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

Summary

• Natural selection was rampant, not quiescent — Reich’s lab (Ali Akbari’s new preprint) finds at least 479 positions in the genome under selection at 99% confidence and ~7,200 at 50% confidence (about 3,600 expected real), overturning decades of consensus that natural selection was quiescent. The catch that hid it: selection is only “2% of the frequency change” — 98% is other factors, especially migration, drift, and population structure — yet “the genome is vibrating with natural selection.”
• The Bronze Age, not farming, was the biological shock. Across trait after trait — immunity (4–5x enrichment), metabolism, depigmentation — selection intensifies 5,000–2,000 years ago, not at the Neolithic transition. Reich’s readout: humans “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… our genome is reacting much more strongly to these events that happened 5,000 years ago.”
• Selection on the genetic predictor of IQ-test performance moved ~1 standard deviation over 10,000 years — the median going to today’s 85th percentile — and it maxed out 5,000–2,000 years ago at 2 SD strength, with “no evidence of natural selection at all in the last 2,000 years.” This points against the collective-intelligence hypothesis (ancients smarter, civilization dumbs us down): “It’s the power of data. It’s not what you expect.”
• But the trait may not be “intelligence” — the schooling predictor correlates with age at first child, obesity, walking pace, and wealth; Reich’s hand-wave is “executive function or a propensity to defer gratification,” and Iceland shows it reversing at −0.1 SD in a single century. The signal is real: European trajectory correlates 5–6 SD with the same variants’ effect on years of schooling in China — populations “essentially completely disconnected.”
• The enabling asset is an industrialized data flywheel: 2015 found 12 selected positions; Copenhagen’s 2024 flagship found 21; this study’s 14-fold data increase (10,000 new ancient individuals, 16,000 total) plus a GWAS-calibration trick — selection statistic >5 lifts trait-association from 15% to 60–70% — found hundreds. Sequencing costs down a millionfold; Reich’s lab alone now generates genome-scale data for 5,000+ ancient individuals a year.
• Reich’s off-script heresy: Neanderthals may be “somehow culturally modern humans, even though genetically they’re mostly Denisovans” — the standard model’s 5% modern-human admixture would have carried their mtDNA and Y chromosome to 100% frequency, a “tracer dye,” like Yamnaya’s 5–10% in India tracking Indo-European language. The standard model, he argues, has accreted Ptolemaic epicycles; his alternative “explains many more things, and it’s no more complicated.”
• AI-relevant aside: intelligence was not the dominant selected trait in the periods discussed — value systems prized strength, courage, religiosity (“If you look at the Hebrew and Christian Bible… intelligence is valued basically not at all”) — implying substantial room at the top for a system optimized purely for it.

Deep dive

1. Ancient DNA’s unrealized dream: biology, not just history

• Reich’s setup: the field’s founding hope 16–17 years ago was tracking biological change through time. It succeeded spectacularly at history — migrations, mixture, sex-biased processes, people not descended from prior occupants of the same land — but “what’s not been successful is learning about biology.”
• The asymmetry, in his framing: one genome contains tens of thousands of ancestors, so a single individual positions “exquisitely” for history. But for a variant affecting skin pigmentation or milk digestion, one person is one or two samples — detecting a few-percent frequency shift requires sample sizes that only now exist.
• The prize: using “the experiment of nature that’s occurred in our history” — frequency changes “too extreme to be due to chance” reveal where environment pushed on biology.

2. The quiescence consensus — and why it was almost right

• Mainstream view for decades: selection flat for hundreds of thousands of years. Evidence: almost no mutations 100% different in frequency between Europeans and East Asians despite 1,500–2,000 generations since their split 40,000–50,000 years ago — enough time for a selected variant to rocket to fixation.
• The new decomposition across ~10 million variable positions: 98% of frequency change is drift, migration, and population structure; only 2% is directional selection. The signal was drowned, not absent — “instead of being quiescent, natural selection is everywhere.”
• Dwarkesh’s clarifying pushback — why doesn’t population replacement itself count as selection? Reich’s answer: it “could count, and may count, and probably should count in some respects,” but replacement is a whole-genome shift, possibly cultural; the test is whether one place in the DNA drives change differently from the rest of the genome.

3. The method: Europe as an archipelago of natural experiments

• The conceptual move: treat Europe and the Middle East as “an archipelago of little populations in space and time” — a pocket in Britain isolated a few centuries, one in Hungary, one in Italy — and ask in each whether a mutation shifts the same direction. “If all the arrows point in the same direction, we win.”
• Migrations are the worst moments to look: 4,500 years ago the steppe migration made 40–80% of European DNA Yamnaya, swinging almost all frequencies — not selection, just different evolutionary histories colliding. The method corrects for changing ancestry and measures the consistent residual push.

4. The haul: immunity screams, behavior whispers — but both are selected

• Headline counts: 479 positions at 99% confidence of real selection, ~7,200 at 50% (so ~3,600 real), with tens of thousands more at lower thresholds. Against ~100 GWAS trait sets, selected sites show a four- to five-fold enrichment for immune traits, strong enrichment for metabolic traits (obesity, Type 2 diabetes), and almost none for behavioral or psychiatric traits.
• Reich’s crucial caveat — the behavioral null is a power artifact, and “we have evidence that it’s a wrong conclusion”: immune traits ride few genes of strong effect; behavioral traits are spread across very large numbers of genes of weak effect below single-site detection. “It is not at all the case — and we can prove it’s not the case — that behavioral traits are not selected.”

5. The Bronze Age is the inflection point — farming wasn’t

• Selection on immune and metabolic traits accelerates in the last 5,000 years versus the previous 5,000. The best specimen: the TYK2 variant, today’s major risk factor for severe tuberculosis, rockets up to 9–10% frequency from 8,000–6,000 years ago, then rockets down in the last 3,000 — clear selection both ways, plausibly as TB became endemic 2,000–3,000 years ago. Similar reversals: a multiple sclerosis risk variant, hemochromatosis; ABO blood type B rising to 10% at A’s expense; depigmentation strongest 4,000–2,000 years ago.
• The mechanism sketch: Bronze Age density, animals, urban disease exchange created evolutionary mismatch — hunter-gatherer variants dropped into “an agricultural, Bronze Age, high-population-density, urban environment.” The surprise, per Reich: “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.”
• Dwarkesh’s puzzle — why would FADS1/2 (converting plant fatty acids for a cereal diet) peak in the Bronze Age rather than at agriculture’s dawn? Reich notes it’s ancient — “you see copies in archaic humans too” — with variants fluctuating back and forth across eras.

6. Why 3,000 years detects what 300 cannot

• Dwarkesh cites Reich’s own Bhatia et al. 2014: ~30,000 African American genomes, ~80% West African / 20% European, no locus deviating from expectation despite the extreme environmental shock of chattel slavery — apparently zero selection.
• Reich’s reconciliation: five generations is too short — at 2% selection per generation you’d see only a 10% compounded effect, undetectable. “But the Bronze Age is not 300 years, it’s 3,000 years. It’s the power of compound interest.”

7. Cognitive selection peaked when the Old Testament was being written — then stopped

• Sliding a 2,000-year window through the data: selection on the genetic predictor of intelligence-test performance maxes out 5,000–2,000 years ago at two standard deviations of strength — and “the impact in the last 2,000 years is almost nothing.” Reich flags the bias violated: you’d expect industrialization to intensify it, not the Bronze Age.
• The migration overlay is even bigger: European hunter-gatherers sit three standard deviations below the modern mean on the predictor; incoming farmers sit at zero; steppe pastoralists lower again. The selection test isolates the consistent within-population push from those jumps.
• Dwarkesh raises the collective intelligence hypothesis — specialization means “the ancients were actually much smarter than us.” The data point the other way, and against Henrich-style intuitions that hunter-gatherer information demands should have maxed the trait. Reich, twice: “This is the power of data. It’s not what you expect.”

8. But what is the trait? Schooling genes, Iceland’s reversal, and the China test

• Reich’s deconstruction: the years-of-schooling predictor correlates with age at first child, obesity, walking pace, household wealth — “if you think you’re actually measuring the genetic prediction of intelligence, or actual studiousness, you should think again.” Control for age at first birth and the schooling signal vanishes. His hedge, explicitly waved: “some kind of general trait that you could maybe think of as executive function or a propensity to defer gratification.”
• Direction can flip fast: a 2017 Icelandic study found a 0.1 SD decrease in one century in the schooling predictor — “an absolutely huge effect over a short period.”
• The validation that killed their incredulity: variants’ effect sizes on years of schooling in Chinese people in China correlate 5–6 SD with those variants’ 10,000-year trajectories in Europe — “just as strong, actually” as the European effect sizes — in populations with no shared history. “We just could not see a way this could happen by chance.”

9. Why wasn’t intelligence maxed out long ago?

• Reich’s first answer is cultural relativism about our own moment: “it’s unprecedented in human history that we live in a time like this.” Dwarkesh’s sharp interjection — the Bible was written exactly when selection for intelligence peaked — and Reich’s concession: “But there it’s about strength or courage or religiosity… If you read Homer… it’s beauty and other things.”
• His speculative mechanism (self-flagged: “I’m waving my hands”): a toggle between reproductive strategies — many kids with little investment versus few kids heavily invested — like mammals versus spawning fish, flipping with conditions. Iceland’s 20th-century plenty may favor quantity.
• On schizophrenia and bipolar: maybe “subclinical versions” — anxious, imaginative, neurotic — were advantageous in shamanistic or religious traditions valuing visions. Dwarkesh invokes Julian Jaynes’ bicameral mind (“up until Homer, basically everybody was schizophrenic”); Reich stays agnostic but notes risk variants are low-frequency and small-effect, so mostly disadvantageous.

10. The thrifty gene, and why famine timescales matter

• Clear selection against body fat, BMI, waist-to-hip ratio, and Type 2 diabetes risk — about one SD over 10,000 years. The thrifty-gene story: stable agricultural food stores made fat-banking less adaptive; Europeans are now better protected against T2D than African Americans or Native Americans with less agricultural exposure.
• Dwarkesh’s objection: hunter-gatherers supposedly had more stable, varied diets. Reich’s resolution is tempo: hunters gorge after a kill and fast for days — boom-and-bust fat storage — while farming famines hit every three or five years. “Your fat store from the latest hunt is not going to carry you through to the famine three years later.”

11. All the clay was already there — and population size isn’t the Bronze Age answer

• Reich’s view of standing variation: the ~10,000-person out-of-Africa population contained “the clay that’s needed to make almost any trait” — set all height variants one way and a person is “as tall as a tall building.” For complex traits, no new mutation is needed to shift the adaptive set point within hundreds or thousands of years; only single-mutation traits (lactase persistence, sickle cell protection) wait on mutation supply. Today, with 8 billion people and ~240 billion new point mutations per generation, “every mutation that can occur does occur about 100 times every generation.”
• Dwarkesh’s hypothesis — was the Bronze Age special just because populations hit ~50 million and weak selection became visible? Reich: “not likely to be true, but it’s an extremely interesting thing to think about.” A 1% coefficient works fine even in a population of 1,000; coefficients needing large populations take hundreds of thousands of years. “Once you hit a given threshold of population, the dominant factor is time span, not population size” — “actually not widely understood.”
• The AI hook Dwarkesh drives at: if immunity, not intelligence, dominated selection, is there “more room at the top” for a system trained purely on intelligence? Reich: “I think there’s more room at the top for a lot of these traits” — with trade-offs, but the mean would move.

12. No selective sweeps, no farming before the Holocene — “a crazy set of things we’re asked to believe”

• Mallick 2016: scanning for sweeps ancestral to all living humans, “we couldn’t find anything more recent than four or five hundred thousand years ago” — no key genetic switch behind the 100,000–50,000-year cognitive quickening. Any adaptation was polygenic; and there is no evidence of any key mutation lacking in any group — Khoisan and rainforest hunter-gatherers have substantial ancestry from groups that diverged maybe 200,000 years ago, and “all of these groups today are able to go to college and do everything everybody else does.”
• The farming puzzle: genetically capable populations distributed worldwide 50,000+ years ago, yet agriculture ignites independently — maize in the New World, cereals in the Old — only after 12,000 years ago. Climate scientists tell Reich the Holocene is uniquely stable on a two-million-year scale: “It’s a crazy observation that people just accept, but it’s unbelievable.”
• Against Dwarkesh’s “maybe early farmers just vanished from the record”: Dwarkesh points to the Americas’ entirely Stone Age achievements — Teotihuacán, “without metal… without animals and without wheels”, as impressive as ancient Egypt, and Reich agrees. “Take any person who has an old world superiority to these places, and they will not have it anymore.”

13. The whiteboard heresy: Neanderthals as culturally modern humans

• Captured after the recording ended: Reich is “very troubled” that mitochondrial DNA and Y chromosomes cluster Neanderthals with modern humans (300,000–450,000 years) while the rest of the genome clusters them with Denisovans — “a crazy result that is not seen in any other species.” The standard model asks us to believe that 5% modern admixture introduced mtDNA and Y that both swept to 100% frequency — “maybe 5% times 5%, a very small number. It’s what we actually all believe, but it’s a very surprising event.”
• His alternative: a population inventing Middle Stone Age / Levallois technology 300,000–400,000 years ago expands into Europe; wave-front dynamics (known from mammals and birds) mean pioneers get genetically swamped by local archaics — 95% replaced — while retaining culture and, via matrilineal or patrilineal transmission, mtDNA or Y. Sima de los Huesos in Spain fits: Neanderthal-like nuclear genome, Denisovan-like mtDNA and Y. Result: Neanderthals are “somehow culturally modern humans, even though genetically they’re mostly Denisovans.”
• The same expansion into Africa mixes ~20/80 with an archaic lineage 1.5 million years diverged — forming anatomically modern humans at the very same moment. “In one place, it forms Neanderthals. In another, it forms the ancestors of everybody living today.”

14. Tracer dye, incompatibility math, and Ptolemy’s epicycles

• Why 5% matters — the Yamnaya analogy: their ancestry in India today is mostly under 10%, “but it is the tracer dye” for Indo-European languages and culture. “So this 5%, you shouldn’t sneeze at it.”
• Why outcomes differed: biological incompatibilities grow roughly as the square of separation distance (pairs of interacting genes). Eurasian archaics were ~400,000 years diverged from the expanders; the African archaic lineage 1.2 million — “at the edge of not being able to produce children.” Khoisan–Bantu mixing today, ~200,000 years apart, is “totally compatible.”
• Reich’s meta-framing: the consensus model has accreted patches like Ptolemaic epicycles; his alternative “is probably wrong, but it’s much more plausible. It explains many more things, and it’s no more complicated.” Dwarkesh’s Aristarchus challenge — what’s the “stars must be impossibly far away” objection here? Reich’s honest answer: nobody has connected the modern-DNA African-substructure literature with the ancient-DNA archaic literature — “if you put them together, you realize they line up.”

15. The engine: 12 hits in 2015, hundreds now — and a scientist who keeps being wrong

• The trajectory that motivated the method: 2015 (Mathieson et al., ~200 ancient samples) found 12 selected positions; Copenhagen’s 2024 study, with far better data, found only 21 — suggesting an asymptote. Akbari’s approach broke it: predict each genotype from the relatedness matrix (absorbing drift, bottlenecks, admixture), then ask whether a constant selection coefficient explains the data better.
• The calibration trick: sliding the selection statistic above ~5 lifts the fraction of hits that affect a UK Biobank trait from the 15% baseline to 60–70%, plateauing — independent evidence the signals are real, robust to background-selection and frequency-matched controls. The plateau even yields calibrated per-site probabilities.
• The industrial base: in-solution enrichment targeting 1M+ positions rescued samples that are often <1% human DNA (the rest microbial); sequencing costs fell a millionfold since the late 2000s; Reich’s lab now produces genome-scale data on 5,000+ ancient individuals per year, with the field passing 20,000 reported sequences.
• Reich’s closing epistemic note — the through-line of his career: “I’ve really been almost traumatized by this.” He helped establish there was no Neanderthal admixture, then found the opposite in the Neanderthal genome data and “spent several years trying to make these results go away… They just kept getting stronger.” The selection result got the identical treatment — and the identical outcome.