Steve Hsu - Intelligence, Embryo Selection, & The Future of Humanity

Genomic Prediction and the Future of Human Genetics

• Genomic Prediction utilizes AI and machine learning to correlate genomic data (genotype) with phenotypic traits such as height, IQ, and disease risk.
• The core scientific insight is that complex human traits are governed by a "polygenic" architecture where thousands of genetic variants (SNPs) act additively rather than through complex non-linear interactions.
• A 2012 mathematical proof established that predicting traits like height requires approximately 200,000 genomes, a threshold that the company validated with real data around 2017 using half a million genomes.
• The company currently operates with 200–300 IVF clinics across six continents, utilizing embryo biopsies to calculate polygenic risk scores for specific diseases before implantation.
• Forecasts suggest that selecting the best embryo out of 10 can increase a child's "disability-adjusted life years" (health span) by approximately four years.
• While the ultimate predictive model is a simple additive sum of effect sizes, the algorithmic process involves L1 penalized optimization (compressed sensing) to ensure sparsity and avoid overcounting correlated variants.
• The "genetic space" for human variation is estimated to contain roughly 1,000 independent parameters, allowing for significant modification of traits without inherent trade-offs due to pleiotropy in many cases.
• Current predictive models are most accurate for individuals raised in favorable environments; environmental factors like nutrition significantly impact the heritability of traits like height in populations with historical malnutrition.
• A major current limitation is the "portability" of polygenic scores across different ancestral populations due to differences in linkage disequilibrium (tagging), which causes prediction accuracy to drop for non-European groups.
• Future solutions involve using machine learning to identify causal variants that remain significant across diverse populations, independent of local tagging structures.

Market Dynamics and Technological Trajectories

• Genomic Prediction holds a first-mover advantage through established trust and channels with IVF clinics, though the long-term moat is debated against data-rich competitors like 23andMe.
• The business model is shifting from a "wet lab" service (shipping samples for genotyping) to a "bit and cloud" service as in-clinic sequencers become standard and data is uploaded for immediate analysis.
• Advanced capabilities such as predicting facial features or reconstructing faces from DNA are theoretically feasible and will likely become operational once specific heritability data for facial parameters is aggregated.
• Potential future applications include "genome-based matchmaking" where compatibility is assessed based on complementary genetic profiles for traits like intelligence, a concept previously patented by 23andMe.
• Critics and the public may eventually demand free access to these technologies via national healthcare systems to prevent inequality, similar to current IVF funding models in Denmark and Israel.
• The industry faces potential regulatory hurdles, particularly regarding non-medical trait selection (e.g., height, appearance), though the global nature of the data flow may make strict bans difficult to enforce.
• The "first mover" advantage may be transient if governments or competitors in China, Singapore, or other regions accelerate data collection for specific ancestry groups.

Evolutionary Theory and Intelligence

• The additive nature of genetic traits aligns with Fisher's Fundamental Theorem of Natural Selection, which posits that populations evolve most efficiently when traits are controlled by independent additive variants.
• Evolution has not "optimized" humans for maximum longevity or intelligence because modern environments (e.g., agriculture, high-calorie diets) are recent changes, and natural selection pressures historically acted only on reproductive success before age 50.
• Intelligence (g-factor) remains under-studied in genomic data due to the "woke" cultural climate and fear of social implications, creating a "data black hole" where cognitive metrics are missing from biobanks.
• Theoretical projections suggest that with 1 million+ phenotyped individuals, an IQ predictor with a standard error of roughly 10 points could be developed.
• Geneticists argue that high-dimensionality (10,000+ variants for intelligence) allows for optimization of cognitive traits without the necessary trade-offs (e.g., autism, mental illness) often cited by critics.
• Educational attainment (EA) predictors currently conflate intelligence with conformity and conscientiousness; separating these requires large-scale personality surveys and matrix diagonalization.
• Historical "geniuses" who were physically normal serve as existence proofs that high intelligence can be decoupled from other traits, suggesting a viable path for genetic engineering.

Physics, Skills, and Society

• Theoretical physicists transition to fields like finance and genomics because their training in handling noisy data, building robust models from imperfect information, and visualizing complex systems is highly transferable.
• Physicists uniquely possess the ability to distinguish between elegant mathematical models and "messy" real-world data, a skill set often lacking in pure mathematicians or computer scientists who may rely more on abstract logic.
• Entrepreneurial success (e.g., Jeff Bezos) is attributed to a combination of high general intelligence (g) and distinct soft skills like risk tolerance, hard work, and multi-band communication, rather than abstract theoretical prowess alone.
• The "brain drain" from India and China to the US has historically slowed scientific progress in those nations, though this gap is closing as top talent begins to stay domestically.
• Immigration of foreign scientists creates a dual effect: it benefits the US economy and research output while potentially suppressing wages for native-born engineers in traditional fields and reducing the long-term talent pool in developing nations.
• The concept of "spatial ability" and "mathematical ability" are distinct but correlated psychometric traits; while programmers may vary in their visual vs. logical processing styles, engineers generally require higher spatial ability than pure software developers.
• Cultural shifts have led to a reversal in "progressive" stances on eugenics; early 20th-century progressives (e.g., Margaret Sanger) supported genetic screening to improve society, whereas modern progressives often oppose such technologies due to fears of inequality and discrimination.
• China is culturally and politically open to "eugenics" (viewed as "healthy production"), potentially giving them a head start in genomic optimization and public health management compared to Western nations.