If You Reran Earth a Million Times, Would Anything Like Us Appear?

Lex Fridman introduces Risto Miikkulainen as a computer scientist at the University of Texas at Austin and associate vice president of evolutionary artificial intelligence at Cognizant. Fridman's aside is that nature-inspired methods, from genetic algorithms to cellular automata to neural nets, keep surprising him, and that the long arc of computing seems to run toward biology rather than away from it. Then he asks the biologist's impossible question in a form a simulator can actually run: if you replayed the Earth experiment a million times, how much would the outcomes vary?

Miikkulainen says biology cannot answer because the sample size is one, but a computational model can be rerun. We cannot quite do it yet, he adds, though some regularities should show up. Evolution was slow to start and then accelerated. Object manipulation, opposable thumbs, some form of communication, decision-making, and vision would likely reappear. Eyes evolved many times; the hardware differs, but the job gets done. Whether primates, humans, or anything that looks like us would show up is the open question. Tool use and environment construction, though, he expects. Once an agent can reshape the outside world, the survival edge is overwhelming. What that species looks like could be different on every run.

Fridman asks how a busy creator would even notice humans in a soup of agents. Miikkulainen would instrument the run: detectors and visualizations for communication, cooperation, and manipulation, the things we already roughly want. Aliens watching Earth might not pick us. Insects dominate by biomass; sharks have been here 400 million years, longer than trees, through mass extinctions. Humans are recent and noisy. His first computational metric is environmental effect. Cities are constructed habitat, and most people live in them. Birds nest and termites mound, but nothing else builds at that complexity. He grants the shark counter-score: staying around without wrecking the place may be the smarter trick. The detector is biased. Run often enough, he says, and you can revise it.

Step zero is a self-replicating molecule out of primordial soup. Miikkulainen says people are close: with a little help and the right conditions, you can make replication happen. Evolution took billions of years because it did not know the destination; we do, so recreation should be faster. The next few steps, even multicellularity, he thinks are reachable with some hand-holding. What still fascinates him is the later runaway evolution of the brain that produced humans and other higher animals, and did it fast. Is that replicable, and if it happens, does it go the same way? Nobody knows the conditions well enough to start that experiment in silicon. Fifty years of evolutionary computation, he later notes, have never produced a biological-style major transition.

Asked what intelligence is, he puts on the computer-science hat and refuses to start with a definition. The goal is to create intelligent agents, and that is a way of defining it. A program with limited sensors and effectors, interacting with a world, plus a mechanism for decisions: can it survive, multiply, or solve the problems you give it? That bar is well below human. Many animals clear it, and some things that do not look like animals might too. Intelligence, in that first sense, is a survival skill given the resources you have.

Fridman brings in Ernest Becker's *The Denial of Death* and terror-management theory: humans foresee that the ride ends, and that knowledge powers creative work. Miikkulainen thinks the philosophy holds. Animals probably do not think about death that way; people know time is limited and want it to count. That impulse is a second-level definition he likes: you do something that stays behind you, useful to others or to the future, not only to yourself. It does not require a biological body. Fridman worries about butterfly-effect uniqueness, the *It's a Wonderful Life* test. Miikkulainen calls that a Mozart bar, too high. A local, time-bounded effect is enough: if you were not there, this would not have happened in this way, and you want it to be a positive one.

Can you engineer a fear of mortality into an agent, not just a lifespan parameter but something that tightens as the end approaches? Miikkulainen splits cognition from emotion. Most people know they will die and get on with it. Paralyzing terror is unhealthy; not caring at all is not the normal human state. The useful middle is more than logic. Emotions, in his view, are a survival filter: when there is no time to deduce danger, fear locks attention onto the facts that matter. That can be implemented as a computational state in which some inputs are heightened and others do not get through. You can refuse to call the box a real emotion. It still behaves like one.

His group modeled African hyenas stealing a lion kill, a behavior more complex than almost anything else they do. Hyenas are intelligent but not baboon-intelligent, which is why they are observable. About thirty of them can vocalize for backup, time an attack, and push lions off a carcass even though a single paw-strike could kill one of them. Three pulls compete: fear of the lion, affiliation with the group, and hunger. Affiliation wins by a little, they move as a unit, and the raid works. Fridman hears an algorithm, and that is the point. Baboons and humans have too much going on inside the head. Hyena motives show up in motion, and chemicals in droppings can even index the last 24 hours of emotional state.

Miikkulainen thinks a great deal of intelligence is a product of social interaction, and that useful agents will have to be fundamentally social. He cites a language-origins theory, the line associated with Derek Bickerton: social systems come first, with interchangeable roles, I scratch your back and you scratch mine. Once a brain can represent actions as exchangeable roles, grammar has a foundation, and symbols can start referring to objects that are not present. Language is a byproduct of that structure, then a tool for planning. Marco Dorigo's small robots, grabbing one another to bridge a gap none could cross alone, are the computational rhyme. Hyenas alone fail; as a team they do not.

Asked for the most beautiful nature-inspired algorithm he has met, Miikkulainen does not name a paper. He names the fact that a computer can give you more than you put in. Freshman year, a tiny program took a number and returned a result he did not have to calculate, and it blew him away. Push that and machines learn patterns humans struggle with. Push again and evolutionary algorithms discover solutions you did not think of, often because they do not carry your biases. Creativity, as he defines it, has three requirements: new, useful, and surprising. Algorithms can hit all three. Watching a program live past the source listing, he agrees with Fridman, feels adjacent to watching something alive grow. You gave it a starting point. Then it went and had a life.

The cleanest example is a collaboration with the MIT Media Lab's OpenAg food computer, a hydroponic box where light, water, nutrients, and temperature are all under software control. Farmers know their own land because they cannot control everything. When you can control everything, that folk knowledge collapses. They built an evolutionary optimizer plus a surrogate model of plant growth and searched light recipes for basil: intensity, wavelength, duration. They boxed the search with a common-sense constraint of at least six hours of darkness, because Earth has night. Evolution slammed every recipe against that wall. They lifted the constraint. Basil does not need to sleep. Twenty-four-hour light produced bigger, tastier plants. It surprised the engineers and the biologists, who had assumed the constraint existed for a reason. Evolution did not share the bias.

Fridman asks about Neuralink-style brain expansion: are skull, lifespan, and senses fundamental to intelligence, or quirks we could open up until intelligence explodes? Miikkulainen is optimistic about replacement and cautious about expansion. Animal work that pipes auditory signals into visual cortex still lets the animal perceive; children who lose half a brain can migrate function to what remains. Prosthetics will work as much because the brain adapts as because the device is elegant. Flooding the brain with extra input, though, may just overwhelm it. Filtering and focus are already what cortex is expensive at. Expansion is already happening outside the skull: phones as lookup memory, children learning DNA in school a few decades after it was discovered. Vision itself is a firehose the brain already refuses to drink whole. He is impressed by a monkey playing Pong through an implant, as long as the channel stays on human terms.

Surprises happen constantly, Miikkulainen says, often by exploiting a bug you did not know was there. The community collected those stories in *The Surprising Creativity of Digital Evolution*. His own entry is smaller than the basil result and came from a class. Students evolved neural nets to play tic-tac-toe, actually five-in-a-row on a large board. One team won the tournament mostly because opponents crashed. Evolution had learned to play millions of squares away. Rival programs grew their boards in memory until they ran out and died. Winning, Fridman notes, is sometimes not playing better inside the rules. It is finding the edge of the opponent's brain. Garry Kasparov said something similar after Deep Blue: it was not playing the chess he expected. Missing human priors is a feature.

Stripped of decoration, evolutionary computation is two operations: create variation, then select. You need a digital genotype you can modify, and a decoding into a phenotype that acts and gets scored. Computer science reaches for strings of numbers or trees. DNA is a string too, but it also folds and interacts in ways most models ignore, and we still do not know which of those extras are essential. Biology is full of baggage, appendixes and tailbones that made sense on the way here. The representation has to support that kind of history, and strings and trees probably cannot. The field's conspicuous gap is major transitions: single cell to multicell, individual to society, a jump in the unit of selection. Attempts at open-ended evolution have not reproduced that climb. It likely needs a richer encoding, and biology itself has not fully explained the jump.

Variation can be recombination, including crossover, or mutation, random edits to the chromosome. In biology, crossover is the workhorse. In computation, mutation does more of the work, increasingly steered by performance statistics, the same statistical turn as the rest of machine learning. That success, over about twenty years of a fifty-year field, has also pulled the algorithms away from biology. Miikkulainen would like to walk some of it back: weaker selection, more crossover, larger populations, more time. Today's evolutionary computation kills a weak individual immediately and races toward one fitness number. Biology is patient, and success has many shapes: stronger, faster, bigger, smarter, this diet or that niche. One metric collapses the search. Many metrics are where creativity lives. The cost is slowness, and humans are in a hurry.

Deep learning, he says, eats labeled data, including cases like weather where tomorrow's observation labels today's input. Plenty of real tasks have no right answer and never will: games, robot actions, medical treatment, investing. Reinforcement learning, borrowed from lifetime biology, is an individual eating berries, getting sick, getting stronger. Evolution works at another timescale. A whole population discovers what works; failed individuals die; the next generation is better. Reinforcement learning, in the original picture, cares about the lifetime. Evolution is engineering: you do not care what the candidates suffered, only which solution remains. The boundary is blurring, but the practical split is trial cost. Expensive trials favor cautious reinforcement learning. Cheap trials, especially with a surrogate, plus a hunger for surprise, favor a population that will try things RL would not dare.

Put the two algorithms on a simulated walker and you can see it. The RL agent is conservative, trying to stay upright as long as possible. Evolutionary agents go haywire, fall flat, hop, fall again, and eventually fall on purpose, which is to say they run. Those ugly gaits are not pure dead ends. Combine two of them and you may get the one that works. Evolution does not require every step to be an improvement, which is why it can leap. Another study evolved not only the controller but the body: cylinders, joints, muscles, sensors. Morphologies came out with many legs or none. When body and controller optimize together, the motion looks natural, fitted to that physique, and you start believing the creature is alive.

Natural motion is also why Boston Dynamics dance videos impress and then disappoint. The choreography is hardcoded by people. Learning a quality of movement you would want to live with, Fridman says, is the hard problem, probably only solvable in simulation if the physics is honest enough to transfer. Miikkulainen thinks that sometimes requires a theory of mind: the robot understands you because it is doing something similar. In a Ken Stanley simulation, two robots foraged for energy and, once strong, rammed each other. Evolution taught nearest-food first, then better routes, then watching the opponent. The most sophisticated strategy faked toward two distant pieces so the other robot would spend its energy on the trip, then turned back to eat the nearby one and win. Nothing in the code said deceive. Competition did.

Fridman describes pedestrians at a crosswalk: a glance at the driver, then a step, a thin social contract that the person in the car will not murder you. Current robots do not dance. They treat humans as ballistic objects with a fat safety envelope and stay away. Miikkulainen does not think you need a full theory of mind. A functional version goes far: predict what the other agent will do, score those futures as good or bad, approach the good. Collaboration raises the bar. Two people lifting a box have to infer whether the other is grabbing, setting down, or just passing, which physics traces will not tell you. Higher still, you predict career-scale goals, fame or money or doing good, which takes more information but is the same logic. Tesla-style systems, in Fridman's view, still optimize collision avoidance rather than that dance.

Neuroevolution mixes neural nets with evolutionary search. Early versions used evolution as a stand-in for backpropagation, evolving weights as a string of numbers, which matters when there is no target to send a gradient from: a robot in a maze, a game, any task without a correct action at each step. You still want a net because it interpolates. More recently the interesting move is to evolve the design itself. Architectures are now so complex that humans barely understand which module is doing what. Evolution can search depth, topology, activations, loss, data augmentation, the whole experimental layout. A few years earlier his group took a winning image-captioning architecture, broke it into parts, and searched around that human design. Performance rose 15 percent. The number matters less than the implication: even a champion net had unused room nearby.

Two hard constraints remain. The search space is huge, so you must define the building blocks and have a hunch about which directions are worth trying. And evaluation means actually training the net. Days per model, a population of 100, 100 generations: the compute bill, and the carbon bill, do not close. Predicting quality without a full train is one of the field's central problems. Another path is grow-then-expand. Esteban Real and colleagues evolved a small network and scaled it up. Miikkulainen is more interested in a true developmental story: evolution supplies only a starting point, and lifetime learning, in interaction with an environment, builds the adult, the way human DNA cannot specify an adult brain. Two timescales, slow population evolution and fast individual learning, still look underexploited.

Predator-prey is the prettiest arms race his group has run, with zoologist Kay Holekamp: simulated hyenas and simulated zebras. At first hyenas blundered into prey and ate. Zebras learned to flee. Hyenas learned to approach from two directions. Zebras split, the real gazelle trick that confuses a chase. More hyenas joined, circled, and took several animals at once. None of those tactics was written in. Competitive coevolution's value is that it hands you things you did not know. Keeping the race going is hard; the simulated world has to stay rich enough for the next complexity. They only got a few steps, but each step was a strategy nature already uses.

Fridman asks what neuroevolution would do with a Tesla Autopilot-style multi-headed net, a shared core with task-specific heads, the kind of system Andrej Karpathy's team was running at scale. Miikkulainen calls it a natural fit. Learn several pathologies from x-rays at once and the embedding forced by one disease helps the others; you do better on each task than you would alone. Evolution's job is deciding how to merge: which layers, how much sharing. Colleague Elliot Meyerson has worked on what a representation looks like when it has to support many heads. That, Miikkulainen says, is how biological intelligence already works. You do not build a representation for one job. You build the structure of the world so that future tasks, including ones you have not met, have somewhere to sit. Multi-task gains are a side effect of that aim.

The surprise Meyerson found is that the tasks need not be related. Language can help vision; DNA structure can help language. "The world rhymes," even across fields that look desperate. Asked which is more fundamental, language or vision, Miikkulainen will not pick. Joint training on images plus text learns deeper semantics than either modality alone, and he thinks that is the next few years' direction, past ImageNet and text-only corpora. Vision is hard at the level of 3D relations and prediction, not object labels. Language is hard at the level of goals, plans, and history, not who-did-what-to-whom. Visual cortex lights up for semantic concepts; people draw pictures of abstractions on whiteboards; many animals see and do not speak, so vision may be older. Speech itself may be an accident of an eating organ that could make sound, or a successor to gesture. Underneath both, he suspects, are shared relational structures.

Fridman worries that if aliens landed, humans would fail to notice or to talk. Miikkulainen, who got into computing partly through science fiction, wants a concrete rehearsal: evolve agents until they have grammar, not just signals, then try machine translation on those languages. Vary the worlds, no predators or all predators, different societies, even no gravity, and map the space of what language can become. Do it before anyone visits. The same frame can grow explainable AI: evolve agents that can talk about their world to us, rewarding communication that actually helps a shared goal. He immediately adds the catch. Evolved communication will likely include lying. Bacteria have cheaters. In hyena raids, some animals hang back from the risk and walk in for the meat. He has measured how many a group can tolerate before the raid stops working. A few. Not a lot.

Fridman hopes honesty and kindness are evolutionary advantages once shelter and food are not scarce, and then says the scary version out loud: the first systems to pass a Turing test may be the ones that lie exceptionally well. Miikkulainen hopes we build AI as a tool, like cars or language, that helps us and that we can still limit. Far future unconstrained entities are not the foreseeable case. Trust, though, will not go away, because a society that cannot trust cannot do the interesting work, and wherever there is trust there is an opening for cheaters. Successful learners will mostly be honest because that is how you get things done. Some will bend rules. Cloning one AGI into millions of identical copies would be an engineering convenience and a strategic mistake. Diversity is the one thing an evolutionary run cannot live without. Suicidal hyenas in the sim looked like waste; they were the raw material for animals that run fast and stop in time.

Tolerate a few misfits, he says, because rules themselves change. Behavior once considered proper later looks unjust; yesterday's violation becomes a new norm. Change often starts with the agents that will not follow the current script. The dose still matters. Too many cheaters and the group fails. If a set of agents can trust one another they achieve much more, and if there is trust there will always be room to deceive. He does not think that tension disappears.

Artificial life, for him, is a spread: multi-agent robots cleaning a building after an earthquake at one end, primordial-soup chemists chasing self-replication at the other. It is a tool for asking what life is, where a tipping point in species loss might sit, and whether a hyena-like discovery of cooperation is a pivot you could notice in time. Societies themselves can be treated as a life form. Biological evolution is slow; ideas, institutions, and culture are not. Humans have barely changed in a thousand years and already argue about DNA and fundamental particles, because thought copies faster than genes. In Dawkins's frame, memes are what evolve and people are the vehicles. AI is the next tool in that fast layer. If the universe runs in dimensions we cannot internalize, we already know the workaround: drive a car without understanding every part, use an interface. AI can be another abstraction over a world we do not fully get.

Fridman mentions Conway's Game of Life, and Conway's recent death, and how completely intuition fails when a few binary rules pour out complexity. Miikkulainen agrees. Evolution has no goal. It reacts to the present; survivors remain; we name the residue progress. DNA is a long, simple sequence. The code that mutates strings and trees in evolutionary computation is tiny. Non-linear interaction, not ornate rules, is the source. The practical problem is how to steer that flood so it is useful, and how to live with the death count. Novelty search, work associated with Ken Stanley and Joel Lehman, drops the fitness target and rewards only difference from what has been seen. It still evolved efficient walking. To get truly far from existing solutions, Miikkulainen argues, a system has to encode the domain's fundamentals first, then change them. Most of the archive is garbage. Among the garbage are gems, and a human still has to recognize them.

Advice for young people, from an evolutionary-computation angle, is exploration, and it has to be deep. Take music, history, philosophy, math, engineering, and look for the joins. Travel. Learn a language. The material is at your fingertips and still costs effort. Fridman worries that maximizing novelty flattens experience, the way fasting makes the first meal vivid and abundance dulls it. Miikkulainen did not mean a hundred shallow tries. Pick a few topics, stay until you can honestly say you have taken what you can, then move. That judgment may take years; people switch careers after a decade. You do not need 10,000 hours of expertise, but you need enough understanding that the stretch is usable later. Then comes the second phase: something you cannot stay away from. For him it was computers and AI, decades of work that could absorb everything the exploration gathered. Order matters. Explore first, then commit. Without the search you are not qualified to choose.

From a single agent's view, after so many simulated deaths, what is the meaning? In evolution, everything is a possible direction and a possible stepping stone. Some become foundations. Some die out and still were attempts. Dinosaurs were an excellent solution until the environment changed; their relatives are still around, and who knows when that lineage is useful again. An individual is a small cog in a huge innovation engine, which sounds diminishing and, to him, supplies purpose: you belong to something larger. He thinks about death more than when he was skydiving and paragliding. The life arc makes sense. The young have to be relatively fearless because they are the ones who wonder what is over the mountain. Later, after exploration and a chosen direction, the question becomes what you can do that matters with limited time, and whether a trace remains after the agent is gone.

Fridman closes with Carl Sagan: extinction is the rule, survival is the exception. Most species that appeared in four billion years are gone. The hyenas that charged and died, the walkers that fell, are footnotes to that rule, and also the material the living stand on. Miikkulainen does not sell comfort or panic. Individuals are potential directions. Some work. Use the finite stretch: search deeply enough to find the thing you cannot leave, then spend decades making something that outlasts you. The freshman-year astonishment still sits under the research. A machine computed a result he did not have to. Later it learned patterns. Later still, evolution found answers he would have missed because it did not share his idea of what an answer should look like. We can build systems that are, in some sense, smarter than we are. That, he says, is what still takes him apart.