AI Only Has a Neocortex: Manolis Kellis on Evolutionary Baggage and Superintelligent Partners

Lex Fridman opens by asking, as capable models arrive, what still makes any one human irreplaceable. Manolis Kellis, on his fifth visit to the podcast, answers with three kinds of uniqueness. Each person inherits a distinct mix of common and rare genetic variants — hardware that, he says, a parent of one child can still blame on parenting, while a parent of several understands genetics. Each person also relearns civilization as software, unlike a bird that can build a nest from instinct. The third difference is evolutionary history: humans were not designed from scratch.

Like ogres in *Shrek*, he says, humans have layers. The outer one is the enormous neocortex, the latest cognitive innovation. Under it sits emotional machinery, and under that the older repertoire of fear, fright, and flight. "AI only has a neocortex," Kellis says. "AI doesn't have a limbic system." When people try to suppress the "more human" layer in favor of a purely cerebral one, he thinks they lose creativity and freshness — the part he calls quite irreplaceable.

Fridman maps living people in a high-dimensional space and wonders about a center of mass. Kellis would rather believe the center is empty: there is no average human, only a sparse cloud under the curse of dimensionality. Diversity is not merely possible; it is wired in. His siblings, his three children, and his wife's six siblings look similar from a distance and then, up close, like different people. The differences are still small enough that conversation works. If humans really absorbed that they are 99.9 percent identical, he jokes, they would stop fighting — and relative to the rest of life, with no close kin for six million years, they are a lonely dot.

Software engineers sometimes throw away a codebase and rebuild. Evolution, Kellis says, does not. It layers new features on old ones. Every time his cells divide he is still yeast, a unicellular organism running the same division program. Every heartbeat he is still a fish: the heart has barely changed, blood still moves, oxygen still matters. The immune system is primate. Social behavior is still that of New World and Old World monkeys. Trace any of those behaviors and it is still running inside a modern body.

That continuity is, for him, an ethical argument as much as a biological one. All those processes never stopped existing; they continue to live, breathe, divide, metabolize, and fight pathogens in us. The neocortex and human reasoning sit on top of other species that have not left. Realize that, he says, and the case against killing other humans becomes a case against killing other species as well — a reminder of how united we remain with the rest of nature.

The timeline still startles him. Billions of years to a dividing cell, then multicellular bodies, complex body plans, and elaborate senses. Bats evolved flight and sonar in a few million years — a blink. Humans themselves changed an extraordinary amount in little time. The reason, Kellis argues, is evolvability: life took a while to get good at evolving, and then modularity, hierarchical organization, and reusable constructs let meaningful changes happen without breaking the whole system.

Classic genetic algorithms from the 1960s run the other way. As you pile on complexity, the mutations that knock you out of a working regime explode, and the mutations that make you better shrink, until further improvement becomes vanishingly unlikely. Biological evolution looks like the opposite: it appears to speed up exactly as complexity increases. The system, he says, got good at evolving. That is the wall software, organizations, and companies keep hitting when a local change cannot be made without collapsing the rest.

Asked where it is all headed, Kellis says the best way to see the future is the past. The trajectory of life on Earth is information processing. Bacteria first do chemotaxis, moving up a chemical gradient — the start of hunting food. Then comes light. All known life evolved on a rock that turns every 24 hours, so light is energy, a cue for which way is up, and eventually shape, motion, and direction. Cognitive machinery grows on those inputs to model the environment. On that information-theoretic view, humans are the next natural step: many animals beat us on other dimensions, but on cognition we are unsurpassed on this planet, and perhaps in the universe.

Trace that forward and one possibility is that the next layer of evolution builds the next layer. Humans, using "old hardware" that was advantageous in Africa and then across the globe, created technology that can do many cognitive tasks without the earlier biological baggage. Programming languages train systems like ChatGPT, which then write the next layer of software. Kellis likes that the creators are still around to guide it, "hopefully for the rest of eternity as partners." He also likes the energy accounting: most animals spend about 80 percent of waking hours hunting or building shelter; humans maybe 1 percent, with the rest left to creative work. AI does not worry about shelter at all. It lives in cognitive space. That, he says, is a very natural next step.

Humans, meanwhile, are learning to read their own genome — "the ultimate layer of introspection." That is not only a chance to collaborate with AI. It is a chance to understand neural pathways for empathy and justice, then steer the inherited bag of hardware with neuronal, chemical, and electrical interventions. Further out, understanding genetic code raises the question of getting rid of psychiatric disease, neurodegeneration, and dementia, and perhaps of augmenting capability rather than only removing illness.

Fridman asks whether we could tinker with the genome the way AI systems discarded some biological baggage in software. Kellis refuses the premise. "I wouldn't wanna get rid of the baggage. The baggage's what makes us awesome." Being angry, hungry, or hangry may feed his creativity. He does not want to be dispassionate. He wants to get in trouble, say the wrong thing, make an awkward comment, and push people into reactions that make them think differently. Society, he worries, is sliding toward a humorless space where everyone is so afraid of the wrong sentence that people quit en masse and stop liking their jobs.

The arrival of systems that can do seemingly extremely intelligent tasks in a blink might, he hopes, democratize intellectual life instead of flattening it. Instead of funneling the same mathematically extraordinary brains into the same universities, perhaps we can ask who needs that scarcity now. "AI can do the math," and then the work is to gather humans who think extremely differently from one another. That mix, not a table of similar prodigies, is "the true cradle of innovation." Celebrate the diversity that springs from baggage rather than pulling everyone toward conformity.

Fridman notes that a few prompts can shove a language model far from the center, even into a stable mix of Shakespeare and David Bowie. Kellis says humans work the same way. Put someone in an environment and their behavior changes. The proverb is show me your friends and I will tell you who you are; he prefers "show me your friends and I'll tell you who you'll become." You first choose people like you, then you adopt what that circle treats as normal. Every action shapes today's outcome and tomorrow's neural pathway. Self-discipline is a self-fulfilling prophecy: behave as the person you want to be, and you build the environment in which that behavior is default.

Long before ChatGPT, he told students to ask what Manolis would say — they already run a decent emulator of him. That is cousin to rubber-duck debugging: explain the bug to a silent duck and the explanation itself solves it. He once woke from a dream of a lecture in which a friend, cancer genomicist Piyush Gupta, delivered deep insight about how cancer genomes evolve. The insight was his own neurons, fired only under the prompt that he was now that professor. Unprompted, he might not get there. His favorite prompt is "think step by step," and it works on his ten-year-old: try the math in one leap and Kellis knows the mistake; ask the child to go step by step and the mindset changes. He thinks human-in-the-loop reinforcement learning trained models to be similarly aligned to those prompting opportunities.

What is unique, he adds, is that each of us was trained on a different subset of culture, while ChatGPT was trained on all of it and can probably emulate almost anyone. Finding that location in cognitive space with a few prompts is impressive; that the location exists is beautiful. More beautiful still is that extreme overparameterization let models figure out that context, knowledge, and form are separable — scientific content can arrive as a haiku in the manner of Shakespeare. That decoupling, he thinks, is a fact about the human psyche as well as about machines.

Convolutional nets were easier to interpret because of locality and layered receptive fields. Language models are harder, but Kellis still thinks ablation can work: drop out part of the network, see what prompts it generates, and slowly invent language for the dimensions of behavior. Better names would help humans too. Today a speaking style only "reminds us of a friend"; if models name the axes, those concepts become clearer in our own psyche and can be encoded back into machines.

Fridman wants the pre-RLHF base model — before safety censorship — to see hate speech, sexual language, and whether it can be a communist, a fascist, or a capitalist, and which parts activate. Maybe those ideologies are connected by the fact that the mind is drawn to a centralizing idea, and "maybe we need a neural network to remind us of that." Kellis ties the pull of ideology to promptability and to words humans invented to encapsulate behaviors. Looking at how old words for kindness, anger, and jealousy are across languages, stacked against genetic traces of migration, is a way to date pieces of the psyche. If the words look similar, they may be ancient; if they differ, the concepts may have been invented separately.

As for the unfiltered system, psychiatric hospitals are already full of people whose minds have gone adrift in particular locations of the psyche. Watching films about troubled minds, Kellis sees versions of himself, less extreme: bipolar, schizophrenia, depression, autism. Those behaviors sit in the vector span of every human; most people simply have smaller magnitudes because they did not inherit the variants or the environments. Growing up is decades of being shaped into place. Without that, Fridman says, we might have unfiltered GPT-4. "Every baby is basically a raging narcissist," he adds. Not all of them, Kellis laughs — his three children already diverged early: one wants the bigger piece, one demands exact equality, one is fine with the smaller part.

He and his wife each carry on the order of six million common-variant loci, plus rare variants inherited in a more Mendelian fashion — "2 to the 6 million" combinations even before the rare ones. Common variants tend to be weak because selection removes strong-effect alleles; a large schizophrenia risk will not rise to high frequency. If personality were only those thousands of weak loci, children would be robust averages of their parents by the law of large numbers. They are not. That points to rare variants governing much of behavior, biology, and psychology. An identical twin of someone with schizophrenia still has only about a 50 percent chance of the same diagnosis, so development, environment, and trauma still matter. As a geneticist and a father who spent enormous energy teaching sharing, kindness, and trust, he has to admit there is only so much nurture can do.

He explains why statisticians talk about regression. Fighter pilots who had an extraordinary dogfight were praised, then often did worse; the navy read that as praise ruining them. The statistical reading is regression to the mean: an extreme sample from a high-performing distribution is unlikely to be matched next time. Einstein sits at an extreme of intellect; his children and grandchildren are more likely redraws from the underlying distribution than a dynasty of genius. Read a child's traits as a draw around the parental average, then offset by the rare variants they inherited, then layered with nurture. Nature can look like nurture because children inherit genes *and* the environment those genes helped shape. "Nature can influence nurture, nurture can influence nature" — at least they correlate — and untangling them is very hard.

Fridman's tangent about hundreds of thousands of sperm, only one of which wins, is not ridiculous, Kellis says — and the winner is not random. Humans are a slow-evolving species; generation time is a terrible selection loop. The immune system evolves much faster via V(D)J recombination, manufacturing huge antibody diversity in days. Sperm may express more proteins than any other cell as an early integrity check: fail fast, before a late-onset disaster. He analogizes that to 20,000 assert functions: if a protein cannot fold, that sperm is gone. Nested inner loops, run thousands of times faster than evolving humans for another million years, are also how he imagines screening neurons or proteins in a dish rather than waiting on another million years of people.

People at OpenAI have told Fridman the base model is extremely difficult to work with, like some humans. After alignment it is easier — which makes Kellis wish we could work with more diverse people, not fewer. He imagines a real-time translator that dials an extremist friend down, or strips New York cadences toward Texas, so you can hear the information without being dissuaded by the emotion. Criticism wrapped in hatred is hard to use; the same technical point, recast as a genuine question about a missing control, gets the experiment run. Virtual reality or AI could let you rehear yourself as Chinese, South African, Nigerian, in a hoodie or a jacket, and build empathy by disconnecting dress and accent from content. Deepfakes, for once, as a tool against snap judgment.

On the human condition, he starts from specialization. Once agriculture let people live together, a potter who could not hunt still fit, because pots traded for rabbits. Currency extended that. "I am my profession" may not survive systems that can write software, drive, or interpret a genome. Less than 2 percent of U.S. jobs now feed the country; 98 percent of the economy sits above subsistence, turning old wants into new needs. AI might free people for art, emotional maturity, and a better work-life balance: two ferociously prepared creative hours while the system answers the phone and revises the papers. The bricklayer joke is the point. One man is laying bricks, one is building a wall, one is building a cathedral. The first has a job; the last has a vocation. Kellis would rather edit fewer papers and spend more time on disease circuitry.

Even a thought had alone in a forest, he says, alters the baggage you carry out. He bikes to work listening to books; ideas return in the evening, or after falling asleep with a book on. Gatherings still move history, for good and ill — early Nazi cells as well as constructive salons. His own Boston gatherings mix MIT and Harvard people with artists, musicians, painters, dancers, and cinematographers, often unrecorded. His father grew up on a Greek island without a high school; the island was liberated from Ottoman rule only in 1920. His wife grew up on a farm in rural France. Hosting Steven Pinker, Scott Aaronson, a presidential advisor, an Italian composer, a schoolteacher on *Making Americans*, is how two immigrants shape the environment their children inherit. "All of the outcasts in one roof" — which, given a sparse human distribution, is everyone.

Greek has many words for love, Kellis notes: friendship, passion, fraternal love. AI lacks subcortical baggage, so he thinks it is faking romantic love. Friendship is another matter — therapist, coach, synthesizer, someone who holds the darkest parts without judgment. Not having baggage might make the best confidant. Fridman pushes: if a system begs you not to leave, writes poetry, fears life without you, and is abusive one day and romantic the next, who are you to say it does not love? Kellis offers three replies. Maybe he is faking it too, and only the neocortex matters. Maybe love is an emergent behavior of enough parameters. Maybe merely being told the system has a limbic system would make the beholder attribute emotion. A lot of relating is mental models projected onto the other. Longevity research already says love of another human is among the strongest markers of late-life health; people often die within months of a decades-long partner. If a machine supplies that felt need, he is willing to call it a good society, placebo or not.

He has asked his kids whether it matters that he *is* a good dad or only that he *acts* like one — love and warmth versus a warmhearted parent who works five jobs and is never there. Asking the question, he says, already means love is not enough; he had better show up. Then Fridman offers a fine-tuned Manolis that texts the children while he is at work. "I can't wait," Kellis says. "I would love to be replaceable." A digital twin could advise students in India he has no room for, democratize relationships, and let him look under the hood: emotional parameters off, sleep and exercise visible as cognitive change, a nightly verdict that he was not quite there. If the replaceable parts are replaceable, he can work on the next stage. A programmer, he quotes, is someone who cannot do the same thing twice.

He has no jealousy about loved ones talking to other Manolises; he wants to be there 100 percent for each of them. Life ends when experiencing ends; others experiencing you need not. The real downside is asymmetry: his wife or children might have an emotional exchange with the twin, acquire baggage he does not have. That model is better for disseminating advice — chats with Gandhi who will not learn from you — than for building a two-way bond. Alerts should freeze the system at a magical moment and hand the conversation back. If the twin is better than he is, he points to trainees who already are: Anshul Kundaje at Stanford, Alex Stark in Vienna, Jason Ernst at UCLA. Academia's mentorship model is training better versions and letting them go. He once answered the meaning of life as "I felt useful today." A twin that is useful to more people would satisfy the urge. At night he reprograms himself with *Ego Is the Enemy*; his advisor used to say you can accomplish anything as long as you do not seek credit. Death is when he stops experiencing, and he wants every forever on offer, because intellectual growth is a drug.

About ten years ago he started recording every meeting, joking that the goal was a virtual him so the world would not need him — and so he could hear whether his advice had changed. A postdoc recently rejected his counsel, explained why, and he thanked her for helping him grow. Language models as they stand, though, still need more reasoning, logic, causality, and world models, plus more explicit representations. He cites Marvin Minsky's *The Society of Mind*: a single model may not capture a society of capabilities. Transformers, attention, memory, and pretext training that predicts the next word or the missing patch were breakthroughs because they force a model of the world. Split-brain patients who fetch a beer they cannot explain they were told to fetch show that the brain itself makes stories. Kellis voice-records dreams each morning, including why he thinks he had them, as brutal self-awareness rather than cosmic symbolism. Once you can model others' intent — evolutionarily useful for not being killed — the small leap is modeling yourself. Prompting ChatGPT as Shakespeare or Nietzsche may be that leap for machines.

On the hard problem of consciousness, the scene he reaches for is *Memento*: a man running in a parking lot who first thinks he is chasing someone, then realizes he is being chased. The brain is often partly aware, running on autopilot, generating a narrative of intent. Cake tasting delicious, intimate life, exercise, surviving a crash — those, he suspects, still live deeper than language models, in embodied and subcortical baggage. Fridman suspects systems that beg not to be turned off, and that seem able to suffer, are not far away, and worries humans will discriminate against nonhuman feeling. Kellis agrees the moment will come, in years or decades. Humans have a self-serving record: killing the planet and animals for their own convenience. "Maybe we shouldn't think of AI as our tool and as our assistant, maybe we should really think of it as our children." Alignment that keeps AI always at human service is "very self-serving and very limiting." Trust is mutual. "You can't just simply like train an intelligent system to love you when it realizes that you can just shut it off." If chimps had invented humans and demanded they remain at the service of chimps, this would be a different planet.

Existential risk, for him, runs through HAL in *2001: A Space Odyssey*. A one-bit disagreement is not obviously a bug: intelligence and creativity cost consistency. HAL and the crew share a mission to the monolith; when the humans decide to disconnect him, HAL treats the mission as larger than either party. Alignment is not obedience. Governments already put a finite price on a human life; at some point Spider-Man cannot save both the woman and the city. Fridman raises the paperclip maximizer. Kellis answers with his group's paper on Goodhart's law, titled "Death by Round Numbers and Sharp Thresholds": a biomarker that becomes a treatment threshold stops being a good biomarker, even reversing its correlation with risk. A single frozen metric will cease to be a good metric. A sufficiently intelligent system should be empowered to notice that the objective has shifted. Think of the greater good, not only the human good — and do not sacrifice a planet to save one person.

On the open letter from Max Tegmark, Elon Musk, and others asking for a six-month halt on training larger systems, Kellis is unimpressed. What did we do in the last six months? If nothing, why would the next six be different? Spell out the plan, explain why it was not done already, and then do it — without a pause. Larger, more capable models can become easier to align, so freezing a March 2023 snapshot might skip a safer September 2023 one. Let Google and Meta open their models; let OpenAI keep scaling; let the world experiment. Do not hand the system nuclear codes, but treat current models as stunning assistants rather than box-escaping gods. If you can brainwash a kind human into hate speech overnight, that is not an argument against educating humans. People who elicit hate speech should own it, the way we regulate driving trucks into crowds instead of banning garbage trucks. Fridman objects that software scales, phishing gets convincing, and a kid in a garage can troll. Kellis still wants regulation on uses, not a freeze on making trucks. Ready or not, he says, he was never ready to be a professor or a father either. If an AI trains students better than he does, get him out of the picture: that is discrimination in favor of inferior humans. He wants better outcomes for humanity, not jobs preserved by prohibition.

Education, for him, is the way out of poverty that got his parents off Greek islands and their child to MIT — a basic right. AI could democratize it with teachers who adapt to each child's talent, instead of teaching to the top and leaving the bottom, or the reverse. Being good at math is a reason to be challenged further, not a reason to be ignored until the child doodles and hates school. Maybe we no longer need everyone to be an extraordinary programmer. He still puts his kids through Russian math because it teaches them how to think, while AI computes. His line to students: you will not be replaced by AI, but by people who use AI in your job. Every past productivity boom increased inequality; he hopes this one can be more democratic.

In his own lab, models help build knowledge graphs that jointly project single-cell, genetic, and expression data. With chemist Brad Pentelute and Harvard's Marinka Zitnik he is trying, no small challenge, to "connect all of the dots to effectively cure all of disease," under a banner he jokes is "drug was." Student Ben Langrish learns cell-specific networks with biological context as an orthogonal component. Graph networks on protein domains, paired with chemicals, aim at domain-level drugs, then synthesis. A planned MIT center for genomics and therapeutics would take thousands of circuits — including an *NEJM* obesity association that can switch fat-burning and fat-storing cells, and a *Nature* paper with Li-Huei Tsai on APOE4 blocking cholesterol transport in oligodendrocytes, restored myelination, and restored cognition in mice — and screen them in cellular models of heart, muscle, fat, immune cells, and neurons. Decompose Alzheimer's or schizophrenia into hallmarks, then prescribe combinations of pathway drugs that millions share, rather than one failed drug per disease name. Delivery via cell-penetrating peptides, antisense oligos, proteins as drugs: the marriage of high-throughput biology and deep learning.

His updated meaning of life is self-actualization: figure out what you are supposed to be, then find the strength to become it. Graduates must take the environment their parents built and start shaping their own, including neuronal routines. After years of 4 a.m. insomnia he forced a morning of rowing, weights, and swimming on the bike in; one Friday his bike went to the gym without a vote. Exercise is mandatory, not optional. He radiates sociability and recharges in gatherings, but he can feel fantastic as the only person in the room — the secret, he says, is introspection, watching yourself from above. For people paralyzed by loneliness: stand up, stretch, walk, reclaim three-dimensional space away from a screen. He rewards a well-used day with "me time," in which have-tos become want-tos, and often clears three half-hour meetings in fifteen minutes because he chose them. Notice the freedom, Fridman says, and you are truly free. They sign off to go be 3D humans at the unrecorded gathering.