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

Millions of IVF cycles happen globally each year; clinics routinely biopsy embryos and send samples to labs like Genomic Prediction, which provide polygenic risk reports so parents can avoid embryos with outlier risks for major diseases.

Traits like height, disease risk, and likely cognitive ability are influenced by thousands of variants; using high-dimensional regression and compressed sensing, simple additive models (weighted sums of SNPs) can capture much of the heritable variance once datasets reach hundreds of thousands to millions of genomes.

Because traits depend on thousands of loci, changing the equivalent of only √N variants can shift a trait by a full standard deviation; this implies a deep “well” of unused variation that breeding, embryo selection, or future editing could exploit to extend healthspan or alter traits substantially beyond current human norms.

Polygenic scores trained on Europeans degrade in accuracy for other ancestries due to different LD/tagging structures; combining diverse biobanks and cross-group machine learning can progressively localize truly causal variants and improve portability.

Face-recognition embeddings are highly heritable; inverting such models using genomic data plus training could allow reconstruction of likely faces from embryo DNA, alongside independent control of hundreds to ~1,000 trait dimensions (health, morphology, psychology) given the information content in human genetic variation.