"David Reich – Bronze Age shock, the Neanderthal puzzle, & the sudden spread of farming" — Dwarkesh Patel

Episode summary

David Reich (Harvard professor of ancient DNA) returns to walk through a new preprint with Ali Akbari that overturns the prevailing view that natural selection has been quiescent in humans since the agricultural revolution. By scaling ancient DNA sequencing into the thousands of samples and developing a new statistical method that corrects for ancestry shifts from migration, they found that selection has actually accelerated — most dramatically around the Bronze Age (~5,000-3,000 years ago), affecting immune traits, metabolic traits, pigmentation, and the polygenic predictors of cognitive performance / years-of-schooling. The Bronze Age "wrenching" appears to be a bigger biological shock than the original farming transition.

Key arguments / segments

[00:00:00] Cold open: "humans were wrenched into a way of living so different from how hunter-gatherer ancestors lived that the organism had to adapt very strongly. The genetic data is saying our genome is reacting much more strongly to events 5,000 years ago than to the initial farming transition."
[00:02:00] The original ancient-DNA dream (16-17 years ago) was to learn biology by tracking gene-frequency changes over time. That dream wasn't realized — sample sizes were too small. A single genome reveals tons about history (thousands of ancestors captured in one sample) but only ~2 alleles for any specific selection question. Now sample sizes are finally large enough.
[00:04:00] Why frequency changes matter: when environment shifts, genetic variants helpful in the new environment systematically rise in frequency. Detection requires big samples to distinguish real selection from chance / drift.
[00:06:00] The mainstream view being overturned: comparing Europeans and East Asians (split ~40-50k years ago), there are almost no 100%-different mutations between them — evidence that selection has been quiescent since ancestral populations hit an "optimum" hundreds of thousands of years ago.
[00:08:00] What they actually found: 98% of frequency changes are non-selective (migration + drift), 2% is directional selection — but that 2% is "everywhere," tugging nearly every position. Genome is "vibrating with natural selection."
[00:10:00] Methodology: divide Europe + Middle East data into an "archipelago of little populations in space and time" (a few hundred years of relative isolation each between migrations). In each pocket, check whether candidate variants slowly blow in the same direction. If all arrows point the same way, selection is real.
[00:12:00] 7,200 positions identified at 50% confidence = ~3,600 real selected positions in the last 10,000 years. Trait enrichment: 4-5x for immune traits, strong for metabolic; "no detectable enrichment" for behavioral / psychiatric — but Reich is explicit that this is a statistical-power artifact, NOT evidence that behavioral traits aren't selected. They are; the gain-of-weak-effect architecture of behavioral traits just dilutes the signal across many more loci.
[00:17:00] Bronze Age intensification: immune-trait selection signal is much stronger 5,000 years ago to present than 10,000 to 5,000 years ago. The intensification happens after farming starts spreading, not during it.
[00:20:00] Specific examples: TYK2 variant for severe tuberculosis rocketed UP 8-6k years ago, then rocketed DOWN in the last 3k years — possibly reflecting tuberculosis becoming endemic. Hemochromatosis variant also reversed direction in the Bronze Age. Depigmentation (lighter skin) strongest between 4,000-2,000 years ago, then much weaker.
[00:23:00] Dwarkesh sidebar (note this for tracked-author signal): mid-episode he describes using Cursor to spin up multiple LLM instances against the relevant paper, having models critique each other, generating flashcards from the synthesis. "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." This is a real research workflow note from a domain-respected interviewer.
[00:27:00] African-American natural-selection negative result (Reich's 2014 study with Batia): no detectable selection in 30k African-American genomes after 200-300 years of intense environmental change post-Middle-Passage. Explanation: 200-300 years is just too short — only ~5 generations of compounding. Bronze Age is 3,000 years, "the power of compound interest."
[00:30:00] Polygenic predictor of cognitive performance / years-of-schooling: hunter-gatherers ~3 standard deviations BELOW modern mean (this is the migration effect, not selection). Then selection drove a 1-2 SD increase, peaking 4-2k years ago. Last 2,000 years: NO evidence of further selection.
[00:36:00] Pushback against Henrich's collective-intelligence hypothesis (that ancestral hunter-gatherers were smarter because they had to be self-sufficient). The data instead suggest selection FOR the predictors of years-of-schooling intensified in the Bronze Age.
[00:38:00] The "years of schooling" genetic predictor is correlated with: age-at-first-child, body-mass-index, walking pace, household wealth. Controlling for age-at-first-child kills the years-of-schooling signal entirely. Reich's interpretation: the underlying trait is probably "executive function / delay-of-gratification / planning," not intelligence per se, manifesting differently in different eras.
[00:42:00] Cross-population replication trick: the trajectory in 10,000-year European data correlates 5-6 standard deviations with effect-sizes for years-of-schooling in modern Chinese people in China. These populations have been disconnected for millennia — the correlation is implausible by chance. Strong evidence the signal is real.

Notable claims

98% of genetic-frequency change in the last 10,000 years is migration + drift; 2% is directional selection.
~3,600 confirmed selected positions in human DNA in the last 10,000 years (out of ~10M analyzed positions).
Immune-trait selection is 4-5x enriched vs background.
Bronze Age (~5,000-2,000 years ago) is a stronger biological inflection than the original Neolithic farming transition.
Polygenic predictor of cognitive performance / years-of-schooling moved ~1 SD over 10k years; effectively zero further selection in the last 2,000 years.
TYK2 tuberculosis-risk variant reversed direction (selected UP, then DOWN) around the Bronze Age, plausibly tracking the rise of endemic TB.
Iceland (Kong et al. 2017): 0.1 SD decrease in years-of-schooling polygenic score within one century. Reich's interpretation: this is selection on whatever the underlying executive-function / age-at-first-kid trait is, not on schooling per se.
African-American ancestry-mixing study (Reich + Batia, 30k samples): no detectable selection signal over 200-300 years — too short a window even for strong environmental change.
0.1% mention: Jane Street's internal "hive bucks" compute-auction market is a sponsor mid-episode read (not Reich-related; flagged for sponsor-context awareness).

Guests

David Reich — Professor of ancient DNA at Harvard Medical School. One of the founders of the field; runs the largest ancient-DNA lab focused on industrializing the data-generation pipeline. Previously interviewed by Dwarkesh (one of Dwarkesh's most popular episodes).
Ali Akbari — Co-author of the preprint. Started as a postdoc in Reich's lab 7 years ago, now permanent staff scientist. Built the methodology to detect selection while controlling for ancestry change from migration.

Mapping against Ray Data Co

Medium-strong, methodology-side. Three RDCO-relevant veins:

"Power of compound interest" framing for biological adaptation. Reich uses the compound-interest analog to explain why the Bronze Age window is detectable but the post-Middle-Passage window is not. This is a clean version of an RDCO-foundational claim — small selection coefficients over long horizons yield massive net effects, but you need the time window. Maps directly to the small-bet-portfolio thesis where compounding is the substrate, not the headline. Worth a concept-article candidate: "Compounding requires time-windows long enough to register."
Methodology of detecting weak signals against drift-dominated noise. Reich's 98%-drift / 2%-selection ratio is structurally identical to many decision problems where founders are trying to detect a weak intentional signal inside a noisy environment. The "archipelago of independent pockets, ask if all arrows point the same way" methodology is a beautiful template for noisy-signal detection — could be useful for any RDCO measurement work where you're trying to disentangle migration-like effects (audience churn, market churn) from selection-like effects (durable preference change).
The "executive function vs intelligence" reframe. Reich's argument that the years-of-schooling predictor is actually capturing "delay-of-gratification / planning / age-at-first-child" rather than "intelligence per se" is a strong reframe — the apparent trait is a manifestation of a deeper latent trait. Useful as a thought-tool when RDCO is measuring downstream metrics that may actually be tracking something different from what they claim.

Tracked-author candidates:

David Reich is already implicitly tracked via the prior 2024 Dwarkesh episode that was filed. Reinforces the case for explicit tracked-author entry.
Ali Akbari is the co-author behind the methodology — worth surfacing as a tracked-author candidate if Reich's lab continues to produce papers at this cadence. Lower priority than Reich.

Dwarkesh's Cursor workflow note (00:23:00) is a small but useful production-pattern signal: he's openly using multi-model parallel research with critique-and-flashcards. Confirms the workflow is mainstream-credible.

[[2025-10-04-dwarkesh-sutton-interview-thoughts]]
[[2025-10-17-dwarkesh-karpathy-ghosts-not-animals]]
[[2025-12-23-dwarkesh-what-are-we-scaling]]