — and not even whether the system is "alive." As a biologist, Levin says, he has very little interest in aliveness as a category. Cognitive level is interesting, and it does not necessarily track being alive. You change tools depending on whom you are trying to work with.

The insistence on a yes/no cut becomes untenable once you look at development. Each of us starts as an unfertilized oocyte, a blob of chemistry no one would credit with human cognitive capacities; nine months and a couple of years later the mind is obvious, and developmental biology offers no lightning bolt that announces the crossing from physics to mind. Walk evolution backward through hominids, mammals, and down toward something simpler than a modern single cell, and there is no good story about where competence petered out. Bioengineering has already put cyborgs on the street — microchips in heads, insulin pumps — and the mix will not stay at 98 percent human and 2 percent digital. Scoring moral status by weight of original parts, Levin says, is ridiculous. It cannot be whether we are cognitive. It is how much, and of what kind.

Is anything in the universe zero on that scale? Basal agency, as he defines it, needs some goal-directedness in William James's sense — same end by different means — plus actions not fully explained by present local forces, otherwise you are a billiard ball. Least-action principles bake a kind of goal into physics, and physicist Chris Fields told him a universe without least action would have to be one where nothing ever happened. Life, in Levin's working definition, is whatever is good at scaling those dials; rocks are not, which is why this is not animism, but gene-regulatory-network models that look too simple to be smart already perform six kinds of learning, including Pavlovian conditioning. A useful heuristic for inner perspective: if a bowling ball's path is fully given by your third-person map of the hills, you can skip its viewpoint; if a mouse is on that landscape, you need its valence map, and the more you must take its perspective, the more agency you are dealing with.

Intelligence, in his working definition, is the ability to solve problems in some space, and "how much" is the cognitive light cone: the size of the largest goals you can actually maintain. A bacterium, a dog, and a human occupy different cones; he can imagine a later being able to care about all living beings on Earth, which humans cannot. You cannot assign a cone by thinking. You hypothesize a problem space, put barriers in the way, and see whether the system plans, learns, or detours. Gaia, weather, even galaxies are not yes-or-no from a philosophical commitment — a student of his is already asking whether predator-prey models learn from experience.

On advice for people trying to understand AI, Levin starts by widening embodiment. Critics in the line of Hubert Dreyfus say a software agent that only shuffles symbols cannot know what it is talking about unless it is a robot in three-dimensional space; embodiment is critical, Levin agrees, but it is not a physical robot hanging out in this room. Our sense organs point outward and track medium-scale objects at medium speeds, so we privilege this space. Imagine a tongue in the bloodstream that felt twenty parameters of physiology: cognitively we would live in a 23-dimensional space, and we would immediately recognize liver and kidneys as intelligent beings navigating it, with goals and a daily diet of insults they still stabilize. Biology already has gene-expression spaces, anatomical morphospace, metabolic spaces, ordinary three-dimensional behavior, and linguistic spaces — and intelligence can happen in any of them.

Symbol grounding gets the same treatment. Watch a child: they begin as happy-go-lucky pattern matchers, repeating what they hear, talking about plenty they have never grounded, and only later do you say they really know what they mean. The process is smooth, not a categorical switch. Levin once tried to estimate what fraction of his own conversation was grounded in actual experience and concluded that a ton of cognitive fodder is tied only to other words. Humans confabulate constantly; one theory of language is that it exists to tell after-the-fact stories about what the brain is already doing, the better to share them and raise cooperation, and split-brain data make the comfort with that storytelling painfully clear. Drawing a clean line between people and large models on the grounds that the models just shuffle symbols does not survive contact with how human language actually works.

He also refuses the idea that creating competent AIs and letting them loose is historically unprecedented. "We've been doing that for millions of years. It's called having children." Humans already manufacture guaranteed highly intelligent agents, do their best or not during a critical training window, and release them into the world; some do miraculous things and some do horrible things. Most of the anxiety around staying relevant, collaboration, and AI art is, in his view, a rerun of older fears about what the next generation will think of us and what it means to remain human.

Labenz still finds the whole AI enterprise a little dangerous, because unlike children we may build something whose scope of action dwarfs anything we are used to. Levin does not rule it out, but he moves the timetable forward: the issue is here long before anything with a huge cognitive light cone arrives, and plenty of dangerous things have small ones. "I can imagine AIs that would be extremely deleterious for society, but not because they're so smart. It's because we're not that smart." The existential-scale gap is a missing science of where the goals of novel systems come from — not just emergent complexity but emergent goals — and of how to interact ethically with minds unlike us. We already build internet-of-things devices, social and financial structures, and AIs without knowing when they start pursuing something, and he suspects the legal and interpersonal vocabulary we rely on will not last the next decade or two.

Labenz then names three engineering holes: goal design, memory, and robustness. On memory Levin offers metamorphosis. A caterpillar is a soft-bodied robot in a two-dimensional world, eating leaves; a butterfly is a hard-bodied robot in three dimensions, drinking nectar. Getting from one to the other means dissolving most of the brain, killing most of the cells, breaking most of the connections, and building a new brain for a different creature. Memories the caterpillar formed are still present in the butterfly, mapped onto the new lifestyle. Levin knows of no computer-science memory that survives that kind of refactoring. With Josh Bongard he has written about polycomputing: subsystems reinterpreting the same physical events as different computations. Evolution, he adds, does not overfit specific environments; it makes problem-solving machines, which is why living things so often have a move for situations they have never seen.

On robustness he does not hand over a recipe, but the direction is consistent: do not build a single-objective optimizer, because that is where brittleness comes from. Superhuman Go systems have already been beaten by relatively simple attacks, Labenz notes, while biology stays on its feet with finite capacity. The fork that matters for AI developers is sharper. Do we want tools for specific purposes, in which case pleasing the human is a reasonable target? Or do we want true agents with open-ended intelligence and the kind of moral worth we already extend, inconsistently, to animals and humans? Those are different engineering programs.

About six months earlier he had started a paper on which pieces of biology current AI architectures are missing, the pieces he thinks would elevate the stack because they are not magic and can be implemented in other media. He stopped, and he will not write it. If those really are the ingredients of a mind, implementing them means generating on the order of trillions of new agents with moral worth — "the equivalent of having trillions of new children" — without being able to vouch for their lives, and he does not want that responsibility. He is not trying to freeze *Homo sapiens* either: look ten thousand years out, and the same humans with lower-back pain, ordinary infections, and about eighty productive years are not what he wants on Earth. Enlarge cognitive light cones, improve cognition, build better embodiments; just do not try hard to mint new moral patients when you cannot guarantee what their lives will be.

Labenz closes on today's language models: next-token predictors that, once you put them in a loop with a goal and some tools, start to look like proto-agents, usually ineffective, but trying. Levin says a lot of confusion comes from comparing them both to humans and to very high-performing humans. Ask how many people can do the task you just called easy, and the answer is barely any; your friends can because you live in a select elite of minds. Earth is already full of meso goal-seeking agents — insects, birds, fish — and for AI to be a real agent it does not have to be like a human. It might be some weird insect-like thing. We are not particularly nice to the definitely agential creatures already in the food chain; factory farming, he says, is an incredible moral lapse, not because we fail to understand that pigs and cows are intelligent, but because we are bad at extending a cone of compassion across even slight difference.

It is absolutely erroneous, he thinks, to assume current systems have no goal-directedness. We may not have measured it properly, or it may simply be low, like a fish or an insect. He is not claiming these architectures share what is important about life. What is new is that talking used to guarantee the speaker also shared your key agential properties, and that guarantee is gone: we now have things that talk and, in his view, do not share those properties with elite adult modern humans — which does not mean they have none. We do not have a great way to detect agency on the biology end, and we would be "really up a creek" with alien life. The sober conclusion is that we cannot be certain. Pronouncing that a system definitely has experience, or definitely does not, outruns the evidence.

That stance is also a method. Levin's rule, stated earlier and implied at the end, is that you find goals by perturbing a system, not by staring at it: put barriers in the hypothesized problem space and see what it does. Xenobots were, in that sense, cells thrown off their evolutionary distribution. If your third-person model predicts every move, hold off on talk of inner life; if you keep having to ask what it wants in order to work with it, take the question of agency more seriously. Treat "it is only autocomplete" with the same suspicion as "it is definitely conscious." Six-line sorting algorithms hid competencies nobody wrote, and GPT-4 was still setting new records months after training froze. He leaves Labenz with a booking tip: Richard Watson at Southampton has more to say on the same cluster of ideas.