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

Dwarkesh Patel opens with a deliberately offbeat question: what did Richard Feynman tell Steve Hsu about picking up girls? Hsu answers with a specific memory from the spring of his senior year at Caltech. He was a physics major and, in his words, a linebacker on the football team at about 200 pounds and slightly over six feet tall. Feynman, whom he had known since his first year and regarded as a hero, stopped him on campus and remarked on his size.

The conversation quickly turns into an illustration of how Feynman thought. Hsu says Feynman, born in 1918 and in graduate school when the Second World War began, had no concept of a gym even in the 1980s, when Gold's Gym was becoming a national franchise. Feynman asked whether people went there to meet girls or to train, and then questioned Hsu at length. Hsu describes this as typical: when Feynman did not understand something he wanted to get to the bottom of, he would focus on the person in front of him until he did. Only after this interrogation did he offer advice in return.

The advice itself was to be rational and treat approaching women as a numbers game. Feynman noted that big men might not appeal to women, but added that Hsu had a nice smile. In a bar or airport lounge on a Saturday night in Pasadena or Westwood, Hsu recalls him saying, you will probably never see her again, so do what you have to do and move on if she is not interested. Hsu adds that Feynman had a reputation at Caltech as a womanizer, but that he knows this mostly from the secretaries, who at that time were almost all women.

Hsu draws out what he thinks mattered most. He compares the approach to the pick-up-artist idea of following an algorithm so that rejection does not reflect on your self-esteem, and says the main thing he remembers is that Feynman had operationalized the problem. Rejection hurts, and walking across a loud bar to talk to a woman and her friends was terrifying, so Feynman was telling him to train himself out of that fear. The reasoning offered is that the "face space of humanity" is so large that you will probably never meet the person again.

the face space of humanity is so big that you’ll probably never re-encounter them again

Patel wonders whether such behavior existed in the 1940s. Hsu is confident it did, pointing to literature from that period or slightly earlier, such as Hemingway and John O'Hara, which describes men and women meeting in New York bars. He says this was much more common than for Patel's generation, and admits he cannot work out how boys and girls meet now, including among his own children. In his telling, the man used to do all the work and had to overcome the fear of doing it.

Patel asks for context on Hsu's physique, since he looks muscular in a photo on his website, and wonders how common that was at the time. Hsu says Feynman found gyms strange partly because of how bodybuilding entered the United States: bodybuilders were first regarded as freaks and homosexuals. He picked up advanced training techniques in swimming from the Russians and East Germans, but his football coach in junior high told players not to lift weights because it would make them slow. This changed around his high school years, which Hsu attributes mainly to the University of Nebraska's well-known strength program popularizing weight training.

Hsu gives two pieces of evidence for how rare knowledgeable training was in the 1980s. People who knew how to train used what counted as advanced knowledge then, such as a split routine alternating squats with upper-body days. And when he pulled a quadriceps muscle, a trainer at the Caltech athletic facility looked at his well-developed thigh and said it was swollen. Hsu says he could squat about 400 pounds at the time and that her failure to recognize well-developed anatomy surprised him. He closes the thread by saying the field has come a long way and that you do not have to be old to understand how it evolved over 30 to 40 years.

Patel pushes on whether this was specific to that era, citing stories of Roman soldiers carrying 80 pounds for 10 or 20 miles a day and the muscular figures in ancient sculpture. Hsu replies that the Greeks were special because they were the first to think about the gymnasium, including the Palaestra, where people trained in wrestling and boxing. He calls them the first to take specific training for athletic competition seriously. He then returns to the twentieth century: old photos from the 1970s show skinny men, and soldiers of World War Two, whether German or American, were around 5'8" to 5'9" and 130 to 140 pounds. His conclusion is that physical culture was largely lost after antiquity and was only just coming back to the US as he grew up, so a lean 200-pound man who could bench over 300 pounds was rare.

Asked to introduce Genomic Prediction, Hsu offers two views, the scientific and the IVF view. Scientifically, more and more genomic data exists. Given the genomes of many people plus information about each, such as whether they have diabetes, how tall they are or their IQ score, finding which features of DNA variation predict a given variable is a natural machine learning problem.

He frames this as the old question of relating genotype, the specific DNA pattern, to phenotype, the expressed characteristics of the organism. Since the molecular revolution showed that DNA stores the information passed on, and since evolution selects on DNA variation as expressed in phenotype through its effect on fitness and reproductive success, he calls this the whole ballgame for biology. As a physicist trained in mathematics and computation, he says he is lucky to have arrived when brute force, AI and machine learning can solve this problem. He then poses the entrepreneur's follow-up himself: working with postdocs on computers is one thing, but what is it for?

His answer is IVF. Millions of families worldwide go through it each year, typically because of fertility problems and mainly because the mother is in her 30s or 40s. Hormone stimulation produces more eggs: instead of one per cycle, anywhere from five to twenty depending on age, and sixty to a hundred for young, hormonally stimulated egg donors. Fertilizing them is easy, so a family can end up with more embryos than it will use and face the embryo choice problem, for example picking among 20 viable embryos.

The application he describes is genotyping those embryos from a small biopsy and reporting on them. His examples are an outlier in breast cancer risk for the fourth embryo, which would give him pause about using it, and an outlier in cardiovascular risk for the tenth, while the others are fine. He states that the company works with 200 or 300 IVF clinics on six continents.

Patel observes that disease risk of this kind is polygenic, with thousands of single nucleotide polymorphisms (SNPs) potentially involved, and asks how Hsu's mathematics and physics helped him make sense of the data. Hsu begins to answer by describing the underlying genetics: individual locations in the genome where people differ, typically a few million places between any pair of individuals, which controls why he looks a little different from Patel. This sets up the technical discussion of how predictors are built that follows.

Hsu explains how he chose this problem. Theoretical physicists, he says, often feel they have spare horsepower and dabble in other fields, but he has to pick problems where he can have an impact rather than spending a lifetime on one molecule or one biological system. About ten years earlier he confronted the falling cost of sequencing and could anticipate that millions of genomes with good phenotype data would become available, with a typical training run involving half a million to a million genomes. The mathematical question was which algorithm builds the best predictor from genomes and phenotypes, which he says reduces to a well-defined machine learning problem.