David Reich – Bronze Age shock, the Neanderthal puzzle, & farming's sudden spread

ELI5 / TLDR

Until very recently, geneticists thought human evolution went quiet a few hundred thousand years ago — that we’d basically settled into our final form and the rest was history, not biology. David Reich’s lab just blew that idea up. By analysing DNA from 16,000 ancient skeletons, they show our genes have been rapidly adapting throughout the last 10,000 years, and especially during the Bronze Age (5,000 to 2,000 years ago). The trigger seems to be that the shift from hunting to farming-with-cities was a bigger biological shock than the shift from hunting to farming itself. In a long aside, Reich also floats a heretical new theory about Neanderthals: maybe they were less like distant cousins and more like culturally modern humans wearing mostly archaic genes — a kind of genetic Trojan horse for the first big toolkit revolution.

The Full Story

The dream that didn’t pan out — until now

When the ancient DNA field started in the late 2000s, the dream was to watch human biology change in real time. Pull DNA out of skeletons spaced across thousands of years, line up the frequencies of various genetic variants, and you should see selection at work — diet adaptations, immune adaptations, the lot. That dream basically did not happen for fifteen years. The field made its name on history (mass migrations, surprise replacements, sex-biased mixing) but kept failing on biology.

The problem was sample size. A single ancient genome is amazingly informative for history — Reich points out that one genome is effectively many people, because it’s a record of two parents, four grandparents, sixteen great-grandparents, and so on, all collapsed into one body. But for tracking how the frequency of one specific variant shifts over time, one body gives you one or two data points. To see selection, you need armies of skeletons. Until very recently, no one had them.

“We just didn’t have that until the last few years.”

The 2% that explains everything

Imagine the genome as a noisy ocean. Most of the waves you see are from migrations — one population walking into another, suddenly the local variant frequencies shift dramatically, but not because anything was being selected. People just brought different luggage. Reich’s group ran the numbers and found that 98% of all the frequency changes they see in ancient DNA are this kind of luggage swap (genetic drift, migrations, mixture). Only about 2% is actual directional selection — natural selection blowing a variant in a particular direction because it’s adaptive.

That 2% sounds tiny. But because the genome is so big, and because the team finally had enough samples (16,000 ancient individuals, 22,000 with moderns), they could detect it. And what they found is that the genome is “vibrating with natural selection.” Almost every position is being tugged in some direction. Where the previous biggest study (Copenhagen, 2024) had found 21 positions under selection, Reich’s group found 479 they’re 99% sure about, plus thousands more at lower confidence.

Concept first, name later: think of selection like a slow current under a choppy sea. The waves on top (migrations) hide it, but if you average enough water for long enough, you can prove the current is real and figure out which way it’s flowing.

The Bronze Age was apparently insane

Here’s the surprise. If you’d asked any reasonable person where the biggest biological shock to humans came from, they’d have said “farming.” Hunters become farmers, diet flips, sedentism arrives, animals get domesticated — surely the genome reacts.

The data say no. The genome reacts much more strongly to what came after — the Bronze Age, roughly 5,000 to 2,000 years ago. That’s when populations got dense, people started living right next to their animals (and their animals’ diseases), urban environments emerged, and trade networks connected previously isolated groups. The body of a Bronze Age farmer was being asked to do something its hunter-gatherer ancestor had never been asked to do.

“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.”

A few concrete examples of what got selected:

Lactase persistence (the variant that lets adults digest milk) — intensifies hard in the Bronze Age, when people start using cattle for milk and wool, not just meat.
TYK2 — a variant that increases your risk of severe tuberculosis. It rocketed up in frequency from 8,000 to 6,000 years ago to about 9-10%, then rocketed back down in the last 3,000 years. The likely reading: it was protective against something earlier, then tuberculosis became endemic in dense populations and the variant became a liability. Selection went into reverse.
Skin pigmentation — Europeans get lighter throughout the period, but the strongest depigmentation pulse is between 4,000 and 2,000 years ago.
Body fat — clear selection against obesity-related variants, consistent with the “thrifty gene” idea: once food becomes more reliable, banking fat is no longer worth it.
Hemochromatosis (iron overload) — selection reverses around the same Bronze Age window.

Imagine the body as a piece of software written for one operating system (hunter-gatherer life) being suddenly forced to run on a totally different one (urban Bronze Age). The genome is patching itself in real time, and you can watch the patches roll out.

The intelligence finding (handle with care)

The most loaded result in the paper involves variants that, in modern white British people, predict performance on intelligence tests and number of years of schooling. Reich’s group asked: did selection on these variants change over time?

The answer, in their data, is yes — strongly. Between 2,000 and 4,000 years ago, the polygenic score for “this thing that today predicts schooling and IQ” was being pushed upward by selection at roughly two standard deviations of strength. In the last 2,000 years, the signal vanishes — no selection at all. So the inflection was right at the dawn of complex society, not during industrialisation.

Reich is careful here, and Dwarkesh pushes hard on what “intelligence” even means in this context. A few things to keep in mind:

“Years of schooling is connected to so many other things genetically. 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.”

So the polygenic score isn’t measuring “smart” in any clean sense. It’s measuring something more like a general propensity to defer gratification, plan, and invest more per child rather than have many children. Think of it as an executive-function dial. In some environments the dial wants to go up (Bronze Age complex societies, apparently), in others down. There’s a famous Iceland study showing the same score has decreased by 0.1 standard deviations in just one century — selection pulling the other way as soon as conditions changed.

To prove this wasn’t a statistical artifact of how the European studies were done, the team checked whether the same variants predict years of schooling in Chinese people in China today. They do — with a five-to-six standard deviation correlation. Two populations disconnected for tens of thousands of years still show the same biology underneath. Whatever the score is measuring, it’s a real biological thing, not a quirk of European GWAS methodology.

The opposite finding (the “collective intelligence hypothesis” — that ancient hunter-gatherers were smarter because they had to know everything themselves) is not supported. Hunter-gatherers in the data sit three standard deviations below the modern mean on this predictor. Of course, “predictor of years of schooling” is not the same thing as “what hunter-gatherers needed to be good at” — but it’s a striking number.

Why no farming before the Ice Age ended?

This part is properly mind-bending. Genetically, all the ingredients for farming were in place at least 50,000 years ago, probably longer. The common ancestral population that fanned out across the world had everything they needed. And yet farming pops up nowhere on Earth until about 12,000 years ago — and then independently appears in multiple disconnected places (the Middle East, China, Mesoamerica, the Andes, New Guinea) within a few thousand years of each other.

What changed? According to climatologists, just one thing: stability. We are living in a freakishly calm climatic interval on a 2-million-year timescale. Year-to-year and decade-to-decade temperature variation dropped dramatically about 12,000 years ago, and farming requires that kind of predictability — you need to know that next spring will be roughly like this spring.

“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.”

This is one of those “oh that’s interesting” facts that quietly rewrites a chunk of your worldview. The agricultural revolution wasn’t really a human innovation. It was a planetary weather front lifting, and humans (who’d been ready for tens of thousands of years) finally got the green light.

The Neanderthal puzzle Reich can’t stop thinking about

After the formal interview ended, Reich grabbed a whiteboard and walked through a heretical theory he’s been chewing on. It’s worth taking slowly because it pushes against the standard story.

The standard story (post-2010, when the Denisovan genome was sequenced): Modern humans split from the lineage leading to Neanderthals and Denisovans about 700,000-800,000 years ago. Neanderthals and Denisovans then split from each other about 500,000-600,000 years ago. That makes Neanderthals and Denisovans “sister” archaic populations, with us as a more distant cousin. Genome-wide, the data clearly support this.

The puzzle: But there are two big things shared between Neanderthals and modern humans that are NOT shared with Denisovans:

The mitochondrial DNA (passed mom-to-child) clusters Neanderthals with us. Same time-depth as if we shared a common ancestor only ~300-450,000 years ago, not 700,000+.
The Y chromosome (passed dad-to-son) does the same.
The Middle Stone Age / Levallois stone-tool toolkit — a genuinely new way of making tools from carefully prepared flint cores — appears in both Neanderthals and modern humans, but not in East Asian Denisovans.

The current accepted explanation is that there was a small (~5%) interbreeding event from modern humans into Neanderthals 200-300,000 years ago, and somehow that 5% pulse swept the Y chromosome and mitochondrial DNA to 100% frequency in Neanderthals. Reich finds this explanation increasingly hard to swallow. The probability of both Y and mitochondrial sweeping from a 5% starting frequency is, on first principles, vanishingly small.

Reich’s alternative (which he stresses might be wrong): Imagine a single innovating population somewhere around the Caucasus or Northeast Africa about 300,000 years ago. They invent the Levallois toolkit. They expand outward. Two things happen:

They push north into Europe, where they meet local archaic humans (the people who’ll become Neanderthals). They interbreed extensively at the wave front, but because the locals outnumber them massively, the genetic signature gets diluted to about 5%. The locals’ DNA largely wins. But the toolkit and some lineage markers (mtDNA, Y) survive — possibly because the expansion was matrilineal or patrilineal, with the incoming group’s social structure preserving one parental line.
They push south and west across Africa, where they meet other archaic humans (who’d been separated for ~1.5 million years). Here the genetic incompatibilities are too strong — these archaic Africans are too distant to interbreed easily. So the modern signal stays much higher: about 80% modern, 20% archaic, producing the ancestors of all of us.

In Reich’s view, what we call “modern humans” and what we call “Neanderthals” might both be products of the same innovating population colliding with different archaic locals. We are 80% modern / 20% archaic African. Neanderthals are 5% modern / 95% archaic European. Same revolution, different recipes.

He compares the current standard model to Ptolemy’s epicycles — a once-elegant theory that’s been patched and patched to fit awkward new data, until you start to suspect that a different framework might explain everything more cleanly.

“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.”

The clincher (or at least the strong supporting evidence) is a 300,000-400,000 year old site in Spain called Sima de los Huesos. The skeletons there have nuclear DNA that looks Neanderthal-ish, but mitochondrial and Y chromosome DNA that looks Denisovan-ish. Exactly what you’d expect if a modern-ish population had pushed in and replaced the parental lineage markers but mostly got swamped by local DNA.

Two cousins, three relationships

One of the wilder side-facts: the average pair of human chromosomes shares a common ancestor 1-2 million years ago. This is older than the split between modern humans and Neanderthals. So it’s perfectly normal that, at many positions in your DNA, you’re more closely related to a Neanderthal than to your own father. Think of it like having a sister: at some chromosomes you and she share parental DNA, at others you don’t, and at those latter ones a stranger could be closer to her than you are.

Reich keeps emphasising that we’re at the early-data stage. Every time he comes into a project with a confident prediction, the data refute it. Non-Africans turned out to have Neanderthal DNA after he’d previously argued they didn’t. Natural selection turned out to be rampant after he’d previously thought it was quiet. The honest position is that the field is still very much in motion.

Key Takeaways

The genome is “vibrating with selection” — about 2% of all observed allele frequency change is directional selection, but it touches almost every position. The other 98% is migration and drift.
Bronze Age > Neolithic as biological shock — the shift to dense, urban, animal-cohabiting life was a bigger evolutionary forcing event than the original transition to agriculture.
Immune traits dominate the selection signal — about 4-5x enrichment. Metabolic traits also enriched. Behavioural/psychiatric traits are NOT obviously enriched, but only because they’re underpinned by many genes of small effect, which the method can’t yet resolve.
TYK2 is a textbook reversal — a tuberculosis-risk variant rose in frequency, then crashed once TB became endemic. Selection can flip direction within the same gene over a few thousand years.
The “intelligence” polygenic score is really executive-function-ish — correlates with years of schooling, household wealth, age at first child, walking pace, BMI. It’s measuring a deferred-gratification dial, not pure cognitive horsepower.
The score moved up most strongly 4,000-2,000 years ago — not during industrialisation. In the last 2,000 years there’s no detectable selection on it. In Iceland in the last century, it’s gone the other way (-0.1 SD).
Same predictor works in China — the variants associated with years of schooling in Britain show a 5-6 sigma correlation with the same trait in Chinese people. So the underlying biology is real, not an artifact of European study design.
Climate stability, not human innovation, may have triggered farming — the Holocene’s freakishly calm climate (relative to the last 2 million years) appears to be the actual unlock.
Selection coefficients of ~1% double a variant in dozens of generations — over 3,000 years (Bronze Age), even modest pressures compound into huge effects. The Bhatia 2014 paper on African Americans found no detectable selection, but only ~5 generations had passed — too short for compound interest to bite.
The Sima de los Huesos skeletons (300-400 kya, Spain) have Neanderthal-like nuclear DNA but Denisovan-like mtDNA and Y. Strong evidence that one population can push in and replace the parental-lineage markers while getting genetically swamped.
At many chromosomes, you’re closer to a Neanderthal than to your own father — because the typical time-to-common-ancestor for any two human gene copies is older than the modern-Neanderthal split.
Population size doesn’t really matter for selection above ~1% coefficient — what matters is time. The Bronze Age had time (3,000 years), African Americans post-slavery did not (5 generations).
All the cognitive equipment for farming was probably in place 200,000-300,000 years ago — which means roughly 200,000+ years of cognitively modern humans waited around doing essentially Neanderthal-level stuff before something (climate?) lit the fuse.
Reich’s heretical Neanderthal theory: Neanderthals may be the same Levallois-inventing population we descend from, just heavily diluted (~95%) by local European archaics, while we are the same population diluted ~20% by African archaics. Same Copernican-style simplification, fewer epicycles.

Claude’s Take

This is one of the more dense, useful Dwarkesh episodes — partly because Reich is in the middle of a paradigm shift in his own field and is willing to say the unsettling things out loud. The headline finding (the Bronze Age was when our genome really started screaming) is solid and well-supported by the new methodology. The intelligence stuff, which will get the most attention online, is presented honestly — Reich is transparent that the polygenic score is measuring “something” rather than measuring intelligence, and he resists the obvious overclaims. That’s refreshing.

The Neanderthal theory at the end is the most fun part. He’s clearly thinking out loud, says repeatedly that it’s probably wrong, and offers it more as a worked example of “the standard model has too many epicycles” than as a finished hypothesis. The Ptolemy comparison is a nice rhetorical move and the Sima de los Huesos data point is a genuinely strong piece of corroboration. Whether this lands or not, the meta-point is the right one: paleogenomics is at a stage where the data keeps refuting the models, and the field’s standard narrative deserves more skepticism than it gets in popular accounts.

Score 9/10. Marked down only because the conversation jumps around (the Bronze Age core paper, then a long Neanderthal aside, then the methodology at the end) and could have been organised more tightly. But the density of new and load-bearing facts is exceptional. The climate-stability-as-farming-trigger detail alone is worth the watch. The honest reckoning with how often Reich himself has been wrong is better than most of what passes for science communication.

The one section worth being a little wary of: the cognitive-traits discussion is going to get quoted out of context everywhere. Reich frames it well in person, but a single screenshot of “selection on intelligence peaked in the Bronze Age” will travel without the seven layers of caveat he wraps around it. Read the actual Akbari et al. paper before forming strong views.