Steve Hsu - Intelligence, Embryo Selection, & The Future of Humanity
What this covers
Steve Hsu, a theoretical physicist and geneticist, discusses the feasibility and implications of embryo selection and eventual genetic editing for complex human traits. The conversation centers on a mathematical argument: because traits like intelligence, height, and disease susceptibility are polygenic—controlled by thousands of largely additive genetic variants—the machinery for moving these traits across genetic space is far simpler than conventional wisdom suggests. Hsu walks through the statistical foundations, real-world examples from agriculture and animal breeding, and the current state of SNP-based prediction in IVF clinics, arguing that the primary barriers are now practical (data availability, especially across ancestry groups) and ideological rather than biological.
The discussion ranges across multiple domains. Hsu explains how Fisher's Fundamental Theorem of Natural Selection accounts for why genetic effects are predominantly additive, details the math behind Square-Root Scaling of Polygenic Selection, and describes how his group used compressed sensing theorems (credited to Terence Tao) to predict when height prediction would saturate—predictions that proved accurate. He contrasts the political history, noting that early progressives like Margaret Sanger supported eugenics while today's progressives largely resist it, whereas East Asia views the enterprise more positively. Hsu addresses contentious territory: whether genetic architectures are sufficiently disjoint to "drive around" pleiotropy, whether induced pluripotent stem cells could eventually supply thousands of embryos for selection, and whether historical allele frequency shifts between European populations suggest traits have been genuinely selected over millennia—all presented as open questions grounded in genomic evidence rather than settled fact.
Hsu argues that most complex human traits—including disease risk, height, and cognitive ability—are governed by thousands of largely additive, disjoint genetic variants, which means embryo selection (and eventually editing) can reliably move traits across a vast high-dimensional genetic space, and the main barriers are data availability and ideological resistance rather than biological complexity.
- Polygenic traits are controlled by ~10,000+ additive variants, so flipping the square root of n variants yields a standard-deviation shift, implying a quasi-infinite well of variation to select or edit on.
- Genetic architectures for different traits are largely disjoint rather than pleiotropic, so a skilled engineer can 'drive around' undesirable correlations in high-dimensional space.
- The science is solved for selection in IVF clinics today; remaining obstacles are cross-ancestry data, IQ phenotype data suppressed by ideology, and editing technology maturing over ~10 years.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Because a trait controlled by ~10,000 variants follows the statistics of coin flips, shifting it by one standard deviation requires editing only the square root of the number of variants (about 100), so flipping 500 of the 10,000 height variants could in principle make someone five standard deviations taller—revealing a quasi-infinite well of variation available to selection or editing.
“if I could come to this little embryo, and I want to give it one extra standard deviation of height, I only need to edit 100. I only need to flip 100 minus variance to plus variance.”
Fisher's Fundamental Theorem of Natural Selection explains why genetic architecture is additive: additive variants dominate the rate of evolutionary response to selection because in sexually reproducing species where DNA is scrambled each generation, independent additive switches can be inherited without breaking, whereas nonlinear mechanisms get broken apart during recombination.
“He showed that it's the additive variants that dominate the rate of evolution. It's easy to understand why if it's a nonlinear mechanism, you need to make the rat bigger. When you sexually reproduce, and that gets chopped apart, you might break the mechanism.”
Most common diseases we care about manifest at age 50-60, beyond the typical reproductive lifespan in our evolutionary past when humans died early, so there was never strong evolutionary pressure to optimize against them; we only see them now because good conditions let people live to 80 or 90.
“most of the diseases we care about now—not the rare ones, but the common ones—manifest when you're 50-60 years old. So there was never any evolutionary advantage of being super long-lived.”
The existence of a vast well of selectable genetic variance is already established by animal and plant breeding: modern dairy cows, bioengineered chickens that lay nearly an egg a day (vs. one a month in the wild), and crops have all been dramatically transformed, increasingly using sophisticated machine learning genotyping over the last decade.
“The egg you ate for breakfast this morning, those bio-engineered chickens that lay almost an egg a day… A chicken in the wild lays an egg a month. How the hell did we do that? By genetic engineering.”
The political valence of embryo selection has inverted: early-20th-century progressives like Margaret Sanger embraced eugenics as a Darwinian way to keep society healthy, whereas today's progressives largely oppose it; in East Asia, including among Chinese communists, the view remains positive (the Chinese characters for eugenics mean 'healthy production').
“most progressives like Margaret Sanger, or anybody who was the progressive intellectual forebears of today's wokeist, in the early 20th century, were all that we would call today in Genesis”
Entrepreneurial success depends on several at-least-partly-uncorrelated traits—abstract intelligence, generalist intelligence, ability to work hard, take risks, function under pressure, and communicate across audiences—so being top-10% in each simultaneously is rare; thus a single unidimensional measure of intelligence (general factor G) is convenient for discussion but misleading, since high-G people can still be unusually helpful on technical problems outside their training.
“You have to distinguish between the very abstract intelligence—which is helpful in physics and math... versus a more generalist intelligence. Those are correlated, but they're not the same thing.”
Polygenic predictors trained on one ancestry group (mostly European data) lose accuracy when applied to a more distant ancestry group, likely because the predictor uses non-causal SNPs that merely tag the truly causal variant via population-specific correlation structure, and these tags decay across populations even if the underlying causal architecture is shared—a hypothesis solvable with machine learning once non-European biobanks come online.
“if you train a predictor in this ancestry group and go to a more distant ancestry group, there's a fall-off in the prediction quality.”
DNA is so robust that a person can be identified from a few molecules of skin cells left on a coffee cup, enabling intelligence agencies to identify a suspected spy and even determine details like that he has a daughter at a particular school, making it impossible for public figures to control their DNA.
“You can just take a few molecules of skin cells and phenotype them and know exactly who they are.”
Mathematical ability and spatial ability are distinct but correlated factors in psychometric principal-component analysis, and people solve identical problems through very different cognitive styles (visual vs. verbal-logical stepping); spatial ability was historically measured by standardized tests and manifested in shop-class machining skill, and is more central to engineering than to software programming.
“when you do principal component analysis and psychometrics, they turn out to be different but correlated.”
China and India are underrepresented in Nobel Prizes and top research per capita primarily because such metrics are lagging indicators of past development; their best undergraduates historically came to the US for PhDs (a powerful elite brain drain), but as those countries developed in Hsu's adult lifetime, their professors are becoming world-class, so the gap is gradually evening out.
“many of the things you mentioned are lagging indicators. So they reflect how the West was developed and had a solid scientific and engineering tradition, while China and India were desperately poor”
Young egg donors (around 19-23) can produce 60 to 100 eggs in a single IVF cycle, vastly more than older women in their 30s-40s who struggle to get a few viable embryos, which means a wealthy tech-savvy family could in principle select the best embryo from hundreds—dramatically increasing the returns to selection.
“I was told that 60 to 100 eggs per cycle is not unknown.”
Even if some intelligence-boosting variants have undesirable correlated effects (autism, dullness, weakness), the sheer number of variants (~10,000-20,000) means a skilled genetic engineer can simply select a different cluster of variants to 'drive around' any obstacle, as evidenced by historical geniuses who were also athletic and socially capable—proving such combinations exist.
“If I find out you can make someone smart using this stuff in this cluster, but it makes them dull or makes them autistic... I'll just go round. I don't need to use those, I have plenty more, look over here!”
The truly causal SNPs can be identified via machine learning by finding the variant that remains consistently significant across multiple ancestry groups as the surrounding tagging correlations vary—the SNP appearing at the intersection of all these different sets is likely the causal one—though this requires substantial non-European data.
“If there's an SNP, which is always coming up as very significant when I use it across multiple ancestry groups, maybe that one's casual.”
Using compressed sensing theorems (L1-penalized high-dimensional regression with performance guarantees, proved in part by Terence Tao), Hsu's group predicted in 2012 that a few hundred thousand genomes with phenotype data would be needed to 'solve' a trait like height, and the predicted phase transition from low to high predictor performance occurred exactly where the math said it would when they implemented it around 2017.
“We showed that you would need at least a few hundred thousand individuals and their genomes and their heights to solve for height as a phenotype. We proved that in a paper using all this fancy math in 2012.”
A simple calculation using the current polygenic longevity predictor shows that the best person out of a billion would be predicted to live about 120 years—matching the observed one-in-a-billion human maximum lifespan—suggesting that with editing of the top disease variants, lucky individuals could potentially live to 200 or even 300 years.
“If you asked, out of a billion people, “What's the best person typically, what would their score be on this index and then how long would they be predicted to live?”’ It's about 120 years. So it's spot on.”
The alleles activated in height predictors show signals of selection between North and South Europe over the last 5000 years, suggesting average height potential may genuinely differ between, e.g., Italians and Nordics—and by extension other traits could have been differentially selected over the same period, which is the politically dangerous implication few people understand well enough to discuss.
“The alleles, which are activated in height predictors, look like they've been under some selection between North and South Europe over the last 5000 years for whatever reason.”
Induced pluripotent stem cell technology—reverting a skin cell to a pluripotent state and turning it into an egg cell—has been more or less mastered in mice and rats across multiple generations in Japanese labs, so getting it working in humans is likely just a matter of years of lab work, which would remove eggs as the limiting factor and allow selection from a thousand embryos.
“There are a few labs in Japan where they seem to have fully mastered this for multiple generations of rats using induced pluripotent C to make the eggs.”
Genomic prediction could model complementarity between partners' genomes—if the regions where one partner has favorable variants for a trait like intelligence complement the other partner's favorable regions, the couple is likely to produce children scoring well above the parental mean (especially when combined with embryo selection), enabling advice on mating that 23andMe filed an early patent on.
“because of the complementarity structure of your genome in the regions that affect intelligence, you're very likely to have some super intelligent kids way above your, the mean of your you and your girlfriend's values.”
Because face-recognition parameters (hundreds of them) are highly heritable—as shown by twins mapping to nearly identical values—it will soon be possible to solve the inverse problem of predicting a person's face from their DNA and reconstructing what an embryo will look like at age 16 or 32; this is now purely an AI/ML problem since the basic biology clearly works.
“When you do face recognition in AI, you're mapping someone's face into a parameter space on the order of hundreds of parameters, each of those parameters is super heritable.”
By genotyping embryos from a small biopsy during IVF, Genomic Prediction can identify which of a set of viable embryos are statistical outliers for elevated disease risk (e.g., breast cancer, cardiovascular disease), allowing families to avoid selecting high-risk embryos.
“I could tell you things like your fourth embryo being an outlier. For breast cancer risk, I would think carefully about using number four.”
Humans aren't fully optimized on traits because evolution has had limited time and the human environment changed radically over the last 10,000 years (agriculture, modern life), so 'optimization for the environment' is ill-defined; in high-dimensional fitness landscapes you move on nearly-flat tilted surfaces and never reach a global maximum.
“evolution has not had that much time to optimize humans. The environment that humans live in changed radically in the last 10,000 years.”
Height predictions apply to people born and raised in a favorable, nutritionally adequate environment, where genes cap achievable height and environmental variation matters little; in poor environments (as seen with shorter Asian grandparents or current data from Asia) environmental effects are large, so heritability is high only conditional on a good environment.
“we can predict height for people who will be born and raised in a favorable environment... the height phenotype becomes super heritable because the environmental variation doesn't matter very much.”
High-skill immigration is good for the US on average but creates losers: it drains developing countries of talent and makes it harder for native-born Americans to compete at elite levels, with some analysts observing native-born (non-immigrant) kids increasingly absent from competitions like the IMO, suggesting open-door talent recruitment may discourage the native talent pool—a second-order effect.
“It's terrible for native-born Americans who have to compete against the best brains from all over the world; so much harder for an American kid”
Because modern science is so specialized and resource-intensive, it is increasingly rare to find polymathic individuals who can make recognized frontier contributions across multiple fields—far rarer than in Feynman's and von Neumann's era when science was smaller—even though the internet has created a golden age for general intellectual exchange.
“it's getting less and less common to find polymathic people who can do things at the frontier, where they make a significant contribution, and it's recognized by the natives in that sub-specialty”
Relating an organism's genotype (specific DNA pattern) to its phenotype (expressed characteristics) is fundamentally a machine learning problem: given enough genomes paired with phenotype data, AI can identify which DNA variations predict a given trait, which constitutes solving the central problem of biology by brute force.
“It's a natural AI machine learning problem to figure out which features in the DNA variation between people are predictive of whatever variable you're trying to predict.”
Contrary to expectations a decade ago that trait genetics would be highly nonlinear with many interaction terms, the genetic architecture of individual differences turned out to be largely additive—predictors that simply sum effect sizes across variant states work well—reflecting a deep and under-appreciated truth that the ways humans differ are mostly a matter of having more or less of additive switches rather than complex interactions.
“the ultimate predictors, which are very simple and addictive—sum over effect sizes and time states—work well. That’s related to a deep statement about the additive structure of the genetic architecture of individual differences.”
Dawkins-era antagonistic-pleiotropy stories about aging (a gene helping when young but harming when old) are true for some specific variants like sickle cell, but they are not the general story; the general finding from the last five years is that thousands of largely disjoint variants control almost every trait, so old-school evolutionary theorists weren't wrong but lacked the big picture.
“that's not the general story. The general story you only discovered in the last five years is that thousands of variants control almost every trait and those variants tend to be disjoint from the ones that control the other trait.”
The most scientifically interesting phenotype, cognitive ability, is lagging because researchers fear the controversy, leaving enormous genomic datasets without attached cognitive measurements even though a 12-minute test (like the NFL's Wonderlic, correlated 0.8-0.9 with full IQ) could cheaply provide them; with one to two million well-phenotyped genomes, a decent IQ predictor with a ~10-point standard error could be built.
“the most interesting phenotype of all is lagging because everybody's afraid, even though there are very good reasons for medical researchers to want to know the cognitive ability of people in their studies.”
Because humans passed through a tight evolutionary bottleneck and are not very genetically different from each other, the complex nonlinear subsystems are 'fixed' (identical across all humans), so the small ways individuals differ are mostly simple additive switches rather than nonlinear differences.
“humans have been through a pretty tight bottleneck, and we're not that different from each other.”
Contrary to naive expectations of widespread pleiotropy, the genomic regions controlling different disease conditions show very modest overlap—they are largely disjoint—because with 3 billion base pairs there is enough information for roughly 1000 independent traits, each controlled by its own disjoint set of variants.
“It's the opposite of what naive biologists would say when they talk about pleiotropy. They're just disjoint! Disjoint regions of your genome that govern certain things.”
Testing polygenic predictors on 20,000 UK Biobank sibling pairs—who share the same family environment—shows only a minor fall-off in predictive power versus unrelated pairs, demonstrating predictors capture true genetic signal rather than environmental confounds; notably the educational-attainment predictor degrades much more for siblings than the general-intelligence predictor, revealing that EA captures parental/environmental effects that G does not.
“There are 20,000 sibling pairs in the UK Biobank... let’s look to see how well we can predict which of the two brothers who experienced the same environment will be taller.”
Elite schools like Harvard produce overrepresented outcomes (top jobs, fundraising) beyond what SAT scores predict because attending exposes students to ambitious, wealthy, and well-connected peers who teach them how the elite world works and provide networks—a combination of selection, social learning, and connections rather than pure ability.
“When he gets to Harvard, he will meet many super ambitious, aggressive, smart kids... You just get a better view of what's possible in the elite sector of society from that exposure.”
Within roughly 10 years, CRISPR gene-editing technologies will be ready for massively multiplexed accurate edits, allowing genetic engineers to navigate the high-dimensional space of human traits at will, on roughly the same timescale that induced-pluripotent egg production matures and that the true causal variants are identified.
“Pretty confident in 10 years or so, with the CRISPR gene editing technologies will be ready for massively multiplexed edits. We'll start navigating in this high-dimensional space as much as we like.”
Gwern predicts modest total uptake of embryo selection because much of the population is hostile to new fertility technology and deterred by the cost, length, and pain of IVF; Hsu counters that since 3-10% of babies in developed countries are already born via IVF and already screened for chromosomal abnormalities, those families will almost certainly also use cheap additional disease-risk reports.
“That 10% of the population that's using IVF is going to look at the report, which can be generated by the cost of running some bits through the AWS server.”
Reversion to the mean normally prevents runaway genetic divergence between families, but if embryo selection becomes heritable in effect and there are increasing returns (wealthy families afford more treatments), society could shift from a normal distribution toward a bimodal one—a Morlock/Eloi split—that keeps diverging, a legitimate inequality concern though dependent on social policy.
“instead of a normal distribution for society, you can have a bimodal distribution that keeps getting further and further apart. That is a potential possibility.”
Unlike jiu-jitsu, which lets a smaller person reliably defeat a larger one through trainable technique, there is no analogous structured training system (dojo) that systematically amplifies raw cognitive ability; while tools and information channels (like skilled Google searching) can amplify capability somewhat and are highly G-loaded themselves, no method reliably lets someone punch above their cognitive weight.
“there are tools, but nobody uses them. There’s no dojo where you can go where they start teaching you immediately... There isn't anything analogous to this for cognition.”
Physicists, unlike mathematicians or most computer scientists, are uniquely trained through lab courses to connect noisy, imperfect real-world data to elegant theoretical models—iteratively debugging whether anomalies come from the model or the apparatus—a skill that transfers to finance and other empirical fields, whereas mathematicians tend to accept data as given.
“we spend a lot of time looking at bad noisy data. So even if you're a theorist, you had to go through these lab courses.”
A genomic prediction company's defensible advantage lies primarily in the business-channel relationships and trust built with IVF clinics worldwide rather than in a permanent scientific moat, since a competitor with much more data (like 23andMe) could in principle build better predictors—though sustained focus and expertise still make a market leader hard to beat.
“There is a huge first mover advantage because they're important in the channel relationships between us and the clinics. Nobody will be able to get in there very easily when they come later.”
The genomic prediction business will evolve to be purely cloud/software—clinics will have on-site sequencers, upload data, and receive a report seconds later—which makes the service extremely difficult to regulate because data can simply be uploaded to a vendor in a permissive jurisdiction even if a repressive country bans certain trait selection.
“we are going to be out of the wet part of this business and only in the cloud and bit part of this business.”
If embryo selection yields large gains accessible only to the wealthy, the resulting outcry over inequality could push countries to make it a free part of national healthcare—as Denmark and Israel already fund IVF aggressively, with one in ten Danish babies born through IVF.
“the main outcry will be if it turns out that there are massive gains to be had, and only the billionaires are getting them. But that might have the consequence of causing countries to make this free part of their national health care system.”
Physical culture and weight training are recent phenomena in the US; bodybuilders were initially regarded as freaks, football coaches discouraged lifting as making players slow, and the WWII generation was notably small (around 5'8", 130-140 lbs), with strength training only popularizing in the US around the 1980s.
“the way body-building entered the United States. They were regarded as freaks and homosexuals at first.”
Highly successful founders are selected for the ability to communicate effectively across very different audiences—investors, tech-heavy venture funds, engineers—by modeling how each thinks about a problem, which explains why someone like Sam Bankman-Fried can fluently reframe any challenging question for a given audience.
“Being a super successful founder selects for the ability to figure out how to effectively communicate with a person according to their background.”
Feynman advised approaching social rejection as a numbers game governed by an algorithm: since you'll likely never see the person again, you should not internalize rejection as a reflection on your self-esteem but simply move to the next attempt.
“the main thing I remember was the operationalization of it as an algorithm. You shouldn’t internalize whatever happens if you get rejected, because that hurts.”
Genomic Prediction works with 200 or 300 IVF clinics across six continents.
“We work with 200 or 300 different IVF clinics in six continents.”
When future superintelligent AI looks back on the early 21st-century genomics revolution, it will find Hsu's archive papers proving polygenic prediction was possible and give him credit, even though the work is currently under-appreciated.
“the future AI––that Roko's Basilisk AI–will look back and will give me a little credit for it.”