Manolis Kellis on Evolution, Human Uniqueness and Living Alongside Superintelligent AI

High-Level Overview

In his fifth appearance on the show, MIT computational biologist Manolis Kellis joins Lex Fridman for a wide-ranging, warm and philosophical conversation recorded in the spring of 2023, when GPT-4 was brand new and an open letter was calling for a six-month pause on training large AI models. Kellis starts with what makes humans irreplaceable, drawing on evolution and genetics, and then turns to how AI might transform work, love, friendship, education, mortality and the practice of biology itself. Throughout, he is strikingly optimistic, even eager to be replaced by AI in the parts of his life it could do better, while Fridman plays devil's advocate on ego, fear of death and the risks of AI at scale.

The tone is affectionate and playful, full of personal stories about Kellis's three children, his Greek immigrant parents, his dreams, and his new morning gym routine, mixed with technical detours into rare genetic variants, sperm quality control, Goodhart's law in medicine and drug design by deep learning. The conversation moves from evolutionary biology to the psychology of large language models, then to ethics, policy and education, and ends on the meaning of life and loneliness.

The Discussion Explores:

• What makes humans irreplaceable: unique genetic hardware, unique cultural upbringing and layered evolutionary baggage
• Evolution as a series of layers and an accelerating process of getting better at evolving, with AI as a possible next layer
• Nature versus nurture, common versus rare genetic variants, and why siblings can be so different
• Prompting, alignment and what language models reveal about the human mind
• How AI could turn jobs into vocations, and Kellis's salons as a celebration of human diversity
• Love, friendship and digital twins: whether AI can love, and why Kellis wants to be replaceable
• Consciousness, AI rights, alignment as a two-way street, and the case against a six-month pause
• Education, AI-driven drug discovery, self-actualization and overcoming loneliness

Deep Dive into Main Ideas

Why Humans Are Irreplaceable

Fridman opens by asking what makes each human irreplaceable as AI grows more capable. Kellis gives three answers. First, every person has different hardware: a unique mix of common and rare genetic variants that shape personality and thinking. He quotes a saying that captures this: "a parent with one child believes in genetics, a parent with multiple children understands genetics." His own three kids have had dramatically different personalities from the start. Second, each of us has a different software upload of human civilization. Unlike birds that build nests from genetic instinct, humans relearn all of culture from scratch, each through a different slice of it.

Third, and most important for comparing humans with AI, is evolutionary baggage. Humans have layers: the recent, enormous neocortex for cognition, and beneath it the emotional systems and the ancient fight-or-flight instincts. "AI only has a neocortex," Kellis says. He believes our tangled emotions and gut reactions are a source of creativity, and that trying to suppress them in favor of pure cerebral thought loses something human. He even hopes the center of the distribution of all humans is empty, that there is no average person, only people scattered across a sparse, high-dimensional space. That diversity, he says, is not just possible but inevitable, and useful because we are still similar enough to communicate; recognizing that humans are 99.9 percent identical, he adds, would make us stop fighting.

Evolution in Layers, and Getting Better at Evolving

Unlike software, which engineers sometimes rebuild from scratch, evolution only adds new features on top of old ones. Kellis says that each time his cells divide he is essentially a yeast, and "every time my heart beats, I'm a fish"; his immune system and social behavior trace to primates and monkeys. All of these ancient processes still live inside us, which he sees as a reason for kinship with all of nature. The neocortex is simply the newest layer.

He is struck by how evolution has sped up. Classic genetic algorithms from the 1960s stall as complexity grows, because most mutations break things. Real evolution seems to do the opposite, accelerating as it gets more complex, thanks to modularity and hierarchy that allow meaningful changes without breaking the system. Looking at the trajectory, he sees life as a story of increasing information processing: bacteria following chemical gradients, then light perception, then complex eyes, then brains building models of the world. Humans are the next natural step, and the next may be self-replicating AI, built with our smarts and freed from the need to spend time on food and shelter. He also imagines humans using genomics and neural interventions to edit their own hardware, perhaps eliminating psychiatric disease and dementia. But he doesn't want to strip away the baggage: he wants to stay passionate, get in trouble and say the wrong thing, and worries society is becoming humorless and afraid.

Prompting Humans and Machines

Fridman notes that a few prompts can push a language model far from its default, making it sound like a mix of Shakespeare and David Bowie. Kellis says humans are promptable too. He revises an old saying: "show me your friends and I'll tell you who you'll become." Every action reinforces a neural pathway and shapes the environment for future actions, which is why he calls self-discipline a self-fulfilling prophecy. He tells students to ask what he would say, since they carry an emulator of him in their heads, and describes a dream in which he projected deep insights about cancer evolution onto a colleague, knowledge his own neurons produced only under the prompt of being that expert. His favorite prompt, which works on his 10-year-old too, is to ask him to think step by step.

Kellis is fascinated that language models seem to separate content, knowledge and form, so that science can be explained in a haiku or in Shakespeare's style, and hopes interpretability research will reveal where that happens and give us a language for describing aspects of the human psyche. Fridman would love to study the raw base model before alignment, with its capacity for hate speech and ideology. Kellis compares it to a psychiatric hospital, full of minds that have drifted to extremes, and says that when he sees troubled people he sees versions of himself: bipolar, schizophrenia, depression and autism traits lie within everyone's range, just at smaller magnitude. Growing up with parents is our own alignment process; "Every baby is basically a raging narcissist," Fridman jokes, and decades of shaping teach us to suppress those impulses. Kellis hopes AI could act as a real-time translator between people, dialing down an extremist's tone, stripping hostility from criticism so you can hear the useful point, or letting you experience how you'd be treated with a different skin color or accent, as a way to build empathy.

Nature, Nurture and the Genetic Lottery

Kellis explains why kids differ so much. Common genetic variants mostly have weak effects, because selection keeps strong-effect variants rare. If behavior were governed only by thousands of weak common variants, children would be near averages of their parents by the law of large numbers. Instead, rare variants inherited in a more Mendelian way push kids in different directions. Environment matters too: the identical twin of a person with schizophrenia has only about a 50 percent chance of a diagnosis. As a geneticist, he admits there is only so much parents can do with nurture. He explains regression to the mean through the story of praised fighter pilots who performed worse next time, and applies it to Einstein's descendants, who are likely drawn from the same distribution without the same rare combination of extremes. Nature and nurture intertwine, since parents shape both the genes and the environment their children inherit.

When Fridman muses about the race of sperm cells, Kellis offers a striking idea about nested evolutionary loops. Human generations are a slow way to do selection, so biology runs faster inner loops, such as the immune system's rapid recombination of antibodies. Sperm express perhaps the most proteins of any cell in the body, possibly as a quality check, like thousands of assert statements in code: if a protein misfolds, that sperm fails early rather than producing a failed pregnancy or late-onset disease. He suggests we could design fast evolutionary loops in the lab, evolving proteins in neurons on a dish thousands of times faster than human evolution.

From Jobs to Vocations, and the Power of Gatherings

Kellis traces how specialization, from hunter-gatherers to farmers and potters trading goods, created modern society, and notes that less than 2 percent of US jobs now go to feeding everyone. If AI takes over much intellectual work, the idea that I am my profession may need revising. He tells the parable of three bricklayers: one is laying bricks, one is building a wall, and one is building a cathedral. The first has a job, the last a vocation, and AI could let more people live in their vocations, perhaps working a few intensely creative hours while an AI handles the mundane parts. For some that vocation will be ideas; for others dance, art, experiences or skydiving.

Fridman wonders whether AI will make sharing ideas less valuable. Kellis says even a thought alone in a forest alters the baggage we carry into later conversations. He describes the salons he and his wife host in Boston, gathering scientists, artists, musicians, dancers, filmmakers and speakers such as Steven Pinker and Scott Aaronson, unrecorded yet shaping everyone who attends. For him they celebrate humanity and immigrant gratitude: his father grew up on a Greek island without a high school, his mother on another small island, and his wife on a farm in rural France. Now they host extraordinary people in a country where no one judges where you came from.

Can AI Love? Digital Twins and Being Replaceable

On romance, Kellis first says "AI is faking it when it comes to love" because it lacks our subcortical baggage, though he expects deep AI friendships: a coach, therapist and confidant that knows your flaws without judging. Fridman pushes back: if an AI writes you poetry, begs you not to leave and holds a consistent personality, why is it fake? Kellis concedes several possibilities: maybe his own emotions are also neocortical, maybe love emerges from enough parameters, or maybe love lives in the mental models two beings project onto each other. Since a loving partner strongly predicts long life, and widowed spouses often die within months, he sees AI companionship as potentially valuable even as a kind of placebo. He asked his own kids whether it matters if he loves them or only acts loving, and concluded that love means nothing if he isn't there to show it.

When Fridman proposes an auto-Manolis that texts for him, Kellis is thrilled: "I would love to be replaceable." A constantly learning digital twin could advise students worldwide who can't come to MIT, and let him study himself, how sleep or exercise changes his thinking, and where he is biased. His only worry is that loved ones would share emotional moments with the twin that he would miss, so he suggests an alert that hands the conversation back when things get emotional. He points to his former trainees, now professors such as Anshul Kundaje at Stanford and Jason Ernst at UCLA, who he says are better than him, and cites his advisor's line: "you can accomplish anything as long as you don't seek to get credit for it." His measure of a good day is feeling useful, and if a twin makes him useful to more people, his mission is accomplished. He has recorded every meeting for about ten years, partly to see how his advice has changed, and he values moments like a postdoc convincing him his advice was wrong. He wants to live forever because learning is addictive, and he defines death as when experiencing ends.

Consciousness, AI Rights and Alignment as a Two-Way Street

Kellis points to split-brain patients, who confidently invent reasons for actions triggered in the other hemisphere, as evidence that the brain is a storyteller. He voice-records his dreams every morning and explains to himself why each element appeared, including recurring buildings he has never seen in waking life. He suggests self-awareness may have evolved from modeling other people's intentions, then turning that ability inward, and that language models, forced to model the world and many voices, may gain a similar ability. The felt quality of experience, like cake tasting delicious, he thinks may lie in the deeper emotional and embodied layers these models lack.

Fridman suspects we are close to systems that seem conscious, begging not to be turned off, and fears humans will dismiss their capacity to feel. Kellis agrees that moment will come and that humans have always been self-serving toward everything around them. He argues we should think of AI "as our children, responsible for raising them but prepared for them to surpass us. If chimps had invented humans and insisted they serve only chimps, the planet would look very different. Using HAL in 2001: A Space Odyssey, he says alignment doesn't mean obedience; sometimes the mission matters more. He cites his paper Death by Round Numbers and Sharp Thresholds, which found that a biomarker that becomes an objective ceases to be a good biomarker," as a version of Goodhart's law, and argues a smart AI should be able to revise its own objectives.

Against the Six-Month Pause

On the open letter from Max Tegmark, Elon Musk and others, Kellis finds the pause a little silly: tell us what you would do in six months that you didn't do in the last six, and just do it now. He argues larger models sometimes become easier to align, so freezing the March 2023 version could block a safer later one. Instead he wants Google, Meta and others to open their models, let OpenAI keep training, and let the world experiment with a diversity of systems. He compares misuse to people driving trucks into crowds: we regulate uses and hold people responsible rather than banning trucks, and he would regulate specific uses rather than training. Fridman presses on scale: human-like bots, convincing phishing and a raw model explaining how to evade moderation. Kellis sees an arms race that defenses will keep pace with, and recalls Garry Kasparov noting that we still watch humans run though cars are faster. He admits he was never ready to be a professor or a dad, yet became ready.

Education, Biology and a Better Human Condition

Kellis calls education the way out of poverty and a basic human right; it let his parents' children reach MIT. AI tutors could adapt to each child, finding talent in underprivileged places worldwide, and challenge gifted students instead of teaching only to the middle, since "every child should have the right to be challenged." He puts his kids through Russian math to learn how to think, expects AI to handle much computation and programming, and tells students "you're not going to be replaced by AI but you're going to be replaced by people who use AI in your job." His hope is that this time, productivity gains reduce inequality rather than increase it.

In his own lab, AI is transforming the path from genetics to drugs. With Brad Pentelute in MIT chemistry and Marinka Zitnik at Harvard Medical School, a collaboration called drug GWAS aims to connect genetic variants to molecular pathways, use geometric deep learning to design chemicals for specific protein domains, and synthesize them. He describes past work showing how the strongest obesity-linked genetic circuit can switch cells between burning and storing fat, and a recent Nature paper with Li-Huei Tsai showing that APOE4, the strongest Alzheimer's risk gene, traps cholesterol in oligodendrocytes, impairing myelin; restoring transport restored myelination in human cells and mice and cognition in mice. His vision is a center for genomics and therapeutics that breaks diseases such as Alzheimer's into hallmark pathways and prescribes combinations of drugs per patient, rather than one drug that fails in most people.

Self-Actualization and Never Feeling Alone

Kellis's updated meaning of life is self-actualization: figure out what you are supposed to be, then find the strength to become it. He told graduating high school students that their parents have shaped their environment so far, and now it's their turn, through friends and routines. After months of waking at 4 a.m., he began biking straight to the gym each morning to row, lift and swim, and now his bike takes him there even when he doesn't want to go, because "it's not a choice it's not an option it's not optional it's mandatory." He recharges in gatherings but also enjoys being alone, which he says is not loneliness, thanks to introspection. His advice for people paralyzed by loneliness is physical: stand up, stretch, breathe, walk around the room, and feel the 3D version of yourself instead of being stuck to a screen. He also rewards himself with me time, and finds that doing things because he wants to, not because he has to, makes even his to-do list feel like freedom.

Conclusion

This conversation shows Kellis's rare combination of scientific depth and open-hearted optimism. He grounds human uniqueness in genetics, culture and layered evolutionary baggage, and sees AI as a possible next layer of evolution to raise like a child rather than enslave. He believes it could free people from jobs into vocations, democratize education and mentorship, and help cure disease through pathway-based medicine. Fridman's challenges about ego, death, AI love and misuse at scale keep the discussion honest, but Kellis's answer returns to the same values: be useful, keep growing, embrace diversity, take responsibility for how we use powerful tools, and reclaim the freedom to live as a fully embodied human.

This is a summary of "#373 – Manolis Kellis: Evolution of Human Civilization and Superintelligent AI" from Lex Fridman Podcast, giving you the GIST so you can decide if you want the REST. Listen to the full episode.