Steve Hsu and Dwarkesh Patel discuss genomic prediction, embryo selection, inherited traits and the limits of statistical scores. Seven research notes distinguish measured prediction accuracy from modelled health gains and technological forecasts. Hsu describes his company’s commercial work, while the wider conversation considers inequality, scientific careers, mathematical models and the roles of education and networks.
HostDwarkesh Patel
GuestSteve Hsu
About these timings. Publisher-supplied VTT cue times for the official audio. Automated captions contain recognition errors; summaries use clear passages cross-checked against the editorial transcript. Complete spoken closing ends at 8416.702 seconds; publisher audio metadata is 8448.549 seconds, rounded up here. No audio audit or cross-platform offset is claimed.
Hsu describes using genomic predictors to help choose among embryos and introduces his company’s commercial application.
Background and context
Lello and colleagues’ August 2018 study constructed predictors for height, heel bone density and educational attainment using UK Biobank data and additional validation datasets. The height predictor explained about 40 percent of total variation in held-out data, with a correlation of roughly 0.65 between predicted and observed height. These are substantial predictive results, but they are explicitly imperfect.
The experiment tested predictions against measurements in existing people. It was not a prospective trial showing improved health or other outcomes in children selected as embryos. Predicting one measured trait also does not establish the accuracy of every disease score or its usefulness within every family. The distinction is particularly relevant because Hsu is discussing a business he co-founded. The technical achievement can be evaluated on its own terms without treating it as clinical validation of the commercial service or a guarantee about an individual child.
Held-out prediction in existing participants is not a prospective embryo-selection trial or a guarantee of an individual outcome.
Patel asks about a four-year gain from a polygenic health index; Hsu treats it as evidence for the potential of selection.
Background and context
Widén and colleagues’ October 2022 publication, following the preprint discussed in the interview, combined scores for 20 diseases into a health index. It tested retrospective selection among UK Biobank participants, including sibling pairs and trios. The roughly four disability-adjusted life years figure came from selecting the highest-scoring person in groups of ten unrelated participants, using disease-burden weights.
That calculation is not a measurement of ten embryos producing children who subsequently lived four additional healthy years. The authors identify limitations involving omitted diseases, survival to recruitment, rare variants and future environmental conditions. They also report no statistically significant antagonistic trade-off across the studied disease set; this does not exclude every biological trade-off. The paper therefore supports an observed relationship between scores and disease records plus a modelled health metric, while the application to future children requires additional assumptions and evidence.
Retrospective adult comparisons and disease weights cannot establish a guaranteed lifespan benefit from embryo selection.
Hsu acknowledges that predictors trained mainly on European-ancestry data perform less well in more distant ancestry groups.
Background and context
Martin and colleagues’ March 2017 study examined population history and polygenic scores for eight phenotypes, then used simulations to investigate transfer between populations. Their analysis showed that scores built from European association studies could become biased and less accurate when applied elsewhere. Differences in allele frequencies, correlations among variants and discovery samples affected the result.
This is a warning about the measurement tool. It does not license ranking populations by a score whose calibration and portability have not been established, nor does it isolate genetic causes of observed group differences. The interview’s discussion of statistical tags touches one part of this problem, but identifying a predictive marker is not the same as proving its biological mechanism. Better representation and appropriate validation are empirical requirements. Hsu’s expectation that additional data will solve the problem on a particular schedule remains a forecast beyond this study.
A difference in score distributions or prediction quality is not, by itself, evidence of an innate hierarchy among populations.
Hsu describes using siblings to test genetic predictors and says educational-attainment predictions weaken more within families than some other traits.
Background and context
Howe and colleagues’ May 2022 study combined sibling data from 19 cohorts to compare population-level and within-family genetic associations across 25 phenotypes. The within-sibship estimates were smaller for educational attainment, height, cognitive ability and several behavioural outcomes. Results for many molecular traits were more similar between the two designs.
The design helps distinguish inherited associations from demographic structure and indirect effects associated with relatives. It does not make every environmental influence disappear: siblings need not experience identical environments, and the interpretation still depends on the statistical model and available measurements. This is a separate primary investigation that provides useful context for Hsu’s account of his group’s work. It also reinforces his distinction between educational attainment and cognitive test performance. A score associated with years of schooling is not simply a direct measure of an individual’s intelligence or an established causal intervention.
Within-family designs reduce particular biases; they do not remove all environment, prove every causal mechanism or equate schooling with intelligence.
Hsu discusses generating eggs from pluripotent cells and forecasts that related technologies could enable much larger selection pools.
Background and context
Hikabe and colleagues’ October 2016 paper reconstructed the female germline cycle in culture using mouse pluripotent stem cells. They generated mature oocytes from embryonic stem cells and induced pluripotent cells, including cells derived from adult tail tissue. The resulting eggs could be fertilised and produce viable offspring; pluripotent cell lines could also be derived again from the generated eggs.
The species and endpoint matter. This experiment was in mice, rather than evidence that the entire process had been demonstrated safely for human reproduction. It also did not test iterated human embryo selection or highly multiplexed editing. Hsu’s discussion shifts from a laboratory achievement to predictions about translation, scale and adoption. Those later steps require their own evidence. The original paper supports a remarkable reproductive-biology result while leaving human safety, efficiency and the proposed timeline unresolved.
A mouse laboratory result does not establish safe human gamete production, reproductive use or the forecast timetable.
Hsu explains the historical attraction of derivatives pricing for physicists by pointing to familiar mathematics.
Background and context
Black and Scholes’ May 1973 paper derived an option-valuation formula from a principle about portfolios of options and their underlying stocks: correctly related prices should not permit a sure profit through the proposed combination of positions. The paper extended this analysis to corporate liabilities that could be represented through options. It is a theoretical construction with a defined pricing problem.
That result helps explain Hsu’s observation about transferable mathematical tools. It does not show that a physics background alone guarantees investment success, predict the direction of a stock price, or establish that every market behaves according to a simple model. The distinction also echoes the later discussion of noisy data and debugging: mathematical reasoning and empirical adequacy are separate tasks. Similar equations can be useful across disciplines while the meanings of their variables, assumptions and tests of success remain different.
An analytical pricing result is not a guarantee of trading performance or evidence that one profession always outperforms another.
Hsu offers networking, family resources and exposure to ambitious peers as possible explanations for elite-college career advantages.
Background and context
A later study by Chetty, Deming and Friedman provides a way to investigate this question beyond anecdotes. The August 2023 working-paper version linked admissions records, test scores and tax data. It used variation in decisions for waitlisted applicants to estimate effects of attending highly selective private colleges, finding larger effects on reaching upper-tail outcomes than on average earnings.
The design relies on assumptions about admissions variation and concerns applicants near the admissions margin. It does not prove that every individual benefits equally, identify one universal networking mechanism, or establish an effect of innate intelligence. The authors also separate advantages in gaining admission from outcomes after attendance. This is subsequent empirical context, not a paper Hsu could have cited in the August 2022 conversation. Keeping the version and comparison explicit helps assess his proposed explanations without treating illustrative personal stories as causal evidence.
A working-paper estimate for marginal applicants does not identify all mechanisms or determine the outcome for every prospective student.
Remember a word from the conversation? Find an approximate listening time without reading through a transcript.
Times follow the Publisher audio edition. All times are approximate. Use up to six words; matching uses 30-second windows, not exact phrases. Transcription can be wrong, so listen to check. Searches stay in your browser.
PodCompanions