We Are the Descendants of the First Digital Computer

Lex Fridman introduces Manolis Kellis as a professor at MIT and head of the computational biology group, then asks for the most beautiful aspect of the human genome. Kellis will not stay inside the human species. Genomes in general are beautiful, he says, and the sentence he uses on the first day of class is that we are not the inventors of the first digital computer. We are its descendants. Life is digital: at every replication step you do not lose information, because the information is digital. If inheritance were analog, a matter of concentrations, it would dissolve after a few generations. That is what the ancients did not understand.

The first person to see discrete inheritance, he says, was Mendel, whose theory sat on a bookshelf for about 50 years while Darwin became famous for natural selection. The missing piece was the mechanism. People treated Mendel's peas as a fluke that would not apply to humans, because what they saw was a continuum of eye color, skin, hair, and height. It took Ronald Fisher in the twentieth century to show that even five Mendelian loci add up to a continuous-looking trait. There are not five genes for height, Kellis says. There are a thousand, and each one contributes less than a millimeter. You change more in height during the day than any one of those variants contributes. By evening you are shorter than you were when you woke up.

Why, Fridman asks, are we not more different, given so much combinatorial room? Selection pushes toward the middle. Extremely tall or extremely short runs into breathing, running, sitting. Natural history shows selection for height in northern Europe and against it in southern Europe. Balancing selection, mate choice, and a simple rule of thumb, average the two parents and adjust for sex, keep the distribution from flying apart. The double helix, which Kellis calls the most noble molecule of our time, holds the secret of that discreteness. The conceptualization of discrete elements, he adds, precedes the structure of DNA.

Asked what is beautiful about the human genome specifically, Kellis says that if more people understood it there would be fewer wars and less anger. Any two people on the planet are 99.9 percent identical. "How can you fight with someone who is 99.9 percent identical to you? It's just counterintuitive." And yet any two siblings of the same parents differ in millions of locations, so each child is, in his phrasing, two-to-the-million unique from a given pair of parents, let alone from random parents. Every person is as unique as a star, and in some ways more so, because stars have only a few parameters while humans have thousands scattered across the genome.

That pair of numbers is the political claim of the episode as much as a genetic one. There is no stretch of sequence that encodes race or friend-versus-enemy. The 0.1 percent is mostly drift, frequency shifts from geographic isolation, and tens of thousands of years of local selection. Kellis puts individuality and similarity on the table at once: you and a stranger are nearly the same molecule, and you and your brother are still not interchangeable.

Most species, Kellis says, inherit vertically: instinct is a huge part of behavior. He and his children have been watching a nest of two eggs outside the window. A lot of bird behavior is hard-coded. A bird born in Boston will be the same bird as one born in California. Take a person from today and place them in ancient Egypt, or the reverse, and they grow up completely normal. That is not genetics. It is the other human inheritance: ideas built up each generation and transmitted sideways.

The word he wants is neoteny, from neo for newborn and teneo for holding: holding a newborn state. The birds were eggs two weeks ago; one has already flown, the other is ready. Humans need six years, 18 years, 24, and he is still learning. Compared with other primates we lose badly at seven weeks. At 18, all bets are off. The wiring of the brain is delayed on purpose. The longer you delay it, the more time parents, kin, and teachers have to pass on knowledge, concepts, and ideals while the connections are still plastic.

Lateral transfer today is not only uncles, aunts, and schoolteachers. It is Wikipedia, review articles, journals that put information out for free, podcasts, and video games. You can learn pretty much everything in a super-advanced textbook in a matter of days, instead of sailing to the Library of Alexandria to read three books and then sailing on to Athens. The speed of that spread, Kellis says, defines the human inheritance pattern. Fridman asks whether he is also afraid, given that younger people may be watching five-second clips rather than papers.

Kellis is excited that ideas can reach anyone and that it is the idea, not who you are, that matters. The danger is the attitude that "my ignorance is as important as your expertise": you spent 10,000 hours on human brain circuitry, why should I trust you, I will make up my own theories. Education has to teach not only knowledge but the means to get to knowledge, the old fish-versus-fishing point. Google it, check what is real, check sources, form a basic opinion. The foundation should include epistemology, not just facts.

Fridman notes that in the coronavirus year, Twitter, a homemade network of experts and kids who plot data, often beat centralized institutions. Kellis treats the pandemic as a case of horizontal transfer at full speed. The first sample was obtained on December 29, 2019; the genome was sequenced in early January; a week after the sequence was public, Moderna had finalized its vaccine design and was moving to production. From not knowing what was killing people in Wuhan to naming SARS-CoV-2, its genes, sequence, and polymorphisms took weeks.

There were plenty of mistakes. Some may have been political, some innocuous, some misleading "for the greater good," such as telling people not to wear masks so supplies would not run out. That last one, Kellis says, was very silly and a very big mistake. Bogus articles snuck onto the front page and were debunked within 24 hours. The beauty of science now is that knowledge is not fixed. Everything is the current best hypothesis and the current best model that fits the current data. Success is how long you were not proven wrong, not a claim that you were exactly right with partial information. The arc toward perfection matters more than how far the first step went.

Where would he send a child to learn about coronavirus? First a textbook: a virology chapter on how coronaviruses work, then basic epidemiology, first and second waves, R0 and Rt, how it spreads, how it enters cells, which cell types express the ACE2 receptor, how the immune system helps or hurts, cytokine storm, why obesity and age are risk factors, why children often do not even know they are spreading it. Then, a little sheepishly, Wikipedia is pretty awesome.

People used to mock Wikipedia as unreliable. Kellis inverts the complaint the way he inverts complaints about Nature. So many Nature papers are wrong, critics say; his answer is that so many Nature papers are scrutinized, and a higher rate of being proven wrong is evidence they are better papers, because they are looked at harder than other journals. Judge Wikipedia by the number of revisions, not the first draft. It is still superficial, scratching the surface; for depth you go to review articles. For most scientific topics it is quite authoritative because it is the place everybody likes to criticize. If your field's page is shallow, he says, you should feel partly responsible: if more people from the community edited it, it would not be. Some fields embraced Wikipedia; others relegated it. You need a few pioneers who say enough, we are going to fix that.

The last uniqueness Kellis wants to name is a reversal of the old lab order. You used to be told not to do that experiment in human: learn it in fly, yeast, mouse, or primate first. Human genetics flipped that. We are now the most studied species on the planet, which he finds a little embarrassing. The old sequence was viruses, bacteria, yeast, fruit fly, worm, mouse, and only then human. In a population of seven billion there is a mutation at almost every nucleotide. If you want to know what a base does, you can search a database for a living person who already carries it. Directed knockouts of three genes at once still belong to model organisms; there are not enough humans, and we have not sequenced all seven billion. But the trend is that it is often simpler to mine human data than to go back to flies.

Human genetics has found thousands of genes involved in cognition, psychology, emotions, and feelings once assumed to be purely learned. Free will, in this telling, is a beautiful concept that keeps getting sliced thinner by chemical reactions, receptor abundances from yesterday's meal, wiring from parents and upbringing. Fridman asks how much freedom we have to escape the genome. Kellis describes what freedom would look like: I resist the apple to prove I choose. But receptors may have produced the urge to prove individuality by refusing the apple. Then I eat chocolate instead, and those receptors too may be in play. You keep kicking the bucket down the road.

The analogy is weather. You can predict ten days well. Is that because weather is free after that, or because the system is deterministic and you cannot model Jupiter's pull, solar input, a bit more carbon from China today affecting Egypt in three weeks? You might need to build a universe to simulate a person. Asked about the simulation hypothesis, Kellis does not hedge: complete nonsense, no empirical evidence. As a thought experiment it is cute, but Occam's razor asks what is more realistic, that you are a machine in someone else's sim or a person whose experiences sit in chemical molecules. He challenges the machines, if they exist, to wake him and terminate him now.

Kellis uses computer science to understand biology. Methods in AI, machine learning, statistics, and algorithms matter to him when they advance that understanding, not as a separate dream of building a superhuman artificial life. Understanding the brain is still coupled to better AI, because so much of AI was inspired by the brain, even if it took decades from early neural nets to deep networks that play Go and chess and synthesize images. That raises his favorite question: why do neural networks perform so well? Not because they can simulate any function. They simulate a very small subset. That subset is closer to the functions of the universe we inhabit.

Brains did not evolve in a vacuum. They evolved in a world with physical constraints. Vision is a slice of the electromagnetic spectrum, hearing is movement in air, and so on. Animal brains, and the AI systems built by studying them, work well on images of that world and fail on noise that matches nothing in it. They also make human-like mistakes: engineered pixel dithers that turn a zebra into a baboon, and undoctored photos that confuse muffins and dogs. The more we learn about the tricks the brain evolved for this world, he argues, the more computational primitives we can bring into AI, including for data we have not been exposed to but that still live in the same universe.

Even our scientific senses cover only a sliver. LIGO, in the news for a year and a half, created a new sense. Gravitational waves have been traversing Earth since its formation. Life evolved photoreceptors, hearing, touch. No organism evolved a way to sense neutrinos or gravitational waves. For the first time in the history of life on the planet, we can capture signals we were previously blind to. Picture 3020: what new sensors might we discover, for physics that is permeating us all the time? Premonitions he mostly files under ascertainment bias, remembering the friend who then calls and forgetting the times they did not. He still wonders whether the brain has unrecognized sensors, including in the emotional transfer of hugging a child, and he is willing to sound romantic about it.

Pressed for a line of work he is proud of, Kellis refuses to rank his children and instead describes a path that started with his first paper: the first comparative analysis of multiple complete genomes. The idea was genome-wide evolutionary signatures. Align human, mouse, rat, and dog, all mammals doing similar work with heart, brain, and lungs. Functional elements shared among mammals are conserved. Only about 1 percent of the human genome codes for protein; the other 99 percent was largely unknown until comparative genomics. He and colleagues applied the method to yeast, flies, fungi, a dozen Drosophila species, 29 mammals, then 200 mammals.

After the meteor that ended the dinosaurs, niches opened and mammals radiated into air and ocean, still remaining mammals. Align those genomes at nucleotide resolution and protein-coding regions show a pattern set by the level at which selection acts. Change the third position of a codon and you often keep the same amino acid; mutations invariant to the final function are tolerated. Evolution, in his two-part model, is random, blind, stupid mutation plus supersmart, ruthless selection. Evolutionary signatures look at the shape of the distribution of sequences that encode the same thing. Scan a region, ignore what it "says," and you can still say this is evolving like a protein, that like an RNA structure, that like a regulatory motif.

They put the same method on SARS-CoV-2 against 44 SARS-related beta coronavirus strains, most infecting bats, only a few ever infecting humans. The last little open reading frame is often not a protein at all but an RNA structure that is not translated. Inside another gene sits a tiny additional gene: DNA can be read in three frames, so there is a protein within a protein that had not been counted. "We don't even know the building blocks of SARS-CoV-2." The comparison also sets the gas pedal on each part. Polymerase and nucleocapsid evolve slowly. Spike's S1, the piece that binds the ACE2 receptor and gives the corona its look, evolves ridiculously fast. S2 is the latch that fuses membranes and gets viral RNA in. Anything that contacts a changing environment evolves faster. There are thousands of bat species, each a different lock.

Fridman is in awe of viruses jumping from bats to humans in a distributed search. Intelligence, Kellis says, is in the eye of the beholder; stupid is as stupid does. He does not think viruses are intelligent. Mutation is blind; selection is the smart-looking half, implemented as huge parallel search across infections. Push back that viruses would then be "smarter" because there are so many more of them, and he will grant we would not be here without them. Early in the mammalian radiation, viruses spread through the genome and created binding sites that now help control mammalian circuitry. Coevolution is real. Caring is not. A virus is not trying to kill you. If you die, that lineage does not spread. Over time, versions that transmit without immediately killing the host are the ones that remain.

The design of the current coronavirus, however lethal, he finds beautiful. Human cells are hardwired to make one protein from each RNA. The virus throws in a single RNA. A giant open reading frame, about two thirds of the molecule, is translated as one huge protein because the cell sees a start codon and assumes the RNA is its own. As that protein grows it is cleaved into on the order of 20 peptides. Early pieces shut off the host ribosome and mark human mRNAs as junk, sparing viral RNAs that carry a particular mark. Between proteins 11 and 12 the ribosome slips reading frame, so about half of the products stop at 11 and the other half continue through 16. Downstream genes never come from that long polyprotein. Between open reading frames a small signal loops the polymerase back to the start, making partial negative strands and then subgenomic positive RNAs that look like ordinary human mRNAs. The cell translates spike, envelope, membrane, nucleocapsid; particles assemble and leave. Human cells do not do this. Only viruses do.

He is not concerned that this is the end of the species. Humans have been nearly wiped out and never fully wiped out. On lab origin, he is technical. If you wanted a virus that kills a lot of people, you would start with SARS or a close relative. The nearest relative of this genome sits on another branch, none of whose members have infected humans. No one in their right mind would start there. Claims that S1 was mixed in a lab because its three parts have different trees miss that recombination of this kind has been happening across the coronavirus family for a very long time. Two viruses in one cell can mix positive and negative strands and produce a hybrid with S1 from one and the rest from another. A virus is not even classified as alive: it does not self-replicate. You do not kill it so much as stop it from replicating. It is not a bacterium living in a puddle, and unlike HIV it does not integrate and go dormant. Unexpressed coronavirus RNA degrades.

The Black Death killed about 30 percent of Europe and left a hole in the immune repertoire. Malaria reshaped Africa and, through that, frequencies in people of African descent: sickle cell maps onto malaria maps because the parasite replicates less well in those cells. Disease associations still show those waves of killing. Outward-facing proteins evolve faster because the environment changes; the genome has also co-opted immune genes for other jobs. Microglia in the brain prune unused synapses, use-it-or-lose-it as an immune mechanism. M1 versus M2 macrophages in fat help determine obesity. Selection is brutal and is not "better." After famine, the survivors are whoever survived, maybe the ones who ran slower. An obesity-associated allele Kellis describes sat around 2 percent frequency in Africa and rose to about 44 percent in Europe through ice-age calorie storage. He inherited a risk copy from his mother and from his father. A hundred years ago it was a great allele. Now it is a bad one, with the same molecular job.

Diversity is the species-level insurance policy. If a hundred million nuclear physicists lived on Earth and no one else, we would be in trouble. You need artists, musicians, mathematicians, and, yes, politicians. Immunologically the same point holds: VDJ recombination, blood type, MHC, layers of repertoire. That is why he does not worry that a natural pandemic will end humanity. Engineered ones worry him less than popular stories suggest, because he thinks human designers are dumber than evolution. Childhood vaccination and dirt exposure still matter: northern Europe's clean environments and allergies versus kids who grew up eating dirt. The more you protect the immune system from seeing non-self, the less it is prepared for the next insult.

A cartoon he likes has two windows: a long line under "pill," nobody under "exercise and diet." That empty window, he says, improves Alzheimer's, cancer, coronavirus, every aspect of health. His group is comparing diet and exercise interventions in humans and mice with single-cell profiling, watching stromal and immune cells and the signals muscle sends through blood and lymph. Fat cells are not the enemy; without them, lipids and calories would float in the blood and kill you. A metabolic switch in the first three days of adipocyte differentiation points a cell toward burning or storing; burning fat cells are closer to muscle. The most personalized nutrition advice, surprisingly, may come from the gut microbiome rather than the genome. You have more known non-human cells than human cells, "a giant bag of bacteria with a few human cells." Effect sizes are not enormous and the science is not mature. Placebo still stuns him: a sugar pill described as an amazing cancer drug raises the chance the cancer stops. That means if you can trick the brain into thinking it is healing, it finds a way. Stress we understand better, heart rate, toxins, damage, so mind-healing-the-body may partly be the mind killing the body less. Either way, the pathway is scientifically almost untapped.

Around the digital core, Kellis says, much of biology is analog: brain, RNA, the cell. The code itself is digital. Think of genes as functions. Old languages called them by going to an address in memory; modern languages wrap that in objects, but underneath there is still an instruction that jumps. Genomes have no `goto`. You do not start transcribing at position 13,527 of chromosome 12. You have content-based indexing. Drink coffee, and cells that metabolize caffeine see the signal and turn on genes marked with a coffee motif. Those regulatory motifs are about eight characters, sequences such as GCAAAC, working in combination, each recruiting a protein. Collections of them make promoters near the start and enhancers that loop DNA to bring polymerase. Then export, then translation in the cytoplasm.

Fault tolerance follows. Flip 20 percent of the bits on a hard drive and you are in trouble. Flip 20 percent of the letters in a genome and you might not notice. Resilience is the first design principle; everything else is later. Engineering teaches compartmentalization, separated functions, modularity, hierarchy. Biology does the opposite and looks dirty on the way to elegance. Compact viral genomes have often been stripped down from larger ancestors: you add stuff, raise complexity, then slim it. Baker's yeast, *Saccharomyces cerevisiae*, the sugar fungus of beer, is the descendant of a whole-genome duplication about 100 million years ago, coinciding with fruit-bearing plants. Animals ate fruit, walked, and deposited seeds plus nutrients, creating a sugar niche that could absorb a messy doubling, eight chromosomes to sixteen, probably via nondisjunction. Then 90 percent of the duplicated genes were lost, because a compact genome plus a random hit plus ruthless selection kills the extra copy. The 10 percent kept in two copies involved environmental adaptation, replication speed, translation speed, and sugar processing. Dual function on one gene is a load; duplication lets one copy keep the ancestral job and the other specialize. New function comes from the same process that breaks things.

In the current pandemic, genes that evolved fast in the past still evolve fast. The mutation that made the news is D614G in spike: a site perfectly conserved across millions of years of these viruses, never changed when they moved bat to bat, now gone from about 1 percent to 90 percent in almost all outbreaks. The old letter was suited to bats; the new one makes human-to-human transmission easier by orders of magnitude. For vaccines, that map says which regions are stable and which newly evolved functions matter because they escaped a bat local optimum. If engineers had designed evolution, we would still be perfectly replicating bacteria, because only by making the bacterium worse do you reach a new optimum. "We as engineers need to embrace robustness as the first principle beyond perfection, because nothing is going to ever be perfect." A Mars probe that dies at the first fault is the wrong lesson; systems that tolerate failure, even get better through it, are the biological one. Deep learning already steps that way, not because it copied the genome but because it copied the brain, while the same genome still encodes every cell type, organs, blood, and a brain that takes about 20 percent of our energy.

Two apes talking on a podcast, Kellis says, are compressing huge neural activity through a funnel of words that the listener expands again. Engineers would love to dump one brain into the other. That might not be better. "In your misinterpretation of every word that I'm saying, you are creating new interpretation that might actually be way better than what I meant." Ambiguity of language may be the secret of creativity. In lab meetings he restates a student's point until by the third time it is not what they said, and they learn something else. That messiness is a feature. On brain-computer interfaces, he thinks output will come first: humans will learn to produce patterns a machine can read, the way Graffiti on the Palm Pilot trained people to write machine-friendly glyphs, or the way a child learns a third limb if the mapping is consistent. Input into thought is another matter. Engrams are combinations of neurons; present resolution is nowhere near reading "7.5 million" into a head. Language will remain the inbound channel for a long time. Translation of Dostoevsky, Fridman says, is deeper than most NLP papers. Kellis cannot hear a new word without etymology: *parallel* is para, side by side, plus allelon. Meaning lives on the receiving end. A French audience and a Russian one need different sentence structures to evoke the same emotional set. Nothing takes on meaning in the act of being created alone.

For his 42nd birthday, the Hitchhiker's Guide number, he ran a meaning-of-life symposium, a drinking-together in the original Greek sense, as he had once celebrated turning 32 as binary 100000. Friends and collaborators each spoke. Everyone began with I do not know, then gave a different answer, mutually consistent. A Greek linguistics professor at Harvard offered a Delphic line with three readings included: become one. Become one year old, with a child's excitement of discovery. In whatever you take on, become the one who is first, who excels. When people are separate, become one, come together, learn to understand each other. The symposium itself, Kellis says, illustrated that the quest for meaning may be the meaning: a constant search for something sublime and human. He leaves open whether life is inevitable, a layer of software on physics, a fight against entropy by compartmentalizing and concentrating rather than diluting.

During the pandemic he worked from 6 a.m. to 7 p.m. every day, including weekends, and told his wife, after a day that made his brain hurt, that he had been useful. He wrote a bedtime prayer for his children: thank you for all you have given me, and give me the strength to give unto others with the same love. We are the only species that worries about the meaning of life, and maybe that is what makes us human. People who say they will not have children, he thinks, are missing half the picture, if not the whole: unfiltered criticism, a heart he did not know he had, hardware rearranging while software is installed, a frontal cortex still growing for another decade. Creating life is easy at conception. Creating human life takes decades of compassion, sharing, love, anger, patience, and impatience. Parenting changed him as a teacher. If a tree falls in the forest and no one hears, has it fallen? If you do awesome research and write an impenetrable paper, it is as if you never did it. Encapsulating complexity in bite-sized chunks is part of the job. That is horizontal inheritance again: the genome's 0.1 percent is not the whole human trick. Twenty years of plasticity exist so the last generation can be received, and a little added.