Cut Off a Planarian's Head. The New Brain Still Remembers

Michael Levin drops the result as if it were lab trivia. Train a planarian, cut off its head, and the tail grows a brand-new brain that still remembers the original information. These flatworms are a centimeter or two long, with true bilateral symmetry, a real brain, and a stack of internal organs. They do not age. There is no such thing as an old planarian, which, he says, already wrecks theories that treat lifespan as a thermodynamic ceiling. The worms in his lab are in physical continuity with animals that were here some 400 million years ago.

Lex Fridman calls Levin one of the most fascinating biologists he has talked to. Levin's lab at Tufts University works on how to understand and control complex pattern formation. Andrej Karpathy is the person who first pointed Fridman at the work, and the host's reason for mentioning that is simple: biology has a lot to teach AI, and AI might have a lot to teach biology. Embryogenesis is the process that makes the point visceral. Each of us starts as a quiescent unfertilized oocyte, basically a bag of chemicals, and months and years later there is an organism with high-level cognition, preferences, and an inner life.

What embryogenesis shows, Levin argues, is that the jump from physics to mind is gradual. There is no lightning bolt that says *now* you have true cognition. DNA does not encode that whole story. It encodes the hardware: proteins, signaling factors, ion channels, the micro-level parts every cell gets to play with. The rest sits in what he calls generic laws, of mathematics, computation, and physics, that are not written in the genome. Try to evolve a particular triangle and you can evolve the first angle and the second; you do not need to evolve the third. Geometry already gave it to you.

Ion channels are transistors: voltage-gated current conductances. Evolve one and you immediately get truth tables and logic functions. You do not have to encode NAND-gate completeness in DNA either. All evolution has to invent is the first little machine that couples you to those laws. Levin admits this is not the standard view in molecular biology. He thinks it is right on the money. Embryogenesis, on this telling, does not build the laws. It creates the capacity to reveal them.

Levin's latest work with xenobots is the simplest test of that claim. Take skin cells off an early frog embryo and remove everything else. In the intact animal those cells are a passive two-dimensional barrier keeping bacteria out. Left by themselves, within about 48 hours they form a little creature that runs around the dish, navigates mazes, and acts both independently and together. Sprinkle loose cells into the dish and the xenobots collect them into piles that become the next generation. It is von Neumann's dream of a machine that copies itself from material in its environment, implemented by wild-type frog skin. No genomic editing, no nanomaterials, nothing added.

The engineering move is subtraction. Other cell types normally bully these cells into being skin. Take the bullies away and their default behavior is to be a xenobot. Ask where a tadpole's form comes from and the textbook says selection over millions of years for froggy environments. Ask where a xenobot's form comes from and there is no answer of that kind. There have never been xenobots. There has never been selection to be a good xenobot. In two days the cells find a proto-organism lifestyle, including kinematic self-replication, that is not how frogs reproduce and is not done anywhere else in the biosphere. A xenobot, Levin says, is both a self-assembling proto-organism and a biological robot; those classes are not distinct.

Working with Josh Bongard's group at the University of Vermont, the team uses simulated evolution to ask how to poke these cells with signals, not new DNA: remove some, add some, give them experiences. The deeper shift is the material itself. Old engineering used wood and metal, passive stuff that kept its shape and did nothing else. These are *agential materials*. The substrate has an agenda, billions of years of evolution, goals, and preferences. Levin's figure, later accepted in a Nature Bioengineering paper, is a tower of Legos versus a tower of dogs. Stack Legos and you have full control, but if someone knocks it over the game is over. You cannot stack dogs, but if you train them they get back up.

The useful engineering question is not what something is made of or whether it was designed or evolved. It is how much you can depend on it to do when you are not around to micromanage, and therefore whether you should place parts, train it, or send it signals. Humanity trained animals for thousands of years before anyone knew what was between their ears. Regenerative medicine, Levin argues, needs the same move. Do you want to control tens of thousands of genes to grow an arm back, or find the high-level control knob that says *build an arm here, you already know how*? A salamander amputated anywhere along the limb grows exactly what is missing and then stops, when a correct salamander arm has been completed. That is a goal-directed loop: know the target, measure the delta, keep remodeling until the error is zero.

People get yelled at in class for saying their cells want things. Levin thinks that taboo leaves capabilities on the table. There is no central dictator in this science, because everything is made of parts. We feel like a unified intelligence, but we are a bag of neurons. A lot of people reserve "real" intelligence for themselves and treat ant colonies as a maybe. That split does not survive contact with the fact that we, too, are a collection of cells navigating the three-dimensional world. Even "central nervous system" is a misleading name. Zoom in and you see cells running around. There are no singular, indivisible diamonds of intelligence anywhere we have looked.

The cheap move, once you see how a collective works, is to say it is just physics and therefore not real cognition. Levin calls that fundamentally flawed. Zoom into anything and you will see physics; it is not going to be fairy dust. What matters is that certain arrangements of physics and chemistry, especially bioelectricity, yield a collective with goals, preferences, memories, and anticipations that do not belong to any subunit. That is the origin of a self with a capital S. Fridman notes how uncomfortable it is to treat human civilization or Earth as one living organism. Levin wants to grow up past the idea that the human-scale self is where selves are first born.

He likes to roll cognition backward. Grant that the two of them sitting there are examples of true cognition if anything is, then walk back day by day through childhood toward a paramecium and an unfertilized oocyte. No one has a convincing discrete step where the cognitive powers disappear. Biology offers a slow continuum, which is what you would expect from evolution and developmental biology. The other story is how many selves are in an amniote egg. A fertilized bird egg holds about 50,000 cells in a flat disc. One region declares it will be the head and tells the others not to be. Scratch that blastoderm so cells cannot talk for a while, and both regions decide they can be the embryo. When they heal you get conjoined twins. You can make two, three, many. How many selves are in there cannot be answered until the process plays out.

A region of cells organizes into a collective that has goals individual cells do not: make a limb, make an eye, make exactly two eyes. Cells do not know what an eye is. The establishment of a boundary that can pursue those goals is the origin of selfhood. Levin does not think evolution produces solutions to specific environments so much as problem-solving machines that work in many spaces. Brains are a late development for pursuing goals in three-dimensional space by commanding muscles. They evolved from a more ancient system in which cell collectives navigated morphospace, the space of anatomies, and before that transcriptional space and metabolic space. Every collection of cells has to figure out from scratch where it ends and the world begins.

Biology, Levin says, uses a multi-scale competency architecture. Molecular networks have goals, cells have goals, tissues, organs, and colonies have goals, and the interplay lets the system solve problems in new ways, xenobots included. Evolution is itself a kind of agential bioengineering. Sometimes it changes hardware, a protein's structure. Much of the time it changes the signals cells give one another, which is what engineers do when the material already wants to do something. He likes competency as an engineering term because it is empirically useful. If calling a system a thermostat, or a better thermostat, helps you control, predict, and build it, the philosophy can wait. A goal is whatever the system spends energy to return to after you deviate it.

He sometimes gives a talk called "Why don't robots get cancer?" Typical robots are dumb parts hoped to add up to a smart machine. Biology is not like that. Every level has an agenda, and the outcome is cooperation *and* competition within and across levels. Tissues and organs compete during embryogenesis for information and limited metabolic resources. Each level also bends the option space for the level beneath, terminology Levin takes seriously from relativity. Lower parts cannot see the big picture. If you deform the space correctly, they just go down their local gradient and end up doing the higher level's bidding. Conversely, the higher level does not need to compute low-level controls. It only has to bend the space.

Tadpoles becoming frogs have to rearrange the face: eyes forward, jaws out, nostrils moved. The old assumption was that every piece simply travels a fixed distance in a fixed direction. Levin's group made Picasso tadpoles, eyes on the back of the head, jaws off to the side, the whole face scrambled. They still make pretty normal frogs. Parts take novel paths until they reach the correct frog-face configuration, then stop. Now imagine a mutation that does something useful and also moves your mouth to the side. Without multi-scale competency, fitness goes to zero because you cannot eat, and you never get to explore the beneficial side of the mutation. With it, the mouth finds its way. Mutations that would have been lethal become neutral. The fitness landscape gets smoother, and evolution can actually run.

The catch, when people compare this to human institutions, is that every component has its own goals, usually with little regard for other levels. You go rock climbing, scrape cells off your hands, and come home happy. Those cells are dead. Levin is not recommending Stalin. Fridman notes that "all men are created equal" is a fight *against* this architecture. The AI contrast is the flatness of most current systems, and the fact that their sensors, effectors, and borders are given from outside. Biology constructs those from scratch. Josh Bongard's 2006 robots that did not know their own shape are a start. The further step is not even knowing what sensors you have, or where you end, only that you must minimize surprise under metabolic constraint and coarse-grain the world into cats, table edges, and a region you call *me* because you have more control over it.

You cannot afford to be Laplace's demon. Under energy and time limits you have to coarse-grain, and you start telling stories about agents that do things: avoid this, chase that, that is food. Primitive organisms already model the world as full of agency. Turn that model inward and you say you have agency too, and then you decide. One way to view free will, Levin suggests, is as that energetically efficient narrative. Any self-constructing agent under energy constraints is going to believe in free will. Whether it *has* free will is a separate argument. The agential view of self and world, he thinks, is inescapable even for primitive organisms, scaled to their metacognition.

Humans are very good at detecting agency in medium-sized objects moving at medium speeds. A bowling ball and a mouse are obvious. From birth, that visual chunk is the training set. We have no high-bandwidth sense of blood chemistry, of what the pancreas or liver is doing. If we did, we would be good at detecting intelligence in physiological space. Facial symmetry as a beauty cue is, in his telling, a half-assed version of the same inference: if your cells could not cooperate enough to keep the organism symmetrical, you can guess what else is wrong. Why didn't evolution give us a higher-resolution health signal? He guesses an arms race. You want your own parts to understand one another, but you do not want to be too readable, or parasites and cheaters hijack you.

The most evolvable system has strong top-down control: easy to say *make another finger* instead of listing ten thousand gene-expression changes. That same openness is a gift to hijackers. Planaria carry a bacterium that interfaces with the system that calculates how many heads the worm is supposed to have and can make it grow two. You want to be understandable to your parts and not too understandable to everyone else. Levin calls it a biological game of Texas Hold'em: some cards shown, some hidden, with bluffing in the mix. That tension, he suspects, is why we do not walk around broadcasting everything about our physiology.

Where did neurons come from? Not from nowhere up in the skull. They evolved from a much more ancient ability of cells to form networks and solve problems such as navigating morphospace. Ion channels, neurotransmitter machinery, electrical synapses: all of it is older than brains, older than neurons, older than multicellularity. Bacterial biofilms already show brain-like dynamics, work Levin flags from UC San Diego. Evolution learned early that electrical networks are good at memory, at integrating information across distance, and at optimization, long before there were brains. Bioelectricity is not a new obsession of 2022. People were thinking about electrical phenomena in life by 1903, probably the late 1800s, with waves of research in the 1930s, 1940s, and again in the 1970s through 1990s.

What Levin's group adds, he says, is treating this not as another piece of physics to track but as a privileged computational layer that gives access to the basal cognition of tissue. A cell membrane holds ion-channel proteins that let potassium, sodium, and chloride through under certain conditions. Charge imbalance makes a voltage gradient, and living cells spend a lot of energy holding a particular voltage. Neighboring cells can share that state through gap junctions, little docking tunnels that let ions flow both ways. Werner Loewenstein rediscovered gap junctions; the first inventors, Levin jokes, probably lived four billion years ago. Many of those channels and junctions are themselves voltage-sensitive. That is a transistor, and the platonic space of mathematical truths opens again.

Neuroscientists already watch this in brains. Transparent zebrafish light up as they decide what to eat. Levin's claim is that your whole body is doing a slower version of the same thing. Embryos are thinking about anatomical space: how many fingers we are supposed to have, how many we have now, what to do about the gap. Gap junctions are magic because they wipe ownership metadata. A chemical signal arriving at a receptor is obviously from outside. A calcium spike that crosses a gap junction looks exactly like one of your own memories. If you cannot tell whose memories are whose, that is the beginning of a mind meld, the scale-up from me-and-you to us. State lives in many layers: gene-regulatory networks that fall into stable dynamical states, cytoskeletal configurations, and electrical flip-flops that hold a bit without any hardware moving, like volatile RAM.

Fridman asks whether the future of programming looks like Karpathy's Software 2.0: set hyperparameters of a biological-like system and watch it correct itself. Levin says that is a good description of their regenerative-medicine program, a kind of somatic psychiatry. Most medicines other than antibiotics fail the moment you stop them. You addressed symptoms. Side effects are huge because you interacted at the lowest level, like trying to program a computer by changing the melting point of copper. Yoga and biofeedback people already talk as if the body has intelligence, and Levin thinks they are pointing at the better locus of control. Figure out what tissues can learn, why drugs habituate, whether chemical pathways show sensitization and associative memory. Somatic cells, not just neurons, form electrical networks that process current shape versus goal shape during development and regeneration.

Planaria, Levin says, hold the answer to pretty much every deep question of life. They resemble our ancestors more than earthworms do. They are immortal. They reproduce by grabbing the dish with the back end and ripping themselves in half, so they never pass through the single-cell bottleneck of an egg. Humans squash down to one cell and expand again, like an autoencoder. These worms just tear. Thomas Hunt Morgan cut one into 276 pieces, and each piece grew a perfect worm. Chop a worm in half and as the missing half grows, the original tissue shrinks so the new tiny head stays in proportion. Starve them and they shrink; feed them and they expand. Anatomical control is, in his word, insane.

They generally do not get cancer. They can learn. Train one, cut off the head, and the regenerated brain still holds the lesson. The tail without a head just sits there until a new brain arrives. They have the same neurotransmitters we do. The experimental question is sharper: cut off head and tail, and the middle fragment must make one head and one tail. How does it know how many, and where? The hypothesis was a somatic electrical network that remembers the correct pattern. Visualizing electrical activity the way people look at brains, they saw a specific pattern that shows where head and tail go. Then they edited it, with drugs that target ion channels, into a two-headed pattern. The cells built a two-headed worm. The genome was wild type.

Cut those two-headed animals into pieces again and some pieces keep making two-headed animals. You can take a two-headed worm back to one head. When they reproduce by tearing, you still get two-headed animals. Levin's thought experiment: throw them in the Charles River, wait a hundred years, scoop up a one-headed form and a two-headed form, call it speciation, sequence the genomes. Identical. DNA produces hardware that defaults to one head, the way a calculator defaults to zero when you turn it on. It is a programmable calculator. Bacteria and parasites can rewrite that memory too. Hardware is essential; without it you never get the physiology of the memory. It does not fully determine the information once the hardware is there.

Most animals do not pass somatic mutations to children because of the egg bottleneck. Planaria have kept every non-lethal mutation for 400 million years. Their bodies are mixoploid: every cell may have a different chromosome count. Look at the genome and it looks like a tumor, an incredible mess. Look at the anatomy and they are the best regenerators on the planet, rock-solid and cancer-resistant. Levin's evolutionary reading is a runaway. Multi-scale competency hides genomic quality from selection: fitness can be high because the genome is great or because the genome is junk and competency papered over it. The more competent the tissues, the harder it is to pick the best genomes. Pressure piles onto the algorithm until what survives is whatever can build a great worm no matter how bad the DNA. Evolution, he says, does not take the past too seriously. It builds problem-solving machines, not exact replays of last time.

The word *xenobot* causes a little drama. Levin thinks a lot of biologists want a sharp line between living organisms and robots: robots come out of factories, are predictable, and are made of metal; living things are magical. Those binary categories were never good and will be useless in a couple of decades. *Xeno* is for *Xenopus laevis*, the frog the first ones were made from, but he treats them as a biobot technology. Once you can communicate with the cells' inputs, you can get them to build useful machines. That is robotics. The first papers stayed away from synthetic circuits on purpose. Engineer the cells heavily and critics will say you made them do it. The point was to show what native cells do that is surprising before you touch the hardware.

*Xenopus* has been convenient since the 1950s: adults in a frog habitat, tens of thousands of eggs, development in a dish at room temperature, organs forming in front of your eyes, decades of tools. Mammalian twins already show the same logic. Cut a mammalian embryo in half and you do not get two half-bodies; you get two normal bodies, a regeneration event. Levin stresses that xenobots are not a frog phenomenon. Unpublished work, he says, has made them from other things. The plasticity is not about being a frog. The coordination question during cleavage is real, but biology does not assume the past is as expected. An old experiment on newts, going back to the 1940s, makes the point.

A newt kidney tubule in cross-section is eight to ten cells cooperating to make a tube. Force early divisions so cells are gigantic while the whole animal stays the same size, and the cross-section may be four or five cells, then three, until one enormous cell wraps around itself and still gives you the same large-scale structure. The molecular mechanism has been swapped. Cell-to-cell communication becomes one cell using its cytoskeleton to bend. If you were an embryonic cell with a fixed script of who does what, you would be screwed. Life is highly interoperable: change part size, DNA content, stick in electrodes or weird nanomaterials, and it still works. William James defined intelligence as getting to the same goal by different means. This is that. *C. elegans* is the counterexample, a numbered cookie-cutter of identical cell lineages. Most organisms are plastic.

Levin asks students to imagine shrinking to cell scale in the middle of an embryo. Cells run around, some die, the count changes every time you look. Without knowing what development is, you would have no clue that this activity always builds the same thing, let alone what that thing is. Even with the full genome we have no way to predict the anatomy they are building. Evolution's trick is hardware with a stable baseline mode, very good at doing the same thing left alone, plus enough problem-solving that wrong cell sizes, wrong cell numbers, or an electrode halfway through the body still get most of the job done. The human brain is incredible. Levin's complaint is not that it gets too much credit. It is that treating it as a sharp binary, you are a brain or you aren't, creates pseudo-problems. Other architectures, chimeras, plants, embryos, and things that are sort of brains can have cognitive status too.

Asked how many intelligent alien civilizations are out there, Levin says he has no expertise, then adds that he has met the ones his lab made. Standard model systems are an N of one: one course of evolution on Earth. Inferring the laws of biology from that sample is testing a theory on the data that generated it. Xenobots have no history of selection to be good xenobots. The lab makes frog–axolotl chimeras and hybrids of living tissue with robots. Until we find actual aliens, we need those constructs. Even synthetic biology is locked to Earth's components and lab environments, and Levin thinks it is very hard to create something truly different. What might be universal is not carbon or the genetic code. It is resource limitation, a hostile world, a boundary between self and world that nobody hands you, coarse-graining of experience, and the drive to minimize surprise. Selves, in his framework, are demarcated by the scale of goals they can pursue.

If plant cognition already weirds you out, he says, you are not ready for exobiology. People can take robots two ways: expand the category of mind, or define anything designed as fake, a magic trick that stops being fun once you see the method. Levin finds it strange to insist that real intelligence can only come from a blind hill-climbing search and never from a smart engineering team that can also use evolutionary techniques. Unconventional cognition, a term he did not coin, is whatever today's landmarks cannot classify: no frontal cortex to measure, no familiar brain structure, because the thing is synthetic, engineered, or alien. He is interested in mind in all its implementations. Andy Adamatzky's work on slime molds sits in that literature. Communicating with such a system means learning its space, not only whether it sees or feels vibration but what is salient, and then running a spectrum of persuadability instead of assuming where it sits.

A gene-regulatory network looks like dumb hardware you can only control by rewiring. Simulations, and now lab work, show those networks can have habituation, sensitization, and associative memory. Charles Abramson and Levin wrote about behaviorist protocols for a synthetic organism you know nothing about, starting simple and moving up, making no assumptions. If you can find the currency of reward and get the system to do something it was not doing, you have communicated. If it gets you to do something you were not doing, it has answered. Levin's son suggested a slime-mold experiment: *Physarum* grows toward an oat; you razor off the advancing piece. That fragment faces a game-theory problem with a non-fixed number of agents. Eat the oat alone, or rejoin, after which there is no *me*, only *we*. Empirically they tend to merge first. Actions can change payoffs and also change what you are.

Levin almost never talks about consciousness as such. Cognition is easier to handle experimentally. The test he insists on is engineering: does the claim help you control, predict, and design the next experiment? Some people will reject a theory that attributes tiny goals to thermostats on philosophical principle. He will not. First-person consciousness is different because we do not even know what format a correct theory would output, besides perhaps a poem that puts you in a similar state. Intelligence, he agrees with Fridman, is publicly observable. If a shiny thing gets off a craft, hands you a poem, and you cannot look in a skull for a frontal cortex, existing tests are primitive. Knowing something is as cognitive as you are is also no guarantee of kindness. People with the same anatomy kill each other daily. Long before frog skin, there is the way we already treat mammals we know are like us.

Where did evolution come from? Levin thinks it is inevitable once you have heredity, imperfect heredity, and selection. Those three things and you are off to the races, in computers, in chemical systems, on cosmic scales. The standard story says the process is completely blind, a reaction against creationism when the only other option was human-like foresight. He is not sure. Everything is a continuum. Multi-scale competency might bake in a little look-ahead without being smart the way we are. Whether rewinding the tape of life, as Simon Conway Morris and others have argued, would hit a deep attractor for something like us, or whether frozen accidents send every run somewhere else, he does not know. We are bad at predicting attractors in complex systems. What we urgently need is a science of the cognitive goals of composite systems: swarms, cellular collectives, internet-of-things, social and financial structures. We build them constantly and cannot predict what autonomous goals they will have, of which we are now cogs.

People email him that the work is not natural. Evolution, he replies, optimizes for biomass. It is not optimizing for your happiness, or necessarily for intelligence or fairness. The Garden of Eden assumption that natural is for the best is long gone. On a practical level, what he hopes to understand is the embodied intelligence of organs and tissues well enough to change regenerative medicine. The endgame is an anatomical compiler. You sit down, draw the body or organ you want, not molecular details, a six-legged frog with a propeller, a heart like this, a leg like this, and the system converts that description into stimuli that convince cells to build it. He probably will not live to see it. He thinks it is achievable. Except for infectious disease, that would dissolve birth defects, traumatic injury, cancer, aging, and degeneration. Instead of heroic, expensive patches on a sinking 90-year-old ship, you regenerate early, before the system degrades.

Xenobots might clean environments or joints. More important is using synthetic systems to learn how to detect and manipulate the goals of cellular collectives. The ethical payoff is retiring the old tap test: metallic clang, take it apart; squishy and warm, be nice. 3D printing already makes bladders and ears because you place every cell, but CRISPR-style reverse engineering, *this is the knee I want, which genes do I edit*, is the hard inverse problem. The better move is to rewrite what cells think they are supposed to be building and take your hands off the wheel. AI can help the way AlphaFold helped folding, and Levin's group has used it to model repairs of frog brain defects and melanoma. Today's popular approaches, scaffolds plus cells, stem-cell transplants, work for relatively simple structures. For an eye or a hand, he thinks the only hope in a reasonable time frame is to understand how the thing was motivated to get made in the first place, so we can all be like planaria.

Adult frogs, unlike salamanders, do not regenerate legs. Levin's group more or less solved that. After amputation you put on a wearable bioreactor, work with David Kaplan, loaded with a silk gel and drugs such as ion-channel drugs, telling the cells they should regrow what normally goes there. The device stays on 24 hours. Then you take it off and do not touch the leg again. We do not know how to make a frog leg. The frog does. Eighteen months later you have a pretty good leg. They have a spin-off, Morphoceuticals, now in mouse trials. The proof of concept is the window right after injury, when cells are deciding what to do. Impact them then. Once they have decided to make a leg, they do not need you.

The wrong question, Levin says, is why there is cancer. The right question is why there is ever anything but cancer. In the normal state, cells cooperate toward a large-scale goal. If a cell isolates from the electrical network that remembers that goal, it reverts to a unicellular lifestyle. The border between self and world changes. Connected by gap junctions, the cells are one self with tissue-level goals. Disconnect, and the self is tiny. Cancer cells are not more selfish. They are equally selfish with a much smaller self. The goals of a tiny self are proliferate, and migrate to wherever life is good: tumor and metastasis. People noticed years ago that converting cells close their gap junctions. Until those junctions close, as in the slime-mold cut, there is no *you* that can entertain leaving. Afterward the rest of the body is just environment.

The lab worked out a way to control cells' bioelectric state and physically force them to remain in the network. Nasty oncogenic mutations such as KRAS then greatly reduce tumorigenesis, or cells that had begun to convert go back to being normal. Sequence the nucleic acids and you will see the KRAS mutation and predict a tumor. There is not a tumor, because bioelectrics kept the cells connected and they are making skin and kidneys. The work has started moving to human glioblastoma cells. Levin thinks a real cure is possible. Immunotherapy is a promising technology if the immune system can be taught to recognize enough of the cancer. Chemotherapy, hoping to kill more tumor cells than you, takes an unbelievable toll because the cells are yours. Keeping cells harnessed toward organ-level goals is, he argues, more fundamental. Nobody is then making a tumor.

Viruses, in this picture, are mostly passive. Levin does not see much reason to attribute a mind to them, though they hijack other minds. Fridman pushes the thought that both virus and cell are transformed by the encounter. Levin has not found the living/non-living binary useful. Level of cognition as a continuum is interesting; after you stamp something living or not, he does not know what you do next. The actionable questions are how you relate to it: take it apart, listen to its advice, or something in between. That is what *agential material* is for. *Teleophobia* is his word for the unwarranted fear of granting too much agency to a new system, the fear of anthropomorphizing. He thinks anthropomorphism is a holdover from a pre-scientific age in which humans were magic. A cognitive claim is an engineering claim. If you say this thing knows, hopes, wants, or predicts, hand over the protocol that hypothesis implies, and we will see if it helps.

The cognitive light cone came out of a 2018 challenge from the diverse-intelligence community: a framework that could analyze a human, an octopus, and a plant together. What cognitive agents share, whatever their origin or architecture, is some competency to pursue a goal. Levin draws something like an inverted Minkowski cone, space collapsed to one axis, time on the other, and semi-quantitatively estimates the spatial and temporal goals a creature can work toward. A bacterium maximizing a local chemical has a tiny cone. A dog remembers a little backward and predicts a little forward, but will not care about the next town in four weeks. A human might work toward world peace after death. Somewhere there may be a Buddha-like intelligence that can really care about everyone's welfare the way we cannot. It is not a map of sensing range or acting range. It is the size of the goals you can envision. That lets you put AIs, aliens, and swarms on one diagram. *Target morphology* names the fact that biology often navigates toward a region in anatomical space. You know because you deviate the system and it goes back.

Levin is wary of advice because things change so fast, but one distinction has held in academia. If someone smart, successful, and competent gives you specific critique of what you have done, that is gold: a chance to hone the craft and find mistakes. If they tell you what you ought to be studying, how you ought to approach things, what the right way to think is, ignore most of it. Successful people are well calibrated on their own ideas and their own field. They are not calibrated on yours. "This is a dumb idea, don't do that" is generally worse than useless. It is demoralizing and limiting. Most of what he considers interesting and useful in his own work, very smart people said was a terrible idea. At best we know what we ought to be doing. We very rarely know what anybody else should be doing.

A crowd saying something is stupid is not data. It might mean it is stupid, or it might mean there is a chance to do something new, and you cannot tell which. You have to build your own intuition by taking risks that make sense to you, making mistakes, hitting dead ends, and learning to trust your gut even when the standardized approach says slow down. Life is hard and science is hard. You are going to sweat and bleed. You should do that for ideas that fire you up. Death, climbing the competency hierarchy, is a factor that promotes change and turnover. The first thing that could be said to die is already a fuzzy question; a cabbage in the fridge does not announce the moment. An aquatic animal that dies in a pond still has living cells. Some could crawl off as amoebas, some could join into a xenobot. Planarian cells do not obey the Hayflick limit. HeLa cells have had a life far larger than Henrietta Lacks's body. Molecules cycle through ecosystems. Even cells need not die when the organism does.

A human cognitive light cone, perhaps for the first time, can hold goals longer than a lifetime and guaranteed not to be achieved. That may be a perennial thorn in our psychology. Another puzzle is giving up. Insects crawl until they physically cannot. Mammals, including rats, can decide a situation is hopeless and die while they still could have kept going. In unpleasant drowning experiments, rats normally last a couple of minutes; if you teach them that treading water brings rescue, they can last about an hour. They literally give up. Evolutionarily it looks like a bad strategy: one time in a thousand you get saved if you never quit. It suggests metacognitive controls in which survival is no longer the top drive. Suicide is a fascinating evolutionary question. Levin is not particularly worried about death itself. He has helped people pass. Sitting in a facility that stretches you out as long as possible does not seem good, and modern society offers few chances to sacrifice yourself for something useful. The ride began after billions of years he was not around. That is okay.

Asked the meaning of it all, he is not sure the question is well posed. Scientific discoveries about consciousness will not alter the primal first-person fact. If he opened himself and found cogs, his conclusion would not be that he lacks true cognition. It would be: wow, cogs can have true cognition. Fake memories, Boltzmann brains, a non-corporeal soul: here he is now. Experience is what matters from the inside; the backstory is a third-person curiosity. He is also unimpressed by *is* to *ought* shortcuts, including "that is not natural." We are in a position to do better if we can. When he says science and nothing else, he does not mean today's textbooks, which will look hilarious. He includes first-person investigation. Unlike a third-person experiment that barely changes the scientist, studying consciousness makes you part of the apparatus. "We are our own experiment." The conversation itself, two systems interacting, is that kind of experiment. Fridman closes with Darwin: from the war of nature, endless forms most beautiful and most wonderful have been and are being evolved.