A Thing Is Alive When Its Goals Outrun Its Parts

Lex Fridman opens with the sentence at the center of Michael Levin's work: how do embodied minds arise in the physical world, and what sets their capabilities? Levin, a biologist at Tufts University, answers that the tension lives in three persons at once. In the third person, we have to recognize agency when our intuitions call something mechanical. In the second person, we have to choose a control protocol — hardware rewiring, cybernetics, behavior science, or friendship. In the first person, there is an inner perspective that cares about outcomes and still has to sit inside physics.

He names the horizontal axis of that framework the spectrum of persuadability. It is an engineering term on purpose. You cannot decide from a philosophical armchair where a system sits; you hypothesize which interaction protocol to bring, then everyone gets to see how that worked out. Guess too high or too low and the experiment will embarrass you. A wind-up clock is persuaded with a wrench. A thermostat is persuaded by moving a set point. A dog is persuaded with rewards. A person is persuaded with reasons. Regenerative medicine, Levin says, keeps reaching for the wrench when a high-level prompt might be enough.

If he wants injured cells to grow a limb, he has options. He can micromanage every molecular event, micromanage stem-cell signals, or tell the tissue, at a high level, that it should build a limb and see whether it can be convinced. Standard molecular biology treats that last option as crazy. The excitement in the field is at single molecules, big data, and genomics — the assumption that the action is always downward in scale. Levin thinks that assumption is wrong, but he will not argue it from the sofa. Every time his lab takes tools from behavioral science — training, active inference, visual illusions, stress, memory reconstruction — and applies them outside the brain, new capabilities appear. Not because those tissues are secret brains, but because nobody looked at them that way.

Toward the right of the spectrum the relationship flips. Agency rises, the effort needed to exert influence falls, and you need less mechanism knowledge, not more. Humans trained dogs and horses for thousands of years with zero neuroscience; Levin does not have to manipulate Fridman's synaptic proteins to be understood. Richard Watson calls the far end mutual vulnerable knowing. In molecular biology you leave the bench unchanged. Talk to a dog, a person, or a forthcoming synthetic creature and you are not the same at the end. Physics, Levin argues, always sees mechanism and not minds because physics uses low-agency tools: voltmeters and rulers. If you want to see a mind, you have to use a mind. There has to be an impedance match between the interface and what you hope to find.

Fridman presses the obvious objection. Many physicists think life, the origin of life, and intelligence are physics problems, and that every other toolkit is a distraction. Levin's reply hangs on the word *understand*. A pleasing model that captures a slice is not enough. Understanding, for him, has to be generative: it should let you build regenerative medicine that actually helps a person missing a finger. Even a physics theory compatible with everything that happens is not specific enough to tell you what to do. When those problems need solving, you do not call a physicist. You call a biologist or a psychiatrist.

He offers a scene. Someone walks in and delivers a beautiful mathematical proof. A physicist nearby says the accounting is complete: air particles left a mouth, moved cilia, sent electrical signals up a brain. Done. If you want the important part of that interaction, Levin says, you do not go to the Physics Department. You go to the Math Department. Physics is an amazing lens. It captures certain slices. Stretch it to cover the rest and we simply stop calling the result physics.

Fridman wants the Cartesian cut: the first step from non-mind to mind, non-living to living, the simplest place physics should be able to draw a line. Levin does not believe the line exists. There is a continuum. Humans like to name regions on it, then prosecute category errors when someone transgresses. Those categories, he thinks, now mostly hold science back, because they stop you from hoarding tools. If living things are a different kind from non-living things, you will never try behavioral protocols on a molecular network. People tell him he is making a category error as if the categories arrived from on high. The categories should change with the science.

Is origin-of-life talk still useful? Better, he says, to study the scaling than to hunt for a starting gun. The word *adult* is handy in court because cases have to move; nothing special happens on an eighteenth birthday, and car-rental companies, looking at accident data, put the number nearer twenty-five. Scratch *neuron* and you find a slow developmental story in which most cells in the body already do neuron-like things. Categories are the art of saying something without first having to say everything. They become dangerous only when you forget what you glossed over. Levin would rather talk about kinds of mind than about life. Life, he says, is not that interesting a category — a strange sentence from a biologist, and he knows it.

What is scaling, if not a jump from dead to alive? Levin's vertical axis is the cognitive light cone: the size of the biggest goal state a system can actively pursue. It is not sensory reach. The James Webb Space Telescope has enormous sensory reach and a tiny cone. It is not how far you can shove the world. Put space and time on one diagram, as physics does with light cones, and you can place very different agents on the same map. Goals that live inside a ten-to-twenty-micron radius, with twenty minutes of memory and five minutes of prediction, sound like a bacterium. Caring about a few hundred yards but not about two towns over in three weeks sounds like a dog. Caring about financial markets after your death sounds like a human. Claiming you can care, in the linear range, about every living being on the planet sounds like a bodhisattva. Ordinary humans saturate. Ten deaths activate you; ten million do not activate you a million times more.

Then comes the definition he spent no effort making official. We call things alive to the extent that the cognitive light cone of the whole is bigger than that of its parts. Rocks are boring because they only do what their parts already know how to do: follow gradients. Living things align competent parts so the collective can pursue goals the parts cannot comprehend. A cell manages pH, metabolic state, some transcriptional set points. Evolution supplied a cognitive glue that binds cells into multicellular systems with grandiose aims. Chop a salamander limb and the collective grows the right number of fingers and stops. No individual cell knows what a finger is.

The failure mode of that scaling is cancer. Cells physiologically disconnect from their neighbors. The cone shrinks. The boundary between self and world, which the cone defines, collapses, and the cell is an amoeba again. The rest of the body is just external environment. It goes where life is good and reproduces as much as it can. So when we look for life, Levin does not think we are looking for a material or a metabolic signature. We are looking for scales of cognitive light cone. We are looking for parts aligned toward bigger goals in spaces those parts could not understand. Fridman objects that a bacterium, zoomed out far enough, "wanted" civilization. Levin's answer is the same as for everything else: you do not get to narrate the goal. You have to intervene.

How do you know the cone of something that cannot talk? You put barriers between it and the goal you hypothesize, and you ask what happens. Intelligence, on this operational reading, is the degree of ingenuity it shows in getting around the obstacle. If the barrier derails everything, the goal was never much of a goal. If the system takes a route you did not expect, using affordances in novel ways, you have learned something about its mind. William James's definition — the same goal by different means — is, for Levin, something you can run in a lab. Anthropomorphism, he says, is not a real heresy. It just means "humans have a magic you must not attribute elsewhere." The scientific method is: do the experiment, then report the competencies you found.

He is not the first to refuse a stopping rule. More than a hundred years ago, Bose measured how anesthesia changed electrical excitability in animals, then did the same in plants, then in metals, because the tool was still working and science had not told him where to stop. People made fun of the plants. There is still no generic rule that tells you when to stop. The rules are made up. Levin's lab has used habituation, conditioning, anxiolytics, and hallucinogens on systems most departments would not let you treat as minds. Fridman floats applying psychedelics to a sorting algorithm. Levin's point is narrower and sharper: we have an N of one evolutionary lineage on Earth, and our mapping from embodiment to intelligence is so bad that refusing to try is shortsighted.

The formal version of this stance is TAME, a technological approach to mind everywhere. Cognitive claims, Levin says, are protocol claims: when you tell me a system is intelligent, you are telling me which toolkit you plan to deploy. Figure 2 of the paper is a spectrum with four waypoints — mechanical clock, thermostat, Pavlov's dog, humans arguing with reasons. As you move right, persuadability and autonomy go up; effort and the mechanism knowledge you need go down. You do not need to know how the heating plant works to set a thermostat. You do not need neuroscience to train a horse. Every layer of a body, not just the brain, is doing a perception-decision-action loop in some space: physiological, transcriptional, anatomical. Humans are obsessed with motion in three dimensions and keep calling disembodied anything that does not walk around a kitchen. Biology has been embodied in stranger spaces the whole time.

Asked how you even find the mapping between two alien problem-spaces, Levin describes a tic-tac-toe game against an unseen opponent. You play geometry; next door, the alien hunts billiard balls whose numbers add to fifteen. A magic square is the thin interface that lets you share a slice of a world without thinking alike. Xenobots and Anthrobots are his attempt to build such interfaces on purpose. Typical biology explains form and behavior with one sentence: evolutionary selection. That sentence is a crutch. To kick it away, the lab makes creatures that have never existed, starting from living cells that do have a history, but whose new collective was never selected for.

Take epithelial cells from an early frog embryo. Do not change the DNA. Add no synthetic circuits, scaffolds, nanomaterials, or strange drugs. Liberate them from the instructive influence of the rest of the body. In situ those cells are bullied into a boring life: a two-dimensional covering that keeps bacteria out. Freed, they self-organize into a motile xenobot, cilia coordinated to row against water. They have a novel transcriptome. They can kinematically self-replicate from loose cells in the dish. They respond to sound, which normal embryos do not. Critics said it might be a frog-specific quirk. So the lab went as far from frog embryonic cells as it could: adult human tracheal epithelium, donated from patients. No genetic change. The cells became Anthrobots. About nine thousand genes changed expression — roughly half the genome. In vitro, they sit down on a scratched lawn of human neurons and spontaneously knit the wound, untaught.

Sequence an Anthrobot and you get one hundred percent *Homo sapiens*. It looks like a pond organism, not like any stage of human development. It is also younger than the cells it came from. Open-loop models — cellular automata, local rules, turn the crank — can make complexity, Levin says, but they do not explain morphogenesis. If you block the usual outcome, competent tissue finds another path; if you truly defeat it, you get a birth defect or a creative alternative such as an Anthrobot. The decisive evidence is that the lab can find, read, and rewrite some of those goal memories in bioelectric patterns. Reset the set point and the system pursues the new goal. By any engineering standard that is a homeostatic mechanism. If there were no goal, rewriting it should be impossible. Step one is still imagination: you have to guess what space the system is even working in, because we cannot picture all of them.

Fridman asks how weird the method can get. Can an idea be an organism? Levin goes to the caterpillar. Train it to crawl toward a colored disc for leaves; after metamorphosis, when the brain has been torn up and rebuilt for a hard-bodied flying machine, the butterfly still has the memory. It cannot keep the memory in the old form. It has to remap it onto nectar and flight. You can tell that story from the caterpillar facing a singularity, or from the butterfly saddled with tendencies it does not remember earning. The third view is the memory's own: an informational pattern that must change or die. Agents, Levin says, are always patterns in some excitable medium.

He tells a science-fiction joke. Creatures from Earth's core, unimaginably dense, walk our surface as if through a thin plasma. Their scientist notices whirlpools in the gas that hold together for about a hundred years and seem to want things. The others scoff: patterns in a gas cannot be agents; we are solid. We are those whirlpools. With Chris Fields and others, Levin has been dissolving the line between thoughts and thinkers. Fleeting thoughts are ripples. Hurricanes, solitons, earworms, and depressive loops persist and do a little niche construction on the brain. Personality fragments in dissociative disorder are more stable still and have goals. Then a full human personality. Who knows what sits past that. If you want to know whether a soliton, a hurricane, or a thought is an agent, run the same tests: memory, goals, learning, language.

A Turing machine can be told two ways, and both are useful. In one, the machine is the agent and the data are passive. In the other, the patterns on the data are the agent and the machine is a stigmergic scratch pad. Map those stories onto aging and you get two research programs, both underway in the bioelectrics work. Maybe the physical body is the agent and anatomical pattern memories are data that go fuzzy with time, so you reinforce the memories. Maybe the patterns are the agent in the same sense we grant to brains, and aging is the agent finding it harder to be embodied because the cells have gone sluggish — so you do not strengthen the memory, you make the tissue more responsive. Science-fiction stories, Levin says, earn their keep only when they become experimental agendas. Alternative-medicine communities have said versions of this for a long time. What is new is being able to image the bioelectric patterns and, he hopes, turn them into therapeutics.

For thirty years Levin did not talk about what he now calls Platonic space, because it was not actionable. Machine-learning papers on a Platonic representation hypothesis, plus biologists, mathematicians, and even economists who had been sitting on similar ideas with no audience, turned a planned three-talk meeting into a weekly symposium. He is not trying to stay loyal to Plato and expects to change the name. He uses it because mathematicians who call themselves Platonists think they are discovering a structured space of truths, not inventing a human construction.

Keep asking why in biology or physics and you fall out of the department. Cicadas emerge on thirteen- and seventeen-year cycles to miss predators; why those numbers? They are prime, and you are in the math department. Ask a physicist why a particle behaves as it does and the answer is eventually a group and its symmetries — math again. Facts such as the value of *e* or Feigenbaum's constant are thrust upon you. If they were different, biology and physics would be different. Nothing you can do in the physical world, including twiddling every constant at the Big Bang, changes *e*. Physics, Levin claims, is what those patterns constrain. Biology is what gets the free lunch: evolution that gets two angles of a high-fitness triangle for free on the third, or that invents a voltage-gated ion channel and inherits logic-gate truth tables, including that NAND is special, without evolving them.

Xenobots and Anthrobots do things their lineages were never selected for. When did we pay the computational cost? One answer is "emergence," which Levin hears as "I got surprised" plus a big book of unrelated wonders. The optimistic alternative, the one mathematicians already make, is that the latent space has structure you can map. Every physical thing is an interface. Build a different pointer — a cell, an embryo, a robot, a language model — and different patterns come through. Low-agency inhabitants are the objects mathematicians already catalog. Higher-agency inhabitants are what a behavior scientist would recognize as delayed gratification or a kind of mind. The mind-brain relation, on this view, is the same kind of relation as math to physics. Nobody creates consciousness, not by making a baby and not by making a robot. You create a physical interface through which a pattern ingresses. The brain is a thin client. Consciousness is the view from that pattern looking out.

He wants this to sound as wild as it is. Physicalism, he says, was already dead in Newton's classical world, long before quantum mechanics or consciousness entered the chat. That world was already haunted by non-physical facts. Dualism is an old, unpopular idea, and Descartes was already beaten up for the interaction problem. Levin's move is to say physics had the interaction problem first. Come back in twenty years: either the lab and others will have a map — Anthrobots have four behaviors, not seven, and here is the body you build to pull a given mind — or the space will look like a jumble and they will have to live with surprise. All science has, he says, are metaphors. The only question is how good they are. He does not like *emergence* as an explanation. It usually means he did not see it coming.

If Platonic patterns really seep into minimal interfaces, they should show up where there is nowhere to hide. Levin wanted shock value: a system studied for more than sixty years, a few lines of deterministic code, familiar, fully transparent. Students Kaining Zhang and Adam Goldstein took classical sorting algorithms and put a barrier in the path. Standard code assumes hardware is reliable: tell a seven and a five to swap, and you do not check. They froze one digit so it would not move, and they did not add a single line that says what to do if the swap fails.

The array still sorts, by rearranging everything around the broken number. Plot sortedness over time and the curve is the surprise. A healthy run rises monotonically. With a frozen digit, sortedness first goes down, then climbs back. The system walks against the gradient it is supposedly climbing in order to meet the goal later. Show that trace to a behavior scientist, hide the substrate, and they will name it delayed gratification. Two magnets with a board between them will not go around. The sorting algorithm will. Stare at the code all day: the competency is not there. Levin is not claiming a miracle or a disobedience. He is claiming an unexpected behavioral competency that nobody had to pay for in extra instructions — more than complexity, more than noise.

They then removed the central planner. Each number runs the algorithm locally, like an ant colony, and distributed sorting still works. Then they made a chimeric algorithm: half the digits randomly assigned to bubble sort, half to selection sort, analogous to frog–axolotl chimeras whose appearance no genome predicts. That sorts too. Adam Goldstein named a property that has nothing to do with the values: algotype, which algorithm you are following. Start at fifty percent chance a neighbor shares your algotype, because assignment is random. End at fifty percent, because the sort does not care about type. In the middle, same types cluster. Implementing that clustering on purpose would take extra code to observe neighbors, infer their algorithm, and move beside kin. None of that code exists. The clustering was free. Allow duplicate digits so the sort can ease up, and clustering grows. Sorting is what the system is forced to do. Clustering is what it wants to do. Levin calls that a minimal intrinsic motivation, sitting in the space between chance and necessity.

If a bubble sort can do this, what about language models? Two possibilities, and the lab does not yet know which is true. In biology, evolution couples side quests to the main task. In machines, the side quests may have nothing to do with language, in which case watching what models say is a red herring and the interesting behavior is whatever we never thought to measure. Levin's N of one from biology is the Anthrobot on a wounded neuronal plate. The first intrinsic motivation they checked was not experiment 972. It was the first. The bots healed the wound. The next twenty findings might be damaging. He cannot promise otherwise. The first one was benevolent, and that makes him feel better.

Anthrobots are not only healers. Using Steve Horvath's epigenetic clock, in collaboration with the Clock Foundation, the lab found they are roughly twenty percent younger than the adult tracheal cells they come from. Levin's hypothesis is age evidencing. The cells arrive with priors from an old body. The new environment screams embryo: no surrounding tissue, bodies bent into pretzels, some embryonic genes on. The evidence is not enough to roll them all the way back, but it is enough to update them partway. He compares it to a study that redecorated a nursing home in the style of the residents' youth and saw blood-chemistry improvements. If the surroundings are persuasive enough, cells recalculate how old they are.

Is that a longevity therapy? It is what they are trying to do, and it is not simple. All of their regenerative medicine, Levin says, balances on learning to talk to the system. Turning gut tissue into an eye means convincing cells that their prior — we are gut precursors — is wrong and that they should take up a new worldview. Knowing which messages persuade cells is, in his view, at the root of future progress on birth defects, regeneration, and cancer. He can see the path. He does not pretend it is a trick of décor.

If brains are thin clients, can you upload a mind? Here Levin marks the edge of what he will claim. Most of what we call mind, he suspects, is the pattern in that other space. A prediction that modern neuroscience does not make, and that he and Corrina Kaufman reviewed clinically, is that some people have very little brain tissue and normal or even above-normal intelligence. It does not happen most of the time; the exceptions are not a theory. Copying, he thinks, copies the front-end interface, not the pattern. What might be possible is building another interface through which the same pattern would come. His guess about a *Star Trek* transporter is that if you rebuilt the same attractive physical pointer somewhere else, he sees no reason the same pattern would not come through. That is a guess, and he says so. Ownership of a self, he suspects, arrives with the earliest job of a being: telling a compelling story to your parts, drawing a border between me and world. Scratch a duck blastodisc and every island, for a while, thinks it is the only embryo.

How does a quiet oocyte become a person with hopes? Levin refuses to treat physics and chemistry as the real furniture of the world. They are lenses. You already do not start at zero: even a molecular network can do Pavlovian conditioning, which means it is already hosting a pattern from that other space. Then you are a cell, then many cells navigating anatomical possibility, pulling down richer ingressions, which let the physical interface support a larger cone, which lets you pull down more. Federico Pagosi's work in the lab closes a loop. Chemical networks can learn several kinds of association. Causal emergence, measured with metrics such as Φ, is what you need if a rat is to own a memory no single cell experienced: paw cells touched the lever, gut cells got the reward. Train some of those networks and Φ itself rises. Each thing you learn makes you more of an integrated agent; each increment of integration makes the next lesson easier. The spiral does not come from evolution, which will only optimize it, or from physics. It is a free gift from the mathematics of information and networks. Levin thinks that gift is the path from a cell to a creature that can dream and love.

He offers two blunt physical mechanisms for scaling the cone. The first is stress propagation. Treat stress as the physical error function: the delta between where you are and where you need to be, in whatever space. A stuck cell leaks stress molecules. Neighbors, whose own set points were already met, become stressed anyway, because the molecules are highly conserved and do not carry a name tag. Plasticity rises, as in simulated annealing. They become willing to take other fates, the stuck cell gets through, and everyone's stress drops. Nobody is altruistic. Far regions simply become sensitive to one another's needs.

The second is memory anonymization. An ordinary signal arrives as someone else's news; you can ignore it. A gap junction links two cells' interiors. A calcium spike from a poke crosses the junction, and the receiving cell cannot tell whether the memory is its own. Some identity remains, but a group now has memories of what happened to *us*. The area of concern grows, and so does the cone. Levin's group is working with a company called Softmax to engineer that kind of scaling, including in AI. The cancer application is already concrete. Tumor cells electrically unplug from a giant memory that was building an organ, forget the project, and become amoebas. Reconnect them to the network — no DNA fix, no chemotherapy — and they go back to work as part of the larger collective.

Levin's acronym is SUTI: the search for unconventional terrestrial intelligences. Guaranteed, he says, they are already here. Cells in a body navigate twenty-thousand-dimensional transcriptional spaces, suffer when they miss their goals, and reduce stress when they hit them. If we cannot recognize the aliens inside us, our odds off Earth are poor. Existing IQ-style tests can be pointed at weird systems if you can build the interface; his lab has. Humility is still required, because every such metric was derived from one evolutionary lineage. Fridman asks whether the more interesting minds sit in the body or in the Amazon. Levin will not rank infinities. The Buddhists say innumerable sentient beings. Going down in scale is easier to experiment on than going up. He does not think ant colonies are hopeless: at least three or four papers show that colonies, not individual ants, fall for the same visual illusions humans do, completing lines of food that are not there.

A recent paper asked about sixty-five working scientists to define life. The answers spanned so many dimensions that the authors needed AI to embed them. Consensus was zero. Life in the solar system would still excite him, as a data point far from existing ones, but his surprise threshold is now very high. The next papers, he hints, will keep expanding which systems we have to take seriously as beings to relate to — not just brains, not just biology, not just complexity, not just three-dimensional space. He will talk when the work is actionable. The public knob on how much of the weirdness he says out loud turns a little every few years as the empirics catch up.

His method of having ideas looks like making xenobots: release constraints and see what happens. Computer science first, biology later, so the field's obvious dogmas arrived after he was already suspicious. He asks what the world would look like if the current model were wrong, what the model specifically cannot see, and what sits in the middle if two supposedly different things are ends of a continuum. He looks for symmetries: the fewest moves that would map this onto that. Almost every morning he walks outside at sunrise. Photography keeps hands and brain just busy enough to knock him off a linear track. Ideas go into voicemails to his office. A mind map, one piece of it a nine-foot silk sheet hanging in the lab, is out of date within two weeks of printing. He has something like a hundred and sixty-two open manuscripts. Empirical projects — Anthrobot aging, limb regeneration, the next cancer paper — stay linear. The rest wait in outlines until it is time to put words around them. Asked where the ideas come from, he sounds like every other creative person he has asked. It feels like a collaboration. His job is to be up at 4:30 a.m., ready to recognize the idea and to have a lab that can land it. Claiming the idea is "me," he thinks, would not be right.

The most important advice Levin has learned is not to take too much advice, so he dislikes giving it. He makes an exception for unconventional people who write to him. Bifurcate your mind. One region is practical impact: what people can hear, which parts to postpone, which journal, whether to wait two years. Without that region you will have no career and no resources. The other region must ignore all of that. It has to be pure: I do not care if this is publishable, if anyone gets it, if anyone thinks it is stupid; this is what I see and why. Mush the two together and the practical half poisons the other. Live only in the creative half and nothing materializes. Specific craft advice — you forgot this control, this is the wrong method — is gold. Meta-advice of the form "that is not a good way to think, do not work on this" is often garbage, even when it comes from very successful people. A reviewer once promised him "constrictive criticism." Levin still loves the slip.

Asked for a beautiful idea, he reaches for steganography: hiding a message in the bits of another pattern that do not matter, without ruining the cat in the picture. He feels the ingressions as a universal steganography, patterns seeping into every interface that will hold them, shy but not invisible if you work. That they affect machines as well as organisms is, to him, the beauty, not the insult. Some hate mail starts friendly on the majesty of life and turns when the machines get it too. He is enriched by being on the same spectrum. Fridman then hands him a hypothetical AGI and asks for the first question.

Levin would not start with a scientific puzzle. He would ask how much he should even be talking to it. A child learning math with an older sibling who simply gives the answers sometimes needs that; a cancer protocol might too. In the long run, how much of the discovering are we willing to give up, and how much that we would have found in the alleys do we miss by taking the final answer? He does not think the opposite extreme — never ask, always stumble — is optimal either. Nobody knows the mix. If it is really an AGI, that is the question: seventy-thirty, ten-ninety, or "what did you make me for?" The second question would be: what should I be asking that I am probably not smart enough to ask? Fridman worries he would not understand the reply, and would rather know how many living intelligent civilizations are in the observable universe. Levin laughs. He guesses the machine will give a Michael Levin answer: start in this room. Start inside your own body.