Risto Miikkulainen on Neuroevolution, Evolutionary Computation and Creative Machines

High-Level Overview

In this conversation, Lex Fridman talks with Risto Miikkulainen, a computer scientist at the University of Texas at Austin and Associate Vice President of Evolutionary Artificial Intelligence at Cognizant. Miikkulainen specializes in evolutionary computation, the branch of AI that borrows the logic of biological evolution (variation plus selection) to discover solutions nobody programmed in. The talk ranges widely: what would happen if we re-ran Earth's evolution a million times, how aliens might recognize intelligence, the role of emotion and social structure in intelligence, how evolutionary algorithms actually work, neuroevolution (evolving neural networks), multitask learning, evolved language, whether AI would learn to lie, artificial life, cellular automata, and finally advice for young people and the meaning of life.

The tone is curious, gentle and philosophical, with Miikkulainen patiently grounding Lex's big speculative questions in concrete experiments from his lab: basil that thrives without darkness, a tic-tac-toe program that won by crashing its opponents, simulated hyenas learning to mob lions, and robots that learned to fake out rivals. The conversation swings between thought experiments, lab anecdotes and reflections on mortality, diversity and what it means to leave a trace.

The Discussion Explores:

• What a million re-runs of Earth's evolution might reveal, and which solutions (vision, manipulation, communication) would likely reappear
• Definitions of intelligence, from survival with limited resources to leaving a lasting, unique contribution
• Emotions as attention filters, and simulated hyenas that balance fear of lions against loyalty to each other
• How evolutionary computation works: representation, variation, selection, crossover versus mutation
• Why evolution is creative, with examples like 24-hour-light basil and a memory-bomb tic-tac-toe player
• Neuroevolution: evolving neural network weights, architectures, activation functions and loss functions
• Multitask learning, the link between vision and language, and evolving languages we might translate
• Cheating, lying and diversity in agent societies, artificial life, cellular automata, and advice to explore before committing

Deep Dive into Main Ideas

Re-running Earth's Evolution and Detecting Intelligence

Lex opens by asking how much variation we would see if we re-ran the "Earth experiment" a million times. Miikkulainen points out that this is precisely the kind of question computer scientists can approach, since they build simulations of evolution and can run them repeatedly. He expects regularities: evolution took a long time to get started and then accelerated, and certain solutions are likely to be rediscovered again and again, including manipulation of objects, opposable thumbs, some way to communicate (maybe not speech, but sounds of some kind), decision making and vision, which has evolved many times independently. The open question is whether primates and humans, or something like them, would emerge, and whether they would become an apex.

On how to detect intelligence, Miikkulainen says that as the designer of a simulation you would build in detectors and visualizations for the things you care about, such as communication, cooperation and manipulation, while recognizing that detectors may miss forms you did not anticipate. Lex pushes back from an alien's perspective: might they notice insects, dolphins, trees or sharks (which he notes have been around roughly 400 million years, longer than trees) instead of humans? Miikkulainen's first measure would be how much an agent changes its environment. Humans build cities, which stand out even compared with termite mounds and bird nests, although he acknowledges sharks might look smarter for not destroying their surroundings. He is optimistic we can get past that by building in ways that enhance nature's resilience.

They also touch on the origin of life. Miikkulainen notes researchers are getting close to producing self-replicating molecules from a primordial soup with some help, and that the early steps up to multicellular organisms might be simulated. The truly hard part is the runaway evolution of the brain, which happened fast, and nobody yet knows the conditions that produce it or whether it would go the same direction twice.

What Intelligence Is, and Why Mortality Matters

Miikkulainen says his goal is to create intelligent agents rather than define intelligence, but his working definition is an agent with limited sensing and acting abilities plus a mechanism for decisions: can it survive, multiply and solve the problems you give it? That definition includes animals. Lex brings in Ernest Becker's The Denial of Death and terror management theory, the idea that awareness of our own mortality drives human creativity. Miikkulainen finds this plausible and offers a second-level definition: intelligence means doing something that stays behind you, creating something useful for others and for the future, which could apply to artificial agents as well.

Lex worries that this makes everything intelligent, since a butterfly's wing flap can ripple through history. Miikkulainen refines the idea with a what-if-you-were-not-here test: did you contribute something unique that would not have happened otherwise? That need not be on the scale of Mozart; a local, positive contribution during your lifetime counts.

On engineering a fear of death into machines, Miikkulainen frames emotions as functional. Fear focuses attention on what matters when there is no time for logical deduction, and becomes unhealthy only when it paralyzes you in normal life. Agents can be given filter mechanisms that put their computation into a different state, heightening some signals and blocking others, which acts very much like an emotion, whatever label you put on it.

Hyenas, Social Structure and the Origins of Language

A recurring example is Miikkulainen's work modeling spotted hyenas stealing a kill from lions. Individually, hyenas cannot stand up to a lion, but when about 30 of them band together and precisely time their approach, the lions give way. The behavior seems to rest on competing emotional states: they vocalize to call reinforcements, fear the lions, want the kill, and feel a strong affiliation with each other. When affiliation wins, they move together as one unit. Miikkulainen chose hyenas because they are intelligent but relatively rigid, so the behavior can be studied like a breakthrough in evolution about to happen. Researchers can even estimate their emotional states from chemicals in their droppings.

Lex suggests much of intelligence may be an outgrowth of social interaction, and Miikkulainen strongly agrees. He describes a theory of language origins in which social roles came first, roles became exchangeable (I scratch your back, you scratch my back), and brain structures for understanding exchangeable roles became the basis for grammar. Language then lets you refer to things not present, which enables planning and building. He also recalls a demonstration by Marco Dorigo in which small robots crossed a gap none could cross alone by gripping one another to form a longer robot, a mechanical echo of the hyenas' teamwork.

Why Evolutionary Algorithms Are Creative

Asked for his favorite nature-inspired algorithm, Miikkulainen picks evolutionary computation. What fascinates him about computers is that you can get more out than you put in, and evolutionary algorithms take that furthest because they discover solutions humans did not think of. Humans carry biases about what solutions should look like; algorithms do not, so they explore more freely. He defines creativity with three requirements: it has to be new, it has to be useful and it has to be surprising.

His favorite example is a collaboration with Caleb Harper's lab at the MIT Media Lab on a computer-controlled hydroponic "food computer" that regulated nutrients, water, light and temperature. The team paired an evolutionary optimizer with a surrogate model of plant growth to explore growing recipes for basil. They had imposed at least six hours of darkness, assuming plants need night, but evolution pushed every recipe to that limit. When they removed the constraint, the system discovered that basil does not need to sleep: with 24 hours of light it grew bigger and tastier, surprising even the biologists on the team.

Lex and Miikkulainen also discuss the paper collecting surprising anecdotes from the evolutionary computation and artificial life communities, where evolved systems often exploit bugs or loopholes. Miikkulainen's contribution came from a student project in his class: teams built game-playing AIs for tic-tac-toe or five-in-a-row on a large board, and one team's evolved neural network won the tournament largely because opponents crashed. It had learned to make moves millions of squares away, forcing other programs to expand memory until they ran out. Lex calls it a lesson in winning by breaking the opponent's expectations, and Miikkulainen links it to Kasparov's observation that Deep Blue did not play the way he expected.

How Evolutionary Computation Works

Miikkulainen describes the algorithms as caricatures of biology that keep what seems essential: creating variation and then selecting from it. You need a genotype (an encoding, usually a string of numbers or a tree), a decoding into a phenotype that performs a task, and a fitness measure to decide who reproduces. DNA involves folding and interactions beyond the sequence, and biological evolution carries baggage like the appendix and tailbone, so representations are a big open question. One thing still missing in the field is what biologists call major transitions, such as single cells becoming multicellular organisms and then societies, where the unit of selection itself changes.

Variation comes from recombination (crossover) and mutation. In biology crossover is the workhorse, while evolutionary computation relies more on mutation, increasingly guided by statistics about which changes help. Over the field's roughly 50 years, this statistical turn has driven much of its recent success but has pulled it away from biology. Miikkulainen suggests revisiting weaker selection, more crossover, larger populations and more time, because the field is impatient: if something does not perform, it is killed immediately. Biology also rewards many ways of succeeding (niches, size, speed, diet), which breeds creativity, whereas algorithms usually chase a single metric.

He contrasts evolution with deep learning and reinforcement learning. Deep learning excels where there is lots of labeled data. For games, robotics, healthcare treatments or investing, the optimal answer is never known. Reinforcement learning, which also comes from biology, learns within an individual's lifetime, while evolution works over a population and only cares about the best final result. His rule of thumb: if every trial has a real cost, reinforcement learning may be the choice; if trials are cheap (for example via a surrogate model) and you want surprises and broad exploration, population-based evolution is often better.

Learning to Walk, Virtual Creatures and Theory of Mind

Walking robots make the contrast vivid. A reinforcement learning agent tends to be conservative so it can stay upright. Evolution can afford agents that fall flat on their face, jump and fall again, and out of that comes something like controlled falling, and eventually running. These apparent dead ends serve as stepping stones that combine into better solutions. Miikkulainen also describes evolving virtual creatures' bodies (cylinders, muscles, joints, sensors) together with their controllers; the resulting movement looks natural because it fits the body, so observers start to believe the creatures are alive.

Lex raises the dance of communication between humans and dogs or robots, noting that Boston Dynamics' dances are choreographed. Miikkulainen thinks natural interaction may require a theory of mind. He recounts a simulation by Ken Stanley in which two robots foraged for food to gain energy and could win by bumping into a weaker rival. Evolution produced increasingly complex behavior, culminating in a robot that feinted toward distant food so its opponent would travel far, lose energy and lose the contest, something like an emergent theory of mind. For autonomous vehicles, Lex argues pedestrians and drivers share a simple social contract and that cars need more than collision avoidance. Miikkulainen suggests a functional approach of predicting what other agents will do and steering toward desirable states goes far, though genuine theory of mind becomes important when agents must collaborate.

Neuroevolution: Evolving Neural Networks

Neuroevolution combines neural networks with evolutionary computation. Early versions used evolution instead of backpropagation to set a network's weights, useful when there are no targets to learn from, such as an agent running a maze or playing a game. More recently, evolution optimizes deep learning design itself: hyperparameters, depth, topology, activation functions, loss functions and even data augmentation. In one experiment, his team took the architecture that won an image captioning competition, broke it into pieces and searched over recombinations, achieving 15% better performance.

The challenges are defining a search space that is constrained yet expressive, and the cost of evaluation, since every candidate must be trained; a population of a hundred run for a hundred generations is not yet feasible for large networks and has a carbon footprint. Researchers are working on better representations and ways to evaluate designs without full training. Miikkulainen also highlights two timescales: evolution produces a good starting point, a baby neural network, which then learns during its lifetime. Because DNA cannot specify every detail of a brain, biology relies on development and learning, and this interaction can be a powerful way to construct brains and behaviors.

He describes predator-prey coevolution in simulation: hyenas learned to team up and approach from different directions, zebras learned to scatter in different directions like gazelles, then more hyenas circled and herded them. This competitive arms race can produce behaviors nobody built in, though keeping it going requires rich simulations.

Multitask Learning, Vision and Language

Lex asks about Tesla Autopilot's multi-headed network, where a shared core feeds many task-specific heads. Miikkulainen says this is a very good problem for neuroevolution. When tasks are learned together, such as classifying different diseases on X-rays, each forces useful internal representations that help the others. Evolution can decide where and how much representations are shared. His team member Elliot Meyerson found that the tasks do not even need to be closely related: learning language can improve vision, or learning about DNA structure can improve language, which Lex sums up as "the world rhymes."

On vision versus language, Miikkulainen thinks learning both together is the exciting direction. Recognizing objects or parsing sentences is manageable; the hard parts are understanding the 3D world and its relationships, and understanding meaning, which ultimately requires knowledge of society and history. Humans are visual animals who draw abstract concepts on whiteboards, and the visual cortex lights up even during language tasks, so the two share structures. He suspects language itself rests on the more fundamental social structures discussed earlier.

Evolved Languages, Aliens and Whether AI Would Lie

Miikkulainen would like to run simulations in which agents evolve genuine language, with grammar built on social structures, under varied conditions (different gravity, no predators, everyone a predator) to see what conditions are necessary. Then one could ask whether those languages can be machine-translated, giving a sense of the space of possible alien languages. Lex notes it would be good practice before the aliens arrive.

On whether AI agents would learn to lie, Miikkulainen points out that even bacteria have cheaters. Hyena groups include individuals that hang back from the risky confrontation and join the feast afterward. Any society built on trust creates opportunities for cheaters, and his hyena simulations studied what proportion of cheaters a group can tolerate while still functioning. He expects the most successful autonomous agents to be honest because that is the best way to get things done, but some will bend the rules. He stresses that AI is being built as a tool to help us, like cars or language, rather than an entity left to run on its own. Lex argues that deploying millions of identical AGIs might be worse than having diversity, and Miikkulainen agrees that diversity is the bread and butter of evolution, even when some variants look destructive: suicidal hyenas that charge and die form the basis for faster ones that stop just in time. Rule-breaking can sometimes lead to better rules, as society's changing norms show.

Artificial Life, Memes and Cellular Automata

Miikkulainen describes artificial life as a range of work, from multi-agent robots cleaning up after an earthquake to simulations of the origin of life and questions about species loss and tipping points. Societies themselves can be seen as a form of life. Lex brings up Richard Dawkins' memes, suggesting ideas may be what is really evolving, faster than biology. Miikkulainen agrees our minds have extended enormously without much biological change, and he sees AI as another step in that evolution of ideas, a tool we can use through abstractions without understanding every detail, much as we drive cars.

On Conway's Game of Life, both marvel at how much complexity emerges from simple rules. Miikkulainen notes that evolution, like cellular automata, has no goal; progress is our perception. The challenge is guiding emergent complexity toward something useful. He points to novelty search, work by Ken Stanley and Joel Lehman, where instead of a fitness goal you reward solutions that are simply different from what has been seen before. In a robot walking study this produced lots of garbage but also gems, including a very efficient walk, because being truly different forces the system to use the fundamentals of its domain.

Advice for Young People and the Meaning of It All

Translating evolutionary lessons into life advice, Miikkulainen recommends diversity and exploration: take classes in music, history, philosophy, math and engineering, see connections, travel and learn a language. Lex counters that abundant choice can dilute appreciation, noting how the first taste of food after a multi-day fast is incredible. Miikkulainen clarifies he does not mean shallow exploration; you can go deep on a few topics, and it takes real effort (he cites the 10,000 hours idea). Eventually you may find something you cannot stay away from, as computers and AI were for him. Lex sums it up: exploration precedes commitment.

On meaning, Miikkulainen says that in evolution everything is a potential stepping stone. Many attempts fail, like the dinosaurs, which were a good solution for a long time until the environment changed, yet even failures serve as raw material. Being part of a giant, innovative engine adds purpose. He admits he thinks about death more now than when he went skydiving and paragliding in his youth, and sees a logic in the life arc: the young are fearless explorers, and later you ask what you can do that matters with limited time, leaving a trace after the agent is gone. Lex closes with Carl Sagan's words: "extinction is the rule survival is the exception."

Conclusion

Risto Miikkulainen makes a warm, persuasive case that evolution is the most creative algorithm we have, precisely because it lacks our biases. His examples, from sleepless basil and memory-bombing game players to feinting robots and coordinated hyenas, show how variation plus selection can find new, useful and surprising solutions, while neuroevolution brings that power to the design of neural networks themselves. Throughout, he returns to the same themes: intelligence is deeply social, diversity is essential even when it includes cheaters and failures, and patience and exploration pay off. For listeners interested in AI beyond deep learning, or in how the logic of evolution might guide a life of exploration before commitment, this episode offers a thoughtful and accessible tour.

This is a summary of "#177 – Risto Miikkulainen: Neuroevolution and Evolutionary Computation" from Lex Fridman Podcast, giving you the GIST so you can decide if you want the REST. Listen to the full episode.