David Reich Bronze Age Shock The Neanderthal Puzzle And Farmings Sudden Spread

TITLE: David Reich – Bronze Age shock, the Neanderthal puzzle, & farming’s sudden spread CHANNEL: Dwarkesh Patel DATE: 2026-05-08 ---TRANSCRIPT--- Humans, at least in this part of the  world, were wrenched into a way of   living that was so different from how  their hunter-gatherer ancestors lived   that the organism had to adapt very strongly. Maybe the degree of that wrenching process moving   into the Bronze Age was qualitatively greater  than the degree of the wrenching process that   happened from the initial transition to growing  plants, which is surprising, because our cartoon   picture is that the big transition is farming. But the genetic data, the biological readout, is   saying our genome is reacting much more strongly  to these events that happened 5,000 years ago.  I am back with David Reich, who is a  professor of ancient DNA at Harvard.  How do you describe what it is that you study? I’m a geneticist, and I work on human history   and how ancient people relate to  each other and people living today.  We did an interview two years ago,  which ended up being one of the   most popular interviews I’ve ever done. I think people found it really compelling   that there’s so much about human history we don’t  know and are just learning about now as a result   of the kinds of techniques your lab is using. You have a new preprint that’s very exciting,   and I wanted to talk to you about it. Can you give me a little bit of context   on what we’re talking about today? The dream was that when this ancient   DNA field started, more than 16 or 17 years  ago, we were going to learn a lot about   biology —about how people’s biology changed  over time— by getting DNA out of ancient   human remains and tracking changes over time. And that dream has really not been realized   since the beginning of this field. The field has been a big success   with regard to learning about human history. It’s resulted in surprising findings about   human migrations —people not being descended  from the people who lived in the same place   hundreds or thousands or tens of thousands of  years before— and mixture being common in human   history, and sex-biased processes being common. And there have been things that were   not expected from archaeology. The field’s been a big success from   that perspective, but what’s not been successful  is learning about biology and biological change.  One big reason has been that the  sample sizes have been too small.  When you have a single person’s DNA, it provides  a tremendous amount of information about history.  That’s because when you look at one person’s  DNA, it’s not a single person. It’s many   people. It’s your two parents, your four  grandparents, your eight great-grandparents,   16 great-great-grandparents, and so on. Going back in time, thousands,   tens of thousands, even hundreds of thousands  of ancestors are contributing to people today.  When you look at the DNA of a single  person’s genome or a Neanderthal genome,   you have effectively tens of thousands of  ancestors all represented in your data.  And you can position that  individual exquisitely with   respect to other people from whom you have data. But when you are interested in how a particular   genetic variant—that affects something like  your skin pigmentation, or your ability to   digest cow’s milk into adulthood, or a behavioral  trait—changes over time, a single person gives   you only one sample, or maybe two samples: the  one in their mother and the one in their father.  To get a high-resolution picture of how the  frequency changes over time, you need very big   sample sizes, truly very large numbers of people. We just didn’t have that until the last few years.  What motivates the study we’re talking about  today, and the work that hopefully a number   of groups will be doing in the coming years, is  the fact that we now finally have those numbers.  We can do something with the data to  see how frequency changes over time.  Can I ask a question? I’ll be asking a lot of  naive questions through the next few hours, but   why are frequency changes especially interesting? What we’re interested in is using the experiment   of nature that’s occurred in our history, over  the last tens of thousands of years, to understand   what’s biologically significant in our DNA. If there has been a change in environment   that a population has experienced—for example,  people shifted to agriculture, began living close   to domesticated animals, or moved from a cold  place to a warm place, or a low place to a high   place—then there’s pressure on the population  to adapt to these new stresses and new needs.  The way you’re going to detect that  is by seeing that the frequency of a   genetic variant—that for example might  allow you to live at higher altitude,   or that might nudge you to have a different  behavioral pattern advantageous in the new   situation—pushes systematically in some direction  in a way that is enough for you to detect.  It’s very hard to detect slight shifts in  frequency by a few percent or ten percent   unless you have a very big sample size. What we’re looking for are those changes   in frequency that are too  extreme to be due to chance.  That will tell us there have been pushes against  the biology as a result of the changes in   environment that people have experienced. Interesting. What did you guys find?  Seven years ago, Ali Akbari, who at the time was a  postdoctoral scientist in my laboratory and a few   years later became a permanent staff scientist,  set out to use the data we were producing to   learn about biological change over time. I think the reason he was interested in   our laboratory rather than other places was  that a focus of our lab has been generating   truly large amounts of data from ancient humans. We’ve been trying to industrialize the process,   make it very inexpensive, make it high quality,  and generate large numbers of samples with lots   of good data for this purpose. There’s been this large amount   of data we’ve generated. And it made it possible to   conceive again of asking whether there  have been frequency changes over time.  The mainstream view in human evolution in the  last several decades has been that natural   selection has been pretty quiescent over the last  several hundred thousand years of human history.  There are several lines of evidence that  have been deployed to document this.  One is that if you compare diverse populations  from different continents around the world,   for example Europeans and East Asians, and  you look at mutations that differ in frequency   between these groups—all mutations differ a little  bit in frequency, sometimes a lot—you can say,   “What are the most different mutations in terms  of frequency between Europeans and East Asians?”  And there are almost no genetic  changes that are 100% different   in frequency between Europeans and East Asians. Europeans and East Asians descend from a common   ancestral population 40,000 or 50,000 years ago  that came out of Africa and the Middle East.  This population had a set of gene frequencies,  and these variants bopped around randomly—a   process known as genetic drift—or perhaps  under selection in one direction or another.  The time that’s passed since 40,000 or  50,000 years ago is sufficiently small on   an evolutionary timescale that there’s  just not much genetic differentiation   on average between these two groups. However, if there’s been natural selection, for   example to help people in one place digest alcohol  better, or digest milk better, what you might   expect is that there would be some mutation that  would have rocketed up to very high frequency.  Forty or fifty thousand years is a lot of  time, it’s maybe 1,500-2,000 generations.  That might easily be enough time to  see a 100% difference in frequency.  Yet you don’t see any more than  what you would expect by chance.  This combination of things made it seem  that selection has just been quiescent.  Maybe a few hundred thousand years ago, the  ancestral human population got to some kind   of optimum, and after that there hasn’t been  much genetic change in one way or the other.  There have been small amounts of natural  selection, or selection to remove bad mutations   that are constantly raining down on the genome,  but not what we call directional selection.  That would be newly arising mutations, or  mutations being pushed in a systematic direction,   to help the population get to a different  adaptive set point more favorable for the   conditions that population is living in. We were able to partition how much of the   changes in frequencies of all the mutations that  we’re seeing in the DNA—we’re looking at about 10   million positions that vary—is due to directional  selection (adaptation) versus other factors,   especially genetic drift. And 98% of it is other   factors, especially genetic drift. It’s overwhelmingly migrations and population   structure causing fluctuations in frequency. As a result, it’s super hard to detect the signals   of adaptive natural selection because they’re  a tiny fraction of the total frequency change.  The vast majority of it are  these migrations and mixtures.  Nevertheless, there’s so much natural  selection, as our study has shown,   that it’s actually been rampant in the genome. Can I ask a clarifying question here?  Why are we discounting population  admixture or replacement as selection?  If you think about it at a group  level, if one population replaces   another population, isn’t that selection? I remember from the last episode you were   explaining how there have been huge changes in  what kinds of people are in a specific area.  One population came in and replaced  the previous one, and then a new   population came in and replaced that one. To the extent that the genetics are relevant   to why that population replaced the other one, why  should that not count towards what we understand   to be selection over the last 10,000 years? It could count, and may count, and probably   should count in some respects. But it could also be that this   population replacement is due to some  cultural phenomenon —technology held by   one of these groups and not others. And maybe there are some genetic   mutations that are contributing to  this. Who knows? It’s possible. But   what you’re seeing is a whole-genome shift. What we’re looking to see is whether there’s one   place in the DNA that is driving the change in a  way that’s different from the rest of the genome.  From a statistical point of view, what happens  at these times of migration is there are just   huge fluctuations in frequencies. These are extremely uninformative   times for detecting natural selection. The best moments to detect natural selection   are when migrations and population admixtures  are not happening for a few hundred years.  During these times, you can  actually see the mutation   slowly blowing in one direction as a result. The way we think about the history of Europe and   the Middle East for the purpose of this study is  as an archipelago of little populations in space   and time, each pretty isolated from each other. You have a little population in Britain isolated   for a few hundred years, or a little population  in Hungary isolated for a few hundred years,   between big events of migration and mixture. In each of those little experiments of nature,   we can ask: does this mutation  slightly increase in frequency?  Does that same mutation  slightly increase in frequency?  If all the arrows point in  the same direction, we win.  They’re telling us that  natural selection is occurring.  For example, 4,500 years ago in Europe, almost  all mutations went through huge frequency changes.  That’s not because of natural selection. It’s because of the steppe migration from   north of the Black and Caspian Sea. 40-80% of  the DNA becomes Yamnaya from steppe pastoralists.  Their frequencies of mutations were different  not because of selection necessarily, but just   because they had evolved in different places  for thousands and tens of thousands of years.  When you look at the descendant populations,  there are huge changes in frequency.  What you need to do is see if natural  selection is explaining a shift more   than you would expect by chance. Ok, in this next section David   explained the nitty-gritty of the methodology  of this paper. It’s honestly a bit technical,   and I wanted you to get a sense of the results  first, so I’ve moved that section to the end.   If you want to understand the methodology  just stick around for the full episode.  So you found these locations that seem to be  under selection. I have another clarifying   question. You say you found 3,800 locations  which you’re 50% confident have been under   selection in the last 10,000 years. It’s 7,200 where we’re 50% confident.  We’re getting about 7,200 positions in the  DNA that have 50% confidence of being real.  Only half of those are real—we don’t know  which ones—so 3,600 of them are real.  Does that also mean that outside of  those 7,200, you’re confident the other   locations in the genome are not under selection? No. If you look at the 25% probability cutoff,   there will be tens of thousands, and  there will be many real ones there too.  In fact, multiple analyses we do suggest that  the genome is vibrating with natural selection.  There are all sorts of weaker  effects that would be picked   up in even larger studies than we’ve done. In fact, almost every position in the DNA is   correlated to another position that  is being dragged in one way or the   other by natural selection. Instead of being quiescent,   natural selection is everywhere. Even though it’s only 2% of the   frequency change, it’s tugging the positions  in one direction or the other everywhere.  So we analyzed these positions that we  had identified, the hundreds of positions   we were super confident about. We looked to see whether they   were randomly distributed in the  DNA or whether they had patterns.  We looked at maybe 100 or so traits where  there had been genome-wide association   studies for all sorts of different traits,  associated with immunity or autoimmunity or   behavior or metabolism, and other things. For each of these we could ask: do the   genetic variations that are known to affect these  traits from genome-wide association studies have   an unusual number of genetic selection signals? What we found is that there was a vast enrichment,   by about four or five-fold, for immune traits. There was a super concentration of selected   signals in immune traits. We also saw a strong enrichment   for metabolic traits—things that might impact  obesity or fat traits or Type 2 diabetes—and   almost no detectable enrichment, as far as we  could tell, for behavioral or psychiatric traits.  Just to make sure I understand. This is not to say that behavioral or psychiatric   or cognitive traits are not under selection. It’s just that the individual sites where such   traits are controlled are not especially  likely to be among the locations you’ve   identified as under selection. That’s exactly right. It might   seem from the results of that analysis that  immune traits are highly selected and that   there’s been no selection for behavior in the  last 18,000 years in this part of the world.  But that’s a wrong conclusion, and we have  evidence that it’s a wrong conclusion.  There’s clear evidence of selection  also on behavioral traits.  The reason we think we see much weaker signals  for behavioral traits is that behavioral traits,   we know from medical studies, are underpinned by  much larger numbers of genes than immune traits,   which are underpinned by relatively  small numbers of genes of strong effect.  Behavioral traits are shaped genetically by  a very large number of genes of weak effect,   and we just don’t have the statistical  power to detect these very weak signals.  When we do an analysis looking at our very  strong signals of selection, that collection   of very strong results is very effectively  querying the immune traits, but is not very   effectively querying the behavioral traits. It may still be the case, and I guess it is,   that immune traits are the most selected category. But it is not at all the case—and we   can prove it’s not the case—that  behavioral traits are not selected.  Interesting. We’ve been able to   prove that there are two ways to reconcile the  previous observations with our new observations.  Remember, the previous observation is that  natural selection seems to have been quiescent   over a timescale of hundreds of thousands or  many tens of thousands of years. Reason? That   you don’t see 100% difference in frequency  variance across Europeans and East Asians.  Now we’re seeing hundreds of positions  that are rocketing up in frequency with   selection rates of 1% or more in a lot of cases. A 1% or more selection rate will mean a rapid   doubling over periods of dozens of generations. Over the 1,500 or 2,000 generations separating   Europeans and East Asians, shouldn’t you see  many genetic variants that are 100% different   in frequency across populations? We were able to show that this   is explained by at least two factors. One is that in this part of the world—Europe   and the Middle East—we are actually in a  period of accelerated natural selection.  One way to see this is to look at the enrichment  pattern we’re observing, where immune traits are   unusually associated with these selection signals. We could compare the last 5,000 years of our   time period, what’s called the Bronze Age and  further onward, to the previous 5,000 years.  What we see is that this intensification of  selection around immune traits, and similarly   the intensification around metabolic traits,  has accelerated over this time period.  It’s not like natural selection has been  at the same rate over all places and times.  It’s increasing over the  time period we’re analyzing.  Plausibly the whole time period has  increased compared to previous periods.  We’re in a period of intensified selection. That’s not implausible, because this is a   population that went through a huge shock in  terms of the way people live and the culture.  Almost everyone we’re analyzing are farmers  or food producers in one way or another.  Farming was invented for the first time  anywhere in the world in the Middle   East 11,000 or 12,000 years ago. The people who invented farming   exploded into Europe after 8,500 years ago,  spread across the continent, and expanded rapidly.  In the Bronze Age, there was an  intensification of how people lived,   with much higher population densities. People were living more and more next to   their animals and getting their  diseases, and exchanging their   diseases with the animals and with each other. This is a period of rapid change in how people   are living, resulting in different  biological needs of this population.  It’s not surprising, perhaps, that in the context  of these dramatic changes, the biology of the   population might not be ideally adapted. There might be what some people call an   evolutionary mismatch, where you take  a genetic variation that evolved in   hunter-gatherers and put it into farmers or  pastoralists, and it’s not exactly right.  What you’re seeing is the DNA of this population,  which descended from hunter-gatherers only 10,000   years ago, reacting to the shock of having  been moved into an agricultural, Bronze Age,   high-population-density, urban environment. A hypothesis is that what we’re seeing is   the adaptation that occurs as a result. In the paper you have many examples of   this intensification of  selection around the Bronze Age.  It might be helpful to go through some of these. One of the things we do in this work is look   carefully at many of these positions in the DNA. We actually have an internet browser called the   AGES browser, which Ali and a colleague of  his—who’s a co-author of our paper—built.  It allows you to query each of these 10 million  positions and see the trajectories at each   position and the evidence for selection. One of the things we see is that,   while for the most part the signals of natural  selection we detect are consistent with constant   natural selection over time, in a handful of  them we’re able to see that there’s been a   reversal or a radical change in natural selection. Very often that occurs in the period between 5,000   to 2,000 years ago, which is the Bronze Age  and the Iron Age, a period of rapid population   growth and rapid movement to intensive use of many  technologies that were not used that way before.  An example of this is the TYK2 genetic  variant that is a major risk factor for   severe tuberculosis, which is the most important  infectious disease killer in the world today.  If you look at this major risk factor for  tuberculosis, this variant rockets up in   frequency from 8,000 or 6,000 years ago to  maybe 9% or 10% in this part of the world.  Then it rockets down in frequency  in the last 3,000 years.  In both cases, there’s very clear  evidence of natural selection,   in the first case to increase in frequency,  and in the next case to decrease in frequency.  A possible reason is the spread of tuberculosis. It maybe becomes endemic in the   population 2,000 or 3,000 years ago. That’s potentially consistent with pathogen   sequence data and other lines of evidence. And maybe this variant was protecting against   something before then, but then tuberculosis  became significant after that point,   and it was so bad that it pushed in the  opposite direction. That’s speculative.  The thing it was protecting against  was probably another disease?  Maybe. Prepping for   this episode required a full lit review. I needed to understand why other methods   had failed to find evidence of natural  selection over the last 10,000 years.  What exactly did Reich and Akbari do differently? Honestly, this was quite subtle because the most   important points were distributed  across a bunch of different papers.  And it was frustrating to talk to LLMs  about it because they kept getting confused.  One of them would fail to  understand an important crux.  And so I’d switch over to a different  model, and that one would get tripped   up on the very next point. I ended up using Cursor to   kick off a handful of models at the same  time and compare their results after.  I could have one model critique  the response of another.  This was super useful because  while I’m not a geneticist,   I do have enough taste to be able to say, “hey,  this answer makes sense, these ones don’t.”  I also had Cursor turn this work into  flashcards so I could retain what I learned.  Cursor started as a programming tool, but I  found it really great for this kind of research.  There’s no other interface where I can get answers  from a bunch of independent LLMs all while reading   the relevant paper on the same screen. Go to cursor.com/dwarkesh to try it out.  One of the big takeaways for me from  the paper was just that something   weird happened in the Bronze Age. As you said, across trait after trait,   the selection intensifies during the Bronze Age. This makes sense for some things.  For example, why do we see lactase  persistence, where adults can process   milk, intensified during this period? This is the time when we start using   cattle not just for the meat, but also for  milk and wool and other secondary products.  So it makes sense why lactase  persistence would matter more.  But then there are other things  that seem like they should have   been relevant since the dawn of agriculture. I forget the exact name of the allele, but was   it FADS1, which helps convert plant fatty acids  into long-chain fatty acids that your body needs?  That’s obviously relevant when  you move from a diet of meat as   a hunter-gatherer to a diet of cereals. That is also one I think you found was   under especially high selection 5,000  to 3,000 years ago. So what’s going on?   Why is the Bronze Age so special across all  these different traits that you’re observing?  So this FADS1/2 variant is a  vegetarian/meat-eating adaptation.  Already in work prior to this, Ian Mathieson,  who worked with me in 2015, identified this as   a very strongly selected variant. It’s actually  ancient. You see copies in archaic humans too.  One of the findings of our  paper is the ABO blood system.  You get your blood typed as A, B, and O. The B variant has increased up to 10% at   the expense of A, but previous work has shown  that A and B were both already present in the   ancestor of humans and gibbons and other apes. Some of these mutations have been going back and   forth and fluctuating over different time periods. But we’re talking about changes in the Bronze Age.  The TYK2 variant for tuberculosis risk, a multiple  sclerosis risk variant, inflected and increased in   frequency before the Bronze Age, and then 2,000  or 3,000 years ago reversed in that period.  There are differences in Northern Europe  where this process is super strong,   very strong positive selection,  very strong negative selection.  And then in Southern Europe, only a little bit,  and not even very strong negative selection.  For hemochromatosis, which is pathogenic  iron buildup that causes problems in Europe,   that too has reversed around this period. In some of the complex traits that maybe   we’ll talk about later, these traits too have  periods of intensification of natural selection.  For example, depigmentation: Europeans have  gotten lighter skin over the last 10,000 years.  You can see it in our data. The period of strongest   depigmentation is between about 4,000 to 2,000  years ago, and then after that it’s much less.  This seems to be a very impactful, eventful,  important period where a lot of the processes   we are seeing become very powerful. It’s  surprising on first principles. You might think,   before you walked into this genetic data, that  the big change is going to be starting to grow   plants and maybe farm animals. That happens in the Neolithic,   beginning 10,000-12,000 years ago, and  spreads into Europe after 8,500 years ago.  But actually, the intensification happens  5,000 years ago, 4,000 years ago. It’s   really interesting. This observation of that  being an inflection point tells us something   about when humans, at least in this part  of the world, were wrenched into a way of   living that was so different from how  their hunter-gatherer ancestors lived   that the organism had to adapt very strongly. It may be that the degree of that wrenching   process moving into the Bronze Age was  qualitatively greater than the degree of   the wrenching process that happened from  the initial transition to growing plants.  That’s surprising, because our cartoon  picture is that the big transition is farming.  But the biological readout is saying our  genome is reacting much more strongly to   these events that happened 5,000 years ago. You did some work with Bhatia and many other   colleagues in 2014 where you were looking at  20,000 or 30,000 African American genomes today.  You were saying, “Look, there’s 80% West  African DNA and then 20% European DNA.  Can we look at their genomes today and see that  their allele frequencies are much different than   we’d expect from this admixture?” Correct me if I’m wrong,   but you found that they weren’t. That is to say, over 200 or 300 years of extremely   intense environmental change—going from chattel  slavery to a completely new environment—there’s   no effect of natural selection. So we see episodes like this where   we don’t see natural selection, but then  the Bronze Age apparently must have had an   even stronger effect, where the change in  environment is even stronger than what we   see from Africans in Africa being migrated  to the New World and living under slavery.  That may be the case. It also may be the case that  that period is just too short to see much effect.  In the Bhatia et al. paper, where we looked at  about 30,000 African Americans, we looked to   see whether—instead of the average percentage of  around 80% West African ancestry—there were some   places in the DNA with significantly more  than 80%, or significantly less than 80%.  That’s what you would expect if there  were natural selection for some genetic   variant from Europeans or from Africans. We didn’t see any place in the DNA that   was significantly different from  what you would expect by chance.  One possible explanation is just that there’s  only a handful of generations, maybe five,   over which natural selection would operate. So if the selection was 2% a generation,   you would still only see a 10% compounded effect,  and there’s just not enough time to detect it.  But the Bronze Age is not  300 years, it’s 3,000 years.  It’s the power of compound interest, and you  have enough time to begin to see a strong effect.  This really, really does seem to be a very  impactful time in terms of human history,   and you can see it in our complex traits. Look at pigmentation, for example,   which is the strongest signal of selection  for a complex trait in our data set.  You look at genetic mutations that  are known to affect pigmentation.  You add up their effect across all of the  DNA, there’s dozens or hundreds of them.  You look to see when natural  selection is strongest, and the   time period is really 2,000 to 4,000 years ago. For some of these other traits as well, you see   again that the time period over which selection  is strongest is 2,000 to 4,000 years ago.  For example, if you look at genetic variants  that affect measures of cognitive performance,   such as performance on intelligence  tests in white British people today.  This is of course a very strange trait to  measure in the past because there were no   intelligence tests and there was no school. But it is a predictor today, and you can   look at how it’s changed in the past. We see very strong natural selection   for this combination of genetic variants  that predicts people’s performance on IQ   tests and is also highly correlated to the  predictor of the number of years of school   or the household wealth of people. All crazy traits in the past because   there was no wealth in the past,  there was no school in the past.  But if you look at the predictors today, there  is a strong movement in a systematic direction,   a large effect, about a standard deviation  on the scale of modern variation.  We can do this trick of looking to see  whether there are periods of time when   this natural selection has occurred  more intensely or less intensely.  We drag a 2,000-year window through our data, and  we repeat our whole analysis, not on 18,000 years,   but just on a short 2,000-year window. We can measure the strength of selection   in each of these 2,000-year windows. What you see when you look at intelligence   is that this maxes out in the Bronze  Age, between 5,000 and 2,000 years ago.  The impact in the last 2,000  years is almost nothing.  There’s no evidence of natural selection at all. Your bias coming into this, my bias   perhaps, might be that if there’s any signal of  natural selection on this trait at all, that it   would be unusually strong in the last 2,000 years. Maybe this is a time of industrialization.  Maybe this is a time of greater  need for this particular trait.  But in fact, there’s no evidence of natural  selection at all in the last 2,000 years.  There’s very strong evidence between 2,000  and 4,000 years ago, where instead of a one   standard deviation strength of selection,  it’s a two standard deviation strength,   averaged over this time period. The standard deviation here is how much   the polygenic score for the trait itself moves? How much the polygenic score for the trait moves   over a 10,000-year period within a population  that is held constant in terms of its ancestry.  What we’re actually doing is looking in our  data set at a heterogeneous group of people.  There’s Southern Europeans and Northern  Europeans and hunter-gatherers and farmers.  At different times in the past, those  groups are more or less represented.  The whole strength of the methodology Ali  Akbari developed is that it corrects for that   changing ancestry over time. Really what’s being asked here   is that we’ve divided up our whole data set  into an archipelago of little populations   in different places in space and time. We’re asking in each place in space and time:   a little pocket of people in Britain from 4,000  years ago to 3,500 years ago, a little pocket of   people in Hungary, a little pocket of people in  Italy from 2,000 years ago to 1,500 years ago.  In each of these places, where the ancestry is  relatively similar without being too disrupted in   that short period by migrations, we watch to see  if the genetic changes blow in the same direction.  We’re measuring the strength of selection at  each point in time after correcting for the big   population changes that have occurred. The effect here is huge then.  One standard deviation above the median  would be somebody in the 85th percentile.  You’re saying the effect of selection has been  so strong that comparing 10,000 years ago to now,   the median has gone to the 85th percentile. That’s just a huge effect over the last   10,000 years on something like intelligence  or the thing that predicts household income.  Especially given that this is only 2% of the  change in allele frequencies, and the 98% is   coming from migration… It’s stupendous to think  about what the impact of migration is, if this   alone is driving a standard deviation change in  these kinds of qualities, at least among the kind   of variation we see in the world today. One thing you can see in the data   is that the migration impact is huge. For example, if you look at the trajectory   for measures of cognitive performance—scores  on intelligence tests in white British people   today—but you look at the predictor of that  in people in ancient times, the estimate for   the hunter-gatherers of Europe is three standard  deviations below the modern mean. So that’s hugely   different. Then you see a huge jump from them to  the farmers, who are at the mean, at zero. That’s   migration. What you’re seeing is that those two  groups had different set points for those traits.  And then the steppe pastoralists  have a lower set value.  You see huge fluctuations in the predictor  of this trait over time. That doesn’t prove   selection. That’s just migration. But what our  test is telling you is: in addition to those   fluctuations due to migration, is there  a consistent effect of natural selection   blowing the trait in the same direction over all  places and times? That’s what we’re detecting.  There’s this theory called the  collective intelligence hypothesis,   which is the idea that selection for intelligence  has actually been in the opposite direction.  As society has developed,  there’s been more specialization,   and if there’s more specialization,  each person only needs to understand   a smaller and smaller part of the world. Therefore, the ancients were actually much smarter   than us, and we’ve evolved down in intelligence. Your results seem to point   in the opposite direction. Although there hasn’t been selection   in the last 2,000 years as society has gotten  more complicated, at least when society began,   there was more need for the kind of  thing that predicts intelligence today.  The reason that’s surprising is, if you  think about hunter-gatherers—reading   your colleague Joseph Henrich’s book—the amount of  information they needed to hold onto and assess,   everything from how to process food, to how  to build shelters, fire, et cetera, compared   to my world, where I just need to know how  to set up mics and ask questions… It seems   like the demands on intelligence should have  been way higher in the ancestral environment.  So it’s very surprising that the  beginnings of civilization increased   the selection on intelligence. This is the power of data.  I think if you asked Joe prior to this work what  the hunter-gatherer selection would be and where   their set point for this particular trait would  have been… I think he probably wouldn’t have   made a very strong prediction, but he would  have said, “Maybe you would have expected   it to have a high predicted value of this trait  because these people were really having to do a   lot of things and figure a lot of stuff out. Maybe once you have more complex societies,   there would be more of a collective brain, and  maybe there’d be selection against this trait.”  In fact, it’s the opposite in some ways.  It’s the power of data. It’s not what you   expect. It’s actually the value of data  to try to make sense of all these things.   It’s very interesting. The genetic predictor of  intelligence, there are lots of things that are   confusing about it, so it’s worth talking about. Or the genetic predictor of years of schooling,   which is highly correlated to  it and is measured even better.  If you look at the genetic predictor of years  of schooling, there’s another amazing study from   2017 from a group in Iceland that looked at this  measure over the last hundred years in Iceland.  It looked at older people and  younger people born more recently.  There’s an estimated 0.1 standard deviation  decrease in the genetic predictor of intelligence   in Iceland just within one century. It’s an absolutely huge   effect over a short period. This is selection against years of schooling.  If I said intelligence, I didn’t mean to. It’s selection against the genetic   predictors of the number of years of school. One possible interpretation of this—hand-wavy—is   that what’s being measured here is not selection  for years of schooling or for real intelligence,   but for another trait altogether  that’s correlated to both of them.  For example, the predictor of the number of years  of schooling is very strongly correlated to the   age at which women have their first kid. If you control for that, all of the   signal of years of schooling goes away. So maybe what you’re measuring is women’s   decision about when to have children. If you have children earlier,   you don’t go to school as much. If you have children later, you go to school more.  Maybe it’s some kind of measurement of delaying  gratification or putting things off or planning.  The same trait is correlated to body mass  index, to obesity, and to walking pace.  So is this really intelligence as we think about  it, or is it something else that manifests itself   differently at different times in the past? Obviously, a trait like years of schooling was   not itself a meaningful thing in the past. The underlying things for it seem   to have been under strong selection. Whatever in the genome predicts years of schooling   seems to have been under strong selection. How should we think about this?  What’s the actual thing  that’s changing in the genome?  There are two things going on  that you need to think about.  Years of schooling is connected to  so many other things genetically.  If you look at the genetic predictor of  years of schooling—this trait has been   measured in millions of people now—it’s  correlated to really surprising things.  It’s correlated to the age at which women have  their first kid. It’s correlated to people’s   obesity. It’s correlated to people’s walking pace. It’s correlated to people’s household wealth.  It’s correlated to a variety of other  traits that seem quite different from it.  If you think you’re actually measuring the  genetic prediction of intelligence, or actual   studiousness, you should think again because  there are many things that it’s correlated to.  There seems to be some kind of general trait that  you could maybe think of as executive function   or a propensity to defer gratification—I’m  just waving my hands—that is under selection.  It pushes all these traits in the same direction  one way or the other, and at different times in   the past, it’s advantageous or disadvantageous. When we found this signal of the genetic   propensity to go to school for more years as it  manifests itself in white British people today,   we were incredulous. How could this be? Maybe  this is a problem. So we did a few tests to   try to figure out whether this was real. One of the tests we did was that we looked   for a study where this measurement of the  number of years of school was done not in   Europeans, but in Chinese people in China. We looked at the effect size of many variants   as they affected the number of years of school in  China, and we saw whether they had a correlation   to the trajectory of those same genetic variants  in Europeans over the last 10,000 years.  These are two parts of the world  where the populations have been   essentially completely disconnected. There’s no way by chance that the   trajectory in Europeans over the last 10,000  years would have anything to do with the effect   on years of schooling in China today. But there’s actually a huge statistical   correlation, a five or six standard deviation  correlation between the effect size of variants   on the number of years of school in  China today and the trajectory in Europe.  It’s just as strong, actually, as the effect size  of variants in Europeans on years of school to   the trajectory in Europeans. We just could not see a way   this could happen by chance. Once we saw that, we felt quite   convinced that this was a real signal and that  somehow there has been natural selection to   increase the genetic changes that today manifest  themselves as predicting more years of schooling.  Just to make sure I understood, you’re  looking at this ancient DNA in Europe.  You’re saying it seems to predict years  of schooling for modern people in Europe,   or at least selection on that ancient DNA seems to  predict more years of schooling in modern Europe.  You also find that the same variants predict more  years of schooling for Chinese people in China.  So this is not just some weird artifact  from the way these GWAS were done in Europe.  These parts of the genome seem to robustly predict  the kind of thing that actually leads to more   years of schooling, at least in people today. Correct.  Jane Street is pretty secretive, but I did learn  about one internal mechanism which illustrates   how high trust and weird their culture is. Researchers aren’t given compute allocations.  Instead, Jane Streeters use an internal  currency called “hive bucks” to bid for   compute in real-time auctions. Everybody can spend as many   hive bucks as they want. But your hive buck bid is   meant to represent the real dollar value  of the experiment that you want to run.  Now notably during the auction,  anybody can change anybody else’s bid.  And after the auction, people  can even kill each other’s jobs.  People just trust each other to do this  in a way that benefits the whole firm.  As a result, Jane Suite’s allocations  reflect a near real-time consensus   on the highest priority uses of compute. As Axel, one of their ML engineers, put it:  “I think Jane Street is like pretty  bottom-up in terms of we have lots   of different researchers who are all  training their own models, sequence models,   all sorts of other weird and wonderful things.” By the way, with their new compute deal,   they’ve just added a six billion dollar hive  buck stimulus to their internal economy.  Jane Street is hiring researchers,  engineers, and interns.  Go to janestreet.com/dwarkesh to learn more. Stepping back, I want to understand what   this tells us about what actually changed in  our environments over the last 18,000 years.  We talked a little about what  happened after the Bronze Age.  We were talking about this during the collective  intelligence part of the conversation.  It’s surprising to me that things  like intelligence, or lack of   schizophrenia—things that just seem robustly  good—were not maxed out before the Bronze Age.  The diversity among different populations was so  big that you have the European hunter-gatherers   having three standard deviations less  predicted value for what they would score   on an intelligence test if it existed. But they were existing in the real   world in a place where intelligence matters. How can it be that this was not a trait… You   just look at the human body or any animal, and  evolution has been acting on it so strongly to   make it functional for the things it needs to do. And this one thing, which seems so   relevant—especially to what human  hunter-gatherers needed to do—doesn’t seem   to have been under that strong selection  in the Mesolithic or Paleolithic eras?  I think that’s a great question. As we talked about before,   selection is very effective. It can move the mean value of   traits within hundreds or thousands of  years in one direction or the other if   that’s adaptive in a particular environment. So you might wonder, isn’t intelligence   good in all contexts and places in time? There are a number of ways to think about that.  First of all, we are speaking from the  point of view of a society which intensely   values this particular trait, the ability to  score well on IQ tests or things like them,   or to go to school for a long time. I think it’s unprecedented in human   history that we live in a time like this. If you look at the Hebrew and Christian Bible,   and you look at how much intelligence  is valued, it’s basically not at all.  But when the Bible was being written,  especially the Old Testament, that’s   exactly when selection for intelligence is at  the highest point it’s apparently ever been.  Exactly. But there it’s about strength or courage  or religiosity. Those are the values. If you read   Homer or the texts of other religions,  it’s not intelligence. It’s beauty and   other things. This value system which has  a hyper-focus on smarts is not obviously a   trait value that’s been common in the past. You might think that in certain communities   there might be valuation of things that  are more proximate to years of schooling.  But really broadly, it’s not been  a high value in the population.  Obviously, the thing we care about is not direct  performance on an IQ test, especially in the past.  The thing I’m trying to understand  better is intelligence more broadly.  Maybe IQ-test intelligence is just not  that correlated with, “Here is a new-world   environment, go figure out how to process food  there and make shelter and everything else.”  Your colleagues like Joseph Henrich  have talked about how modern people   underestimate the difficulty of doing this  kind of thing with a small band of people.  Maybe that’s not IQ-test intelligence,  and that’s why we don’t see that strong   a selection effect on this thing. But intuitively, regardless of   the value system, it just seems very  valuable to have this trait maxed out.  I’m being very speculative. Let me give you two  examples of how I’m thinking about this, not that   I’m a particularly good authority on these things. As I mentioned, a lot of these traits,   which are quite disparate, are  highly correlated to each other.  Obesity, years of schooling, walking pace,  performance on IQ tests, household wealth,   all these crazy traits seem to be  governed to a substantial extent by   a shared combination of genetic variants. Let’s think about what this might mean.  In Iceland in the last hundred years, there’s been  selection against this combination of variants.  One possible interpretation is that  it’s basically selection for two   ways of investing in your children: having  many kids and not investing a lot in them,   or having few kids and investing more in them. If you invest in deferring having kids,   having more wealth, having more resources,  and putting more into each kid, you’re going   to have lower fertility and fewer kids. That’s going to result in lower fertility,   but those kids might survive  more and do better in society.  Alternatively, you can just have as many  kids as you can and invest less in them.  They might individually have less good outcomes,  but in a time of plenty—which is potentially   Iceland in the 20th century—it might make sense  to have more kids and invest less in them.  There’s a toggle between having more kids and  investing less in them, and having fewer kids   and investing more in excelling in various ways. You can imagine that at different times and in   different places… In ecology,  there are different ways.  Mammals often invest a lot with a pregnancy and a  small number of children, whereas fish will spawn   huge numbers of offspring into the river,  the great majority of whom will be eaten.  But that is an effective way to produce  offspring in certain conditions.  So there will be a toggle depending on  the environmental conditions back and   forth between investing in large numbers of  offspring with less investment, or smaller   numbers of offspring with more investment. Maybe we’re just seeing that move back and   forth over different places and times. Similarly, for schizophrenia and bipolar   disease, how could this ever be advantageous? Maybe what we’re seeing with these diseases is   a readout of some spectrum of traits that  in some contexts might be advantageous.  Maybe being anxious, imaginative, or neurotic  might be helpful in a shamanistic tradition or   a religious tradition which values people  who can have visions or be creative.  Maybe these are subclinical versions of  schizophrenia or bipolar disease that in   certain times may be advantageous and  in other times may be disadvantageous.  You might just be seeing selection for different  types of creativity or other thinking that can be   valuable in different contexts. I’m waving my hands here,   but my sense is that these complex traits  have not pushed in one direction because   there are advantages to both ends of the  spectrum, and there are multidimensional   impacts of these different traits. Julian Jaynes has this famous theory   in The Origin of Consciousness in  the Breakdown of the Bicameral Mind.  I’m butchering this, but fundamentally, the  way I understand it is that up until Homer,   basically everybody was schizophrenic. People genuinely thought that gods were real   people that you were communicating with. His claim is that ancient texts seem   to show people behaving in this way. You’re being asked to believe in visions.  Even today, there’s valuation in some religious  communities in communicating with God,   having visions, and having supernatural  communions. So I just don’t know. But I   think it’s super interesting to ask the question  of why certain traits are not always advantageous.  For schizophrenia and bipolar  disease, there is a sense in   which most of the mutations are disadvantageous. We can see that from the patterns of variation,   where the variants that are  risk factors tend to be low   frequency and they tend to be small effects. So another trait you find under selection is   the trend away from body fat since the  agricultural revolution. Why is that?  What you see is a reduction in the  combination of genetic mutations   that make you at risk for obesity, body mass  index, and similarly very correlated to it,   higher fat mass, higher waist-to-hip  ratio, and higher type 2 diabetes risk.  There is clear selection, by about a standard  deviation on the scale of modern variation   for these traits, reducing over the last  10,000 years in this part of the world.  What can be going on there? Why wasn’t there selection for   this combination of traits before? There’s a longstanding idea   known as the thrifty gene hypothesis. The idea is that once you have hunter-gatherer   populations that move into a farming environment  where there’s plentiful food, there is no longer   a need to the same extent to be able to build  up body fat to survive in times of stress,   because there are more constant stores of food. As a result, there will be natural selection   against body fat once you  move into an agricultural   environment and into periods of food plenty. Maybe what you’re seeing is that this group   of people in Europe and the Middle East over  the last 10,000 years has moved into a period   of relatively more stable food, where building up  stores of fat is not as advantageous, and there’s   been selection against this combination of traits. Europeans are actually relatively better protected   genetically against type 2 diabetes than some  other populations around the world, like African   Americans and Native Americans, that have perhaps  not been exposed to agriculture for as much time.  So you may be seeing the effect of more  exposure to more stable food accessibility.  This is also another way in which  the data go against a common story.  The common story is that hunter-gatherers  actually had much more stable diets because   they were more varied, and they weren’t reliant  on a single cereal or crop for their calories.  If one game went away, they had  other things they could scout for.  They could move locations more easily because  they weren’t tied down to the land. So they were   more food-stable. But if there’s been selection  against storage of body fat, that suggests that   as unstable and as common as famines might have  been in agricultural societies, it’s at least   more stable than what the hunter-gatherers had. There’s a timescale issue. You’re absolutely   right. As I understand it, I’m no anthropologist,  when there’s a hunt in traditional societies   or communities that hunt, people will often  gorge themselves, eat a huge amount, build up   a temporary store of fat, and then go multiple  days without eating meat until the next hunt.  There is this boom-and-bust access to  high-value nutrition that is not true   to the same extent in farming communities. On the flip side, famines are something that   occurs more commonly in agricultural societies,  but the timescale and the tempo of them is   very different from the hunting tempo. Maybe there’s a famine every three years.  Indeed, if you look at the bones of  farmers, at least in some communities,   there’s more stress in them, maybe due to a  famine every three years or every five years.  But selection might not be acting  on that three-year time period.  Your fat store from the latest  hunt is not going to carry you   through to the famine three years later. Survival of famines is a different thing   than building up body fat to be  able to survive two weeks later.  A random question I have. You were  mentioning that compared to these   other things which matter much more for fitness  in the ancestral environment—the immune system,   especially after the Bronze Age—all these other  things have mattered more than intelligence.  They’ve been under much more  selective pressure than intelligence.  That makes you wonder whether there’s much  more room at the top for intelligence.  If humans had been selected especially for  intelligence, they could have been much smarter.  The reason that’s relevant is that  we’re currently building AI systems,   which we’re trying to make as smart as possible.  In fact, the only goal of the  training process is intelligence.  We don’t have to worry about at the same  time making their immune systems powerful—  We have lots of energy to spend on it. And at the same time making   sure they’re not schizophrenic. I guess we kind of do worry about that.  But if intelligence has not been the dominant  trait under selection for humans over the last 10,   20, or 100,000 years, does that mean  there’s more room at the top for this trait?  I think there’s more room at the  top for a lot of these traits.  You can move height extremely in one  direction, much more than it is today.  You can move any of these traits much  more extreme in the other direction.  There are probably very strong  negatives to doing that.  You’re probably sacrificing other  things, and there are trade-offs.  But it’s highly likely that if natural  selection pushed any of these traits more   in one direction than it is, the mean would move. So all of this evolution since “Out of Africa” is   acting on alleles that already existed in the  pool of human variation from that first group   we were talking about last time, on the order  of 10,000 people, that exploded out of Africa.  Is it surprising that across all these different  traits, from cognitive profiles to disease   resistance to height, that one pool of people  contained so much latent variation that they   could supply enough stretchiness to accommodate  all of these different traits you’re studying now?  That’s a rich question, and I think the human  population has within it a tremendous amount   of variation for complex traits. There’s a huge amount of variation   that affects height. There’s a huge amount of   variation that affects body mass index. If you take all these mutations and set   them to the high-height variant, a person will be  extremely tall, like as tall as a tall building.  Of course, that will never happen. But if you take all these variants   that affect schizophrenia risk and you point them  all in the same direction, there will be extreme   risk or extreme protection for schizophrenia. For complex traits, ones underpinned by many   mutations, all the variation already exists to  move the population to a different adaptive set   point that’s optimal in the environment it’s in. If you push the population into a new environment,   within hundreds or thousands  of years, the population can   rapidly move to a new adaptive set point. There are some unusual traits, like the ability   to digest cow’s milk or protection against sickle  cell anemia, that require a single very important   mutation that may not yet exist in the population. You have to wait for the mutation to   occur in some people. When the populations are   relatively small, only 10,000 people, you  might have to wait dozens or hundreds of   generations for that mutation to arise. But when the populations are large,   there’s no mutation limit anymore. Every mutation that can occur does occur.  There are eight billion people in the world. There are maybe 30 new mutations every generation,   so that’s 240 billion new point  mutations every generation.  There are only three billion DNA bases in the  genome, so every mutation that can occur does   occur about 100 times every generation. We’re  not mutation-limited anymore. The mutations   can arise again. They do arise again.  But when the population is only 10,000,   you sometimes have to wait dozens or hundreds  of generations for the new mutation to occur.  How likely is it that the thing that changed  with the Bronze Age is just that the human   population was big enough? By 3000 BC, you go to   a population of 50 million-ish people. The population is big enough, and the gene flow   between different areas is high enough, such that  things which don’t have an overwhelming selection   coefficient, which aren’t overwhelmingly favored  by evolution, are finally visible to selection.  I think that’s not likely to be true, but it’s  an extremely interesting thing to think about.  Already when population sizes are on the order  of a million or so, every mutation that can occur   does occur within a few generations. That’s well before the Bronze Age if   you take the population even of a place  like Europe, but also of other places.  Or maybe it’s at the dawn of the  Bronze Age or the farming period.  The question you’re asking is whether,  when the population is small, maybe natural   selection doesn’t work effectively. A common thing people think about with   natural selection, which is true, is that in small  populations selection doesn’t work effectively.  That’s because mutations bop around in  frequency from generation to generation a   lot in a small population, just randomly. If you have a population size of 1,000,   mutations will bop around by a frequency  of one over 1,000 every generation.  If the selection coefficient is less than  that, it will be drowned in the random bopping   around of frequencies due to genetic drift. But that is already for a population of 1,000.  A 0.1% selection coefficient is very weak. We’re talking about 1% effects,   and that’s very strong. It will work very well even   in a population of size 1,000 or 10,000. If you are talking about mutations of the   type that will start rising only in large  populations but not small populations,   those are selection coefficients on the  scale of one over 10,000 or one over 100,000.  Those will take 10,000 or 100,000 generations  to rise in frequency, which is hundreds   of thousands or millions of years. That’s not going to do anything over   the timescale we’re talking about. There’s just  a timescale issue. We’re talking about strong,   measurable selection coefficients on the  order of half a percent or more in this study.  All of those are going to work in  small populations or large populations.  It’s not going to be affected  by the population size.  Interesting. You’re saying that more generally,  once you hit a given threshold of population, the   dominant factor is time span, not population size. Correct. It’s very interesting,   and it’s actually not widely understood. Speaking of data contradicting what you   might have otherwise assumed, one of the  papers you sent me beforehand, Mallick 2016,   found that there are no fixed differences between  modern and archaic humans 50,000 years ago.  We know this is the period in which the  so-called cognitive revolution happened,   and modernity started, and people are making art. Does this suggest that nothing biological   changed to make modern humans modern? The thing that happened was some cultural change?  How do we understand what this data tells us? Right. 100,000 to 50,000 years ago, there’s   a quickening of the pace of change in culture. You see the first extensive representational art,   bead necklaces, drawings on the wall,  and a rapidly increasing pace of   innovation in the types of tools that people use. The thought might be that there would have been   some important genetic switch, a kind of important  genetic change that occurred in the population and   swept to high frequency that everybody soon had. That made it possible to do these things.  Maybe some genes allowed people to have  complex, representational language, for example.  One thing we did in 2016 in this paper  by Swapan Mallick and colleagues was look   across the DNA for places that might  be expected to look like this, where   nearly all people living today share a common  ancestor maybe 100,000 or 200,000 years ago.  We looked really hard, and right  across all the DNA we could look at,   we couldn’t find anything more recent than  four or five hundred thousand years ago.  This is a crazy result because it looks  like there are no key selective sweeps   that have occurred in this period that  are ancestral to everyone living today.  We talked before about no selective  sweeps between Europeans and East Asians,   but there don’t even seem to be any selective  sweeps shared between all humans in this really   important period when a lot of evidence  in the material culture record appears.  It could be that there’s biological  adaptation in this period, but it’s polygenic.  There are lots of mutations that all shift in  the same direction to help the population move   to a new set point, but there’s no key biological  change that rises to high frequency in this time.  This group 50,000 years ago, are  they the ancestors of everybody   out of Africa or also some Africans? This is 100,000 to 50,000 years ago.  This is the population that’s  ancestral to West Africans,   to most East Africans, to all non-Africans. There are a couple of populations in Africa   that have substantial ancestry  coming from more divergent groups.  For example, Khoisan from Southern  Africa or Central African rainforest   hunter-gatherers have substantial fractions  of their ancestry from groups that diverged   maybe 200,000 years ago from the other lineages. But all of these groups today are able to go to   college and do everything everybody else does. There is no evidence that there is any key   mutation lacking in some groups  that is not present in the others.  The differences we see between different  groups of people, especially if this group   50,000 to 100,000 years ago had a very small  population size… I think last time we were   discussing on the order of 10,000 people. So almost everybody in the world,   or the variance we see between different  humans today, was latent in this group.  I get your point that if you just stack  up different things across the genome,   stacking them up really has a big effect. But it’s interesting that we have so many   different groups in the world today, and all that  diversity comes from a very small population size.  A lot of us in human genetics think that  our population contains within it the   clay that’s needed to make almost any trait. And that depending on environmental conditions   or selection conditions, the mean value of  these traits will move in different directions.  There’s an empirical question about  how much selection there’s been   in different human populations over time. One of the things this new work we’re   involved in is showing is that at least in the  last 18,000 years in this part of the world,   there has been significant movement, at  least for a handful of important traits.  We looked at more than 500 traits. About 100 complex traits showed   significant movement in a systematic  direction over this time period.  It really does seem that there is a response  to the environments people are living in that   has occurred over this period, and that is  potentially stronger than in previous periods.  Crusoe has an amazing ML infra team that  keeps finding clever ways to squeeze more   performance out of their hardware. For example, tokenization has become   a real bottleneck for agentic workloads. Agentic prompts are often extremely long.  They tend to have high KV cache hit rates,  which shrink the GPU’s pre-fill work.  This means that the tokenization step,  which is traditionally sequential,   is a much larger fraction of time to first token. To solve this, Crusoe built fastokens,   an open source Rust-based tokenizer which  parallelizes things in order to take   advantage of all the cores on modern CPUs. Crusoe had to get creative here because   the naive approach doesn’t work. For example, for pre-tokenization,   you can’t just split your text into chunks  and run regex because you’d end up with   issues whenever a word straddled the split. Crusoe solved this by giving each thread   an authority zone plus the ability to  read one kilobyte past its own edges.  This one kilobyte buffer guarantees that you  won’t misprocess a token, and the authority zone   guarantees that you won’t end up with duplicates. No cross-thread coordination required.  Crusoe combined this optimization with a  handful of other smart tweaks in order to   get up to 40% faster time-to-first  token on real production workloads.  To learn more, go to crusoe.ai/dwarkesh We were talking earlier about how there   are no fixed differences between humans  30,000 years ago and humans today.  So if there’s no genetic basis for the kind of  thing that allowed humans to have more symbolic   representation, have farming, et cetera—I think  I asked you this question last time we talked,   but especially with this context—why no farming  before the Ice Age? Genetically we were there.  That is such an interesting question.  Genetically we’re there. The common   ancestral population has all of the  ingredients for farming 50,000 years ago.  These people are distributed into different  parts of the world: the Americas 15,000 years   ago or whatever it is, New Guinea 40,000  years ago, East Asia, Europe, West Africa.  No farming developed before  11,000 or 12,000 years ago.  It only developed in the last 12,000  years, the period known as the Holocene,   which is the end of the Ice Age. If you talk to climate scientists   and archaeologists—I keep asking people  this question every time I meet someone   who’s an expert in this—how can it be  that farming develops in all these places?  Are we really living in such an unusual time? People tell me, indeed, we’re living in a very   unusual time on a scale of two million years. That is, 12,000 years ago we switched into this   period of not just warmth, but climate stability. It’s hard to believe that we’re   living in such a special time. But if you look at data from the bottoms   of ponds where you can measure the fluctuations of  temperatures using isotopic signatures, apparently   we’re in a period where it’s fluctuating a  lot less year to year, 10 years to 10 years,   and 100 years to 100 years. It’s a period of relative   stability that we are miraculously living in. When this period of relative stability happens,   it follows that multiple groups independently  turn to agriculture, even though they   all have the same genetic complement that arose  50,000, 100,000, 200,000, 300,000 years ago.  It’s a crazy observation that people  just accept, but it’s unbelievable.  Oh, so you increased the range there. You said 100,000, 200,000, 300,000 years ago.  Based on the genetic differences between modern  people and people from 300,000 years ago.  Do you basically think they’re  modern 300,000 years ago?  I don’t know. This is actively what I’m  thinking about all the time right now.  There’s a big transformation in terms of the  culture of humans 300,000 or 400,000 years ago:   this invention of Levallois technology, the  ability to make stone tools out of cores.  The Middle Stone Age Revolution, or the  Middle Paleolithic Revolution depending   on what you call it in Africa or Eurasia,  is a new way of making stone tools that’s   shared by Neanderthals and by modern humans,  but is not shared in East or South Asia.  It’s a big change, and it presumably  involves a cognitive change in order   to make this sort of technology. Then there’s a further change to   the Upper Paleolithic Later Stone Age, maybe  100,000 to 50,000 years ago, when there’s a   second transition with a new type of tool making,  but it’s not as revolutionary as the earlier one.  So when the cognitive leap happens is unclear. The diversification of the lineages leading to   people living today, like Khoisan Southern  Africans and rainforest hunter-gatherers,   all occurs more on the timescale  of 300,000 or 200,000 years.  All of these people are capable of  going to college and doing everything.  So it’s not obvious that the cognitive toolkit,  the behavioral toolkit, and the genetic abilities   were not all in place 200,000 or 300,000 years  ago, and that even Neanderthals had them.  It’s not obvious that this was not the case.  I just don’t know. You distribute these people   descended from this diversification that  happened 200,000 or 300,000 years ago to   different parts of the world, and then  after 12,000 years ago, you start having   agriculture popping up in different places. It’s an outstanding mystery of human history.  I find it unbelievable that we live in a  time period that climatologically is so   unique on a scale of two million years,  but my colleagues tell me it’s true.  The climate thing seems surprising given there  are so many different environments in which   agriculture was independently developed. I understand that across environments   the variance could have gone down. If it had only happened in one place   at one time, I could have bought that explanation. But the fact that they’re making maize in the New   World and they’ve got cereals in the Old World in  very different environments makes it surprising.  It’s very, very surprising. We accept  it, but it’s a crazy observation that   most normal people don’t realize. The thing that basically everybody   accepts is that the common ancestral  population of almost everybody in the world,   except for rainforest hunter-gatherers  and Khoisan, is around 70,000 years ago.  Everybody accepts that these people all have in  place the cognitive, behavioral, and intellectual   ingredients that are necessary for the farming  revolution and building state societies.  Because when these descendants get distributed  to West Africa, East Africa, the Americas,   Europe, South Asia, East Asia, New Guinea,  and so on, their descendants all do this.  They do it independently, semi-independently,  or demonstrably completely independently in   all these different parts of the world. The cognitive resources for doing this must   have all been in place, but it’s a very long fuse. It delays for 40,000 or 60,000 years in all these   different places after the common ancestral  population splits up, and then ignites into   agriculture and all these other things after  that point. It’s a crazy claim. Then you could   argue about whether the actual fuse is 300,000  years, from when Neanderthals separated and   from when different lineages of extant modern  humans separate, and that’s also plausible.  It’s a crazy set of things that  we’re being asked to believe.  Is it possible that agriculture existed, but  you didn’t have modern metallurgy or whatever   it was that allowed populations to explode  starting in 5000 BC with the Bronze Age?  Population-wise, it doesn’t seem  like much is happening from 10,000   BC to 5000 BC in the early Neolithic. Is it possible that they had farming   but they didn’t have copper or tin, which  you needed to go to the Middle East for,   to develop a civilization that could make use  of bronze at a large scale, and so they just   disappeared from the historical record? I think we would see their archaeology.  There are extraordinary developments in  the Americas which are entirely Stone Age.  You would see them today if  they had completely vanished?  Oh, yeah. We should go for a trip  to Teotihuacán in Mexico. It’s so   impressive. When I went there when I was 20,  it was totally as impressive as ancient Egypt.   It’s huge. It’s massive. It’s without metal. It’s even more impressive because it’s not   only without metal, but without animals  and without wheels, which is crazy.  The marble is just hauled without wheels. Right. Take any person who has an old   world superiority and take them to these  places, and they will not have it anymore.  It’s just extraordinary what’s in these places. These are people who separated 20,000 years ago at   least from the ancestors of East Asians and 40,000  years ago from the ancestors of West Eurasians.  They just had the same biological and cultural  shared toolkit from then, but there’s a long   fuse delay until all this stuff happens. It’s an amazing thing, and we don’t question it.  What are other questions you are  either investigating right now   or want to investigate, these kinds of  big picture questions of human history?  I’m perplexed. I don’t know if we talked about  it before, but I remain very confused about the   relationships between archaic and modern humans. We have genome sequences now from archaic humans   who lived in Europe, West Eurasia, and  Central Eurasia, and the Neanderthals.  We have archaic sequences from these  enigmatic Denisovans, who we now have   a skeleton for since we last talked. There’s now a skull that’s   been shown to be a Denisovan. We have data from lots of modern humans,   and there are really big mysteries about  the relationships amongst these groups.  Genetically, the Denisovans and  the Neanderthals are sisters.  They descend from a common ancestral  population 500,000 or 600,000 years ago.  That group descends 700,000 or 800,000 years  ago from the common ancestors of modern humans.  Genetically, the whole genome data says that  Neanderthals and Denisovans are archaic humans   from a common ancestral archaic population. But there are so many things shared between   Neanderthals and modern humans that  don’t seem to be shared with East Asians.  They both share Middle Stone Age  stone tools, Levallois technology,   this cognitively unique way of making  stone tools that wasn’t used in East Asia.  They both have the same mitochondrial  DNA and Y chromosome sequence.  The Y chromosome sequence of Neanderthals  and the mitochondrial DNA of Neanderthals,   is actually modern human that came through  interbreeding 200,000 or 300,000 years   ago and then shot up to 100% frequency. Neanderthals and modern humans are both the   product of mixture events that happened between  archaic and modern humans 300,000 or 200,000   years ago, demonstrably through patterns  of variation in ancient and modern DNA.  It feels that there’s something shared  between Neanderthals and modern humans   that’s not shared with Denisovans, even  though the vote of the whole genome says   that Denisovans and Neanderthals are related. One wonders whether there’s something connecting   Neanderthals and modern humans that’s different  from Denisovans, even though genome-wide,   Denisovans and Neanderthals cluster. I’m thinking about that all the time now.  Connecting them would be interbreeding  events or being in the same place   at the same time that we missed? There’s a known interbreeding event   from the lineage leading to modern humans into  Neanderthals, but it’s supposed to be only 5%.  I’m interested in the possibility that that 5% is  actually a sign of something much more impactful,   that somehow Neanderthals are in some  sense deeply modern in some ways,   and even though they get swamped by  archaic genes, they actually have   more of a modern impact than one would think. The Middle Stone Age and Middle Paleolithic   Revolution that they share with modern humans  is more fundamentally a part of who they are,   in some sense, than we think. Interesting. Sorry, when was   this interbreeding event? 300,000 to 200,000 years ago.  So the common ancestor between Neanderthals  and most humans alive today is potentially   more recent than the common ancestor  between all humans alive today.  Oh, for sure. Which is crazy.  Well, the divergence to all the archaic humans,  including Denisovans, is within human variation.  Wait, what? Yes. The average time   to the common ancestor of any two human  genes is one or two million years ago.  If you look at the copy of chromosome 3 you get  from your mother and the copy of chromosome 3 you   get from your father, the typical time they share  a common ancestor is one or two million years ago.  That’s before the split from  Neanderthals and Denisovans.  So there are many places in your DNA where  you’re more closely related to a Neanderthal   on your mother’s side than you are to your father. I’m sure there’s a simple explanation, but how?  It’s the same reason that if you  have a sister, in some places in   your DNA you’re more closely related to her  than you are to me because you share a parent.  But in other places you’re more closely related to  me than you are to your sister because you happen   not to share the same DNA from your parents. It’s just that the DNA we get from our common   ancestral population was already  quite variable 500,000 years ago,   700,000 years ago, a million years ago, and some  of us descend from some of those ancestors and   others descend from other of those ancestors. Neanderthals split from our lineage really close   in time on human evolutionary timescales, such  that in some places in our DNA we’re more closely   related to Neanderthals than to each other. Interesting. What are the other big questions?  That’s the main thing I’m  thinking about a lot these days.  I continue to be very obsessed with  questions about the spread of human   populations around the world and trying  to reconstruct that with ancient DNA.  After the recording ended, David started  spontaneously explaining a new theory he’s working   on about Neanderthal genetics on a whiteboard in  the room, which I ended up capturing on my iPhone.  The thing I’ve been thinking about a lot  recently is the possibility that maybe   we’re not thinking in the right way about the  relationship between archaic and modern humans.  The standard model is one where Denisovans—these  archaic humans that were found from ancient   DNA—and Neanderthals descend from a common  ancestral population 500,000 or 600,000 years ago,   and these two separate earlier, maybe 700,000 to  800,000 years ago, from the ancestors of modern   humans, people like us. That’s the big result   of a lot of studies since 2010. But there’s also evidence of an   interbreeding event that happened maybe 200,000 to  300,000 years ago that resulted in modern humans   contributing DNA to the ancestors of Neanderthals. So maybe 5% of the DNA of Neanderthals comes   from this interbreeding event, and  a lot of studies have shown this.  I’m very interested in this because, from the  archaeological record, Neanderthals and modern   humans look quite similar to each other, much more  similar to each other than a lot of them do to   Denisovans, these archaic humans in East Asia. For a lot of history, people have thought that   Neanderthals are our sisters. But in 2010, the sequencing of   the Denisovan genome made it very clear  that on average, Denisovans are closer to   Neanderthals than to modern humans. This was a very confusing result.  Most people now think that Neanderthals and  Denisovans descend from a common ancestral   population that separated earlier  from the ancestors of modern humans.  I’m interested in the possibility that  the right way to think about Neanderthals   is actually as somehow culturally modern humans,  even though genetically they’re mostly Denisovans.  The model I’m thinking about is motivated  by this archaeological phenomenon known   as the Middle Stone Age Revolution. If this is Africa and this is Europe, we know   that the new way of making stone tools—with cores  that were very carefully mined far away from the   locations they were used, made out of high-quality  stone like flint—starts being used 300,000 or   400,000 years ago, first in the Caucasus,  places like Georgia today, or East Africa.  This way of making stone  tools is quite revolutionary.  It is known in Europe as the Middle  Paleolithic and in Africa as the   Middle Stone Age, and is associated with much  more widespread use of fire and moving stone   around at much further distances than before. I’m interested in the idea that this is something   shared between modern humans and Neanderthals. There’s somehow some shared cultural feature   that’s absent in East Asia, and that  might have a relationship in the genetic   data and is somehow related to this 5% DNA. The idea I’m interested in is the possibility   that there is a population here that invents  the Middle Stone Age and the Middle Paleolithic,   sometimes called Levallois technology, and that  people from this population expand into Europe and   mix with the local archaic humans who are there. That is what this 5% interbreeding event is.  It happens 200,000 to 300,000 years ago. It produces a group that, as it expands across   this landscape in Europe, mostly picks up the  local DNA and becomes mostly archaic genetically,   but retains its modern human culture, the way of  making stone tools and some of its traditions.  One of the things that’s super interesting  about this is that if you actually look at   the genetics, across the whole genome,  Neanderthals and Denisovans cluster.  But if you look at the mitochondrial DNA—which  humans and Neanderthals get from their   moms—Neanderthals and modern humans cluster. If you look at the mitochondrial DNA,   Denisovans and modern humans share an  ancestor well more than 700,000 or 800,000   years ago, as you’d expect from the history. If you look at the Y chromosome that you get   from your dad, Denisovans and modern humans  share an ancestor more than 700,000 or 800,000   years ago, which is consistent with this history. But if you look at the Neanderthal mitochondrial   DNA, it’s only 300,000 to 450,000 years. If you look at the Y chromosome,   it’s only 300,000 to 450,000 years. What the current genetic work is asking us to   believe is that even though this is only 5% of the  whole genome, it introduces mitochondrial DNA and   Y chromosomes, and they jump up to 100% frequency. It’s kind of a crazy claim because the probability   of this occurring by chance is low,  maybe 5% times 5%, a very small number.  It’s what we actually all believe,  but it’s a very surprising event.  Somehow it’s accreted to all the findings in the  literature so that we make ourselves believe this,   but it seems unlikely on first principles  that somehow only 5% would introduce both   the Y chromosome and mitochondrial DNA. And it really does look like this.  There’s amazing data from a site in Spain  that’s 300,000 to 400,000 years old,   called Sima de los Huesos. They have a nuclear genome   that looks Neanderthal-like for most of  the genome, but their mitochondrial DNA   and Y chromosome are Denisovan-like. So it really looks like there was a   population related to modern humans that pushed  into this Sima de los Huesos-like population,   displaced its mitochondrial DNA and Y  chromosome, but kept the rest of its genome.  It really looks like something like this happened. The idea I’m playing with—and probably it’s wrong,   who knows—is that there’s a landscape…  This is Europe and you can break up   into a hundred or so demes, little areas. Modern humans get introduced at the bottom   right corner, in the Middle East or  somewhere, and they spread into Europe.  As this population spreads, there’s a  wave front of expansion, and they’re   interacting with the local archaic humans. The theory from simulations and studies of   different species like mammals and birds  shows that even if there’s a small amount   of interbreeding—when there’s an invasion or  a movement of expansion of one group into the   territory occupied by another—there’s  massive introgression of local genes.  The pioneers at the wave front will sometimes  interbreed with the local population.  There are so many of them around that their  DNA will get swamped by the local group,   so by the time they make it to the  other side, they’re largely local.  Maybe this is what we’re seeing. You have a modern human population   that’s matrilineal, for example, where  the transmission of making stone tools   this way is happening from mother to child. That’s why they’re retaining their mitochondrial   DNA, but by the time they get to the other  end of Europe, they’re mostly local archaic.  You end up with a 95% population replacement. This would explain why the mitochondrial DNA   is shared between Neanderthals and  modern humans, and it would also   explain why the mixture proportion is only 5%. The really interesting thing is that there’s   other evidence from studies of modern humans  showing that modern humans are also admixed.  The right way to think about this is  that modern humans are a mixture of   two groups that diverged maybe 1.5 million  years ago, and they come together 200,000   to 300,000 years ago with maybe 20%  ancestry from this archaic African group   and 80% ancestry from this early modern lineage. And that same group then mixes with Neanderthals,   and it’s 5% modern and 95% local. So you actually have this key   population that makes the Middle  Stone Age or Levallois technology.  It appears and expands in all directions—into  Europe and into Africa 200,000 to 300,000 years   ago—bringing this technology, new ideas,  and perhaps some genetic adaptations.  It expands into archaic humans in Europe,  mixes with the local population, and gets   95% replaced but still retains its cultural  features and maybe some genetic features.  It expands in Africa too, but here it’s  not 95% replaced, it’s only 20% replaced.  Probably the reason is that this  group is much more diverged.  It’s 1.5 million years diverged  rather than 700,000 to 800,000 years.  As a result, there are many more genetic  incompatibilities and barriers to gene flow.  But there’s still a lot of mixing, maybe 20%, and  we have evidence that this is a big mixture event.  So what you’re actually seeing is a modern human  expansion both into Europe and into Africa.  In one place, it forms Neanderthals. In another, it forms the   ancestors of everybody living today. But all of these groups are descended from   this key revolutionary event that happens here. We often talk about the revolutionary events 50   to 100,000 years ago, the more symbolic behavior  and so on, that first appears in Africa and   the Middle East and spreads beyond. But there’s also this earlier event,   and it is contemporaneous with the breakup  of all the different groups in Africa today,   the Khoisan Southern Africans and the  Central African rainforest hunter-gatherers.  One wonders whether this is an  equally important formative event.  If that’s true, it makes you think of  Neanderthals as somehow our cousins.  Their shared Y chromosome,  their shared mitochondrial DNA,   they share the formation of this 200,000 or  300,000-year-old event, and their shared toolkit.  Even though the genome is telling us they’re  cousins of Denisovans, the correct way to think   about them may, in an important sense,  be as close cousins of modern humans.  I have so many questions.  Do you have 15 more minutes?  First of all, what is going on with this group  of archaic Africans 1.5 million years ago?  Where in Africa are they, and what happens  to the portion of them that don’t form   modern humans? Do they survive? This is not from ancient DNA,   but from analysis of modern DNA from different  people, mostly in Africa, but also non-Africans.  In multiple studies—there’s at least  three, maybe four or five studies that   I know about—they have looked at the patterns  of variation in people today and say the data   in modern people today, including in Africans,  is not consistent with a homogeneous population.  It looks like a population that split well  more than a million years ago into multiple   groups—at least two, but maybe many—and then  came together a few hundred thousand years ago.  The papers have different models that they fit,  but they all have this feature of a split-up more   than a million years ago, and then on the  order of a few hundred thousand years ago,   a coming together and a remixture event forming  the ancestors of anatomically modern humans.  This includes the Khoisan  and whatever other groups?  Yes. All of these groups have this,  maybe in slightly different proportions.  So you ask, where are these people  living? Who knows. In this scenario,   the 80% is coming from the Caucasus or Northeast  Africa, where the Middle Stone Age forms.  It’s from this population that  the Middle Stone Age comes.  They mix with local groups, and who knows where  they are: Southern Africa, Western Africa,   Central Africa, Eastern Africa. We don’t have any ancient DNA,   but this is a very rich environment. People have been living there for seven   million years at least, and there would have  been different groups of people everywhere.  Probably it’s not just two  groups, it’s probably more.  The important theme here is  there’s evidence of substructure   that’s well more than a million years old. This place would have been a landscape full   of archaic humans that would have been differently  related to these expanding people and would have   admixed with them when they came through. So with the Neanderthals, the first   time around 300,000 years ago, our  ancestors share culture with them.  They share the Middle Stone Age technology,  but they don’t replace the population.  The technology spreads through culture, basically. It spreads through genes too. If you look at   Yamnaya in India, there’s almost no Yamnaya  ancestry in India. It’s just diluted down.   As Yamnaya expanded into Central Asia and into  Europe, it makes the Corded Ware. There’s a 25%   dilution. It expands back across Central Asia. It goes through the Hindu Kush and   gets into northern South Asia. It admixes more with local people.  Today, the most Yamnaya ancestry  you see in India is 20% or 10%.  Most people have less than 10% or 5%. There’s been a lot of mixture on the way,   but it is the tracer dye. It tracks Indo-European languages,   and important aspects of Indo-European  culture are coming through Yamnaya.  So if you know where to look, that tracer  dye is only 10%, 5%, or 2% in some groups.  But it’s the languages people speak,  and important shared cultural elements,   that connect them to people on the other  side of the Indo-European-speaking world.  So this 5%, you shouldn’t sneeze at it. That’s tracing something important in this model.  I understand that if things are transmitted  more through women… Sorry, let me back up.  I don’t understand why the maternal  mitochondrial DNA and the Y chromosome   would be especially privileged as the  spreading is happening. Can you explain that?  The reason I’m talking about these matrilineal  or patrilineal expansions is that I’m really   troubled, and have been troubled for many  years—especially in the last three or four   years—by the fact that the mitochondrial  DNA and Y chromosome cluster Neanderthals   and modern humans, but the rest of the  genome clusters Neanderthals and Denisovans.  This is a crazy result that is  not seen in any other species.  I’m very interested in  patterns that would explain it.  If you assume that there was a matrilineal  or a patrilineal expansion—it could be   either—then modern humans, when they were  expanding across the landscape of Europe,   retained their identity along one of the lines.  If it’s matrilineal, when they incorporate  a male from the local community,   he’s brought into the community, and the kids  are raised based on the culture of the mothers.  If it’s a patrilineal expansion and they  incorporate a female from the community,   she’s raised with the culture of the fathers. If that happens, it guarantees that one of   these two parts of the genome looks like it  does, because it’s a modern human expansion.  If it’s patrilineal, it will  retain the Y chromosome.  If it’s matrilineal, it will  retain the mitochondrial DNA.  So it will solve one of your two problems. But not both.  It won’t solve the other one, so  you need to solve the other one.  You can solve it either by natural  selection or by social selection.  By the way, patrilineality and matrilineality are  the rule, not the exception, in human communities.  Usually communities have continuity  along the male or the female line.  Usually it’s patrilineality,  sometimes it’s matrilineality.  You can also have phenomena like social selection. It could be that once you have kids of someone   whose father, for example, is from the outside  community… Usually in most communities,   females all reproduce. That’s typical  today. Usually women have kids if they can.  But men in traditional societies are actually  very variable in their reproductive success.  A large fraction of men never have kids. Then there’s a subset of men who have   many kids with many women. There’s competition among men for kids.  In this context, where males are competing for  access to females, female mate choice begins to   be an important process. You have a phenomenon   where if your dad is an archaic male, it  could be the case that you’re not going to   be as successful in the competition for local  females as if your dad is a non-archaic male.  Some simple social phenomenon  like that could explain the data,   and we actually see this in human society. For example, if I remember right, in Central   African rainforest hunter-gatherers,  there’s different treatment of boys   and girls depending on whether their  dad or mom is one group or the other.  I guess I don’t understand how  the maternal… The group spreads,   and it gets to the next front. They have kids.  From the humans that have just entered, the kids   will have the mitochondrial DNA from the humans. But from the existing people, they will have the   mitochondrial DNA of the archaic humans. Why are the people with the archaic   mitochondrial DNA not surviving? It’s a question. There are multiple   possible explanations, but it’s much easier  to explain that than both the mitochondrial   DNA and the Y chromosome. One possibility is that the   mitochondrial DNA was less biologically fit. Another possibility is that there’s social   discrimination against people based on whether  their parents are archaic or not, which is not   at all surprising in a human context. It’s the weakest link in this argument.  This argument’s probably wrong, but I’m  just telling you what I’m thinking about.  Okay, the Neanderthals. So 300,000 years  ago our lineage interacts with them,   but mostly their lineage survives, and  there’s cultural and genetic diffusion.  And then is it 70,000 years  ago that we interact again?  Yes. And they don’t survive.  The genetic ancestry doesn’t survive. The genetic ancestry doesn’t survive.   Presumably there was also other contact  between 300,000 years ago and 70,000 years ago.  Probably. But these are the  ones we are detecting currently.  Is it just contingent that one time  there’s this kind of diffusion where   most of the archaic genome survives, and  the other time it’s total replacement?  This is not at all surprising given the context. If you think about this model, this is 700,000 or   800,000 years ago. This is 300,000 years  ago. So this is 400,000 years separated.  You talked about the Bhatia paper with me earlier. That’s two populations 70,000 years separated.  There are no biological incompatibilities  between West Africans and Europeans.  There’s no natural selection against  biological incompatibilities.  We know when Neanderthals and modern  humans met and mixed, there were biological   incompatibilities. That was 700,000 years  ago. As populations become further apart,   biological incompatibilities rapidly develop,  probably as the square of the separation distance,   because you need pairs of interacting genes. Here, it would have been maybe only 400,000 years   separated between this lineage and that lineage. But here, it’s 1.2 million years. That’s a lot.   These are at the edge of not being able to produce  children. These are quite different humans. These   are actually three times closer than these. If you look at mixtures of humans today,   there are mixtures in Southern Africa  of people who are half this distance.  If you look at Khoisan and Bantu people  mixing in Southern Africa, like the Xhosa,   which is the population of Nelson Mandela,  these are groups that are separated by   almost 200,000 years, which is half of this.  Totally compatible. What you’re seeing is a   group that’s actually completely permeable  genetically, or nearly completely permeable.  This other one almost certainly has  substantial biological incompatibilities.  Because 200,000 or 300,000 years later, we  see interbreeding between Neanderthals and   modern humans, or between Denisovans and  modern humans, and there’s clear evidence   of incompatibility at that point. But this would be even bigger.  What you would expect to see is that as this  group spread, they would be moving into a   territory full of archaic humans. There would be some interbreeding,   but the kids would not be very fit. They would  die off. There would be a lot of infertility.  The barriers to gene flow and to  interbreeding would be greater.  To me, it’s not at all surprising that as  this group moves into Eurasia, you have   Eurasian archaics—the ancestors of  Denisovans—who are only 400,000 years   diverged from these people over here. And then you have African archaics,   and these are 1.2 million years diverged. They just don’t interbreed as much,   and you don’t get as much gene flow. But the key thing is the timing. It’s the   same time. It really feels like the signature  of an explosion of people from one place,   interacting with people here and  interacting with people there.  It’s the same cultural or technological  revolution impacting this place and that place,   and creating populations that are impacted by  this cultural revolution, which we know is the   case because they share the same toolkit. Some people argue that Levallois   technology is independently invented. But it’s very similar, and this model would   be a way that it could have the same origin. So there’s a culturally shared thread,   this shared toolkit. There’s a mitochondrial   DNA and Y chromosome thread. And then there is a shared timing thread,   which is they both form by mixture. Because otherwise you’d have to believe   that Neanderthals independently  developed Stone Age tools.  Yes, which is not inconceivable. But it’s a little  bit like believing that farming independently   developed in multiple parts of the world. Right. But it did.  It did. So as I said, this is probably wrong. I’m trying to tell you that we don’t really know   the world we live in. This is not obviously  wrong. In fact, to me, this is much more   plausible than the model we currently write down. It’s probably wrong, but it’s much more plausible.  It explains many more things,  and it’s no more complicated.  Interesting. Do you want to recapitulate  the thing you were saying about the   analogy to Ptolemy and the epicycles? I thought that was quite interesting.  I think the model that we’ve put together  collectively about the relationships between   archaic and modern humans has accreted over time. There was this idea that modern humans are   distinct and that Neanderthals and  Denisovans are sisters of each other.  Over time, we detected additional mixture  events, like this modern human into Neanderthal,   and then these other ones I didn’t even talk  about, like a super-divergent lineage going   into Denisovans and all this other stuff. We still say, “Oh, the whole genome says   Neanderthals and Denisovans are  sisters, so that’s the truth.”  We’ve patched it all together  and gotten it all to work.  You look at the mitochondrial DNA and the Y  chromosome, and they have this odd pattern,   and it’s improbable, but we can get that  to work if we invoke natural selection,   things like this. You patch it all together. It  reminds one of what happened in the ancient world,   where there was this idea that the sun revolves  around the Earth, but it doesn’t quite explain   the movements of the planets properly. In order to get the movements of the planets   to work right, Ptolemy and the astronomers made up  these epicycles, these special extra rotations and   movements to make everything work about right. It was such a convoluted model.  When Copernicus and colleagues suggested instead  that everything is revolving around the sun,   it simplified things ever so much. What was happening is that as astronomical   information accumulated, it kept being  contradictory to the standard model, but it could   be made to work by proposing another complication  and another complication and another complication.  This is not as fantastic as proposing that  everything revolves around the sun rather   than the Earth, but it is much simpler. And it actually explains many things.  What is counterintuitive or unexpected or  hard to accept about this alternative model?  What is the hesitation that  people have for adopting this?  I don’t know. Nobody’s thinking  about this model right now.  It just seems obviously a  very natural model to me.  The reason I ask is that Aristarchus, the  ancient Greek, had the heliocentric theory   because he had deduced how far the Earth  is from the sun and noticed other things.  But it was not adopted, because his fellow  Athenians were like, “Look, if we believe   that the Earth revolves around the sun, for it to  be the case that we don’t see relative movement   of the stars to the Earth, the only possible  explanation is that the stars are so far away that   it is just incomprehensible and implausible.” So the heliocentric theory was dismissed.  What I’m trying to ask is, what is the equivalent  here of “for this to work, the stars have to   be so far away that it’s inconceivable,”  where actually the stars are so far away?  Maybe we should adopt the implausible  implication that this theory gives us.  That’s a great question. I think we have to  assume that there’s a linkage between the   cultural transformations in Africa and Eurasia at  this time, and that’s not something the community   has really put together with the genetic data. There’s this thread in the genetics about   substructure in Africans, and then there’s  this whole world based on ancient DNA,   and they’ve never been put together. Nobody’s put together the now extensive   work on modern human substructure with  the extensive work based on ancient DNA   of archaic human relationships to modern humans. If you put them together, you realize they line up   in terms of their time of substructuring. I don’t know if that’s improbable. It   seems parsimonious to me. It also seems significant   that different groups of humans at this time  were capable of adopting Stone Age technology.  Once one group had figured it out, the  genetic difference between different human   lineages was not so big that you could  not show people how to use stone tools.  Who knows? It could be that  this was genetically driven.  We talked before about the time to  the common ancestor of human genes.  There’s nothing at 100,000 years or 150,000 years,  but there’s a lot at 400,000 or 500,000 years.  If that’s what happens, and you have a mutation  that occurs in the Caucasus, somewhere in the   Middle East, or Northeast Africa, there could  be key genetic mutations that make people able   to do this. Then this population expands. When it  moves into Europe, it’s swamped by local genes,   but there could be retention of those  genes through selection as it expands.  Maybe what you’re seeing is that  there are genetic developments.  Most of the discussion has been focused  on the 50,000 to 100,000-year event,   which is anatomically modern human behavior. But a lot of archaeologists think this is an   equally—if not more—profoundly  significant event in many ways.  Why is that not the event  we should be talking about?  You’re talking about how there are no  fixed differences between modern humans   and the humans 50,000 years ago. Do we know if there are any fixed   differences between the people 50,000  years ago and the people 300,000 years ago?  I think there are. Other than obviously these interbreedings.  If you look at the genetic variation  going back 300,000 or 400,000 years,   there do begin to be places where all  modern humans share common ancestry.  That’s another way of saying there begin  to be fixed differences at that time depth.  That is where you start seeing evidence  for possible fixed differences.  What’s happening, if everybody shares a  common ancestor 400,000 or 500,000 years ago,   is that there’s a single ancestor at that time. If you compared it to another population, they   would descend from a different lineage, so any  mutation that occurred ancestral to that single   ancestor would be a fixed difference. This is the time at which you can   begin to see fixed differences. But anatomically modern, cognitively   modern humans exist by the beginning of the Middle  Stone Age, before we’re breeding with this ancient   group of Africans or breeding with Neanderthals. Anatomically modern humans occur exactly here.   It’s the same moment. This is when they occur.  The people who have skeletal features like ours,   and Neanderthals, appear exactly then. This is when it all happens.  There is this disconnect between  anatomically modern humans in   the skeletal record and behaviorally modern  humans, which is 50,000 to 100,000 years ago.  Anatomically modern humans appear at  this time, and recognizable Neanderthals   appear roughly around this time, too. Interesting. But we don’t know what exactly   happens, if anything, between 200,000 years  ago and 50,000 years ago that goes from just   anatomical modernity to behavioral modernity. My understanding is no. They’re busy making   Levallois stone tools like Neanderthals  for 200,000 years, and they are not   more impressive than Neanderthals in  any obvious way, as I understand it.  Then there begins to be in the archaeological  record a quickening of behavioral traits, which   could be not genetic at all, or could be genetic. There are lots of arguments about this.  We were obsessed with intelligence  earlier in our conversation.  People are obsessed with art and  these things that seem important   to us but who knows what’s important? Interesting. Cool, thanks for the digression.  The work that I’ve been involved in has  consistently shown that I was wrong in   my biases coming into the work, and I’ve  really been almost traumatized by this.  Again and again, I’ve come into a project with  some kind of guess about what the data was   showing, and then the data doesn’t show that. For example, when I got involved in the   Neanderthal genome project helping to analyze  data looking at how archaic Neanderthals were   related to modern humans, I was part of  a group of scientists who had established   that non-Africans were a simple subset of African  variation and that there was no evidence at all   of Neanderthal interbreeding into the ancestors  of modern humans or other archaic interbreeding.  Different analyses that I and many  other people had done made it look   like non-African variation was just a  subset, a small sample of that in Africa,   and that could have fully explained the data. So when I was involved in analyzing the   Neanderthal DNA sequences, what happened  was I found this very strong evidence of   Neanderthals being more closely related  to non-Africans than to Africans.  It was very surprising, and I thought it  must be a mistake. I was quite incredulous.   I thought it was unlikely to be true because  other evidence that had been found before   seemed to point in the other direction. So I spent several years trying to make   these results go away, as did my colleagues, and  we just couldn’t make the results go away. They   just kept getting stronger. And this experience  working on natural selection was the same.  What we were convinced of was that natural  selection had been pretty quiescent in our species   over the last several hundred thousand years. Therefore, if we look at patterns of variation in   non-African people today, or in any people today,  we should see not a lot of selection going on.  Indeed, the first ancient DNA studies, beginning  in 2015 with this paper that we were involved in   with Ian Mathieson and colleagues, seemed  to show relatively small numbers of genetic   positions associated with natural selection. In 2015, we analyzed data from about 200   Europeans and Middle Easterners to try to  understand frequency changes over time.  We compared those ancient people who were the  sources of modern Europeans to people in Europe   today, and we looked at frequency differences  that were too extreme to be due to chance.  We were very excited to find 12 positions  that we were convinced were highly different   in frequency between Europeans today and what  we would expect, based on the history that we   and others had identified as the history  relating modern to ancient Europeans.  Some of these were known and some of these  were not known, and this was very exciting.  We hoped that as the numbers of  samples would increase and we got   higher resolution to be able to appreciate  differences in frequencies over time,   it would make it possible to detect far more. What was quite disappointing over the subsequent   decade is that that didn’t happen. For example, the largest study of   that type in 2024 by a group in Copenhagen  analyzed much better data than we had in   2015 and found only 21 positions that were  highly different in frequency across time.  While that was exciting—it was almost twice  as many as we had found in 2015—in a lot of   ways it was disappointing because the sample  size and data quality had gone up so much,   and yet this is all that was found. That suggested we might be hitting   an asymptote and might not be able to  get beyond where we currently were.  This approach to learning about biology,  which was very promising in theory,   might not produce a high yield. Maybe natural selection was quiescent,   and the reason we’re seeing so  few changes is that there has   not been a lot of adaptive directional selection. That was the situation we found ourselves in until   just a few years ago when we carried out this  study in our research group led by Ali Akbari.  What we did is we deployed a few  innovations to try to improve our   power to detect natural selection. One of them was that we just pumped a   lot of data into the system, increasing  the amount of data by about 14-fold.  The main thing that we do in this study is  report new data from about 10,000 individuals.  This is a very big increase in the  amount of data in the literature.  The total dataset size of ancient individuals  distributed over the last 18,000 years is   about 16,000 people. This is a large dataset.  It’s much larger than was previously possible,   and when you have more data, you can estimate  frequency changes with much more subtlety.  The data comes from only one part of the  world, which is Europe and the Middle East.  It’s not a more important part  of the world than other places,   but it’s the place where maybe 70-80% of  the data in the ancient DNA literature so   far comes from due to historical reasons. It provides us with a natural laboratory   where we can see what happens to the genome  in one place over time as environments change.  It’s really interesting to imagine doing this  type of analysis in other parts of the world,   and the comparative analyses are super  important and interesting, but this study   right now is about this one place in the world  where we have particularly fantastic data.  The other thing we did is that we  developed an entirely new methodology   that hadn’t been used in this area before. The methodology is based on a technique   that had been developed for finding risk  factors for a disease in medical studies.  A simple way to explain it is that we ask how  to predict the genetic type a person has based   on their pattern of relatedness to other people. We have a dataset of about 16,000 ancient people,   and 22,000 people if we include the modern people. Then we look at how closely related each of these   22,000 people are to each other, and we predict  the genetic type at each position in the DNA—at   10 million positions—based on the pattern of  relatedness to all of the other 22,000 people.  Then we ask if natural selection blowing  the frequency of the mutation in the same   direction in all geographic places and at  all times predicts the data a little bit   better than just knowing the relatedness  to all the other samples in the database.  We’re simply asking if the alternative  hypothesis—that selection has been   blowing in the same direction at  all times—explains the data better.  That’s a dumb assumption, because of  course, the truth is that natural selection   will have changed in frequency over time. But we’re just asking the simplest of questions:   whether assuming a constant rate of selection  explains the data more than not doing so.  To summarize to make sure I’ve understood,  you’re trying to make a model that predicts   allele frequency changes over time. You  have two different parts. One part is this   genetic relatedness matrix, which captures how  similar different genomes are to each other.  That should capture the impact of  different bottlenecks, of drift,   of population admixtures, and all those  things which affect the entire genome.  Then you have the separate thing, which is, if we  look at specific locations, can we just say that,   “Oh, this location has been selected  at whatever coefficient over time”?  And if we add some coefficient, does it become  easier to predict the allele frequency changes   than you would have just seen from this other  artifact, which is just looking at, “Oh, if you   look at the whole genome, are these guys in the  same, have they gone through the same bottlenecks?  Have they gone through the same drift,” etc.? That’s precisely right.  Okay, what did we learn? When we analyzed the data this way, we   looked at 10 million positions in the DNA across  these 22,000 people, 16,000 of whom were ancient.  We looked to see if there was more  change in this consistent direction   over time than you would expect by chance. When we analyzed the data, we found many   hundreds of places in the DNA that were  changing too much over time and in too   consistent a way to be explained by chance. There’s a bit of a statistical problem   in figuring out how many there are because  they’re so densely packed that they’re close   to each other and interfering with each other. But when you try to piece them out and say,   “Let’s count only one in each place in  the DNA and blank out the others,” we   find at least 479 positions that are all  independently pushing in the same way.  We are 99% confident that  those positions are real.  By another criteria of being more  than 50% confident that they’re real,   we think that about 3,800 positions  are all pushing in the same direction.  This is a crazy number of results given that  in our previous work and other people’s work,   there were at most a couple of dozen  discoveries coming from a single scan.  So when we got this result,  we were very surprised.  We thought it must be wrong, and we spent the  next couple of years trying to make the results   go away, but they just kept getting stronger. We were trying to look for some independent   type of evidence to tell us  whether these positions were real.  We stumbled on something really powerful for this  purpose that had not been used in this way before.  It relied on the fact that we had  very large numbers of discoveries,   many hundreds of discoveries or even thousands. We took a completely independent dataset, which   was the corpus of genome-wide association studies. These are studies that people have carried out in   hundreds of thousands of people, looking for  whether particular genetic mutations are more   common in people with high blood pressure than  with low blood pressure, or something like this.  We took the UK Biobank, which is about 500,000  people from Great Britain who have been measured   for hundreds of traits. The whole genomes of all   these people have been sequenced. For each of these traits, we could   look at whether each of these 10 million positions  are connected to this trait in a convincing way.  Out of 10 million positions, about 15%—about 1.5  million positions in the DNA—are predictive of   at least one of these several hundred traits. Then we could ask a question: is our natural   selection signal, our statistic, related  to whether a mutation causes high blood   pressure or some other trait? We slid our statistic for natural   selection upward, to a value of  one, two, three, four, or five.  As we did that, the enrichment  for genetic mutations that   affect traits got higher and higher. Whereas it was only 15% when we didn’t   use our selection statistic, when we required  the selection statistic to be above about five,   there was about a five-fold enrichment  for mutations that cause traits.  Sorry, what is a selection statistic? This is the statistic we use to measure   whether a mutation is changing over  time significantly in a non-zero way.  It can be approximately thought of as a normally  distributed statistic, a Gaussian statistic,   which is the number of standard deviations  the statistical value is away from zero,   where zero is no natural selection. It’s not exactly that, but it’s close to that.  If the statistic is above five, we  see about a five-fold enrichment   in mutations that affect a trait. Instead of 15% of the mutations that are   at random affecting the trait, it’s  60% or 70% that are affecting the trait   when we slide our statistic upward. This provides completely independent   evidence that these sites are real, and as you  slide above five, there’s no more enrichment.  Our interpretation of these results—which we  were able to validate and show made sense using   computer simulations of our process—is that once  you slide the statistic above five, essentially   all the signals of natural selection are real. Okay, just to make sure I understand.  You’re saying, in order to figure out what  alleles have been under selection, your model   assigns a statistic saying, “In order to explain  why this allele has a specific frequency,   we’re going to give it a selection statistic.” Independently, we run these studies on modern   populations where we say, “If you look at height,  eye color, intelligence, or whatever trait,   what are the parts of the genome  that are correlated with that trait?”  The higher the statistic you give it in  your study to explain allele frequency   changes over time as a result of selection,  the more probable it is that that region in   the genome is associated with traits that  have some functional thing we can measure.  That’s exactly right. This is  a brilliant idea that Ali had.  It abandons the traditional approach of assigning  statistical significance to mutations that cause   a trait because we’re just using an external  piece of information—the correlation to traits,   measured in a completely different way—to  read off the probability mutations are real.  We can ask how much enrichment for real signal  is there given a particular selection statistic.  If it’s halfway enriched to the plateau, we’re  able to show the correct interpretation is that   50% of the mutations are really selected. If it’s three-quarters of the way toward   the plateau, there’s a three-quarters  probability that the mutation is real.  If it’s 99% of the way to the plateau,  there’s a 99% probability that it’s real.  That gives us a calibrated estimate of the  probability that a particular position is   really under natural selection. A major concern here is that what   we’re actually seeing is not that these mutations  are really under selection, but rather that both   association to a disease and our selection signal  are due to some third thing that’s causing both   of them, which is a type of selection which  is not what we’re after, not selection to   adapt to new environments, but what’s called  background selection: selection against newly   arising bad mutations that are removed from the  population that tend to be concentrated in genes.  Genes are also the parts of the genome  that tend to be associated to traits.  This common process is causing both the  enrichment for trait signals and is also   causing the enrichment for selection signals  that we’re observing. That’s the concern. We   were super concerned about this. So what we did is we repeated this   enrichment analysis in slices of the DNA  that all were affected to the same extent by   background selection, by this rain of slightly bad  mutations, and we get exactly the same pattern.  We also repeated this experiment just  using mutations of the same frequencies   because there’s different statistical power to  detect these signals at different frequencies.  We see the same pattern where above  a value of the selection statistic   of around five, we get this plateau. The thing that changed that allowed   you to increase the amount of sequences  you’re generating by two orders of magnitude   is just the statistical method you’re  using to identify which part is human?  Or what exactly changed in 2014 and since then? There’s been a whole series of improvements.  The big ones have been the huge drop in  sequencing cost, which made it possible to   generate ancient DNA in the first place. The drop in cost has been a millionfold   since the late 2000s, and another maybe  one to two orders of magnitude from 2010   to today. That’s one big change. Another  change has been in-solution enrichment.  It’s been this way of taking a sample that  has very small percentages of human DNA,   but then suddenly creating a process that will  mean that the great majority of the sequences   that one’s analyzing will be useful for analyses. The approach that we used was we took the DNA   samples that we had, most of which were very  low percentages of human DNA—less than 10%,   often less than 1%—which is such a low proportion  that it’s prohibitively expensive to sequence   them and to just brute-force sequencing them given  the technology that we had available at the time.  We took these samples and washed them over  an artificially synthesized set of short DNA   fragments that targeted positions of the  DNA that we were interested in analyzing.  This is more than a million positions that are  highly variable in people, and we picked many   of these to be biologically interesting. We had a whole set of known biological   targets that affected traits in genome-wide  association studies, which is the way that   people look to see if there are particular  genetic variants in modern people that have   particular impacts in phenotypes and traits. And so, what we did is we had this artificially   synthesized set of DNA fragments that we  washed our ancient sample over, and it   bound the parts of the DNA that we targeted. The resulting sequence that we generated was   very enriched for the parts of the genome  that were informative about history.  Even though only 10% or 1% of the DNA was human,  it ended up that a very large fraction was from   the parts of the genome that we were interested  in, and it became economically efficient to do it.  What was the other 99% of the DNA? It’s mostly microbial. It’s from   bacteria and fungi that colonize  a person’s body after they die.  Depending on how they die, there’ll be  more or less of these bacteria and fungi.  When you typically sequence DNA from a person,  it’ll just be full of microbial sequence.  Sometimes the microbial  sequence is very interesting;   it might be pathogens that a person died of. There’s amazing work, for example, about different   plagues of malaria and Black Death and hepatitis  B and so on that have been obtained from the   sequences of these pathogens in people’s teeth  and other parts of their body when they died.  But we’re focusing here on the human DNA. This changed the amount of data that was possible   to produce from tens per year to hundreds  per year, and then we further roboticized   and industrialized the process so that there  were many hundreds or even thousands per year.  Just in our laboratory, we’ve been  generating genome-scale data from   more than 5,000 individuals per year. I know this is true also of several   other laboratories in the world now. This huge jump in data, this semi-exponential   or even super-exponential jump in some cases,  has made it possible to ask and answer questions.  While we were only on the order of 10 genome  sequences from humans in 2010, this year it’s   passed more than 20,000 reported sequences. There are several orders of magnitude increase,   and the questions we were able to ask in 2014 are  just not the same as the ones we can ask today.  Awesome. Excellent. David, thanks for your time. Thank you, Dwarkesh.