{
  "video_id": "sRKBGVFVYAw",
  "channel_slug": "dwarkeshpatel",
  "channel_handle": "dwarkeshpatel",
  "title": "David Reich – Bronze Age shock, the Neanderthal puzzle, & the sudden spread of farming",
  "duration_seconds": 8030.0,
  "url": "https://www.youtube.com/watch?v=sRKBGVFVYAw",
  "upload_date": "",
  "transcript": "Humans, at least in this part of the\nworld, were wrenched into a way of\nliving that was so different from how\ntheir hunter-gatherer ancestors lived\nthat the organism had to adapt very strongly.\nMaybe the degree of that wrenching process moving\ninto the Bronze Age was qualitatively greater\nthan the degree of the wrenching process that\nhappened from the initial transition to growing\nplants, which is surprising, because our cartoon\npicture is that the big transition is farming.\nBut the genetic data, the biological readout, is\nsaying our genome is reacting much more strongly\nto these events that happened 5,000 years ago.\nI am back with David Reich, who is a\nprofessor of ancient DNA at Harvard.\nHow do you describe what it is that you study?\nI'm a geneticist, and I work on human history\nand how ancient people relate to\neach other and people living today.\nWe did an interview two years ago,\nwhich ended up being one of the\nmost popular interviews I've ever done.\nI think people found it really compelling\nthat there's so much about human history we don't\nknow and are just learning about now as a result\nof the kinds of techniques your lab is using.\nYou have a new preprint that's very exciting,\nand I wanted to talk to you about it.\nCan you give me a little bit of context\non what we're talking about today?\nThe dream was that when this ancient\nDNA field started, more than 16 or 17 years\nago, we were going to learn a lot about\nbiology —about how people's biology changed\nover time— by getting DNA out of ancient\nhuman remains and tracking changes over time.\nAnd that dream has really not been realized\nsince the beginning of this field.\nThe field has been a big success\nwith regard to learning about human history.\nIt's resulted in surprising findings about\nhuman migrations —people not being descended\nfrom the people who lived in the same place\nhundreds or thousands or tens of thousands of\nyears before— and mixture being common in human\nhistory, and sex-biased processes being common.\nAnd there have been things that were\nnot expected from archaeology.\nThe field's been a big success from\nthat perspective, but what's not been successful\nis learning about biology and biological change.\nOne big reason has been that the\nsample sizes have been too small.\nWhen you have a single person's DNA, it provides\na tremendous amount of information about history.\nThat's because when you look at one person's\nDNA, it's not a single person. It's many\npeople. It's your two parents, your four\ngrandparents, your eight great-grandparents,\n16 great-great-grandparents, and so on.\nGoing back in time, thousands,\ntens of thousands, even hundreds of thousands\nof ancestors are contributing to people today.\nWhen you look at the DNA of a single\nperson's genome or a Neanderthal genome,\nyou have effectively tens of thousands of\nancestors all represented in your data.\nAnd you can position that\nindividual exquisitely with\nrespect to other people from whom you have data.\nBut when you are interested in how a particular\ngenetic variant—that affects something like\nyour skin pigmentation, or your ability to\ndigest cow's milk into adulthood, or a behavioral\ntrait—changes over time, a single person gives\nyou only one sample, or maybe two samples: the\none in their mother and the one in their father.\nTo get a high-resolution picture of how the\nfrequency changes over time, you need very big\nsample sizes, truly very large numbers of people.\nWe just didn't have that until the last few years.\nWhat motivates the study we're talking about\ntoday, and the work that hopefully a number\nof groups will be doing in the coming years, is\nthe fact that we now finally have those numbers.\nWe can do something with the data to\nsee how frequency changes over time.\nCan I ask a question? I'll be asking a lot of\nnaive questions through the next few hours, but\nwhy are frequency changes especially interesting?\nWhat we're interested in is using the experiment\nof nature that's occurred in our history, over\nthe last tens of thousands of years, to understand\nwhat's biologically significant in our DNA.\nIf there has been a change in environment\nthat a population has experienced—for example,\npeople shifted to agriculture, began living close\nto domesticated animals, or moved from a cold\nplace to a warm place, or a low place to a high\nplace—then there's pressure on the population\nto adapt to these new stresses and new needs.\nThe way you're going to detect that\nis by seeing that the frequency of a\ngenetic variant—that for example might\nallow you to live at higher altitude,\nor that might nudge you to have a different\nbehavioral pattern advantageous in the new\nsituation—pushes systematically in some direction\nin a way that is enough for you to detect.\nIt's very hard to detect slight shifts in\nfrequency by a few percent or ten percent\nunless you have a very big sample size.\nWhat we're looking for are those changes\nin frequency that are too\nextreme to be due to chance.\nThat will tell us there have been pushes against\nthe biology as a result of the changes in\nenvironment that people have experienced.\nInteresting. What did you guys find?\nSeven years ago, Ali Akbari, who at the time was a\npostdoctoral scientist in my laboratory and a few\nyears later became a permanent staff scientist,\nset out to use the data we were producing to\nlearn about biological change over time.\nI think the reason he was interested in\nour laboratory rather than other places was\nthat a focus of our lab has been generating\ntruly large amounts of data from ancient humans.\nWe've been trying to industrialize the process,\nmake it very inexpensive, make it high quality,\nand generate large numbers of samples with lots\nof good data for this purpose.\nThere's been this large amount\nof data we've generated.\nAnd it made it possible to\nconceive again of asking whether there\nhave been frequency changes over time.\nThe mainstream view in human evolution in the\nlast several decades has been that natural\nselection has been pretty quiescent over the last\nseveral hundred thousand years of human history.\nThere are several lines of evidence that\nhave been deployed to document this.\nOne is that if you compare diverse populations\nfrom different continents around the world,\nfor example Europeans and East Asians, and\nyou look at mutations that differ in frequency\nbetween these groups—all mutations differ a little\nbit in frequency, sometimes a lot—you can say,\n\"What are the most different mutations in terms\nof frequency between Europeans and East Asians?\"\nAnd there are almost no genetic\nchanges that are 100% different\nin frequency between Europeans and East Asians.\nEuropeans and East Asians descend from a common\nancestral population 40,000 or 50,000 years ago\nthat came out of Africa and the Middle East.\nThis population had a set of gene frequencies,\nand these variants bopped around randomly—a\nprocess known as genetic drift—or perhaps\nunder selection in one direction or another.\nThe time that's passed since 40,000 or\n50,000 years ago is sufficiently small on\nan evolutionary timescale that there's\njust not much genetic differentiation\non average between these two groups.\nHowever, if there's been natural selection, for\nexample to help people in one place digest alcohol\nbetter, or digest milk better, what you might\nexpect is that there would be some mutation that\nwould have rocketed up to very high frequency.\nForty or fifty thousand years is a lot of\ntime, it's maybe 1,500-2,000 generations.\nThat might easily be enough time to\nsee a 100% difference in frequency.\nYet you don't see any more than\nwhat you would expect by chance.\nThis combination of things made it seem\nthat selection has just been quiescent.\nMaybe a few hundred thousand years ago, the\nancestral human population got to some kind\nof optimum, and after that there hasn't been\nmuch genetic change in one way or the other.\nThere have been small amounts of natural\nselection, or selection to remove bad mutations\nthat are constantly raining down on the genome,\nbut not what we call directional selection.\nThat would be newly arising mutations, or\nmutations being pushed in a systematic direction,\nto help the population get to a different\nadaptive set point more favorable for the\nconditions that population is living in.\nWe were able to partition how much of the\nchanges in frequencies of all the mutations that\nwe're seeing in the DNA—we're looking at about 10\nmillion positions that vary—is due to directional\nselection (adaptation) versus other factors,\nespecially genetic drift.\nAnd 98% of it is other\nfactors, especially genetic drift.\nIt's overwhelmingly migrations and population\nstructure causing fluctuations in frequency.\nAs a result, it's super hard to detect the signals\nof adaptive natural selection because they're\na tiny fraction of the total frequency change.\nThe vast majority of it are\nthese migrations and mixtures.\nNevertheless, there's so much natural\nselection, as our study has shown,\nthat it's actually been rampant in the genome.\nCan I ask a clarifying question here?\nWhy are we discounting population\nadmixture or replacement as selection?\nIf you think about it at a group\nlevel, if one population replaces\nanother population, isn't that selection?\nI remember from the last episode you were\nexplaining how there have been huge changes in\nwhat kinds of people are in a specific area.\nOne population came in and replaced\nthe previous one, and then a new\npopulation came in and replaced that one.\nTo the extent that the genetics are relevant\nto why that population replaced the other one, why\nshould that not count towards what we understand\nto be selection over the last 10,000 years?\nIt could count, and may count, and probably\nshould count in some respects.\nBut it could also be that this\npopulation replacement is due to some\ncultural phenomenon —technology held by\none of these groups and not others.\nAnd maybe there are some genetic\nmutations that are contributing to\nthis. Who knows? It's possible. But\nwhat you're seeing is a whole-genome shift.\nWhat we're looking to see is whether there's one\nplace in the DNA that is driving the change in a\nway that's different from the rest of the genome.\nFrom a statistical point of view, what happens\nat these times of migration is there are just\nhuge fluctuations in frequencies.\nThese are extremely uninformative\ntimes for detecting natural selection.\nThe best moments to detect natural selection\nare when migrations and population admixtures\nare not happening for a few hundred years.\nDuring these times, you can\nactually see the mutation\nslowly blowing in one direction as a result.\nThe way we think about the history of Europe and\nthe Middle East for the purpose of this study is\nas an archipelago of little populations in space\nand time, each pretty isolated from each other.\nYou have a little population in Britain isolated\nfor a few hundred years, or a little population\nin Hungary isolated for a few hundred years,\nbetween big events of migration and mixture.\nIn each of those little experiments of nature,\nwe can ask: does this mutation\nslightly increase in frequency?\nDoes that same mutation\nslightly increase in frequency?\nIf all the arrows point in\nthe same direction, we win.\nThey're telling us that\nnatural selection is occurring.\nFor example, 4,500 years ago in Europe, almost\nall mutations went through huge frequency changes.\nThat's not because of natural selection.\nIt's because of the steppe migration from\nnorth of the Black and Caspian Sea. 40-80% of\nthe DNA becomes Yamnaya from steppe pastoralists.\nTheir frequencies of mutations were different\nnot because of selection necessarily, but just\nbecause they had evolved in different places\nfor thousands and tens of thousands of years.\nWhen you look at the descendant populations,\nthere are huge changes in frequency.\nWhat you need to do is see if natural\nselection is explaining a shift more\nthan you would expect by chance.\nOk, in this next section David\nexplained the nitty-gritty of the methodology\nof this paper. It's honestly a bit technical,\nand I wanted you to get a sense of the results\nfirst, so I've moved that section to the end.\nIf you want to understand the methodology\njust stick around for the full episode.\nSo you found these locations that seem to be\nunder selection. I have another clarifying\nquestion. You say you found 3,800 locations\nwhich you're 50% confident have been under\nselection in the last 10,000 years.\nIt's 7,200 where we're 50% confident.\nWe're getting about 7,200 positions in the\nDNA that have 50% confidence of being real.\nOnly half of those are real—we don't know\nwhich ones—so 3,600 of them are real.\nDoes that also mean that outside of\nthose 7,200, you're confident the other\nlocations in the genome are not under selection?\nNo. If you look at the 25% probability cutoff,\nthere will be tens of thousands, and\nthere will be many real ones there too.\nIn fact, multiple analyses we do suggest that\nthe genome is vibrating with natural selection.\nThere are all sorts of weaker\neffects that would be picked\nup in even larger studies than we've done.\nIn fact, almost every position in the DNA is\ncorrelated to another position that\nis being dragged in one way or the\nother by natural selection.\nInstead of being quiescent,\nnatural selection is everywhere.\nEven though it's only 2% of the\nfrequency change, it's tugging the positions\nin one direction or the other everywhere.\nSo we analyzed these positions that we\nhad identified, the hundreds of positions\nwe were super confident about.\nWe looked to see whether they\nwere randomly distributed in the\nDNA or whether they had patterns.\nWe looked at maybe 100 or so traits where\nthere had been genome-wide association\nstudies for all sorts of different traits,\nassociated with immunity or autoimmunity or\nbehavior or metabolism, and other things.\nFor each of these we could ask: do the\ngenetic variations that are known to affect these\ntraits from genome-wide association studies have\nan unusual number of genetic selection signals?\nWhat we found is that there was a vast enrichment,\nby about four or five-fold, for immune traits.\nThere was a super concentration of selected\nsignals in immune traits.\nWe also saw a strong enrichment\nfor metabolic traits—things that might impact\nobesity or fat traits or Type 2 diabetes—and\nalmost no detectable enrichment, as far as we\ncould tell, for behavioral or psychiatric traits.\nJust to make sure I understand.\nThis is not to say that behavioral or psychiatric\nor cognitive traits are not under selection.\nIt's just that the individual sites where such\ntraits are controlled are not especially\nlikely to be among the locations you've\nidentified as under selection.\nThat's exactly right. It might\nseem from the results of that analysis that\nimmune traits are highly selected and that\nthere's been no selection for behavior in the\nlast 18,000 years in this part of the world.\nBut that's a wrong conclusion, and we have\nevidence that it's a wrong conclusion.\nThere's clear evidence of selection\nalso on behavioral traits.\nThe reason we think we see much weaker signals\nfor behavioral traits is that behavioral traits,\nwe know from medical studies, are underpinned by\nmuch larger numbers of genes than immune traits,\nwhich are underpinned by relatively\nsmall numbers of genes of strong effect.\nBehavioral traits are shaped genetically by\na very large number of genes of weak effect,\nand we just don't have the statistical\npower to detect these very weak signals.\nWhen we do an analysis looking at our very\nstrong signals of selection, that collection\nof very strong results is very effectively\nquerying the immune traits, but is not very\neffectively querying the behavioral traits.\nIt may still be the case, and I guess it is,\nthat immune traits are the most selected category.\nBut it is not at all the case—and we\ncan prove it's not the case—that\nbehavioral traits are not selected.\nInteresting.\nWe've been able to\nprove that there are two ways to reconcile the\nprevious observations with our new observations.\nRemember, the previous observation is that\nnatural selection seems to have been quiescent\nover a timescale of hundreds of thousands or\nmany tens of thousands of years. Reason? That\nyou don't see 100% difference in frequency\nvariance across Europeans and East Asians.\nNow we're seeing hundreds of positions\nthat are rocketing up in frequency with\nselection rates of 1% or more in a lot of cases.\nA 1% or more selection rate will mean a rapid\ndoubling over periods of dozens of generations.\nOver the 1,500 or 2,000 generations separating\nEuropeans and East Asians, shouldn't you see\nmany genetic variants that are 100% different\nin frequency across populations?\nWe were able to show that this\nis explained by at least two factors.\nOne is that in this part of the world—Europe\nand the Middle East—we are actually in a\nperiod of accelerated natural selection.\nOne way to see this is to look at the enrichment\npattern we're observing, where immune traits are\nunusually associated with these selection signals.\nWe could compare the last 5,000 years of our\ntime period, what's called the Bronze Age and\nfurther onward, to the previous 5,000 years.\nWhat we see is that this intensification of\nselection around immune traits, and similarly\nthe intensification around metabolic traits,\nhas accelerated over this time period.\nIt's not like natural selection has been\nat the same rate over all places and times.\nIt's increasing over the\ntime period we're analyzing.\nPlausibly the whole time period has\nincreased compared to previous periods.\nWe're in a period of intensified selection.\nThat's not implausible, because this is a\npopulation that went through a huge shock in\nterms of the way people live and the culture.\nAlmost everyone we're analyzing are farmers\nor food producers in one way or another.\nFarming was invented for the first time\nanywhere in the world in the Middle\nEast 11,000 or 12,000 years ago.\nThe people who invented farming\nexploded into Europe after 8,500 years ago,\nspread across the continent, and expanded rapidly.\nIn the Bronze Age, there was an\nintensification of how people lived,\nwith much higher population densities.\nPeople were living more and more next to\ntheir animals and getting their\ndiseases, and exchanging their\ndiseases with the animals and with each other.\nThis is a period of rapid change in how people\nare living, resulting in different\nbiological needs of this population.\nIt's not surprising, perhaps, that in the context\nof these dramatic changes, the biology of the\npopulation might not be ideally adapted.\nThere might be what some people call an\nevolutionary mismatch, where you take\na genetic variation that evolved in\nhunter-gatherers and put it into farmers or\npastoralists, and it's not exactly right.\nWhat you're seeing is the DNA of this population,\nwhich descended from hunter-gatherers only 10,000\nyears ago, reacting to the shock of having\nbeen moved into an agricultural, Bronze Age,\nhigh-population-density, urban environment.\nA hypothesis is that what we're seeing is\nthe adaptation that occurs as a result.\nIn the paper you have many examples of\nthis intensification of\nselection around the Bronze Age.\nIt might be helpful to go through some of these.\nOne of the things we do in this work is look\ncarefully at many of these positions in the DNA.\nWe actually have an internet browser called the\nAGES browser, which Ali and a colleague of\nhis—who's a co-author of our paper—built.\nIt allows you to query each of these 10 million\npositions and see the trajectories at each\nposition and the evidence for selection.\nOne of the things we see is that,\nwhile for the most part the signals of natural\nselection we detect are consistent with constant\nnatural selection over time, in a handful of\nthem we're able to see that there's been a\nreversal or a radical change in natural selection.\nVery often that occurs in the period between 5,000\nto 2,000 years ago, which is the Bronze Age\nand the Iron Age, a period of rapid population\ngrowth and rapid movement to intensive use of many\ntechnologies that were not used that way before.\nAn example of this is the TYK2 genetic\nvariant that is a major risk factor for\nsevere tuberculosis, which is the most important\ninfectious disease killer in the world today.\nIf you look at this major risk factor for\ntuberculosis, this variant rockets up in\nfrequency from 8,000 or 6,000 years ago to\nmaybe 9% or 10% in this part of the world.\nThen it rockets down in frequency\nin the last 3,000 years.\nIn both cases, there's very clear\nevidence of natural selection,\nin the first case to increase in frequency,\nand in the next case to decrease in frequency.\nA possible reason is the spread of tuberculosis.\nIt maybe becomes endemic in the\npopulation 2,000 or 3,000 years ago.\nThat's potentially consistent with pathogen\nsequence data and other lines of evidence.\nAnd maybe this variant was protecting against\nsomething before then, but then tuberculosis\nbecame significant after that point,\nand it was so bad that it pushed in the\nopposite direction. That's speculative.\nThe thing it was protecting against\nwas probably another disease?\nMaybe.\nPrepping for\nthis episode required a full lit review.\nI needed to understand why other methods\nhad failed to find evidence of natural\nselection over the last 10,000 years.\nWhat exactly did Reich and Akbari do differently?\nHonestly, this was quite subtle because the most\nimportant points were distributed\nacross a bunch of different papers.\nAnd it was frustrating to talk to LLMs\nabout it because they kept getting confused.\nOne of them would fail to\nunderstand an important crux.\nAnd so I’d switch over to a different\nmodel, and that one would get tripped\nup on the very next point.\nI ended up using Cursor to\nkick off a handful of models at the same\ntime and compare their results after.\nI could have one model critique\nthe response of another.\nThis was super useful because\nwhile I'm not a geneticist,\nI do have enough taste to be able to say, “hey,\nthis answer makes sense, these ones don't.”\nI also had Cursor turn this work into\nflashcards so I could retain what I learned.\nCursor started as a programming tool, but I\nfound it really great for this kind of research.\nThere's no other interface where I can get answers\nfrom a bunch of independent LLMs all while reading\nthe relevant paper on the same screen.\nGo to cursor.com/dwarkesh to try it out.\nOne of the big takeaways for me from\nthe paper was just that something\nweird happened in the Bronze Age.\nAs you said, across trait after trait,\nthe selection intensifies during the Bronze Age.\nThis makes sense for some things.\nFor example, why do we see lactase\npersistence, where adults can process\nmilk, intensified during this period?\nThis is the time when we start using\ncattle not just for the meat, but also for\nmilk and wool and other secondary products.\nSo it makes sense why lactase\npersistence would matter more.\nBut then there are other things\nthat seem like they should have\nbeen relevant since the dawn of agriculture.\nI forget the exact name of the allele, but was\nit FADS1, which helps convert plant fatty acids\ninto long-chain fatty acids that your body needs?\nThat's obviously relevant when\nyou move from a diet of meat as\na hunter-gatherer to a diet of cereals.\nThat is also one I think you found was\nunder especially high selection 5,000\nto 3,000 years ago. So what's going on?\nWhy is the Bronze Age so special across all\nthese different traits that you're observing?\nSo this FADS1/2 variant is a\nvegetarian/meat-eating adaptation.\nAlready in work prior to this, Ian Mathieson,\nwho worked with me in 2015, identified this as\na very strongly selected variant. It’s actually\nancient. You see copies in archaic humans too.\nOne of the findings of our\npaper is the ABO blood system.\nYou get your blood typed as A, B, and O.\nThe B variant has increased up to 10% at\nthe expense of A, but previous work has shown\nthat A and B were both already present in the\nancestor of humans and gibbons and other apes.\nSome of these mutations have been going back and\nforth and fluctuating over different time periods.\nBut we're talking about changes in the Bronze Age.\nThe TYK2 variant for tuberculosis risk, a multiple\nsclerosis risk variant, inflected and increased in\nfrequency before the Bronze Age, and then 2,000\nor 3,000 years ago reversed in that period.\nThere are differences in Northern Europe\nwhere this process is super strong,\nvery strong positive selection,\nvery strong negative selection.\nAnd then in Southern Europe, only a little bit,\nand not even very strong negative selection.\nFor hemochromatosis, which is pathogenic\niron buildup that causes problems in Europe,\nthat too has reversed around this period.\nIn some of the complex traits that maybe\nwe'll talk about later, these traits too have\nperiods of intensification of natural selection.\nFor example, depigmentation: Europeans have\ngotten lighter skin over the last 10,000 years.\nYou can see it in our data.\nThe period of strongest\ndepigmentation is between about 4,000 to 2,000\nyears ago, and then after that it's much less.\nThis seems to be a very impactful, eventful,\nimportant period where a lot of the processes\nwe are seeing become very powerful. It's\nsurprising on first principles. You might think,\nbefore you walked into this genetic data, that\nthe big change is going to be starting to grow\nplants and maybe farm animals.\nThat happens in the Neolithic,\nbeginning 10,000-12,000 years ago, and\nspreads into Europe after 8,500 years ago.\nBut actually, the intensification happens\n5,000 years ago, 4,000 years ago. It’s\nreally interesting. This observation of that\nbeing an inflection point tells us something\nabout when humans, at least in this part\nof the world, were wrenched into a way of\nliving that was so different from how\ntheir hunter-gatherer ancestors lived\nthat the organism had to adapt very strongly.\nIt may be that the degree of that wrenching\nprocess moving into the Bronze Age was\nqualitatively greater than the degree of\nthe wrenching process that happened from\nthe initial transition to growing plants.\nThat's surprising, because our cartoon\npicture is that the big transition is farming.\nBut the biological readout is saying our\ngenome is reacting much more strongly to\nthese events that happened 5,000 years ago.\nYou did some work with Bhatia and many other\ncolleagues in 2014 where you were looking at\n20,000 or 30,000 African American genomes today.\nYou were saying, \"Look, there's 80% West\nAfrican DNA and then 20% European DNA.\nCan we look at their genomes today and see that\ntheir allele frequencies are much different than\nwe'd expect from this admixture?\"\nCorrect me if I’m wrong,\nbut you found that they weren't.\nThat is to say, over 200 or 300 years of extremely\nintense environmental change—going from chattel\nslavery to a completely new environment—there's\nno effect of natural selection.\nSo we see episodes like this where\nwe don't see natural selection, but then\nthe Bronze Age apparently must have had an\neven stronger effect, where the change in\nenvironment is even stronger than what we\nsee from Africans in Africa being migrated\nto the New World and living under slavery.\nThat may be the case. It also may be the case that\nthat period is just too short to see much effect.\nIn the Bhatia et al. paper, where we looked at\nabout 30,000 African Americans, we looked to\nsee whether—instead of the average percentage of\naround 80% West African ancestry—there were some\nplaces in the DNA with significantly more\nthan 80%, or significantly less than 80%.\nThat’s what you would expect if there\nwere natural selection for some genetic\nvariant from Europeans or from Africans.\nWe didn't see any place in the DNA that\nwas significantly different from\nwhat you would expect by chance.\nOne possible explanation is just that there's\nonly a handful of generations, maybe five,\nover which natural selection would operate.\nSo if the selection was 2% a generation,\nyou would still only see a 10% compounded effect,\nand there's just not enough time to detect it.\nBut the Bronze Age is not\n300 years, it's 3,000 years.\nIt's the power of compound interest, and you\nhave enough time to begin to see a strong effect.\nThis really, really does seem to be a very\nimpactful time in terms of human history,\nand you can see it in our complex traits.\nLook at pigmentation, for example,\nwhich is the strongest signal of selection\nfor a complex trait in our data set.\nYou look at genetic mutations that\nare known to affect pigmentation.\nYou add up their effect across all of the\nDNA, there's dozens or hundreds of them.\nYou look to see when natural\nselection is strongest, and the\ntime period is really 2,000 to 4,000 years ago.\nFor some of these other traits as well, you see\nagain that the time period over which selection\nis strongest is 2,000 to 4,000 years ago.\nFor example, if you look at genetic variants\nthat affect measures of cognitive performance,\nsuch as performance on intelligence\ntests in white British people today.\nThis is of course a very strange trait to\nmeasure in the past because there were no\nintelligence tests and there was no school.\nBut it is a predictor today, and you can\nlook at how it's changed in the past.\nWe see very strong natural selection\nfor this combination of genetic variants\nthat predicts people's performance on IQ\ntests and is also highly correlated to the\npredictor of the number of years of school\nor the household wealth of people.\nAll crazy traits in the past because\nthere was no wealth in the past,\nthere was no school in the past.\nBut if you look at the predictors today, there\nis a strong movement in a systematic direction,\na large effect, about a standard deviation\non the scale of modern variation.\nWe can do this trick of looking to see\nwhether there are periods of time when\nthis natural selection has occurred\nmore intensely or less intensely.\nWe drag a 2,000-year window through our data, and\nwe repeat our whole analysis, not on 18,000 years,\nbut just on a short 2,000-year window.\nWe can measure the strength of selection\nin each of these 2,000-year windows.\nWhat you see when you look at intelligence\nis that this maxes out in the Bronze\nAge, between 5,000 and 2,000 years ago.\nThe impact in the last 2,000\nyears is almost nothing.\nThere's no evidence of natural selection at all.\nYour bias coming into this, my bias\nperhaps, might be that if there's any signal of\nnatural selection on this trait at all, that it\nwould be unusually strong in the last 2,000 years.\nMaybe this is a time of industrialization.\nMaybe this is a time of greater\nneed for this particular trait.\nBut in fact, there's no evidence of natural\nselection at all in the last 2,000 years.\nThere's very strong evidence between 2,000\nand 4,000 years ago, where instead of a one\nstandard deviation strength of selection,\nit's a two standard deviation strength,\naveraged over this time period.\nThe standard deviation here is how much\nthe polygenic score for the trait itself moves?\nHow much the polygenic score for the trait moves\nover a 10,000-year period within a population\nthat is held constant in terms of its ancestry.\nWhat we're actually doing is looking in our\ndata set at a heterogeneous group of people.\nThere's Southern Europeans and Northern\nEuropeans and hunter-gatherers and farmers.\nAt different times in the past, those\ngroups are more or less represented.\nThe whole strength of the methodology Ali\nAkbari developed is that it corrects for that\nchanging ancestry over time.\nReally what's being asked here\nis that we've divided up our whole data set\ninto an archipelago of little populations\nin different places in space and time.\nWe're asking in each place in space and time:\na little pocket of people in Britain from 4,000\nyears ago to 3,500 years ago, a little pocket of\npeople in Hungary, a little pocket of people in\nItaly from 2,000 years ago to 1,500 years ago.\nIn each of these places, where the ancestry is\nrelatively similar without being too disrupted in\nthat short period by migrations, we watch to see\nif the genetic changes blow in the same direction.\nWe're measuring the strength of selection at\neach point in time after correcting for the big\npopulation changes that have occurred.\nThe effect here is huge then.\nOne standard deviation above the median\nwould be somebody in the 85th percentile.\nYou're saying the effect of selection has been\nso strong that comparing 10,000 years ago to now,\nthe median has gone to the 85th percentile.\nThat's just a huge effect over the last\n10,000 years on something like intelligence\nor the thing that predicts household income.\nEspecially given that this is only 2% of the\nchange in allele frequencies, and the 98% is\ncoming from migration… It's stupendous to think\nabout what the impact of migration is, if this\nalone is driving a standard deviation change in\nthese kinds of qualities, at least among the kind\nof variation we see in the world today.\nOne thing you can see in the data\nis that the migration impact is huge.\nFor example, if you look at the trajectory\nfor measures of cognitive performance—scores\non intelligence tests in white British people\ntoday—but you look at the predictor of that\nin people in ancient times, the estimate for\nthe hunter-gatherers of Europe is three standard\ndeviations below the modern mean. So that's hugely\ndifferent. Then you see a huge jump from them to\nthe farmers, who are at the mean, at zero. That's\nmigration. What you're seeing is that those two\ngroups had different set points for those traits.\nAnd then the steppe pastoralists\nhave a lower set value.\nYou see huge fluctuations in the predictor\nof this trait over time. That doesn't prove\nselection. That's just migration. But what our\ntest is telling you is: in addition to those\nfluctuations due to migration, is there\na consistent effect of natural selection\nblowing the trait in the same direction over all\nplaces and times? That's what we're detecting.\nThere's this theory called the\ncollective intelligence hypothesis,\nwhich is the idea that selection for intelligence\nhas actually been in the opposite direction.\nAs society has developed,\nthere's been more specialization,\nand if there's more specialization,\neach person only needs to understand\na smaller and smaller part of the world.\nTherefore, the ancients were actually much smarter\nthan us, and we've evolved down in intelligence.\nYour results seem to point\nin the opposite direction.\nAlthough there hasn't been selection\nin the last 2,000 years as society has gotten\nmore complicated, at least when society began,\nthere was more need for the kind of\nthing that predicts intelligence today.\nThe reason that's surprising is, if you\nthink about hunter-gatherers—reading\nyour colleague Joseph Henrich’s book—the amount of\ninformation they needed to hold onto and assess,\neverything from how to process food, to how\nto build shelters, fire, et cetera, compared\nto my world, where I just need to know how\nto set up mics and ask questions… It seems\nlike the demands on intelligence should have\nbeen way higher in the ancestral environment.\nSo it's very surprising that the\nbeginnings of civilization increased\nthe selection on intelligence.\nThis is the power of data.\nI think if you asked Joe prior to this work what\nthe hunter-gatherer selection would be and where\ntheir set point for this particular trait would\nhave been… I think he probably wouldn't have\nmade a very strong prediction, but he would\nhave said, \"Maybe you would have expected\nit to have a high predicted value of this trait\nbecause these people were really having to do a\nlot of things and figure a lot of stuff out.\nMaybe once you have more complex societies,\nthere would be more of a collective brain, and\nmaybe there'd be selection against this trait.\"\nIn fact, it's the opposite in some ways.\nIt's the power of data. It's not what you\nexpect. It's actually the value of data\nto try to make sense of all these things.\nIt's very interesting. The genetic predictor of\nintelligence, there are lots of things that are\nconfusing about it, so it's worth talking about.\nOr the genetic predictor of years of schooling,\nwhich is highly correlated to\nit and is measured even better.\nIf you look at the genetic predictor of years\nof schooling, there's another amazing study from\n2017 from a group in Iceland that looked at this\nmeasure over the last hundred years in Iceland.\nIt looked at older people and\nyounger people born more recently.\nThere's an estimated 0.1 standard deviation\ndecrease in the genetic predictor of intelligence\nin Iceland just within one century.\nIt's an absolutely huge\neffect over a short period.\nThis is selection against years of schooling.\nIf I said intelligence, I didn't mean to.\nIt's selection against the genetic\npredictors of the number of years of school.\nOne possible interpretation of this—hand-wavy—is\nthat what's being measured here is not selection\nfor years of schooling or for real intelligence,\nbut for another trait altogether\nthat's correlated to both of them.\nFor example, the predictor of the number of years\nof schooling is very strongly correlated to the\nage at which women have their first kid.\nIf you control for that, all of the\nsignal of years of schooling goes away.\nSo maybe what you're measuring is women's\ndecision about when to have children.\nIf you have children earlier,\nyou don't go to school as much.\nIf you have children later, you go to school more.\nMaybe it's some kind of measurement of delaying\ngratification or putting things off or planning.\nThe same trait is correlated to body mass\nindex, to obesity, and to walking pace.\nSo is this really intelligence as we think about\nit, or is it something else that manifests itself\ndifferently at different times in the past?\nObviously, a trait like years of schooling was\nnot itself a meaningful thing in the past.\nThe underlying things for it seem\nto have been under strong selection.\nWhatever in the genome predicts years of schooling\nseems to have been under strong selection.\nHow should we think about this?\nWhat's the actual thing\nthat's changing in the genome?\nThere are two things going on\nthat you need to think about.\nYears of schooling is connected to\nso many other things genetically.\nIf you look at the genetic predictor of\nyears of schooling—this trait has been\nmeasured in millions of people now—it's\ncorrelated to really surprising things.\nIt's correlated to the age at which women have\ntheir first kid. It's correlated to people's\nobesity. It's correlated to people's walking pace.\nIt's correlated to people's household wealth.\nIt's correlated to a variety of other\ntraits that seem quite different from it.\nIf you think you're actually measuring the\ngenetic prediction of intelligence, or actual\nstudiousness, you should think again because\nthere are many things that it's correlated to.\nThere seems to be some kind of general trait that\nyou could maybe think of as executive function\nor a propensity to defer gratification—I’m\njust waving my hands—that is under selection.\nIt pushes all these traits in the same direction\none way or the other, and at different times in\nthe past, it's advantageous or disadvantageous.\nWhen we found this signal of the genetic\npropensity to go to school for more years as it\nmanifests itself in white British people today,\nwe were incredulous. How could this be? Maybe\nthis is a problem. So we did a few tests to\ntry to figure out whether this was real.\nOne of the tests we did was that we looked\nfor a study where this measurement of the\nnumber of years of school was done not in\nEuropeans, but in Chinese people in China.\nWe looked at the effect size of many variants\nas they affected the number of years of school in\nChina, and we saw whether they had a correlation\nto the trajectory of those same genetic variants\nin Europeans over the last 10,000 years.\nThese are two parts of the world\nwhere the populations have been\nessentially completely disconnected.\nThere's no way by chance that the\ntrajectory in Europeans over the last 10,000\nyears would have anything to do with the effect\non years of schooling in China today.\nBut there's actually a huge statistical\ncorrelation, a five or six standard deviation\ncorrelation between the effect size of variants\non the number of years of school in\nChina today and the trajectory in Europe.\nIt’s just as strong, actually, as the effect size\nof variants in Europeans on years of school to\nthe trajectory in Europeans.\nWe just could not see a way\nthis could happen by chance.\nOnce we saw that, we felt quite\nconvinced that this was a real signal and that\nsomehow there has been natural selection to\nincrease the genetic changes that today manifest\nthemselves as predicting more years of schooling.\nJust to make sure I understood, you're\nlooking at this ancient DNA in Europe.\nYou're saying it seems to predict years\nof schooling for modern people in Europe,\nor at least selection on that ancient DNA seems to\npredict more years of schooling in modern Europe.\nYou also find that the same variants predict more\nyears of schooling for Chinese people in China.\nSo this is not just some weird artifact\nfrom the way these GWAS were done in Europe.\nThese parts of the genome seem to robustly predict\nthe kind of thing that actually leads to more\nyears of schooling, at least in people today.\nCorrect.\nJane Street is pretty secretive, but I did learn\nabout one internal mechanism which illustrates\nhow high trust and weird their culture is.\nResearchers aren't given compute allocations.\nInstead, Jane Streeters use an internal\ncurrency called “hive bucks” to bid for\ncompute in real-time auctions.\nEverybody can spend as many\nhive bucks as they want.\nBut your hive buck bid is\nmeant to represent the real dollar value\nof the experiment that you want to run.\nNow notably during the auction,\nanybody can change anybody else's bid.\nAnd after the auction, people\ncan even kill each other's jobs.\nPeople just trust each other to do this\nin a way that benefits the whole firm.\nAs a result, Jane Suite's allocations\nreflect a near real-time consensus\non the highest priority uses of compute.\nAs Axel, one of their ML engineers, put it:\n“I think Jane Street is like pretty\nbottom-up in terms of we have lots\nof different researchers who are all\ntraining their own models, sequence models,\nall sorts of other weird and wonderful things.”\nBy the way, with their new compute deal,\nthey've just added a six billion dollar hive\nbuck stimulus to their internal economy.\nJane Street is hiring researchers,\nengineers, and interns.\nGo to janestreet.com/dwarkesh to learn more.\nStepping back, I want to understand what\nthis tells us about what actually changed in\nour environments over the last 18,000 years.\nWe talked a little about what\nhappened after the Bronze Age.\nWe were talking about this during the collective\nintelligence part of the conversation.\nIt's surprising to me that things\nlike intelligence, or lack of\nschizophrenia—things that just seem robustly\ngood—were not maxed out before the Bronze Age.\nThe diversity among different populations was so\nbig that you have the European hunter-gatherers\nhaving three standard deviations less\npredicted value for what they would score\non an intelligence test if it existed.\nBut they were existing in the real\nworld in a place where intelligence matters.\nHow can it be that this was not a trait… You\njust look at the human body or any animal, and\nevolution has been acting on it so strongly to\nmake it functional for the things it needs to do.\nAnd this one thing, which seems so\nrelevant—especially to what human\nhunter-gatherers needed to do—doesn't seem\nto have been under that strong selection\nin the Mesolithic or Paleolithic eras?\nI think that's a great question.\nAs we talked about before,\nselection is very effective.\nIt can move the mean value of\ntraits within hundreds or thousands of\nyears in one direction or the other if\nthat's adaptive in a particular environment.\nSo you might wonder, isn't intelligence\ngood in all contexts and places in time?\nThere are a number of ways to think about that.\nFirst of all, we are speaking from the\npoint of view of a society which intensely\nvalues this particular trait, the ability to\nscore well on IQ tests or things like them,\nor to go to school for a long time.\nI think it's unprecedented in human\nhistory that we live in a time like this.\nIf you look at the Hebrew and Christian Bible,\nand you look at how much intelligence\nis valued, it's basically not at all.\nBut when the Bible was being written,\nespecially the Old Testament, that’s\nexactly when selection for intelligence is at\nthe highest point it's apparently ever been.\nExactly. But there it's about strength or courage\nor religiosity. Those are the values. If you read\nHomer or the texts of other religions,\nit's not intelligence. It's beauty and\nother things. This value system which has\na hyper-focus on smarts is not obviously a\ntrait value that's been common in the past.\nYou might think that in certain communities\nthere might be valuation of things that\nare more proximate to years of schooling.\nBut really broadly, it's not been\na high value in the population.\nObviously, the thing we care about is not direct\nperformance on an IQ test, especially in the past.\nThe thing I'm trying to understand\nbetter is intelligence more broadly.\nMaybe IQ-test intelligence is just not\nthat correlated with, \"Here is a new-world\nenvironment, go figure out how to process food\nthere and make shelter and everything else.\"\nYour colleagues like Joseph Henrich\nhave talked about how modern people\nunderestimate the difficulty of doing this\nkind of thing with a small band of people.\nMaybe that's not IQ-test intelligence,\nand that's why we don't see that strong\na selection effect on this thing.\nBut intuitively, regardless of\nthe value system, it just seems very\nvaluable to have this trait maxed out.\nI'm being very speculative. Let me give you two\nexamples of how I'm thinking about this, not that\nI’m a particularly good authority on these things.\nAs I mentioned, a lot of these traits,\nwhich are quite disparate, are\nhighly correlated to each other.\nObesity, years of schooling, walking pace,\nperformance on IQ tests, household wealth,\nall these crazy traits seem to be\ngoverned to a substantial extent by\na shared combination of genetic variants.\nLet's think about what this might mean.\nIn Iceland in the last hundred years, there's been\nselection against this combination of variants.\nOne possible interpretation is that\nit's basically selection for two\nways of investing in your children: having\nmany kids and not investing a lot in them,\nor having few kids and investing more in them.\nIf you invest in deferring having kids,\nhaving more wealth, having more resources,\nand putting more into each kid, you're going\nto have lower fertility and fewer kids.\nThat’s going to result in lower fertility,\nbut those kids might survive\nmore and do better in society.\nAlternatively, you can just have as many\nkids as you can and invest less in them.\nThey might individually have less good outcomes,\nbut in a time of plenty—which is potentially\nIceland in the 20th century—it might make sense\nto have more kids and invest less in them.\nThere's a toggle between having more kids and\ninvesting less in them, and having fewer kids\nand investing more in excelling in various ways.\nYou can imagine that at different times and in\ndifferent places… In ecology,\nthere are different ways.\nMammals often invest a lot with a pregnancy and a\nsmall number of children, whereas fish will spawn\nhuge numbers of offspring into the river,\nthe great majority of whom will be eaten.\nBut that is an effective way to produce\noffspring in certain conditions.\nSo there will be a toggle depending on\nthe environmental conditions back and\nforth between investing in large numbers of\noffspring with less investment, or smaller\nnumbers of offspring with more investment.\nMaybe we're just seeing that move back and\nforth over different places and times.\nSimilarly, for schizophrenia and bipolar\ndisease, how could this ever be advantageous?\nMaybe what we're seeing with these diseases is\na readout of some spectrum of traits that\nin some contexts might be advantageous.\nMaybe being anxious, imaginative, or neurotic\nmight be helpful in a shamanistic tradition or\na religious tradition which values people\nwho can have visions or be creative.\nMaybe these are subclinical versions of\nschizophrenia or bipolar disease that in\ncertain times may be advantageous and\nin other times may be disadvantageous.\nYou might just be seeing selection for different\ntypes of creativity or other thinking that can be\nvaluable in different contexts.\nI'm waving my hands here,\nbut my sense is that these complex traits\nhave not pushed in one direction because\nthere are advantages to both ends of the\nspectrum, and there are multidimensional\nimpacts of these different traits.\nJulian Jaynes has this famous theory\nin The Origin of Consciousness in\nthe Breakdown of the Bicameral Mind.\nI'm butchering this, but fundamentally, the\nway I understand it is that up until Homer,\nbasically everybody was schizophrenic.\nPeople genuinely thought that gods were real\npeople that you were communicating with.\nHis claim is that ancient texts seem\nto show people behaving in this way.\nYou're being asked to believe in visions.\nEven today, there's valuation in some religious\ncommunities in communicating with God,\nhaving visions, and having supernatural\ncommunions. So I just don't know. But I\nthink it's super interesting to ask the question\nof why certain traits are not always advantageous.\nFor schizophrenia and bipolar\ndisease, there is a sense in\nwhich most of the mutations are disadvantageous.\nWe can see that from the patterns of variation,\nwhere the variants that are\nrisk factors tend to be low\nfrequency and they tend to be small effects.\nSo another trait you find under selection is\nthe trend away from body fat since the\nagricultural revolution. Why is that?\nWhat you see is a reduction in the\ncombination of genetic mutations\nthat make you at risk for obesity, body mass\nindex, and similarly very correlated to it,\nhigher fat mass, higher waist-to-hip\nratio, and higher type 2 diabetes risk.\nThere is clear selection, by about a standard\ndeviation on the scale of modern variation\nfor these traits, reducing over the last\n10,000 years in this part of the world.\nWhat can be going on there?\nWhy wasn't there selection for\nthis combination of traits before?\nThere's a longstanding idea\nknown as the thrifty gene hypothesis.\nThe idea is that once you have hunter-gatherer\npopulations that move into a farming environment\nwhere there's plentiful food, there is no longer\na need to the same extent to be able to build\nup body fat to survive in times of stress,\nbecause there are more constant stores of food.\nAs a result, there will be natural selection\nagainst body fat once you\nmove into an agricultural\nenvironment and into periods of food plenty.\nMaybe what you're seeing is that this group\nof people in Europe and the Middle East over\nthe last 10,000 years has moved into a period\nof relatively more stable food, where building up\nstores of fat is not as advantageous, and there's\nbeen selection against this combination of traits.\nEuropeans are actually relatively better protected\ngenetically against type 2 diabetes than some\nother populations around the world, like African\nAmericans and Native Americans, that have perhaps\nnot been exposed to agriculture for as much time.\nSo you may be seeing the effect of more\nexposure to more stable food accessibility.\nThis is also another way in which\nthe data go against a common story.\nThe common story is that hunter-gatherers\nactually had much more stable diets because\nthey were more varied, and they weren't reliant\non a single cereal or crop for their calories.\nIf one game went away, they had\nother things they could scout for.\nThey could move locations more easily because\nthey weren't tied down to the land. So they were\nmore food-stable. But if there's been selection\nagainst storage of body fat, that suggests that\nas unstable and as common as famines might have\nbeen in agricultural societies, it's at least\nmore stable than what the hunter-gatherers had.\nThere's a timescale issue. You're absolutely\nright. As I understand it, I'm no anthropologist,\nwhen there's a hunt in traditional societies\nor communities that hunt, people will often\ngorge themselves, eat a huge amount, build up\na temporary store of fat, and then go multiple\ndays without eating meat until the next hunt.\nThere is this boom-and-bust access to\nhigh-value nutrition that is not true\nto the same extent in farming communities.\nOn the flip side, famines are something that\noccurs more commonly in agricultural societies,\nbut the timescale and the tempo of them is\nvery different from the hunting tempo.\nMaybe there's a famine every three years.\nIndeed, if you look at the bones of\nfarmers, at least in some communities,\nthere's more stress in them, maybe due to a\nfamine every three years or every five years.\nBut selection might not be acting\non that three-year time period.\nYour fat store from the latest\nhunt is not going to carry you\nthrough to the famine three years later.\nSurvival of famines is a different thing\nthan building up body fat to be\nable to survive two weeks later.\nA random question I have. You were\nmentioning that compared to these\nother things which matter much more for fitness\nin the ancestral environment—the immune system,\nespecially after the Bronze Age—all these other\nthings have mattered more than intelligence.\nThey've been under much more\nselective pressure than intelligence.\nThat makes you wonder whether there's much\nmore room at the top for intelligence.\nIf humans had been selected especially for\nintelligence, they could have been much smarter.\nThe reason that's relevant is that\nwe're currently building AI systems,\nwhich we're trying to make as smart as possible.\nIn fact, the only goal of the\ntraining process is intelligence.\nWe don't have to worry about at the same\ntime making their immune systems powerful—\nWe have lots of energy to spend on it.\nAnd at the same time making\nsure they're not schizophrenic.\nI guess we kind of do worry about that.\nBut if intelligence has not been the dominant\ntrait under selection for humans over the last 10,\n20, or 100,000 years, does that mean\nthere's more room at the top for this trait?\nI think there's more room at the\ntop for a lot of these traits.\nYou can move height extremely in one\ndirection, much more than it is today.\nYou can move any of these traits much\nmore extreme in the other direction.\nThere are probably very strong\nnegatives to doing that.\nYou're probably sacrificing other\nthings, and there are trade-offs.\nBut it's highly likely that if natural\nselection pushed any of these traits more\nin one direction than it is, the mean would move.\nSo all of this evolution since \"Out of Africa\" is\nacting on alleles that already existed in the\npool of human variation from that first group\nwe were talking about last time, on the order\nof 10,000 people, that exploded out of Africa.\nIs it surprising that across all these different\ntraits, from cognitive profiles to disease\nresistance to height, that one pool of people\ncontained so much latent variation that they\ncould supply enough stretchiness to accommodate\nall of these different traits you're studying now?\nThat's a rich question, and I think the human\npopulation has within it a tremendous amount\nof variation for complex traits.\nThere's a huge amount of variation\nthat affects height.\nThere's a huge amount of\nvariation that affects body mass index.\nIf you take all these mutations and set\nthem to the high-height variant, a person will be\nextremely tall, like as tall as a tall building.\nOf course, that will never happen.\nBut if you take all these variants\nthat affect schizophrenia risk and you point them\nall in the same direction, there will be extreme\nrisk or extreme protection for schizophrenia.\nFor complex traits, ones underpinned by many\nmutations, all the variation already exists to\nmove the population to a different adaptive set\npoint that's optimal in the environment it's in.\nIf you push the population into a new environment,\nwithin hundreds or thousands\nof years, the population can\nrapidly move to a new adaptive set point.\nThere are some unusual traits, like the ability\nto digest cow's milk or protection against sickle\ncell anemia, that require a single very important\nmutation that may not yet exist in the population.\nYou have to wait for the mutation to\noccur in some people.\nWhen the populations are\nrelatively small, only 10,000 people, you\nmight have to wait dozens or hundreds of\ngenerations for that mutation to arise.\nBut when the populations are large,\nthere's no mutation limit anymore.\nEvery mutation that can occur does occur.\nThere are eight billion people in the world.\nThere are maybe 30 new mutations every generation,\nso that's 240 billion new point\nmutations every generation.\nThere are only three billion DNA bases in the\ngenome, so every mutation that can occur does\noccur about 100 times every generation. We're\nnot mutation-limited anymore. The mutations\ncan arise again. They do arise again.\nBut when the population is only 10,000,\nyou sometimes have to wait dozens or hundreds\nof generations for the new mutation to occur.\nHow likely is it that the thing that changed\nwith the Bronze Age is just that the human\npopulation was big enough?\nBy 3000 BC, you go to\na population of 50 million-ish people.\nThe population is big enough, and the gene flow\nbetween different areas is high enough, such that\nthings which don't have an overwhelming selection\ncoefficient, which aren't overwhelmingly favored\nby evolution, are finally visible to selection.\nI think that's not likely to be true, but it's\nan extremely interesting thing to think about.\nAlready when population sizes are on the order\nof a million or so, every mutation that can occur\ndoes occur within a few generations.\nThat's well before the Bronze Age if\nyou take the population even of a place\nlike Europe, but also of other places.\nOr maybe it's at the dawn of the\nBronze Age or the farming period.\nThe question you're asking is whether,\nwhen the population is small, maybe natural\nselection doesn't work effectively.\nA common thing people think about with\nnatural selection, which is true, is that in small\npopulations selection doesn't work effectively.\nThat's because mutations bop around in\nfrequency from generation to generation a\nlot in a small population, just randomly.\nIf you have a population size of 1,000,\nmutations will bop around by a frequency\nof one over 1,000 every generation.\nIf the selection coefficient is less than\nthat, it will be drowned in the random bopping\naround of frequencies due to genetic drift.\nBut that is already for a population of 1,000.\nA 0.1% selection coefficient is very weak.\nWe're talking about 1% effects,\nand that's very strong.\nIt will work very well even\nin a population of size 1,000 or 10,000.\nIf you are talking about mutations of the\ntype that will start rising only in large\npopulations but not small populations,\nthose are selection coefficients on the\nscale of one over 10,000 or one over 100,000.\nThose will take 10,000 or 100,000 generations\nto rise in frequency, which is hundreds\nof thousands or millions of years.\nThat's not going to do anything over\nthe timescale we're talking about. There's just\na timescale issue. We're talking about strong,\nmeasurable selection coefficients on the\norder of half a percent or more in this study.\nAll of those are going to work in\nsmall populations or large populations.\nIt's not going to be affected\nby the population size.\nInteresting. You're saying that more generally,\nonce you hit a given threshold of population, the\ndominant factor is time span, not population size.\nCorrect. It's very interesting,\nand it's actually not widely understood.\nSpeaking of data contradicting what you\nmight have otherwise assumed, one of the\npapers you sent me beforehand, Mallick 2016,\nfound that there are no fixed differences between\nmodern and archaic humans 50,000 years ago.\nWe know this is the period in which the\nso-called cognitive revolution happened,\nand modernity started, and people are making art.\nDoes this suggest that nothing biological\nchanged to make modern humans modern?\nThe thing that happened was some cultural change?\nHow do we understand what this data tells us?\nRight. 100,000 to 50,000 years ago, there's\na quickening of the pace of change in culture.\nYou see the first extensive representational art,\nbead necklaces, drawings on the wall,\nand a rapidly increasing pace of\ninnovation in the types of tools that people use.\nThe thought might be that there would have been\nsome important genetic switch, a kind of important\ngenetic change that occurred in the population and\nswept to high frequency that everybody soon had.\nThat made it possible to do these things.\nMaybe some genes allowed people to have\ncomplex, representational language, for example.\nOne thing we did in 2016 in this paper\nby Swapan Mallick and colleagues was look\nacross the DNA for places that might\nbe expected to look like this, where\nnearly all people living today share a common\nancestor maybe 100,000 or 200,000 years ago.\nWe looked really hard, and right\nacross all the DNA we could look at,\nwe couldn't find anything more recent than\nfour or five hundred thousand years ago.\nThis is a crazy result because it looks\nlike there are no key selective sweeps\nthat have occurred in this period that\nare ancestral to everyone living today.\nWe talked before about no selective\nsweeps between Europeans and East Asians,\nbut there don't even seem to be any selective\nsweeps shared between all humans in this really\nimportant period when a lot of evidence\nin the material culture record appears.\nIt could be that there's biological\nadaptation in this period, but it's polygenic.\nThere are lots of mutations that all shift in\nthe same direction to help the population move\nto a new set point, but there's no key biological\nchange that rises to high frequency in this time.\nThis group 50,000 years ago, are\nthey the ancestors of everybody\nout of Africa or also some Africans?\nThis is 100,000 to 50,000 years ago.\nThis is the population that's\nancestral to West Africans,\nto most East Africans, to all non-Africans.\nThere are a couple of populations in Africa\nthat have substantial ancestry\ncoming from more divergent groups.\nFor example, Khoisan from Southern\nAfrica or Central African rainforest\nhunter-gatherers have substantial fractions\nof their ancestry from groups that diverged\nmaybe 200,000 years ago from the other lineages.\nBut all of these groups today are able to go to\ncollege and do everything everybody else does.\nThere is no evidence that there is any key\nmutation lacking in some groups\nthat is not present in the others.\nThe differences we see between different\ngroups of people, especially if this group\n50,000 to 100,000 years ago had a very small\npopulation size… I think last time we were\ndiscussing on the order of 10,000 people.\nSo almost everybody in the world,\nor the variance we see between different\nhumans today, was latent in this group.\nI get your point that if you just stack\nup different things across the genome,\nstacking them up really has a big effect.\nBut it's interesting that we have so many\ndifferent groups in the world today, and all that\ndiversity comes from a very small population size.\nA lot of us in human genetics think that\nour population contains within it the\nclay that's needed to make almost any trait.\nAnd that depending on environmental conditions\nor selection conditions, the mean value of\nthese traits will move in different directions.\nThere's an empirical question about\nhow much selection there's been\nin different human populations over time.\nOne of the things this new work we're\ninvolved in is showing is that at least in the\nlast 18,000 years in this part of the world,\nthere has been significant movement, at\nleast for a handful of important traits.\nWe looked at more than 500 traits.\nAbout 100 complex traits showed\nsignificant movement in a systematic\ndirection over this time period.\nIt really does seem that there is a response\nto the environments people are living in that\nhas occurred over this period, and that is\npotentially stronger than in previous periods.\nCrusoe has an amazing ML infra team that\nkeeps finding clever ways to squeeze more\nperformance out of their hardware.\nFor example, tokenization has become\na real bottleneck for agentic workloads.\nAgentic prompts are often extremely long.\nThey tend to have high KV cache hit rates,\nwhich shrink the GPU's pre-fill work.\nThis means that the tokenization step,\nwhich is traditionally sequential,\nis a much larger fraction of time to first token.\nTo solve this, Crusoe built fastokens,\nan open source Rust-based tokenizer which\nparallelizes things in order to take\nadvantage of all the cores on modern CPUs.\nCrusoe had to get creative here because\nthe naive approach doesn't work.\nFor example, for pre-tokenization,\nyou can't just split your text into chunks\nand run regex because you'd end up with\nissues whenever a word straddled the split.\nCrusoe solved this by giving each thread\nan authority zone plus the ability to\nread one kilobyte past its own edges.\nThis one kilobyte buffer guarantees that you\nwon't misprocess a token, and the authority zone\nguarantees that you won't end up with duplicates.\nNo cross-thread coordination required.\nCrusoe combined this optimization with a\nhandful of other smart tweaks in order to\nget up to 40% faster time-to-first\ntoken on real production workloads.\nTo learn more, go to crusoe.ai/dwarkesh\nWe were talking earlier about how there\nare no fixed differences between humans\n30,000 years ago and humans today.\nSo if there's no genetic basis for the kind of\nthing that allowed humans to have more symbolic\nrepresentation, have farming, et cetera—I think\nI asked you this question last time we talked,\nbut especially with this context—why no farming\nbefore the Ice Age? Genetically we were there.\nThat is such an interesting question.\nGenetically we're there. The common\nancestral population has all of the\ningredients for farming 50,000 years ago.\nThese people are distributed into different\nparts of the world: the Americas 15,000 years\nago or whatever it is, New Guinea 40,000\nyears ago, East Asia, Europe, West Africa.\nNo farming developed before\n11,000 or 12,000 years ago.\nIt only developed in the last 12,000\nyears, the period known as the Holocene,\nwhich is the end of the Ice Age.\nIf you talk to climate scientists\nand archaeologists—I keep asking people\nthis question every time I meet someone\nwho's an expert in this—how can it be\nthat farming develops in all these places?\nAre we really living in such an unusual time?\nPeople tell me, indeed, we're living in a very\nunusual time on a scale of two million years.\nThat is, 12,000 years ago we switched into this\nperiod of not just warmth, but climate stability.\nIt's hard to believe that we're\nliving in such a special time.\nBut if you look at data from the bottoms\nof ponds where you can measure the fluctuations of\ntemperatures using isotopic signatures, apparently\nwe're in a period where it's fluctuating a\nlot less year to year, 10 years to 10 years,\nand 100 years to 100 years.\nIt's a period of relative\nstability that we are miraculously living in.\nWhen this period of relative stability happens,\nit follows that multiple groups independently\nturn to agriculture, even though they\nall have the same genetic complement that arose\n50,000, 100,000, 200,000, 300,000 years ago.\nIt's a crazy observation that people\njust accept, but it's unbelievable.\nOh, so you increased the range there.\nYou said 100,000, 200,000, 300,000 years ago.\nBased on the genetic differences between modern\npeople and people from 300,000 years ago.\nDo you basically think they're\nmodern 300,000 years ago?\nI don't know. This is actively what I'm\nthinking about all the time right now.\nThere's a big transformation in terms of the\nculture of humans 300,000 or 400,000 years ago:\nthis invention of Levallois technology, the\nability to make stone tools out of cores.\nThe Middle Stone Age Revolution, or the\nMiddle Paleolithic Revolution depending\non what you call it in Africa or Eurasia,\nis a new way of making stone tools that's\nshared by Neanderthals and by modern humans,\nbut is not shared in East or South Asia.\nIt's a big change, and it presumably\ninvolves a cognitive change in order\nto make this sort of technology.\nThen there's a further change to\nthe Upper Paleolithic Later Stone Age, maybe\n100,000 to 50,000 years ago, when there's a\nsecond transition with a new type of tool making,\nbut it’s not as revolutionary as the earlier one.\nSo when the cognitive leap happens is unclear.\nThe diversification of the lineages leading to\npeople living today, like Khoisan Southern\nAfricans and rainforest hunter-gatherers,\nall occurs more on the timescale\nof 300,000 or 200,000 years.\nAll of these people are capable of\ngoing to college and doing everything.\nSo it's not obvious that the cognitive toolkit,\nthe behavioral toolkit, and the genetic abilities\nwere not all in place 200,000 or 300,000 years\nago, and that even Neanderthals had them.\nIt’s not obvious that this was not the case.\nI just don't know. You distribute these people\ndescended from this diversification that\nhappened 200,000 or 300,000 years ago to\ndifferent parts of the world, and then\nafter 12,000 years ago, you start having\nagriculture popping up in different places.\nIt's an outstanding mystery of human history.\nI find it unbelievable that we live in a\ntime period that climatologically is so\nunique on a scale of two million years,\nbut my colleagues tell me it's true.\nThe climate thing seems surprising given there\nare so many different environments in which\nagriculture was independently developed.\nI understand that across environments\nthe variance could have gone down.\nIf it had only happened in one place\nat one time, I could have bought that explanation.\nBut the fact that they're making maize in the New\nWorld and they've got cereals in the Old World in\nvery different environments makes it surprising.\nIt's very, very surprising. We accept\nit, but it's a crazy observation that\nmost normal people don't realize.\nThe thing that basically everybody\naccepts is that the common ancestral\npopulation of almost everybody in the world,\nexcept for rainforest hunter-gatherers\nand Khoisan, is around 70,000 years ago.\nEverybody accepts that these people all have in\nplace the cognitive, behavioral, and intellectual\ningredients that are necessary for the farming\nrevolution and building state societies.\nBecause when these descendants get distributed\nto West Africa, East Africa, the Americas,\nEurope, South Asia, East Asia, New Guinea,\nand so on, their descendants all do this.\nThey do it independently, semi-independently,\nor demonstrably completely independently in\nall these different parts of the world.\nThe cognitive resources for doing this must\nhave all been in place, but it's a very long fuse.\nIt delays for 40,000 or 60,000 years in all these\ndifferent places after the common ancestral\npopulation splits up, and then ignites into\nagriculture and all these other things after\nthat point. It's a crazy claim. Then you could\nargue about whether the actual fuse is 300,000\nyears, from when Neanderthals separated and\nfrom when different lineages of extant modern\nhumans separate, and that's also plausible.\nIt's a crazy set of things that\nwe're being asked to believe.\nIs it possible that agriculture existed, but\nyou didn't have modern metallurgy or whatever\nit was that allowed populations to explode\nstarting in 5000 BC with the Bronze Age?\nPopulation-wise, it doesn't seem\nlike much is happening from 10,000\nBC to 5000 BC in the early Neolithic.\nIs it possible that they had farming\nbut they didn't have copper or tin, which\nyou needed to go to the Middle East for,\nto develop a civilization that could make use\nof bronze at a large scale, and so they just\ndisappeared from the historical record?\nI think we would see their archaeology.\nThere are extraordinary developments in\nthe Americas which are entirely Stone Age.\nYou would see them today if\nthey had completely vanished?\nOh, yeah. We should go for a trip\nto Teotihuacán in Mexico. It's so\nimpressive. When I went there when I was 20,\nit was totally as impressive as ancient Egypt.\nIt's huge. It's massive. It's without metal.\nIt's even more impressive because it's not\nonly without metal, but without animals\nand without wheels, which is crazy.\nThe marble is just hauled without wheels.\nRight. Take any person who has an old\nworld superiority and take them to these\nplaces, and they will not have it anymore.\nIt's just extraordinary what's in these places.\nThese are people who separated 20,000 years ago at\nleast from the ancestors of East Asians and 40,000\nyears ago from the ancestors of West Eurasians.\nThey just had the same biological and cultural\nshared toolkit from then, but there's a long\nfuse delay until all this stuff happens.\nIt's an amazing thing, and we don't question it.\nWhat are other questions you are\neither investigating right now\nor want to investigate, these kinds of\nbig picture questions of human history?\nI'm perplexed. I don't know if we talked about\nit before, but I remain very confused about the\nrelationships between archaic and modern humans.\nWe have genome sequences now from archaic humans\nwho lived in Europe, West Eurasia, and\nCentral Eurasia, and the Neanderthals.\nWe have archaic sequences from these\nenigmatic Denisovans, who we now have\na skeleton for since we last talked.\nThere's now a skull that's\nbeen shown to be a Denisovan.\nWe have data from lots of modern humans,\nand there are really big mysteries about\nthe relationships amongst these groups.\nGenetically, the Denisovans and\nthe Neanderthals are sisters.\nThey descend from a common ancestral\npopulation 500,000 or 600,000 years ago.\nThat group descends 700,000 or 800,000 years\nago from the common ancestors of modern humans.\nGenetically, the whole genome data says that\nNeanderthals and Denisovans are archaic humans\nfrom a common ancestral archaic population.\nBut there are so many things shared between\nNeanderthals and modern humans that\ndon't seem to be shared with East Asians.\nThey both share Middle Stone Age\nstone tools, Levallois technology,\nthis cognitively unique way of making\nstone tools that wasn't used in East Asia.\nThey both have the same mitochondrial\nDNA and Y chromosome sequence.\nThe Y chromosome sequence of Neanderthals\nand the mitochondrial DNA of Neanderthals,\nis actually modern human that came through\ninterbreeding 200,000 or 300,000 years\nago and then shot up to 100% frequency.\nNeanderthals and modern humans are both the\nproduct of mixture events that happened between\narchaic and modern humans 300,000 or 200,000\nyears ago, demonstrably through patterns\nof variation in ancient and modern DNA.\nIt feels that there's something shared\nbetween Neanderthals and modern humans\nthat's not shared with Denisovans, even\nthough the vote of the whole genome says\nthat Denisovans and Neanderthals are related.\nOne wonders whether there's something connecting\nNeanderthals and modern humans that's different\nfrom Denisovans, even though genome-wide,\nDenisovans and Neanderthals cluster.\nI'm thinking about that all the time now.\nConnecting them would be interbreeding\nevents or being in the same place\nat the same time that we missed?\nThere's a known interbreeding event\nfrom the lineage leading to modern humans into\nNeanderthals, but it's supposed to be only 5%.\nI'm interested in the possibility that that 5% is\nactually a sign of something much more impactful,\nthat somehow Neanderthals are in some\nsense deeply modern in some ways,\nand even though they get swamped by\narchaic genes, they actually have\nmore of a modern impact than one would think.\nThe Middle Stone Age and Middle Paleolithic\nRevolution that they share with modern humans\nis more fundamentally a part of who they are,\nin some sense, than we think.\nInteresting. Sorry, when was\nthis interbreeding event?\n300,000 to 200,000 years ago.\nSo the common ancestor between Neanderthals\nand most humans alive today is potentially\nmore recent than the common ancestor\nbetween all humans alive today.\nOh, for sure.\nWhich is crazy.\nWell, the divergence to all the archaic humans,\nincluding Denisovans, is within human variation.\nWait, what?\nYes. The average time\nto the common ancestor of any two human\ngenes is one or two million years ago.\nIf you look at the copy of chromosome 3 you get\nfrom your mother and the copy of chromosome 3 you\nget from your father, the typical time they share\na common ancestor is one or two million years ago.\nThat's before the split from\nNeanderthals and Denisovans.\nSo there are many places in your DNA where\nyou're more closely related to a Neanderthal\non your mother's side than you are to your father.\nI'm sure there's a simple explanation, but how?\nIt's the same reason that if you\nhave a sister, in some places in\nyour DNA you're more closely related to her\nthan you are to me because you share a parent.\nBut in other places you're more closely related to\nme than you are to your sister because you happen\nnot to share the same DNA from your parents.\nIt's just that the DNA we get from our common\nancestral population was already\nquite variable 500,000 years ago,\n700,000 years ago, a million years ago, and some\nof us descend from some of those ancestors and\nothers descend from other of those ancestors.\nNeanderthals split from our lineage really close\nin time on human evolutionary timescales, such\nthat in some places in our DNA we're more closely\nrelated to Neanderthals than to each other.\nInteresting. What are the other big questions?\nThat's the main thing I'm\nthinking about a lot these days.\nI continue to be very obsessed with\nquestions about the spread of human\npopulations around the world and trying\nto reconstruct that with ancient DNA.\nAfter the recording ended, David started\nspontaneously explaining a new theory he's working\non about Neanderthal genetics on a whiteboard in\nthe room, which I ended up capturing on my iPhone.\nThe thing I've been thinking about a lot\nrecently is the possibility that maybe\nwe're not thinking in the right way about the\nrelationship between archaic and modern humans.\nThe standard model is one where Denisovans—these\narchaic humans that were found from ancient\nDNA—and Neanderthals descend from a common\nancestral population 500,000 or 600,000 years ago,\nand these two separate earlier, maybe 700,000 to\n800,000 years ago, from the ancestors of modern\nhumans, people like us.\nThat's the big result\nof a lot of studies since 2010.\nBut there's also evidence of an\ninterbreeding event that happened maybe 200,000 to\n300,000 years ago that resulted in modern humans\ncontributing DNA to the ancestors of Neanderthals.\nSo maybe 5% of the DNA of Neanderthals comes\nfrom this interbreeding event, and\na lot of studies have shown this.\nI'm very interested in this because, from the\narchaeological record, Neanderthals and modern\nhumans look quite similar to each other, much more\nsimilar to each other than a lot of them do to\nDenisovans, these archaic humans in East Asia.\nFor a lot of history, people have thought that\nNeanderthals are our sisters.\nBut in 2010, the sequencing of\nthe Denisovan genome made it very clear\nthat on average, Denisovans are closer to\nNeanderthals than to modern humans.\nThis was a very confusing result.\nMost people now think that Neanderthals and\nDenisovans descend from a common ancestral\npopulation that separated earlier\nfrom the ancestors of modern humans.\nI'm interested in the possibility that\nthe right way to think about Neanderthals\nis actually as somehow culturally modern humans,\neven though genetically they're mostly Denisovans.\nThe model I'm thinking about is motivated\nby this archaeological phenomenon known\nas the Middle Stone Age Revolution.\nIf this is Africa and this is Europe, we know\nthat the new way of making stone tools—with cores\nthat were very carefully mined far away from the\nlocations they were used, made out of high-quality\nstone like flint—starts being used 300,000 or\n400,000 years ago, first in the Caucasus,\nplaces like Georgia today, or East Africa.\nThis way of making stone\ntools is quite revolutionary.\nIt is known in Europe as the Middle\nPaleolithic and in Africa as the\nMiddle Stone Age, and is associated with much\nmore widespread use of fire and moving stone\naround at much further distances than before.\nI'm interested in the idea that this is something\nshared between modern humans and Neanderthals.\nThere's somehow some shared cultural feature\nthat's absent in East Asia, and that\nmight have a relationship in the genetic\ndata and is somehow related to this 5% DNA.\nThe idea I’m interested in is the possibility\nthat there is a population here that invents\nthe Middle Stone Age and the Middle Paleolithic,\nsometimes called Levallois technology, and that\npeople from this population expand into Europe and\nmix with the local archaic humans who are there.\nThat is what this 5% interbreeding event is.\nIt happens 200,000 to 300,000 years ago.\nIt produces a group that, as it expands across\nthis landscape in Europe, mostly picks up the\nlocal DNA and becomes mostly archaic genetically,\nbut retains its modern human culture, the way of\nmaking stone tools and some of its traditions.\nOne of the things that's super interesting\nabout this is that if you actually look at\nthe genetics, across the whole genome,\nNeanderthals and Denisovans cluster.\nBut if you look at the mitochondrial DNA—which\nhumans and Neanderthals get from their\nmoms—Neanderthals and modern humans cluster.\nIf you look at the mitochondrial DNA,\nDenisovans and modern humans share an\nancestor well more than 700,000 or 800,000\nyears ago, as you'd expect from the history.\nIf you look at the Y chromosome that you get\nfrom your dad, Denisovans and modern humans\nshare an ancestor more than 700,000 or 800,000\nyears ago, which is consistent with this history.\nBut if you look at the Neanderthal mitochondrial\nDNA, it's only 300,000 to 450,000 years.\nIf you look at the Y chromosome,\nit's only 300,000 to 450,000 years.\nWhat the current genetic work is asking us to\nbelieve is that even though this is only 5% of the\nwhole genome, it introduces mitochondrial DNA and\nY chromosomes, and they jump up to 100% frequency.\nIt's kind of a crazy claim because the probability\nof this occurring by chance is low,\nmaybe 5% times 5%, a very small number.\nIt's what we actually all believe,\nbut it's a very surprising event.\nSomehow it's accreted to all the findings in the\nliterature so that we make ourselves believe this,\nbut it seems unlikely on first principles\nthat somehow only 5% would introduce both\nthe Y chromosome and mitochondrial DNA.\nAnd it really does look like this.\nThere's amazing data from a site in Spain\nthat's 300,000 to 400,000 years old,\ncalled Sima de los Huesos.\nThey have a nuclear genome\nthat looks Neanderthal-like for most of\nthe genome, but their mitochondrial DNA\nand Y chromosome are Denisovan-like.\nSo it really looks like there was a\npopulation related to modern humans that pushed\ninto this Sima de los Huesos-like population,\ndisplaced its mitochondrial DNA and Y\nchromosome, but kept the rest of its genome.\nIt really looks like something like this happened.\nThe idea I'm playing with—and probably it's wrong,\nwho knows—is that there's a landscape…\nThis is Europe and you can break up\ninto a hundred or so demes, little areas.\nModern humans get introduced at the bottom\nright corner, in the Middle East or\nsomewhere, and they spread into Europe.\nAs this population spreads, there's a\nwave front of expansion, and they're\ninteracting with the local archaic humans.\nThe theory from simulations and studies of\ndifferent species like mammals and birds\nshows that even if there's a small amount\nof interbreeding—when there's an invasion or\na movement of expansion of one group into the\nterritory occupied by another—there's\nmassive introgression of local genes.\nThe pioneers at the wave front will sometimes\ninterbreed with the local population.\nThere are so many of them around that their\nDNA will get swamped by the local group,\nso by the time they make it to the\nother side, they're largely local.\nMaybe this is what we're seeing.\nYou have a modern human population\nthat's matrilineal, for example, where\nthe transmission of making stone tools\nthis way is happening from mother to child.\nThat's why they're retaining their mitochondrial\nDNA, but by the time they get to the other\nend of Europe, they're mostly local archaic.\nYou end up with a 95% population replacement.\nThis would explain why the mitochondrial DNA\nis shared between Neanderthals and\nmodern humans, and it would also\nexplain why the mixture proportion is only 5%.\nThe really interesting thing is that there's\nother evidence from studies of modern humans\nshowing that modern humans are also admixed.\nThe right way to think about this is\nthat modern humans are a mixture of\ntwo groups that diverged maybe 1.5 million\nyears ago, and they come together 200,000\nto 300,000 years ago with maybe 20%\nancestry from this archaic African group\nand 80% ancestry from this early modern lineage.\nAnd that same group then mixes with Neanderthals,\nand it's 5% modern and 95% local.\nSo you actually have this key\npopulation that makes the Middle\nStone Age or Levallois technology.\nIt appears and expands in all directions—into\nEurope and into Africa 200,000 to 300,000 years\nago—bringing this technology, new ideas,\nand perhaps some genetic adaptations.\nIt expands into archaic humans in Europe,\nmixes with the local population, and gets\n95% replaced but still retains its cultural\nfeatures and maybe some genetic features.\nIt expands in Africa too, but here it's\nnot 95% replaced, it's only 20% replaced.\nProbably the reason is that this\ngroup is much more diverged.\nIt’s 1.5 million years diverged\nrather than 700,000 to 800,000 years.\nAs a result, there are many more genetic\nincompatibilities and barriers to gene flow.\nBut there's still a lot of mixing, maybe 20%, and\nwe have evidence that this is a big mixture event.\nSo what you're actually seeing is a modern human\nexpansion both into Europe and into Africa.\nIn one place, it forms Neanderthals.\nIn another, it forms the\nancestors of everybody living today.\nBut all of these groups are descended from\nthis key revolutionary event that happens here.\nWe often talk about the revolutionary events 50\nto 100,000 years ago, the more symbolic behavior\nand so on, that first appears in Africa and\nthe Middle East and spreads beyond.\nBut there's also this earlier event,\nand it is contemporaneous with the breakup\nof all the different groups in Africa today,\nthe Khoisan Southern Africans and the\nCentral African rainforest hunter-gatherers.\nOne wonders whether this is an\nequally important formative event.\nIf that's true, it makes you think of\nNeanderthals as somehow our cousins.\nTheir shared Y chromosome,\ntheir shared mitochondrial DNA,\nthey share the formation of this 200,000 or\n300,000-year-old event, and their shared toolkit.\nEven though the genome is telling us they're\ncousins of Denisovans, the correct way to think\nabout them may, in an important sense,\nbe as close cousins of modern humans.\nI have so many questions.\nDo you have 15 more minutes?\nFirst of all, what is going on with this group\nof archaic Africans 1.5 million years ago?\nWhere in Africa are they, and what happens\nto the portion of them that don't form\nmodern humans? Do they survive?\nThis is not from ancient DNA,\nbut from analysis of modern DNA from different\npeople, mostly in Africa, but also non-Africans.\nIn multiple studies—there's at least\nthree, maybe four or five studies that\nI know about—they have looked at the patterns\nof variation in people today and say the data\nin modern people today, including in Africans,\nis not consistent with a homogeneous population.\nIt looks like a population that split well\nmore than a million years ago into multiple\ngroups—at least two, but maybe many—and then\ncame together a few hundred thousand years ago.\nThe papers have different models that they fit,\nbut they all have this feature of a split-up more\nthan a million years ago, and then on the\norder of a few hundred thousand years ago,\na coming together and a remixture event forming\nthe ancestors of anatomically modern humans.\nThis includes the Khoisan\nand whatever other groups?\nYes. All of these groups have this,\nmaybe in slightly different proportions.\nSo you ask, where are these people\nliving? Who knows. In this scenario,\nthe 80% is coming from the Caucasus or Northeast\nAfrica, where the Middle Stone Age forms.\nIt's from this population that\nthe Middle Stone Age comes.\nThey mix with local groups, and who knows where\nthey are: Southern Africa, Western Africa,\nCentral Africa, Eastern Africa.\nWe don't have any ancient DNA,\nbut this is a very rich environment.\nPeople have been living there for seven\nmillion years at least, and there would have\nbeen different groups of people everywhere.\nProbably it's not just two\ngroups, it's probably more.\nThe important theme here is\nthere's evidence of substructure\nthat's well more than a million years old.\nThis place would have been a landscape full\nof archaic humans that would have been differently\nrelated to these expanding people and would have\nadmixed with them when they came through.\nSo with the Neanderthals, the first\ntime around 300,000 years ago, our\nancestors share culture with them.\nThey share the Middle Stone Age technology,\nbut they don't replace the population.\nThe technology spreads through culture, basically.\nIt spreads through genes too. If you look at\nYamnaya in India, there's almost no Yamnaya\nancestry in India. It's just diluted down.\nAs Yamnaya expanded into Central Asia and into\nEurope, it makes the Corded Ware. There's a 25%\ndilution. It expands back across Central Asia.\nIt goes through the Hindu Kush and\ngets into northern South Asia.\nIt admixes more with local people.\nToday, the most Yamnaya ancestry\nyou see in India is 20% or 10%.\nMost people have less than 10% or 5%.\nThere's been a lot of mixture on the way,\nbut it is the tracer dye.\nIt tracks Indo-European languages,\nand important aspects of Indo-European\nculture are coming through Yamnaya.\nSo if you know where to look, that tracer\ndye is only 10%, 5%, or 2% in some groups.\nBut it's the languages people speak,\nand important shared cultural elements,\nthat connect them to people on the other\nside of the Indo-European-speaking world.\nSo this 5%, you shouldn't sneeze at it.\nThat's tracing something important in this model.\nI understand that if things are transmitted\nmore through women… Sorry, let me back up.\nI don't understand why the maternal\nmitochondrial DNA and the Y chromosome\nwould be especially privileged as the\nspreading is happening. Can you explain that?\nThe reason I'm talking about these matrilineal\nor patrilineal expansions is that I'm really\ntroubled, and have been troubled for many\nyears—especially in the last three or four\nyears—by the fact that the mitochondrial\nDNA and Y chromosome cluster Neanderthals\nand modern humans, but the rest of the\ngenome clusters Neanderthals and Denisovans.\nThis is a crazy result that is\nnot seen in any other species.\nI'm very interested in\npatterns that would explain it.\nIf you assume that there was a matrilineal\nor a patrilineal expansion—it could be\neither—then modern humans, when they were\nexpanding across the landscape of Europe,\nretained their identity along one of the lines.\nIf it's matrilineal, when they incorporate\na male from the local community,\nhe's brought into the community, and the kids\nare raised based on the culture of the mothers.\nIf it's a patrilineal expansion and they\nincorporate a female from the community,\nshe's raised with the culture of the fathers.\nIf that happens, it guarantees that one of\nthese two parts of the genome looks like it\ndoes, because it's a modern human expansion.\nIf it's patrilineal, it will\nretain the Y chromosome.\nIf it's matrilineal, it will\nretain the mitochondrial DNA.\nSo it will solve one of your two problems.\nBut not both.\nIt won't solve the other one, so\nyou need to solve the other one.\nYou can solve it either by natural\nselection or by social selection.\nBy the way, patrilineality and matrilineality are\nthe rule, not the exception, in human communities.\nUsually communities have continuity\nalong the male or the female line.\nUsually it’s patrilineality,\nsometimes it’s matrilineality.\nYou can also have phenomena like social selection.\nIt could be that once you have kids of someone\nwhose father, for example, is from the outside\ncommunity… Usually in most communities,\nfemales all reproduce. That's typical\ntoday. Usually women have kids if they can.\nBut men in traditional societies are actually\nvery variable in their reproductive success.\nA large fraction of men never have kids.\nThen there's a subset of men who have\nmany kids with many women.\nThere's competition among men for kids.\nIn this context, where males are competing for\naccess to females, female mate choice begins to\nbe an important process.\nYou have a phenomenon\nwhere if your dad is an archaic male, it\ncould be the case that you're not going to\nbe as successful in the competition for local\nfemales as if your dad is a non-archaic male.\nSome simple social phenomenon\nlike that could explain the data,\nand we actually see this in human society.\nFor example, if I remember right, in Central\nAfrican rainforest hunter-gatherers,\nthere's different treatment of boys\nand girls depending on whether their\ndad or mom is one group or the other.\nI guess I don't understand how\nthe maternal… The group spreads,\nand it gets to the next front. They have kids.\nFrom the humans that have just entered, the kids\nwill have the mitochondrial DNA from the humans.\nBut from the existing people, they will have the\nmitochondrial DNA of the archaic humans.\nWhy are the people with the archaic\nmitochondrial DNA not surviving?\nIt's a question. There are multiple\npossible explanations, but it's much easier\nto explain that than both the mitochondrial\nDNA and the Y chromosome.\nOne possibility is that the\nmitochondrial DNA was less biologically fit.\nAnother possibility is that there's social\ndiscrimination against people based on whether\ntheir parents are archaic or not, which is not\nat all surprising in a human context.\nIt's the weakest link in this argument.\nThis argument's probably wrong, but I'm\njust telling you what I'm thinking about.\nOkay, the Neanderthals. So 300,000 years\nago our lineage interacts with them,\nbut mostly their lineage survives, and\nthere's cultural and genetic diffusion.\nAnd then is it 70,000 years\nago that we interact again?\nYes.\nAnd they don't survive.\nThe genetic ancestry doesn't survive.\nPresumably there was also other contact\nbetween 300,000 years ago and 70,000 years ago.\nProbably. But these are the\nones we are detecting currently.\nIs it just contingent that one time\nthere's this kind of diffusion where\nmost of the archaic genome survives, and\nthe other time it's total replacement?\nThis is not at all surprising given the context.\nIf you think about this model, this is 700,000 or\n800,000 years ago. This is 300,000 years\nago. So this is 400,000 years separated.\nYou talked about the Bhatia paper with me earlier.\nThat's two populations 70,000 years separated.\nThere are no biological incompatibilities\nbetween West Africans and Europeans.\nThere's no natural selection against\nbiological incompatibilities.\nWe know when Neanderthals and modern\nhumans met and mixed, there were biological\nincompatibilities. That was 700,000 years\nago. As populations become further apart,\nbiological incompatibilities rapidly develop,\nprobably as the square of the separation distance,\nbecause you need pairs of interacting genes.\nHere, it would have been maybe only 400,000 years\nseparated between this lineage and that lineage.\nBut here, it's 1.2 million years. That's a lot.\nThese are at the edge of not being able to produce\nchildren. These are quite different humans. These\nare actually three times closer than these.\nIf you look at mixtures of humans today,\nthere are mixtures in Southern Africa\nof people who are half this distance.\nIf you look at Khoisan and Bantu people\nmixing in Southern Africa, like the Xhosa,\nwhich is the population of Nelson Mandela,\nthese are groups that are separated by\nalmost 200,000 years, which is half of this.\nTotally compatible. What you're seeing is a\ngroup that's actually completely permeable\ngenetically, or nearly completely permeable.\nThis other one almost certainly has\nsubstantial biological incompatibilities.\nBecause 200,000 or 300,000 years later, we\nsee interbreeding between Neanderthals and\nmodern humans, or between Denisovans and\nmodern humans, and there's clear evidence\nof incompatibility at that point.\nBut this would be even bigger.\nWhat you would expect to see is that as this\ngroup spread, they would be moving into a\nterritory full of archaic humans.\nThere would be some interbreeding,\nbut the kids would not be very fit. They would\ndie off. There would be a lot of infertility.\nThe barriers to gene flow and to\ninterbreeding would be greater.\nTo me, it's not at all surprising that as\nthis group moves into Eurasia, you have\nEurasian archaics—the ancestors of\nDenisovans—who are only 400,000 years\ndiverged from these people over here.\nAnd then you have African archaics,\nand these are 1.2 million years diverged.\nThey just don't interbreed as much,\nand you don't get as much gene flow.\nBut the key thing is the timing. It’s the\nsame time. It really feels like the signature\nof an explosion of people from one place,\ninteracting with people here and\ninteracting with people there.\nIt's the same cultural or technological\nrevolution impacting this place and that place,\nand creating populations that are impacted by\nthis cultural revolution, which we know is the\ncase because they share the same toolkit.\nSome people argue that Levallois\ntechnology is independently invented.\nBut it's very similar, and this model would\nbe a way that it could have the same origin.\nSo there's a culturally shared thread,\nthis shared toolkit.\nThere's a mitochondrial\nDNA and Y chromosome thread.\nAnd then there is a shared timing thread,\nwhich is they both form by mixture.\nBecause otherwise you'd have to believe\nthat Neanderthals independently\ndeveloped Stone Age tools.\nYes, which is not inconceivable. But it's a little\nbit like believing that farming independently\ndeveloped in multiple parts of the world.\nRight. But it did.\nIt did. So as I said, this is probably wrong.\nI'm trying to tell you that we don't really know\nthe world we live in. This is not obviously\nwrong. In fact, to me, this is much more\nplausible than the model we currently write down.\nIt's probably wrong, but it's much more plausible.\nIt explains many more things,\nand it's no more complicated.\nInteresting. Do you want to recapitulate\nthe thing you were saying about the\nanalogy to Ptolemy and the epicycles?\nI thought that was quite interesting.\nI think the model that we've put together\ncollectively about the relationships between\narchaic and modern humans has accreted over time.\nThere was this idea that modern humans are\ndistinct and that Neanderthals and\nDenisovans are sisters of each other.\nOver time, we detected additional mixture\nevents, like this modern human into Neanderthal,\nand then these other ones I didn't even talk\nabout, like a super-divergent lineage going\ninto Denisovans and all this other stuff.\nWe still say, \"Oh, the whole genome says\nNeanderthals and Denisovans are\nsisters, so that's the truth.\"\nWe've patched it all together\nand gotten it all to work.\nYou look at the mitochondrial DNA and the Y\nchromosome, and they have this odd pattern,\nand it's improbable, but we can get that\nto work if we invoke natural selection,\nthings like this. You patch it all together. It\nreminds one of what happened in the ancient world,\nwhere there was this idea that the sun revolves\naround the Earth, but it doesn't quite explain\nthe movements of the planets properly.\nIn order to get the movements of the planets\nto work right, Ptolemy and the astronomers made up\nthese epicycles, these special extra rotations and\nmovements to make everything work about right.\nIt was such a convoluted model.\nWhen Copernicus and colleagues suggested instead\nthat everything is revolving around the sun,\nit simplified things ever so much.\nWhat was happening is that as astronomical\ninformation accumulated, it kept being\ncontradictory to the standard model, but it could\nbe made to work by proposing another complication\nand another complication and another complication.\nThis is not as fantastic as proposing that\neverything revolves around the sun rather\nthan the Earth, but it is much simpler.\nAnd it actually explains many things.\nWhat is counterintuitive or unexpected or\nhard to accept about this alternative model?\nWhat is the hesitation that\npeople have for adopting this?\nI don't know. Nobody's thinking\nabout this model right now.\nIt just seems obviously a\nvery natural model to me.\nThe reason I ask is that Aristarchus, the\nancient Greek, had the heliocentric theory\nbecause he had deduced how far the Earth\nis from the sun and noticed other things.\nBut it was not adopted, because his fellow\nAthenians were like, \"Look, if we believe\nthat the Earth revolves around the sun, for it to\nbe the case that we don't see relative movement\nof the stars to the Earth, the only possible\nexplanation is that the stars are so far away that\nit is just incomprehensible and implausible.\"\nSo the heliocentric theory was dismissed.\nWhat I'm trying to ask is, what is the equivalent\nhere of \"for this to work, the stars have to\nbe so far away that it's inconceivable,\"\nwhere actually the stars are so far away?\nMaybe we should adopt the implausible\nimplication that this theory gives us.\nThat's a great question. I think we have to\nassume that there's a linkage between the\ncultural transformations in Africa and Eurasia at\nthis time, and that's not something the community\nhas really put together with the genetic data.\nThere's this thread in the genetics about\nsubstructure in Africans, and then there's\nthis whole world based on ancient DNA,\nand they've never been put together.\nNobody's put together the now extensive\nwork on modern human substructure with\nthe extensive work based on ancient DNA\nof archaic human relationships to modern humans.\nIf you put them together, you realize they line up\nin terms of their time of substructuring.\nI don't know if that's improbable. It\nseems parsimonious to me.\nIt also seems significant\nthat different groups of humans at this time\nwere capable of adopting Stone Age technology.\nOnce one group had figured it out, the\ngenetic difference between different human\nlineages was not so big that you could\nnot show people how to use stone tools.\nWho knows? It could be that\nthis was genetically driven.\nWe talked before about the time to\nthe common ancestor of human genes.\nThere's nothing at 100,000 years or 150,000 years,\nbut there's a lot at 400,000 or 500,000 years.\nIf that's what happens, and you have a mutation\nthat occurs in the Caucasus, somewhere in the\nMiddle East, or Northeast Africa, there could\nbe key genetic mutations that make people able\nto do this. Then this population expands. When it\nmoves into Europe, it's swamped by local genes,\nbut there could be retention of those\ngenes through selection as it expands.\nMaybe what you're seeing is that\nthere are genetic developments.\nMost of the discussion has been focused\non the 50,000 to 100,000-year event,\nwhich is anatomically modern human behavior.\nBut a lot of archaeologists think this is an\nequally—if not more—profoundly\nsignificant event in many ways.\nWhy is that not the event\nwe should be talking about?\nYou're talking about how there are no\nfixed differences between modern humans\nand the humans 50,000 years ago.\nDo we know if there are any fixed\ndifferences between the people 50,000\nyears ago and the people 300,000 years ago?\nI think there are.\nOther than obviously these interbreedings.\nIf you look at the genetic variation\ngoing back 300,000 or 400,000 years,\nthere do begin to be places where all\nmodern humans share common ancestry.\nThat's another way of saying there begin\nto be fixed differences at that time depth.\nThat is where you start seeing evidence\nfor possible fixed differences.\nWhat's happening, if everybody shares a\ncommon ancestor 400,000 or 500,000 years ago,\nis that there's a single ancestor at that time.\nIf you compared it to another population, they\nwould descend from a different lineage, so any\nmutation that occurred ancestral to that single\nancestor would be a fixed difference.\nThis is the time at which you can\nbegin to see fixed differences.\nBut anatomically modern, cognitively\nmodern humans exist by the beginning of the Middle\nStone Age, before we're breeding with this ancient\ngroup of Africans or breeding with Neanderthals.\nAnatomically modern humans occur exactly here.\nIt's the same moment. This is when they occur.\nThe people who have skeletal features like ours,\nand Neanderthals, appear exactly then.\nThis is when it all happens.\nThere is this disconnect between\nanatomically modern humans in\nthe skeletal record and behaviorally modern\nhumans, which is 50,000 to 100,000 years ago.\nAnatomically modern humans appear at\nthis time, and recognizable Neanderthals\nappear roughly around this time, too.\nInteresting. But we don't know what exactly\nhappens, if anything, between 200,000 years\nago and 50,000 years ago that goes from just\nanatomical modernity to behavioral modernity.\nMy understanding is no. They're busy making\nLevallois stone tools like Neanderthals\nfor 200,000 years, and they are not\nmore impressive than Neanderthals in\nany obvious way, as I understand it.\nThen there begins to be in the archaeological\nrecord a quickening of behavioral traits, which\ncould be not genetic at all, or could be genetic.\nThere are lots of arguments about this.\nWe were obsessed with intelligence\nearlier in our conversation.\nPeople are obsessed with art and\nthese things that seem important\nto us but who knows what's important?\nInteresting. Cool, thanks for the digression.\nThe work that I've been involved in has\nconsistently shown that I was wrong in\nmy biases coming into the work, and I've\nreally been almost traumatized by this.\nAgain and again, I've come into a project with\nsome kind of guess about what the data was\nshowing, and then the data doesn't show that.\nFor example, when I got involved in the\nNeanderthal genome project helping to analyze\ndata looking at how archaic Neanderthals were\nrelated to modern humans, I was part of\na group of scientists who had established\nthat non-Africans were a simple subset of African\nvariation and that there was no evidence at all\nof Neanderthal interbreeding into the ancestors\nof modern humans or other archaic interbreeding.\nDifferent analyses that I and many\nother people had done made it look\nlike non-African variation was just a\nsubset, a small sample of that in Africa,\nand that could have fully explained the data.\nSo when I was involved in analyzing the\nNeanderthal DNA sequences, what happened\nwas I found this very strong evidence of\nNeanderthals being more closely related\nto non-Africans than to Africans.\nIt was very surprising, and I thought it\nmust be a mistake. I was quite incredulous.\nI thought it was unlikely to be true because\nother evidence that had been found before\nseemed to point in the other direction.\nSo I spent several years trying to make\nthese results go away, as did my colleagues, and\nwe just couldn't make the results go away. They\njust kept getting stronger. And this experience\nworking on natural selection was the same.\nWhat we were convinced of was that natural\nselection had been pretty quiescent in our species\nover the last several hundred thousand years.\nTherefore, if we look at patterns of variation in\nnon-African people today, or in any people today,\nwe should see not a lot of selection going on.\nIndeed, the first ancient DNA studies, beginning\nin 2015 with this paper that we were involved in\nwith Ian Mathieson and colleagues, seemed\nto show relatively small numbers of genetic\npositions associated with natural selection.\nIn 2015, we analyzed data from about 200\nEuropeans and Middle Easterners to try to\nunderstand frequency changes over time.\nWe compared those ancient people who were the\nsources of modern Europeans to people in Europe\ntoday, and we looked at frequency differences\nthat were too extreme to be due to chance.\nWe were very excited to find 12 positions\nthat we were convinced were highly different\nin frequency between Europeans today and what\nwe would expect, based on the history that we\nand others had identified as the history\nrelating modern to ancient Europeans.\nSome of these were known and some of these\nwere not known, and this was very exciting.\nWe hoped that as the numbers of\nsamples would increase and we got\nhigher resolution to be able to appreciate\ndifferences in frequencies over time,\nit would make it possible to detect far more.\nWhat was quite disappointing over the subsequent\ndecade is that that didn't happen.\nFor example, the largest study of\nthat type in 2024 by a group in Copenhagen\nanalyzed much better data than we had in\n2015 and found only 21 positions that were\nhighly different in frequency across time.\nWhile that was exciting—it was almost twice\nas many as we had found in 2015—in a lot of\nways it was disappointing because the sample\nsize and data quality had gone up so much,\nand yet this is all that was found.\nThat suggested we might be hitting\nan asymptote and might not be able to\nget beyond where we currently were.\nThis approach to learning about biology,\nwhich was very promising in theory,\nmight not produce a high yield.\nMaybe natural selection was quiescent,\nand the reason we're seeing so\nfew changes is that there has\nnot been a lot of adaptive directional selection.\nThat was the situation we found ourselves in until\njust a few years ago when we carried out this\nstudy in our research group led by Ali Akbari.\nWhat we did is we deployed a few\ninnovations to try to improve our\npower to detect natural selection.\nOne of them was that we just pumped a\nlot of data into the system, increasing\nthe amount of data by about 14-fold.\nThe main thing that we do in this study is\nreport new data from about 10,000 individuals.\nThis is a very big increase in the\namount of data in the literature.\nThe total dataset size of ancient individuals\ndistributed over the last 18,000 years is\nabout 16,000 people. This is a large dataset.\nIt's much larger than was previously possible,\nand when you have more data, you can estimate\nfrequency changes with much more subtlety.\nThe data comes from only one part of the\nworld, which is Europe and the Middle East.\nIt's not a more important part\nof the world than other places,\nbut it's the place where maybe 70-80% of\nthe data in the ancient DNA literature so\nfar comes from due to historical reasons.\nIt provides us with a natural laboratory\nwhere we can see what happens to the genome\nin one place over time as environments change.\nIt's really interesting to imagine doing this\ntype of analysis in other parts of the world,\nand the comparative analyses are super\nimportant and interesting, but this study\nright now is about this one place in the world\nwhere we have particularly fantastic data.\nThe other thing we did is that we\ndeveloped an entirely new methodology\nthat hadn't been used in this area before.\nThe methodology is based on a technique\nthat had been developed for finding risk\nfactors for a disease in medical studies.\nA simple way to explain it is that we ask how\nto predict the genetic type a person has based\non their pattern of relatedness to other people.\nWe have a dataset of about 16,000 ancient people,\nand 22,000 people if we include the modern people.\nThen we look at how closely related each of these\n22,000 people are to each other, and we predict\nthe genetic type at each position in the DNA—at\n10 million positions—based on the pattern of\nrelatedness to all of the other 22,000 people.\nThen we ask if natural selection blowing\nthe frequency of the mutation in the same\ndirection in all geographic places and at\nall times predicts the data a little bit\nbetter than just knowing the relatedness\nto all the other samples in the database.\nWe're simply asking if the alternative\nhypothesis—that selection has been\nblowing in the same direction at\nall times—explains the data better.\nThat's a dumb assumption, because of\ncourse, the truth is that natural selection\nwill have changed in frequency over time.\nBut we're just asking the simplest of questions:\nwhether assuming a constant rate of selection\nexplains the data more than not doing so.\nTo summarize to make sure I've understood,\nyou're trying to make a model that predicts\nallele frequency changes over time. You\nhave two different parts. One part is this\ngenetic relatedness matrix, which captures how\nsimilar different genomes are to each other.\nThat should capture the impact of\ndifferent bottlenecks, of drift,\nof population admixtures, and all those\nthings which affect the entire genome.\nThen you have the separate thing, which is, if we\nlook at specific locations, can we just say that,\n\"Oh, this location has been selected\nat whatever coefficient over time\"?\nAnd if we add some coefficient, does it become\neasier to predict the allele frequency changes\nthan you would have just seen from this other\nartifact, which is just looking at, \"Oh, if you\nlook at the whole genome, are these guys in the\nsame, have they gone through the same bottlenecks?\nHave they gone through the same drift,\" etc.?\nThat's precisely right.\nOkay, what did we learn?\nWhen we analyzed the data this way, we\nlooked at 10 million positions in the DNA across\nthese 22,000 people, 16,000 of whom were ancient.\nWe looked to see if there was more\nchange in this consistent direction\nover time than you would expect by chance.\nWhen we analyzed the data, we found many\nhundreds of places in the DNA that were\nchanging too much over time and in too\nconsistent a way to be explained by chance.\nThere's a bit of a statistical problem\nin figuring out how many there are because\nthey're so densely packed that they're close\nto each other and interfering with each other.\nBut when you try to piece them out and say,\n\"Let's count only one in each place in\nthe DNA and blank out the others,\" we\nfind at least 479 positions that are all\nindependently pushing in the same way.\nWe are 99% confident that\nthose positions are real.\nBy another criteria of being more\nthan 50% confident that they're real,\nwe think that about 3,800 positions\nare all pushing in the same direction.\nThis is a crazy number of results given that\nin our previous work and other people's work,\nthere were at most a couple of dozen\ndiscoveries coming from a single scan.\nSo when we got this result,\nwe were very surprised.\nWe thought it must be wrong, and we spent the\nnext couple of years trying to make the results\ngo away, but they just kept getting stronger.\nWe were trying to look for some independent\ntype of evidence to tell us\nwhether these positions were real.\nWe stumbled on something really powerful for this\npurpose that had not been used in this way before.\nIt relied on the fact that we had\nvery large numbers of discoveries,\nmany hundreds of discoveries or even thousands.\nWe took a completely independent dataset, which\nwas the corpus of genome-wide association studies.\nThese are studies that people have carried out in\nhundreds of thousands of people, looking for\nwhether particular genetic mutations are more\ncommon in people with high blood pressure than\nwith low blood pressure, or something like this.\nWe took the UK Biobank, which is about 500,000\npeople from Great Britain who have been measured\nfor hundreds of traits.\nThe whole genomes of all\nthese people have been sequenced.\nFor each of these traits, we could\nlook at whether each of these 10 million positions\nare connected to this trait in a convincing way.\nOut of 10 million positions, about 15%—about 1.5\nmillion positions in the DNA—are predictive of\nat least one of these several hundred traits.\nThen we could ask a question: is our natural\nselection signal, our statistic, related\nto whether a mutation causes high blood\npressure or some other trait?\nWe slid our statistic for natural\nselection upward, to a value of\none, two, three, four, or five.\nAs we did that, the enrichment\nfor genetic mutations that\naffect traits got higher and higher.\nWhereas it was only 15% when we didn't\nuse our selection statistic, when we required\nthe selection statistic to be above about five,\nthere was about a five-fold enrichment\nfor mutations that cause traits.\nSorry, what is a selection statistic?\nThis is the statistic we use to measure\nwhether a mutation is changing over\ntime significantly in a non-zero way.\nIt can be approximately thought of as a normally\ndistributed statistic, a Gaussian statistic,\nwhich is the number of standard deviations\nthe statistical value is away from zero,\nwhere zero is no natural selection.\nIt's not exactly that, but it's close to that.\nIf the statistic is above five, we\nsee about a five-fold enrichment\nin mutations that affect a trait.\nInstead of 15% of the mutations that are\nat random affecting the trait, it's\n60% or 70% that are affecting the trait\nwhen we slide our statistic upward.\nThis provides completely independent\nevidence that these sites are real, and as you\nslide above five, there's no more enrichment.\nOur interpretation of these results—which we\nwere able to validate and show made sense using\ncomputer simulations of our process—is that once\nyou slide the statistic above five, essentially\nall the signals of natural selection are real.\nOkay, just to make sure I understand.\nYou're saying, in order to figure out what\nalleles have been under selection, your model\nassigns a statistic saying, \"In order to explain\nwhy this allele has a specific frequency,\nwe're going to give it a selection statistic.\"\nIndependently, we run these studies on modern\npopulations where we say, \"If you look at height,\neye color, intelligence, or whatever trait,\nwhat are the parts of the genome\nthat are correlated with that trait?\"\nThe higher the statistic you give it in\nyour study to explain allele frequency\nchanges over time as a result of selection,\nthe more probable it is that that region in\nthe genome is associated with traits that\nhave some functional thing we can measure.\nThat's exactly right. This is\na brilliant idea that Ali had.\nIt abandons the traditional approach of assigning\nstatistical significance to mutations that cause\na trait because we're just using an external\npiece of information—the correlation to traits,\nmeasured in a completely different way—to\nread off the probability mutations are real.\nWe can ask how much enrichment for real signal\nis there given a particular selection statistic.\nIf it's halfway enriched to the plateau, we're\nable to show the correct interpretation is that\n50% of the mutations are really selected.\nIf it's three-quarters of the way toward\nthe plateau, there's a three-quarters\nprobability that the mutation is real.\nIf it's 99% of the way to the plateau,\nthere's a 99% probability that it's real.\nThat gives us a calibrated estimate of the\nprobability that a particular position is\nreally under natural selection.\nA major concern here is that what\nwe're actually seeing is not that these mutations\nare really under selection, but rather that both\nassociation to a disease and our selection signal\nare due to some third thing that's causing both\nof them, which is a type of selection which\nis not what we're after, not selection to\nadapt to new environments, but what's called\nbackground selection: selection against newly\narising bad mutations that are removed from the\npopulation that tend to be concentrated in genes.\nGenes are also the parts of the genome\nthat tend to be associated to traits.\nThis common process is causing both the\nenrichment for trait signals and is also\ncausing the enrichment for selection signals\nthat we're observing. That's the concern. We\nwere super concerned about this.\nSo what we did is we repeated this\nenrichment analysis in slices of the DNA\nthat all were affected to the same extent by\nbackground selection, by this rain of slightly bad\nmutations, and we get exactly the same pattern.\nWe also repeated this experiment just\nusing mutations of the same frequencies\nbecause there's different statistical power to\ndetect these signals at different frequencies.\nWe see the same pattern where above\na value of the selection statistic\nof around five, we get this plateau.\nThe thing that changed that allowed\nyou to increase the amount of sequences\nyou're generating by two orders of magnitude\nis just the statistical method you're\nusing to identify which part is human?\nOr what exactly changed in 2014 and since then?\nThere's been a whole series of improvements.\nThe big ones have been the huge drop in\nsequencing cost, which made it possible to\ngenerate ancient DNA in the first place.\nThe drop in cost has been a millionfold\nsince the late 2000s, and another maybe\none to two orders of magnitude from 2010\nto today. That's one big change. Another\nchange has been in-solution enrichment.\nIt's been this way of taking a sample that\nhas very small percentages of human DNA,\nbut then suddenly creating a process that will\nmean that the great majority of the sequences\nthat one's analyzing will be useful for analyses.\nThe approach that we used was we took the DNA\nsamples that we had, most of which were very\nlow percentages of human DNA—less than 10%,\noften less than 1%—which is such a low proportion\nthat it's prohibitively expensive to sequence\nthem and to just brute-force sequencing them given\nthe technology that we had available at the time.\nWe took these samples and washed them over\nan artificially synthesized set of short DNA\nfragments that targeted positions of the\nDNA that we were interested in analyzing.\nThis is more than a million positions that are\nhighly variable in people, and we picked many\nof these to be biologically interesting.\nWe had a whole set of known biological\ntargets that affected traits in genome-wide\nassociation studies, which is the way that\npeople look to see if there are particular\ngenetic variants in modern people that have\nparticular impacts in phenotypes and traits.\nAnd so, what we did is we had this artificially\nsynthesized set of DNA fragments that we\nwashed our ancient sample over, and it\nbound the parts of the DNA that we targeted.\nThe resulting sequence that we generated was\nvery enriched for the parts of the genome\nthat were informative about history.\nEven though only 10% or 1% of the DNA was human,\nit ended up that a very large fraction was from\nthe parts of the genome that we were interested\nin, and it became economically efficient to do it.\nWhat was the other 99% of the DNA?\nIt's mostly microbial. It's from\nbacteria and fungi that colonize\na person's body after they die.\nDepending on how they die, there'll be\nmore or less of these bacteria and fungi.\nWhen you typically sequence DNA from a person,\nit'll just be full of microbial sequence.\nSometimes the microbial\nsequence is very interesting;\nit might be pathogens that a person died of.\nThere's amazing work, for example, about different\nplagues of malaria and Black Death and hepatitis\nB and so on that have been obtained from the\nsequences of these pathogens in people's teeth\nand other parts of their body when they died.\nBut we're focusing here on the human DNA.\nThis changed the amount of data that was possible\nto produce from tens per year to hundreds\nper year, and then we further roboticized\nand industrialized the process so that there\nwere many hundreds or even thousands per year.\nJust in our laboratory, we've been\ngenerating genome-scale data from\nmore than 5,000 individuals per year.\nI know this is true also of several\nother laboratories in the world now.\nThis huge jump in data, this semi-exponential\nor even super-exponential jump in some cases,\nhas made it possible to ask and answer questions.\nWhile we were only on the order of 10 genome\nsequences from humans in 2010, this year it's\npassed more than 20,000 reported sequences.\nThere are several orders of magnitude increase,\nand the questions we were able to ask in 2014 are\njust not the same as the ones we can ask today.\nAwesome. Excellent. David, thanks for your time.\nThank you, Dwarkesh.",
  "transcript_chars": 124723,
  "ingested_at": "2026-05-15T04:55:19.065204+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": 131305,
    "like_count": 3163,
    "channel_id": "UCXl4i9dYBrFOabk0xGmbkRA",
    "categories": [
      "Science & Technology"
    ],
    "tags": []
  }
}