Dwarkesh PodcastSteve Hsu - Intelligence, Embryo Selection, & The Future of Humanity
CHAPTERS
- 0:00 – 7:58
Feynman’s dating algorithm: rejection, confidence, and being rational
Steve Hsu recounts a memorable campus encounter where Richard Feynman gives him blunt, probabilistic advice on meeting women. The story becomes a broader reflection on fear of rejection, confidence, and how social norms around dating changed from bars to apps.
- •Feynman’s curiosity about gyms as social spaces vs training spaces
- •Dating framed as a ‘numbers game’—don’t internalize rejection
- •Practical tips: confidence, humor, and low-stakes interactions
- •Generational shift from in-person approaches to swipe-based dating
- 7:58 – 12:20
From weight rooms to ‘physical culture’: why everyone used to be skinny
Hsu and Dwarkesh digress into how bodybuilding and strength training spread in the US, and why muscular physiques were once rare. The discussion contrasts modern training knowledge with earlier eras, and how even athletic staff once misread developed musculature as injury.
- •Early stigma around bodybuilding and changing coaching beliefs about lifting
- •Anecdotes from Caltech football and training misunderstandings
- •Historical comparison: WWII-era soldiers vs modern physiques
- •Greeks as early pioneers of systematic athletic training
- 12:20 – 16:21
Genomic Prediction explained: turning genomes into embryo choices
Hsu introduces Genomic Prediction through both a scientific lens (genotype-to-phenotype prediction) and the IVF workflow. He explains embryo genotyping and polygenic risk scoring as a way to reduce disease risk when parents must choose among multiple embryos.
- •Genotype→phenotype as a central machine learning problem
- •IVF creates an ‘embryo choice’ problem when many embryos exist
- •Embryo biopsy + genotyping enables risk ranking for common diseases
- •Company scale: working with hundreds of IVF clinics globally
- 16:21 – 24:54
Why polygenic prediction works: linear models, sparsity, and phase transitions
Dwarkesh challenges why simple weighted-sum risk scores can be strong; Hsu explains the subtle math behind training robust predictors. He highlights sparse regression (L1 methods), correlation structure in genomes, and why performance jumps once datasets pass key thresholds.
- •Human pairs differ at millions of sites; predictors pick informative subsets
- •Compressed sensing/L1 penalization to avoid overcounting correlated SNPs
- •Additive models work surprisingly well despite claims of strong nonlinearity
- •‘Phase transition’ in predictive accuracy as sample sizes scale up
- 24:54 – 34:47
Evolutionary reasons for additivity—and why humans aren’t ‘optimized’
The conversation shifts to why genetic architectures tend to be additive and modular, tying to Fisher’s fundamental theorem and recombination. Hsu argues ‘optimization’ is ill-defined because environments change, and many modern diseases emerge post-reproduction so selection pressure was weak.
- •Fisher: response to selection is dominated by additive genetic variance
- •Sexual reproduction breaks fragile nonlinear ‘mechanisms’; additivity is robust
- •Most major diseases manifest late in life—weak selection historically
- •Changing environments (agriculture, modern life) break naive optimization intuitions
- 34:47 – 43:53
Longevity via embryo selection: polygenic health indices and available variance
Dwarkesh presses on aging and tradeoffs; Hsu describes polygenic health indices and argues there’s substantial variance to select on. He uses square-root-of-N intuition to explain why small numbers of edits could shift traits by meaningful standard deviations, and connects this to animal breeding successes.
- •Polygenic Health Index: selecting among ~10 embryos can add ~4 healthy life-years (DALYs)
- •Square-root logic: large polygenicity implies big selectable/editable variance
- •Evolution didn’t explore the huge high-dimensional space deeply; tech will
- •Animal breeding examples: eggs/day, milk yield, ML-driven livestock selection
- 43:53 – 55:22
First-mover advantage and the IVF ‘stack’: clinics, logistics, patents, and moats
Hsu evaluates whether genomic prediction has a durable moat: business relationships vs raw data advantage. He describes the operational workflow from embryo biopsy shipping to cloud-based reporting, and why regulation may be difficult once it becomes ‘just bits.’
- •First-mover advantage mostly in clinic trust and distribution channels
- •Large-data players (e.g., consumer genetics firms) could compete on models
- •Workflow: clinics biopsy; samples shipped; genotyped in NJ or on-site at large clinics
- •Long-term shift from wet lab logistics to cloud/API-style services
- 55:22 – 1:00:29
Genomics in dating, face reconstruction, and intelligence-community ‘spooky’ uses
The discussion expands beyond IVF into using genomes for matchmaking, verifying traits like height, and even reconstructing faces from DNA. Hsu also notes intelligence agencies’ interest in DNA identification from trace samples, raising privacy and surveillance implications.
- •Dating-app concept: genome upload for compatibility and trait verification
- •Complementarity: matching partners to increase chance of exceptional children
- •DNA→face parameter prediction as an ‘inverse problem’ enabled by ML + data
- •Intel use cases: identifying individuals from cups, stamps, hair—Gattaca-like risks
- 1:00:29 – 1:04:31
Ancestral populations and portability: why predictors degrade across groups
Hsu explains the central technical limitation: most training data is European, so prediction quality drops in distant ancestry groups. He frames this as linkage/tagging decay around causal variants, and outlines how multi-ancestry data can help triangulate causal signals.
- •Predictors trained in one ancestry often transfer poorly to another
- •Hypothesis: causal variants may be similar, but tags differ due to LD structure
- •Using cross-population consistency to identify likely causal SNPs
- •Large non-European biobanks coming online may solve much of the gap
- 1:04:31 – 1:11:53
Taboos, ‘eugenics’ framing, and policy paths: bans vs nationalization
They discuss social backlash risks, inequality, and why cognitive traits are under-measured due to political fear. Hsu contrasts Western taboos with East Asian attitudes, and suggests adoption may outrun regulation—potentially pushing countries toward subsidizing IVF/genetic screening for equity.
- •Cognitive phenotyping is avoided despite scientific value due to controversy
- •Potential for inequality: gains first accrue to wealthy families
- •Policy fork: outright bans vs making IVF/genetic screening public healthcare
- •Historical reversal: early progressives embraced ‘eugenics’ as public health; modern discourse is inverted
- 1:11:53 – 1:34:31
Selecting for intelligence: data bottlenecks, educational attainment, and ‘G’ vs EA
Hsu calls intelligence the most scientifically valuable but politically constrained trait to measure at scale. They explore proxies like educational attainment (EA), how to decompose EA into intelligence vs conformity/conscientiousness with additional surveys, and sibling-pair tests that reveal EA’s environmental contamination.
- •Need ~1–2M well-phenotyped genomes for strong IQ predictors (SE ~10 points estimate)
- •EA is a noisy proxy; captures family/parental forcing and social signaling
- •Proposal: collect personality measures to partial out conformity/conscientiousness
- •Sibling-pair validation: G-like predictors hold up better than EA within families
- 1:34:31 – 1:45:21
Future scaling: egg donors, iPSC-derived eggs, iterated selection, and CRISPR editing
Hsu describes how the limiting factor is eggs, not sperm—and how young donors can yield astonishing numbers, enabling selection among hundreds. They cover iPSC-to-egg technologies (already in rodents), iterated embryo selection concepts, and why multiplex gene editing plus better causal maps could dominate longer-term.
- •Egg donors can produce 60–100 eggs/cycle, enabling extreme selection among embryos
- •iPSC-derived eggs: plausible in humans after rodent successes; adoption may lag due to perceived risk
- •Iterated embryo selection: conceptually feasible though not widely pursued yet
- •Parallel progress: multiplex CRISPR editing and better causal-variant identification
- 1:45:21 – 1:51:48
Wordcels vs shape rotators: spatial ability, programming styles, and ‘visual physics’
Pivoting to psychometrics, they unpack the meme about cognitive styles and the reality of separable-but-correlated abilities (verbal, math, spatial). Hsu argues engineering skews more spatial than pure programming, and shares how physicists’ thought experiments rely heavily on visualization plus estimation.
- •Different cognitive ‘routes’ to the same problem: visual vs logico-verbal
- •Spatial ability was once explicitly tested in shop-class-like settings
- •Engineers tend to have higher spatial ability; great programmers can be non-spatial
- •Physicists often relax by visual-quantitative scenario modeling (e.g., targeting problems)
- 1:51:48 – 2:21:22
Bezos, elite education, and why physics talent spills into startups and finance
The closing sections discuss why physicists move into finance (stochastic math, noisy data instincts), and what elite schools add beyond test scores (networks, ambition exposure, social access). Hsu also revisits Bezos’ famous ‘brilliant friend’ anecdote to argue success is multi-dimensional: generalist intelligence, risk tolerance, focus, and communication across audiences.
- •Finance pipeline: options/random-walk math + physicists’ comfort with noisy data
- •Elite education effects: networks, ambition norms, access to influential circles
- •Bezos story: not just raw ‘physics IQ’—rare combination of traits drives outcomes
- •Founders as ‘multi-band communicators’ tailored to VCs, lawyers, engineers, and operators
