Episode Summary
Executive Summary: Steve Hsu discusses Richard Feynman’s advice on dating, then shifts to genomic prediction: using large genomic datasets and machine learning to predict traits and health risks, especially for IVF embryo selection. He argues most human traits are highly polygenic and largely additive, making prediction and future editing increasingly feasible, while also exploring ethics, inequality, ancestry differences, and the business realities of Genomic Prediction.
Main Topics: Feynman’s dating advice and campus culture (Priority: 4/5): Hsu recounts a memorable conversation with Richard Feynman in which Feynman treated meeting women as a rational, numbers-based problem and emphasized not internalizing rejection. Bodybuilding, physical culture, and generational change (Priority: 3/5): The discussion broadens into how gym culture, muscularity, and training practices evolved in the U.S., with Hsu contrasting modern norms with the 1970s–80s and earlier attitudes toward weightlifting. Genomic Prediction’s core science and IVF application (Priority: 5/5): Hsu explains that the company uses genomic data and machine learning to predict human traits and disease risk, with the main application being embryo selection in IVF when multiple embryos are available. Polygenic risk, additive genetics, and machine learning (Priority: 5/5): Hsu argues that many traits are highly polygenic but still predict well with simple additive models, supported by sparse methods, compressed sensing, and large training sets. Evolutionary theory, trait architecture, and future editing (Priority: 5/5): The conversation covers Fisherian additive variance, why evolution has not optimized humans to the maximum, and how future CRISPR/multiplex editing could exploit many available genetic variants. Ethics, inequality, and regulation of embryo selection (Priority: 4/5): They debate whether embryo selection will widen inequality or become a public-health tool, with comparisons to Denmark, Israel, eugenics history, and possible government support. Talent, education, and cross-disciplinary thinking (Priority: 4/5): Hsu reflects on elite education, why physicists transition into finance, why generalist intelligence matters for entrepreneurship, and how elite schools and immigration shape talent flows.
Key Arguments: Feynman’s advice framed dating as a numbers game: don’t personalize rejection; treat it as an operational process and move on. Bodybuilding and gym culture became normalized only relatively recently; earlier generations viewed weightlifting as unusual or even undesirable. Human traits are often highly polygenic but still predictable because the causal effects are largely additive rather than strongly nonlinear. Sparse, high-dimensional methods like L1-penalized regression/compressed sensing are useful for extracting signal from genomic data. A few hundred thousand to about a million well-phenotyped genomes can be enough to build strong predictors for traits like height or disease risk. Embryo selection in IVF is the most immediate commercial use case because clinics already create multiple embryos and can choose among them. Evolution has not had enough time or stability to optimize every trait; many diseases appear late in life, after reproductive fitness matters most. There is plenty of genetic “room” for improvement: traits like height may be influenced by thousands of variants, so modest edits can shift outcomes materially. Ancestry portability remains a major challenge because predictors trained on European data lose performance in other populations due to tagging differences and LD structure. The company’s business moat may come more from clinic relationships, workflow integration, and patents than from a permanent scientific monopoly. Ethically, the same technologies could either widen inequality or be socialized through public healthcare, depending on policy choices. Physicists are unusually good at moving into adjacent fields because they are trained to handle noisy data, build models, and reason from first principles.
Data Points: Feynman birth year: 1918 - Used to explain his generational unfamiliarity with modern gym culture. Hsu’s size in college: Almost 200 pounds, just over 6 feet tall - He describes his physique when Feynman called him a big guy. Approximate squat strength: 400 pounds - Hsu cites this as his strength level at the time of the trainer-room anecdote. Genomic Prediction clinic reach: 200–300 IVF clinics - Hsu says the company works with clinics across six continents. Genome-phenotype training scale: Almost a million genomes / half a million genomes - He describes the scale of training runs for predictor construction. Height sample size estimate: A few hundred thousand individuals - He says earlier math predicted this would be needed to solve height as a phenotype. Paper timing: Around 2012 - He references the paper proving data requirements for predicting traits like height. Predictive boundary: Phase transition - He says practical performance improved when data crossed the boundary predicted by theory. Human genomic differences: A few million places - He notes that any pair of humans differ at millions of genomic locations. Typical trait variant count: Order of 10,000 variants - He uses height as an example of a trait influenced by about 10,000 variants. One-standard-deviation edit estimate: About 100 variants - He argues that shifting a trait by one standard deviation may require flipping roughly sqrt(n) variants. Top embryo health outcome: About 120 years - He says their health index implies the best embryo among a very large population would be predicted to live to around 120. Denmark IVF birth rate: 1 in 10 babies - He uses Denmark to argue IVF is already mainstream in some countries. US IVF rate: 3% to 5% - He estimates the share of US babies born via IVF. Egg retrieval in young donors: 60 to 100 eggs per cycle - He says this is not unknown for young hormonally stimulated egg donors. Height predictor on ancestry: Falls off on distant populations - He says prediction quality drops when a model trained in one ancestry is applied to another. Evidence from sibling pairs: 20,000 sibling pairs in UK Biobank - Used to test whether polygenic predictions capture true genetic signal beyond shared environment. Wonderlic correlation: 0.8 or 0.9 - He cites this as a rough correlation with fuller IQ measurement.
Pivotal Quotes: "The main thing I remember was the operationalization of it as an algorithm and that you should just not internalize whatever happens if you get rejected because that's what really hurts." — Steve Hsu: Summarizing Feynman’s dating advice and the psychological lesson he took from it. "The ultimate models that are used when you've done all the training and the dust settles, the models are very simple. They have an additive structure." — Steve Hsu: Explaining why polygenic prediction often works with simple linear models despite biological complexity. "We're going to fucking explore it fast now." — Steve Hsu: His emphatic statement about future genetic engineering and embryo selection capabilities.
Implications: Genomic prediction is moving from academic theory to practical reproductive technology. If prediction, editing, and pluripotency scale, IVF selection could become routine, raising major questions about access, regulation, inequality, and how societies manage increasingly powerful human optimization tools.