A billion years of evolution in a single afternoon — George Church

The single greatest increase in one person's destructive capability occurred in the span from ancient humans using bare hands (limited destruction) to the modern era where one person with right connections/technology can blow up a city, representing a 'huge increase in capability' that suggests societal defenses against misuse of biotechnology should be strengthened

“Over the history of humanity, the amount of things that a single person can do has grown very significantly. It used to be, when you had your bare hands, there was kind of a limit to what one person could do. A large number of people could team up and get a mammoth or something like that. Today, one person with the right connections or right access to technology could blow up a city. That's a huge increase in capability.”

Only 3% of children are severely affected by genetic diseases, but people tend to dismiss this risk as low ('I'm in the 97%') when deciding on genetic counseling, despite 97% being acceptable odds for casino gambling; this reflects difficulty with rare event reasoning and the 'trolley problem' framing where not intervening feels blameless even though non-intervention is still a choice

“Only 3% of children are severely affected by genetic diseases and they feel like, "I'm not that unlucky. I'm in the 97%." If those were your odds of winning at the horse races or at the casino, you'd take them. 97% of winning, good.”

Even if a million copies of George Church existed as superintelligences in data centers thinking all the time, they couldn't run biological experiments directly so would only achieve speedup through suggestions/analysis, and the bottleneck of needing real physical experimentation means biological progress can't scale to unlimited parallel thinking

“They can't run experiments directly. They're just in data centers. They can just say stuff and think stuff.”

Library-based screening using real biological experiments avoids the assumptions inherent in computational simulation (quantum electrodynamics → quantum mechanics → molecular mechanics) and thus achieves '100% precision' because it measures actual molecules, not models, making the data harvestable back into conventional AI for further optimization cycles

“It's a hundred percent precision, because you're not simulating. You're not making assumptions. You're not going from quantum electrodynamics, which is an assumption, to quantum mechanics, which is an assumption to molecular mechanics, which are full of assumptions. You're really doing the real thing. So you're doing a kind of natural computing.”

Evolution favors solutions that work incrementally; investment in new base pairs was expensive with unclear payoff, so evolution didn't do it despite multiple technological generations having demonstrated viability (e.g., Romesberg's synthetic base pairs with full replication/transcription/translation)

“I think that evolution has a tendency to go with what works. The investment in making a whole new base pair would have been high. We haven't even articulated what the return on investment would be.”

Laboratory reconstruction of simple abiogenesis pathways from inorganic precursors (cyanide derivatives, reduced compounds → cellular replication) would suggest life can originate readily; however, proving the negative (that life is rare/unique) is harder because Earth encountered ~10^20 liters of water at various conditions favoring emergence

“if you showed, reconstructed in the lab, a very simple pathway from inorganics, cyanide derivatives and reduced compounds, all the way up to some cellular replicating structure, that might lead us to believe that at least life exists.”

Church's lab has produced many spinoff companies and leaders because Boston's geographic/institutional clustering (MIT, Harvard, BU, biotech, pharma) attracts best talent, enables spousal employment, and creates positive feedback loops; clustering advantages cannot be replicated by sudden creation in desert

“Boston is a unique culture. It attracts some of the best and brightest students and postdocs automatically. It is dense enough. Spouses can find other jobs in the same field. Having a concentration of biotech and pharma and MIT and Harvard and BU and so forth, all in one pretty walkable distance, not spread out all along the East or West coast.”

Exponential progress curves (hockey stick) appear flat at first and only become visible once steep; biology/AI/computing are all at inflection points currently, so labs already in virtuous cycles asymmetrically benefit from the exponential turn

“The exponential is pretty much the same in the beginning of the hockey stick and at the end of the hockey stick, but you don't notice it until it gets going. That's what's happening in computing, AI, and biotech. They're all peaking at this point.”

Current gene delivery mechanisms cannot reach all cells in the body with gene therapy, but one company (Dyno Therapeutics) achieved a hundredfold improvement in targeting neurons in the brain by testing millions of different capsid variants with AI, demonstrating that no law of physics prevents whole-body delivery—only practical engineering challenges remain

“There is nothing close to that today. But there's nothing, no law of physics, that would prevent it. Again, there's going to be practical considerations, like how many injections do you need to do to achieve that goal? But we're getting better at targeting tissues. One of my companies, Dyno Therapeutics, showed they could get a hundredfold improvement in targeting neurons in the brain, which is a big deal.”

The difference between an African elephant and a woolly mammoth is approximately one million base pair differences, and de-extinction projects like the dire wolf should be understood not as perfect species recovery but as synthetic biology demonstrations that illustrate the minimal genetic changes required for major phenotypic differences

“The difference between an elephant and a woolly mammoth might be like a million base pairs. How do we think about the kind of thing we're actually bringing back?”

Materials science breakthroughs should accelerate faster than drug approvals because materials don't require regulatory approval in the same way, so the next major breakthrough in biology-derived materials could happen without the 10+ year approval cycles that constrain pharmaceutical development

“Materials actually should go faster though, because they don't require quite as much regulatory approval.”

Transcription factors offer a 'recipe' approach to cell reprogramming where the minimum number of transcription factors required to convert a stem cell to a specific neuron type can be identified by examining which transcription factors that target cell type expresses, then testing whether expressing only those factors converts the stem cell

“What's the minimum number of transcription factors it takes to turn a stem cell into a neuron? There's a bunch of recipes where you can do it with one. Maybe you want a specific neuron, you might need a few more. But then you can kind of quickly go to the answer by looking at each target cell type that exists. You can see what transcription factors does it express at the time that it's the target? Then you say, "Let's just try those on the stem cell and see if they work." That recipe has worked quite well.”

Evidence for life arising independently on other worlds would require either: (1) detection of radio/light signals, (2) laboratory reconstruction of a simple abiogenesis pathway from inorganic compounds to cellular replication, or (3) empirical discovery of multiple independent abiogenesis pathways—but it's harder to prove the negative (that life is rare) because prebiotic conditions are vastly diverse

“With biology, the kind of evidence would be that you show in a laboratory using prebiotic conditions a really simple way to get life. It's harder to prove the negative because we don't know all the possible prebiotic conditions.”

The competition-driven erosion of safety rules in AI development is inevitable, meaning that even if safety rules exist, intense competition undermines them 'and pushed aside' making reliance on rules alone insufficient without deeper shifts in how we approach AGI/ASI development

“What typically happens when there's an intense competition is those safety rules get undermined and pushed aside.”

AGI/ASI represents a 'completely artificial emergency' rather than a natural crisis like COVID-19 because any AGI crisis would be created by our decision to build it, not something we're trying to solve, so there is 'no rush' and development should be slowed deliberately rather than accelerated under false urgency

“This is a completely artificial emergency. This is not like COVID-19, where millions of people were dying if we delayed the science. This is something where, if there ever is a crisis, it's because we created it, it's not because we're trying to solve it. So I think we need to go very slowly on AGI and ASI, and double down on slightly narrower scientific goals.”

We barely know how to pass ethical understanding 'to the next generation of humans' which suggests we're even less prepared to educate 'a completely foreign type of intelligence' like AGI, making it crucial that we slow down AGI/ASI development to allow time to figure out alignment before capabilities exceed human control

“I don't think we understand our own ethics well enough to educate a completely foreign type of intelligence. We barely know how to pass it onto the next generation of humans.”

Genetic counseling is underhyped and underutilized despite being clearly competitive with gene therapy for preventative purposes, with Dor Yeshorim providing proof that counseling can eliminate serious inherited diseases across communities, and the confusion about eugenics stems from forced vs. voluntary choice—eugenics was wrong because it removed choice, not because it selected against disease

“What I would say is genetic counseling is underhyped. It is clearly competitive with gene therapy in a certain sense, clearly not for people that are already born but for people in the future, not even distant future but in the next couple of years.”

Transcription factors can reprogram cell types with minimal changes—a single transcription factor can sometimes convert a stem cell to a neuron, though specific neuron subtypes may require additional factors; this provides a recipe-based approach to cell engineering across all cell types

“There's a bunch of recipes where you can do it with one. Maybe you want a specific neuron, you might need a few more. But then you can kind of quickly go to the answer by looking at each target cell type that exists.”

The most practical protein design workflow is to create libraries inspired by working examples, generate variations, test in real experiments (not simulation), eliminate neutral and lethal mutations, keep game-changing variants, and iterate; this directed evolution approach is faster than natural evolution (millions of base pair changes in a million years vs billions of changes in an afternoon) and more efficient because it avoids wasteful neutral and lethal mutations.

“The way that it's working now—which will get us a long way, won't get us the whole way—is we have something that kind of works and we make libraries inspired by that... we can just keep going. It's kind of like the way evolution worked, except now we can do it at incredibly high speeds... In principle, evolution might incorporate a few base pair changes in a million years. Now we can make billions of changes in an afternoon.”

AlphaFold can predict protein structures with extraordinary precision (fractions of an Angstrom on average) but this structural accuracy does not guarantee functionality—for example, substituting alanine for serine in a serine protease results in correct fold but non-functional protein, requiring additional AI tools that incorporate evolutionary and experimental knowledge to predict functionality

“With AlphaFold—last time I checked anyway—if you substitute an alanine for a serine in a serine protease, it will have exactly the right fold. It will be precise to a fraction of Angstrom overall average. But it won't function. It just won't function.”

The current approach to protein and material design uses iterative library screening where an initial working molecule creates variations, whichever variants work spawn further variations, eliminating neutral and lethal mutations in favor of game-changing changes—this evolution-inspired method works at 'incredibly high speeds' (billions of changes in an afternoon) compared to natural evolution (few changes in a million years)

“The way that it's working now—which will get us a long way, won't get us the whole way—is we have something that kind of works and we make libraries inspired by that. We make variations on it and then whichever of those variations work, we make variations on that. We can just keep going. It's kind of like the way evolution worked, except now we can do it at incredibly high speeds. In principle, evolution might incorporate a few base pair changes in a million years. Now we can make billions of changes in an afternoon.”

Humans could theoretically develop additional capabilities beyond current range (e.g., super-health, 150+ year lifespans, Einstein-level intelligence) if technical barriers were overcome, but the larger opportunity is 'actualizing' current human potential—bringing all 8 billion people up to demonstrated human capability ranges rather than pushing new boundaries

“To some extent, actualizing the people that we currently have would be quite impactful, just getting them all up to whatever speed they want to be up to within the range that's been demonstrated. Some people are going to want to be like Einstein, some people won't.”

The distinction between eugenics and genetic counseling is ethical, not technical: eugenics is forced (government removes choice from people), while counseling enables informed parental choice; psychological and cultural resistance to genetic counseling stems from difficulty thinking probabilistically about rare risks, not principled opposition to genetic medicine.

“The problem with eugenics was that it was forced. The government forced it on people. It wasn't that it enabled people to make a choice. It's that it removed the choice from the people.”

The germline editing moratorium (He Jiankui case) worked for 5 years with only one defector, but the failure to stop it via whistleblower intervention and prevention cost the defector 3 years in prison; the three resulting children are reportedly healthy, representing both a system failure and evidence that the intervention was technically feasible

“It worked for five years with only one defector. That's quite impressive. But all it takes is one. It would have been nice if the whistleblowers could have saved him the three years in prison by getting an intervention. It's not like anybody died. Right. There are probably three healthy genetically-engineered children in the world now. They'll be teenagers soon.”

Reductionism is not inherently problematic in biology—it helps translate findings into medical products and modules that can be reused across species, building a 'tool chest' of biological widgets

“That's a perfect example of how much we can minimize something, sometimes called reductionism. Reductionism isn't all bad. Sometimes it helps us bring a product into medicine. Sometimes it helps us understand or build a tool chest or a module that we could use in other cases and translate it to other species.”

A large fraction of genetic diseases cause developmental delay, often lethal or causing lifetime cognitive deficits; genetic counseling and gene therapy can address these through therapeutic and preventive approaches

“There are a huge fraction of genetic diseases that have as one of their consequences the child being developmentally delayed to such an extent that it's lethal or causes a lifetime deficit. We know the genes involved and we know how to do genetic counseling in some cases, gene therapy and other therapies to deal with it.”

The most achievable near-term goal is eliminating genetic diseases and enabling people to manage their own diseases as they see fit; this is more tractable than enhancing healthy people and more impactful than trying to design superintelligence.

“I think right now, the baby step, or actually the pretty big baby step, is to eliminate diseases or at least make it possible for people to eliminate their own diseases as they see fit.”

No existing gene delivery mechanism today can deliver therapy to every single cell in the body, but there is no law of physics preventing this, and practical improvements in tissue targeting are steadily advancing (e.g., Dyno Therapeutics achieved hundredfold improvement in neuronal targeting)

“There is nothing close to that today. But there's nothing, no law of physics, that would prevent it.”

Genetic counseling is more cost-effective than gene therapy for rare recessive diseases because genomic analysis costs ~$100 per person (soon less) while the lifetime cost of genetic disease (lost workforce participation, caregiving, medical care) is millions, yielding tenfold+ return on investment and making it a public health no-brainer

“You spend a hundred dollars per genome. It'll probably be less soon. You get the whole thing analyzed. Compare that to millions of dollars that will be lost in opportunity costs and them not being part of the workforce, taking care of them and so forth. So the return on investment is tremendous. It's at least a tenfold return on investment.”

Somatic gene therapy rather than germ line therapy is more likely to be the primary path to extreme longevity because 8 billion people have already missed the germline window, and the practical and ethical challenges are substantially lower for somatic approaches that can work on adult tissues.

“Probably there's a lot of forces pushing it towards somatic. For one, there's 8 billion people that have missed the germline opportunity. That's to say, it doesn't apply to us, the two of us and everybody listening to this.”

The COVID vaccine represents a proof-of-concept for rapid deployment of gene therapy at scale: formulated as gene therapy at $20 per dose with 6 billion doses deployed and proven across the entire population, demonstrating that gene therapy can be cost-effective for common diseases.

“In fact the COVID vaccine was formulated as a gene therapy and the cost was in the $20 per dose range. 6 billion people benefited from it, or 6 billion people took it and it was proven over the whole population.”

DNA synthesis cost has dropped a thousandfold over recent decades; when Church's lab published chip-based DNA synthesis in a 2004 Nature paper, it was ignored for a decade despite being 1000x cheaper than prior methods, but now claims exist of 10^17 genes makeable in libraries, representing a much larger efficiency gain if practical

“When we came out with the first chip-based genes in a 2004 Nature paper, basically people dismissed it for about a decade. The only people that used it were collaborators and alumni. It wasn't even listed on the Moore's law curve for DNA synthesis, even though it was thousand times cheaper.”

The brain has 10^11 neurons and 10^14 synapses, making it unclear whether replicating a specific brain in silico or as a biological copy would be easier, but copying is probably harder than synthesis because replicating all synaptic connections requires far more information (10^14 bytes) than the genome (billions of bytes)

“You have 10^11 neurons, 10^14 synapses. If you wanted to reproduce a particular brain, it's speculative as to whether it would be easier to do that by making a copy of it in silico, in some kind of inorganic matrix, or making a copy of it.”

Boston's unique combination of dense clustering (MIT, Harvard, BU, biotech/pharma) plus historical labs building positive feedback loops between basic science and societal application allowed Church's lab to asymmetrically benefit from exponential growth in AI/computing/biotech, making it difficult to replicate this success by starting a new facility in unconnected locations

“Boston is a unique culture. It attracts some of the best and brightest students and postdocs automatically. It is dense enough. Sometimes people want to spread the wealth out evenly all over the universe or the planet. There are advantages to having it clustered.”

Biology is more forgiving than engineered systems because organisms can function with substantial deviations from 'normal' development (e.g., two-headed animals exist and both heads function with independent control and personality), suggesting biology has built-in redundancy and robustness that may make engineering complex phenotypes easier than expected.

“You can have an animal or even a human that has two heads which, evolutionarily, there was no selection specifically to have two heads. But just a little deviation from the normal developmental pattern during fetal development and they both function fine. They control subsets of the body and they have their own personality, their own life.”

Evolution, as a mechanism, selects incrementally—each change must justify itself without knowledge of downstream benefits—whereas technology can adopt changes with non-obvious intermediate steps because engineers can afford non-intermediate-benefit innovations.

“In technology you can jump to things where all the intermediates aren't incrementally useful. But evolution, as far as we know, is generally limited to… You have to justify every change, like some bureaucracy, 'If you're going to put this sidewalk in, you have to justify that before you build a city.'”

Progress in reducing cognitive decline and cognitive enhancement is occurring but is early-stage; treating severe genetic developmental disabilities is more advanced than pushing beyond current human cognitive ranges.

“At the other end, we have reduction of cognitive decline by cognitive enhancement, which is showing some promise. But again, that's kind of like this early stage severe impediment to cognition having a late stage component.”

Church argues against rushing to AGI development, stating there is 'no rush' and it is an 'artificial emergency' unlike COVID-19; pushing hard on AGI/ASI creates safety risks disproportionate to solving actual crises, and we need international consensus on safe AI before proceeding

“There's no rush. This is a completely artificial emergency. This is not like COVID-19, where millions of people were dying if we delayed the science. This is something where, if there ever is a crisis, it's because we created it, it's not because we're trying to solve it.”

Progress to AGI or ASI requires understanding human ethics well enough to educate a foreign type of intelligence, which we have not done—we barely know how to pass ethics onto the next generation of humans, making AGI safety indefinitely challenging without major ethical advancement

“I don't think we understand our own ethics well enough to educate a completely foreign type of intelligence. We barely know how to pass it onto the next generation of humans.”

Intense competition undermines safety rules even if they exist; this pattern is historical and likely to repeat with AGI/ASI development unless deliberate structure prevents it

“What typically happens when there's an intense competition is those safety rules get undermined and pushed aside.”

Creating arbitrary phenotypes (e.g., human wings) requires learning rules of developmental biology at molecular, cellular, and multicellular levels—we can determine morphology at protein/nucleic acid level, but the language of multicellular control remains unknown, with tools like transcription factors and diffusion gradients being partial solutions

“Part of this has to do with just learning the rules of developmental biology. We can determine morphology at the molecular level now: proteins, nucleic acids. Determining at the cellular multicellular level, there's a lot more things you can do and a lot faster. But we don't know the language yet.”

Solving biosecurity requires not just moratoria and voluntary compliance, but active surveillance, clear consequences, whistleblower mechanisms, and international consensus—voluntary approaches alone (as shown by germline editing moratorium breaches) are insufficient

“One of the things I advocated in 2004 is that we stop deluding ourselves into thinking that moratorium and voluntary signups to be good citizens is going to be sufficient. We need to also have surveillance and consequences, and mechanisms for whistleblowers to make it easy for people to report things that they think are out of line.”

Genetic counseling is less expensive than gene therapy (cost per genome ~$100, dropping soon) with tremendous ROI through prevented opportunity costs (parents not leaving workforce, children healthier); genetic counseling should be covered by National Health Service and insurance companies, who would benefit more than disease treatment costs

“it's less than zero because you spend a hundred dollars per genome. It'll probably be less soon. You get the whole thing analyzed. Compare that to millions of dollars that will be lost in opportunity costs and them not being part of the workforce, taking care of them and so forth. So the return on investment is tremendous. It's at least a tenfold return on investment.”

Obtaining superintelligence benefits is hard to calculate—both goods and bads are uncertain; examples like 'instantaneous transport worldwide' could be transformative but might create priorities we don't actually care about (e.g., making travel irrelevant would be simpler than perfecting it)

“It could be a complete game changer. But on the other hand, it's like if we said we could get instantaneous transport all over the Earth. Well, we could say, 'Yes, that could be a game changer.' But do we really need it? Is that really important?”

Height, the most well-studied multigenic trait in humans, is controlled by approximately 10,000 genes out of ~20,000 protein-coding genes, but each individual gene has tiny influence—yet single genes like growth hormone (somatotropin) can produce extreme variation and are used clinically for seven different medical treatments, exemplifying how reductionism can identify high-leverage singular controls

“Take a very multigenic trait in humans. Height is probably the most well-studied one. They tracked it down to something on the order of 10,000 genes, of which we have 20,000 protein coding genes. Some of them are RNA coding genes. They each have a tiny influence on height. But if you take growth hormone, somatotropin, you have extreme examples where you'll get extremely low small stature and extremely high stature due to that one alone. In fact, it's used clinically as well for seven different medical treatments.”

We have little experience with thousands of Einstein-level intelligences operating simultaneously in a generation; pregnancy cannot be parallelized (nine women cannot achieve one-month pregnancy), suggesting some biological processes have irreducible serial structure that more intelligence cannot overcome

“It's like the question of, 'If you have nine women, can you do pregnancy in one month?' No, not at present. But the same thing is that there may be certain things that don't take a lot of people. We just don't know.”

Genetic disease prevalence is ~3% of children, but the public often feels 'I'm in the 97%' lottery-style—however, 3% risk is equivalent to favorable odds at horse races/casinos, and when child welfare is at stake, accepting that risk is ethically problematic

“Only 3% of children are severely affected by genetic diseases and they feel like, 'I'm not that unlucky. I'm in the 97%.' If those were your odds of winning at the horse races or at the casino, you'd take them. 97% of winning, good. But when a child's future is at risk, I think that's not the right solution.”

Biology operates at atomic precision (0.4 nanometer resolution in 3D), which is a billion times higher density than current semiconductor manufacturing (40 nanometers center-to-center in 2D-3D); if biology were applied to materials manufacturing, it could achieve atomic-level manufacturing across the periodic table.

“Biology is already at 0.4 nanometer resolution and it is in three dimensions. Depending on how you count that third dimension, it could be a billion times higher density that biology is already at... Even electrical and mechanical engineering don't typically use the whole periodic table typically, especially not at the atomic level. Biology is just really good at doing atomic precision.”

People resist genetic counseling partly through 'trolley problem' reasoning: inaction feels like not their fault, but actually all reproductive decisions are decisions; doing nothing is a choice with consequences

“I think it has to do with the trolley problem. If you don't influence it, it's not your fault. But actually everything is your fault. Not doing something is a decision.”

With exponential growth, one can appear productive by 'just jumping out of a plane and accelerating steadily'—positioning matters more than effort once exponential kicks in

“To some extent with the exponential, you can really look like you're very productive when really you're just kind of sliding downhill. It's like, 'Yeah, look at how productive I am. I just jumped out of a plane and am accelerating steadily.'”

Church co-authored a warning paper about the dangers of mirror life (organisms using opposite chirality DNA/RNA), noting that if such life could be weaponized, it would be concerning because it could potentially eliminate all competing life if properly weaponized, though weaponizing existing pathogens may be more practical than creating mirror life

“I was a co-author on a paper that warned about the dangers of mirror life.”

Whole-genome engineering to create novel phenotypes not present in human genetic variation (like wings) requires learning the rules of developmental biology at multiple levels: molecular (proteins/nucleic acids), cellular (which we can do faster), and multicellular (where we 'don't know the language yet'), with the transcription factor approach being a near-term tool but many other tools needed

“Part of this has to do with just learning the rules of developmental biology, like I said. We can determine morphology at the molecular level now: proteins, nucleic acids. Determining at the cellular multicellular level, there's a lot more things you can do and a lot faster. But we don't know the language yet.”

Biology operates at 0.4 nanometer three-dimensional resolution in atomic-precision manufacturing, while current semiconductor technology (1nm process node) actually operates at ~40nm with 2D spacing, representing a billion times higher density that biology already achieves, providing a blueprint for future nanotechnology and materials science.

“Biology is already at 0.4 nanometer resolution and it is in three dimensions. Depending on how you count that third dimension, it could be a billion times higher density that biology is already at.”

Scientific AI (tools designed for specific biological problems like protein folding or drug discovery) is more promising than general-purpose language AI because language AI requires AGI or ASI to improve further, which carries extreme safety risks, whereas scientific AI solves concrete problems without requiring artificial general intelligence.

“I'm much more excited about scientific AI than I am about language AI. With languages, we're in pretty good shape already.”

Voluntary moratoria and whistleblower systems have proven insufficient for biosecurity governance, as demonstrated by the He Jiankui germline editing case where moratoria lasted only 5 years before defection, and despite this being 'quite impressive,' all it takes is one defector to cause catastrophe, so surveillance mechanisms with known consequences must replace reliance on ethics alone

“We had essentially moratoria and disapproval for germline editing. Nevertheless, somebody did it and a lot of people knew about it. That was clearly a failure of the whole moratorium and voluntary and whistleblower components. It worked for five years with only one defector. That's quite impressive.”

The human brain contains 10^11 neurons and 10^14 synapses; encoding or replicating a specific brain configuration would require more information than the genome (billions of bytes vs. 10^14 synapses), though whether full replication is necessary versus synthesis is speculative

“You have 10^11 neurons, 10^14 synapses. If you wanted to reproduce a particular brain, it's speculative as to whether it would be easier to do that by making a copy of it in silico, in some kind of inorganic matrix, or making a copy of it.”

Height is controlled by approximately 10,000 genes out of 20,000 protein-coding genes, each with tiny individual effects, but growth hormone (somatotropin) alone can cause extreme variation in stature, demonstrating that complex multigenic traits can sometimes be controlled by single 'master knobs' and this reductionist approach has already been used clinically for seven different medical treatments

“Take a very multigenic trait in humans. Height is probably the most well-studied one, simply because no matter what gene, no matter what medical condition you're studying, you collect information on height and weight and things like that. Anyway, they tracked it down to something on the order of 10,000 genes, of which we have 20,000 protein coding genes. Some of them are RNA coding genes. They each have a tiny influence on height. But if you take growth hormone, somatotropin, you have extreme examples where you'll get extremely low small stature and extremely high stature due to that one alone.”

AlphaFold predicts protein structure with high precision but does not predict function; substituting an alanine for a serine in a serine protease results in correct folding (precise to a fraction of Ångström) but loss of function, requiring additional knowledge from evolution or experiments to determine if a design will actually work.

“With AlphaFold—last time I checked anyway—if you substitute an alanine for a serine in a serine protease, it will have exactly the right fold. It will be precise to a fraction of Angstrom overall average. But it won't function... you need either extraordinary precision or just knowledge of what happens evolutionarily, or happens in experiments, to say that, 'No, alanine won't work.'”