All podcasts / Dwarkesh Podcast / Summary

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

2026-05-08 - 116 min - source - Read full transcript

Dwarkesh Patel (host)David Reich

Key insights

A new relatedness-based statistical method finds orders of magnitude more selection signals than any prior study.

By adding ~10,000 new ancient genomes (18,000-year window, ~16,000 ancient + ~22,000 total individuals) and testing whether assumed constant directional selection improves genotype prediction from a relatedness matrix, Reich and Ali Akbari find 479 positions at 99% confidence and roughly 3,800 at 50% confidence - versus a couple dozen from the best prior scans (12 in 2015, 21 in a 2024 Copenhagen study).

ancient-dna-methods

The method's validity was confirmed using a completely independent dataset: UK Biobank trait associations.

Sites with a high selection statistic are 4-5x more likely to also be known genome-wide-association hits for real traits (60-70% vs. a 15% baseline), and the enrichment plateaus above a threshold - evidence the signals are real biology, not an artifact of background selection, which they ruled out by re-running the analysis controlling for background-selection intensity and allele frequency.

ancient-dna-methods

Only ~2% of allele-frequency change is due to directional selection, yet that fraction is enough to shift complex traits by roughly a full standard deviation.

98% of frequency change in the genome comes from migration, admixture, and genetic drift - not selection. But the small directional-selection component, compounded over thousands of years, still pushed traits like cognitive-performance predictors from the population median to roughly the 85th percentile over 10,000 years.

natural-selection-intensity

Selection signals are strongly enriched for immune and metabolic traits, not behavioral traits - but that doesn't mean behavior wasn't selected.

Single-variant scans show ~4-5x enrichment for immune traits and clear enrichment for metabolic/obesity traits, with almost no signal for behavioral or psychiatric traits. Reich argues this reflects statistical power, not biology: behavioral traits are governed by far more genes of weaker individual effect, so the same total selective pressure is much harder to detect site-by-site. Polygenic-score analysis (aggregating across many sites) does show strong selection on cognitive-performance predictors.

natural-selection-intensity

Selection intensity is not constant - it spikes dramatically in the Bronze Age, more than during the original shift to farming.

Across many traits (depigmentation, TYK2/tuberculosis risk, FADS1/2 fat metabolism, cognitive-performance predictors), the strongest selection window is roughly 4,000-2,000 years ago - the Bronze and Iron Ages - rather than the Neolithic farming transition beginning ~11,000-12,000 years ago. Reich's hypothesis is that rising population density, closer contact with domesticated animals and their diseases, and intensified urban living constituted a bigger biological 'shock' than farming's initial adoption.

agricultural-transition

The genetic predictor of cognitive performance rose sharply during the Bronze Age and shows essentially no selection in the last 2,000 years.

The polygenic score correlated with IQ-test performance and years of schooling shows about a two-standard-deviation strength of selection in the 2,000-4,000-years-ago window (versus one standard deviation averaged over the full 10,000 years) and no detectable selection at all in the last 2,000 years - contradicting the intuition that industrialized, schooling-focused societies should show the strongest such selection.

intelligence-evolution

The 'years of schooling' genetic predictor is validated as real by cross-population replication in China, but likely reflects a broader trait than raw intelligence.

The same European ancient-DNA-derived selection trajectory correlates (5-6 standard deviations) with the effect sizes of the same variants on years of schooling in present-day Chinese people - populations disconnected for tens of thousands of years - ruling out a European-specific artifact. But because the predictor is also strongly correlated with age at first childbirth, obesity, and walking pace, Reich suggests it may capture something like executive function or delayed gratification rather than intelligence per se.

intelligence-evolution

Obesity-risk and body-fat predictors declined by about one standard deviation over 10,000 years, consistent with the thrifty gene hypothesis.

As populations shifted from hunter-gatherer boom-and-bust food access to more stable agricultural food supply, selection favored lower stored-fat set points; Reich notes Europeans are relatively better protected against Type 2 diabetes today than populations with a shorter history of agricultural exposure, such as some African American and Native American populations.

agricultural-transition

Farming arose independently in multiple regions only after ~12,000 years ago because the Holocene is an unusually climate-stable period, not because humans lacked the cognitive capacity earlier.

Genetic data show no fixed differences between humans 50,000 years ago and today, meaning the cognitive/behavioral toolkit for farming was already present tens of thousands of years before farming appeared. Climate scientists tell Reich the Holocene's temperature and year-to-year climate stability is unusual on a two-million-year timescale, which he treats as the leading (if still startling) explanation for the long delay and near-simultaneous independent invention of agriculture worldwide.

agricultural-transition

Reich proposes a new, unpublished model: Neanderthals may be best understood as genetically-swamped modern humans rather than a separate archaic lineage.

He hypothesizes a population that invented Levallois/Middle Stone Age tools near the Caucasus roughly 300,000 years ago expanded in two directions: into Europe, where it interbred with local archaic humans and was genetically swamped down to about 5% ancestry (forming what we call Neanderthals) while retaining its mitochondrial DNA, Y chromosome, and toolkit; and into Africa, where it mixed more substantially (~20%) with more divergent archaic Africans to form the ancestors of modern humans. In this model Neanderthals and modern humans share the same founding population and toolkit, differing mainly in which archaic group absorbed them.

neanderthal-origins

The model is motivated by a genetic puzzle: Neanderthal mitochondrial DNA and Y chromosome cluster with modern humans, but the whole genome clusters Neanderthals with Denisovans.

Whole-genome data places Neanderthals and Denisovans as sister lineages diverging from modern humans ~700,000-800,000 years ago, yet Neanderthal mitochondrial DNA and Y chromosome show only 300,000-450,000 years of divergence from modern humans - the same lineages the standard model attributes to just a ~5% interbreeding event. Reich argues the odds of a 5% admixture event replacing both uniparental lineages by chance are implausibly low, and instead points to a matrilineal or patrilineal population expansion (paralleled by the well-documented dilution of Yamnaya steppe ancestry into India while its cultural/linguistic signature persisted) as a more parsimonious explanation, comparing the standard model's growing list of patches to Ptolemaic epicycles versus a simpler heliocentric alternative.

neanderthal-origins

Because of deep coalescence times, a person can be genetically closer to a Neanderthal than to their own parent at some DNA sites.

The average time to the common ancestor of any two copies of a gene within humans today is 1-2 million years - older than the ~700,000-800,000 year split from Neanderthals/Denisovans - so at many individual genomic positions, a person's maternal and paternal DNA copies coalesce more recently with Neanderthal DNA than with each other, illustrating why genome-wide divergence estimates can diverge sharply from single-locus relatedness.

neanderthal-origins

Books referenced

The Origin of Consciousness in the Breakdown of the Bicameral Mind - Julian Jaynes - Dwarkesh cites Jaynes' theory that ancient people experienced visions/gods as literally real (proto-schizophrenia) while discussing why schizophrenia-linked genes might not be uniformly selected against.

Media referenced

Mallick et al. 2016 (Swapan Mallick and colleagues) - paper - Found no genetic positions where all modern humans share a common ancestor more recent than ~400,000-500,000 years ago, meaning no shared selective sweep coincides with the 'cognitive revolution' 50,000-100,000 years ago.
Bhatia et al. 2014 (African American ancestry study) - paper - Analyzed ~30,000 African American genomes for local deviations from the ~80% West African / 20% European ancestry average; found no significant natural-selection signal over ~300 years of extreme environmental change, used as a contrast case against the much stronger Bronze Age signal.
Mathieson et al. 2015 - paper - The first ancient-DNA natural-selection scan Reich's lab ran, using ~200 ancient Europeans/Middle Easterners; found only 12 confident selection signals, an early data point in the 'selection looks quiescent' consensus this new work overturns.
2024 Copenhagen group study - paper - A larger, higher-quality successor to the 2015 scan that still found only 21 selection signals, reinforcing the pre-2026 view that selection detection was hitting an asymptote before Reich and Akbari's new method.
2017 Iceland study on genetic predictor of years of schooling - paper - Found a 0.1 standard deviation decline in the genetic predictor of years of schooling within one century in Iceland, evidence that this polygenic score is under active, fast-moving selection.
Reich & Akbari 2026 preprint (unnamed in transcript) - paper - The new paper motivating the entire conversation: using ~10,000 new ancient genomes and a novel relatedness-based statistical method, it finds hundreds of confident and thousands of probable natural-selection signals in Europe/Middle East DNA over the last 18,000 years.

Companies

Cursor - Podcast sponsor; Dwarkesh mentions using it to run multiple models in parallel while prepping for the interview.
Jane Street - Podcast sponsor; mentioned for its internal 'hive bucks' compute-allocation auction system.
Crusoe - Podcast sponsor; its ML infra team built the open-source 'fastokens' tokenizer mentioned in the ad read.

Techniques and frameworks

Relatedness-matrix selection statistic (Ali Akbari's method) - The core new methodology: predicts each person's genotype at 10 million DNA positions from their genome-wide relatedness to ~22,000 other ancient and modern people, then tests whether assuming constant directional selection improves that prediction - dramatically more sensitive than prior single-variant frequency scans.
Levallois / Middle Stone Age (Middle Paleolithic) toolmaking technology - A distinctive method of carefully preparing stone cores, shared by Neanderthals and modern humans but absent in East Asia; central to Reich's new hypothesis about a shared origin population ~300,000 years ago.
Thrifty gene hypothesis - Invoked to explain why obesity/body-fat-risk alleles declined after the shift to farming: less need to store fat once food supply became more stable.
Collective intelligence hypothesis - The theory Dwarkesh raises (and Reich's data appears to contradict) that individual intelligence should have been selected against as societies specialized and offloaded cognition collectively.
AGES browser - An internet tool built by Ali Akbari and a colleague that lets researchers query the selection trajectory and evidence at each of the 10 million analyzed DNA positions.

Summary

David Reich returns to discuss a new preprint with his collaborator Ali Akbari that overturns the long-standing view that natural selection in humans has been largely dormant since the agricultural revolution. Using a novel relatedness-matrix statistical method and roughly 10,000 newly sequenced ancient genomes (bringing the total dataset to about 16,000 ancient and 22,000 total individuals across 18,000 years of European and Middle Eastern history), the team finds hundreds of confident and thousands of probable positions in the genome under directional selection - vastly more than the roughly two dozen signals found by any prior scan. They validate the signals independently using UK Biobank trait-association data, showing that high-scoring positions are far more likely to be linked to real traits like immune function, metabolism, and cognitive performance.

A central finding is that selection is not constant: it spikes dramatically during the Bronze Age, roughly 4,000 to 2,000 years ago, more intensely than during the initial Neolithic shift to farming itself. Reich attributes this to a "shock" of rising population density, closer contact with domesticated animals and their diseases, and rapidly changing ways of life. Traits from immune resistance (tuberculosis risk) to skin depigmentation to fat metabolism (FADS1/2) show their strongest selection signal in this window. Most strikingly, the genetic predictor of cognitive performance and years of schooling rose by roughly one standard deviation over 10,000 years, concentrated almost entirely in the Bronze Age, with essentially no detectable selection in the last 2,000 years - a result that runs counter to the "collective intelligence hypothesis" that individual smarts should matter less as societies specialize. Reich is careful to note this predictor correlates with age at first childbirth, obesity, and walking pace, suggesting it may reflect something closer to delayed gratification or executive function than intelligence in the colloquial sense; he validates it as a genuine, non-artifactual signal by showing the same variants predict years of schooling in present-day Chinese populations that have been genetically isolated from Europeans for tens of thousands of years.

The conversation also revisits why farming itself took so long to emerge. Since there are no fixed genetic differences between humans living 50,000 years ago and today, the cognitive capacity for agriculture was almost certainly present long before farming appeared roughly 12,000 years ago in multiple places independently. The leading explanation Reich has heard from climate scientists is that the Holocene is an unusually climate-stable period on a two-million-year timescale, which he still finds hard to accept as sufficient but has come to view as the best available account.

The episode's second half shifts to an unpublished, more speculative model Reich sketched out for Dwarkesh after the main recording, on the relationship between Neanderthals and modern humans. Standard whole-genome analysis places Neanderthals as closer to Denisovans than to modern humans, diverging around 700,000-800,000 years ago, with only about 5% of Neanderthal DNA coming from a later interbreeding event with modern humans some 200,000-300,000 years ago. But Neanderthal mitochondrial DNA and Y chromosome closely resemble modern human lineages - a pattern that shouldn't arise by chance from a mere 5% admixture event. Reich's alternative model proposes that a population which invented Levallois/Middle Stone Age toolmaking technology near the Caucasus around 300,000 years ago expanded in two directions: into Europe, where it was genetically swamped by local archaic humans down to about 5% ancestry while retaining its toolkit and uniparental DNA lineages (becoming what we call Neanderthals), and into Africa, where it mixed more substantially (about 20%) with more deeply diverged archaic African populations to become the ancestors of everyone alive today. Under this model, Neanderthals and modern humans share the same founding population and cultural toolkit, differing mainly in which local archaic group absorbed most of their ancestry - making Neanderthals, in an important sense, closer cultural cousins to us than the genome tree alone suggests. Reich repeatedly stresses how often his data has overturned his own prior assumptions, comparing the standard model's accumulation of special-case patches to Ptolemaic epicycles that a simpler, if initially less intuitive, model could resolve.

Notable Quotes

"Instead of being quiescent, natural selection is everywhere." - David Reich

"It's the power of data. It's not what you expect." - David Reich

"We spent the next couple of years trying to make the results go away, but they just kept getting stronger." - David Reich

"This is probably wrong. I'm trying to tell you that we don't really know the world we live in." - David Reich

"It's a crazy observation that most normal people don't realize." - David Reich, on the Holocene's climate stability