But there is another intuition, familiar from the internet’s favorite civilizational cycle: hard times create stronger, more organized people; good times eventually soften them. The meme is crude, but the underlying question is not stupid. Do pressure and adversity sometimes force societies to become more complex?

Ancient history does not always look like a record of progress under comfort. Major transitions in subsistence, technology, settlement, and social organization often occurred under stress: changing seasons, unstable ecologies, shifting food systems, and population movements that forced human groups to adapt. The question, then, is not whether cold weather was pleasant. It is whether long-term climate change, measured over timescales meaningful for human populations, is associated with changes in archaeological stage.

To test this, I linked ancient DNA samples to reconstructed temperature histories and asked a simple question: when a location became warmer or colder than it had been 1,000 years earlier, did the samples from that location tend to belong to a later archaeological stage?

This is not a study of modern global warming. The data come from deep history, long before industrial emissions, fossil-fuel economies, or modern climate politics. The temperature changes examined here are natural, slow-moving fluctuations across thousands of years of human prehistory.

The point is not to draw a political lesson about the present, but to ask a historical question: did long-run ecological change leave a detectable trace in the way ancient societies were organized?

The answer, in the current data, appears to be yes.

What Exactly Are We Measuring?

I am not measuring IQ, wealth, literacy, or state capacity here. The outcome is much simpler: the archaeological stage assigned to each ancient genome, such as Mesolithic, Neolithic, Bronze Age, Iron Age, and related period categories.

That is obviously cruder than a full database of institutions or technologies. But it has one major advantage: it stays close to the metadata attached to the actual ancient DNA samples, rather than importing a separate historical classification system.

The climate variable is also straightforward. For each sample, I measured how much the local temperature had changed over the previous millennium:

delta1000 = current temperature − temperature 1,000 years earlier

A positive value means the location had warmed over the previous thousand years. A negative value means it had cooled.

The models then ask whether this temperature change predicts archaeological stage after accounting for the obvious confounders: current temperature, time, sample coverage, ancestry principal components, and educational-attainment (EA) polygenic scores.

The EA PGS control is especially useful because it addresses a possible objection directly. If climate change predicts archaeological stage, is that really a climate signal, or is it just picking up genetic differences correlated with educational attainment? Adding EA PGS lets us test whether the climate-stage association survives after that source of variation is included.

The Main Result: Linear Temperature Change

To get the cleanest possible result, I didn’t just rely on default database labels. I used a manually verified subset of the data, meaning rows where the archaeological period assignment was rigorously cross-checked and verified against the original published supplementary materials to filter out noise.

It’s also vital to clarify the mechanics of the model: this does not estimate a separate, isolated cooling slope. Rather, it evaluates continuous temperature change across the board.

To make the direction intuitive and readable, Table 1 reports the coefficient as the implied effect of 1°C of cooling. Because the underlying model was fit as current - lag1000, the reported cooling effect is simply the sign-flipped coefficient of that linear model.

Table 1. Linear Temperature Change, Displayed as the Implied Cooling Effect

In this linear temperature-change model, the negative-delta direction implies that cooling over the previous 1,000 years predicts a higher ordered civilization stage.

Is this effect driven by genetics? How large is this shift in reality? And does this mean cold weather mechanically caused civilization?

Read more

In one world, it can mean a chain of neighboring communities, regular marriage contacts, shared routes, seasonal movement, trade, and overlapping political horizons. In another, it can mean a hard break: different worlds separated by mountains, forests, water, language, subsistence, and time.

Ancient DNA gives us a way to ask whether that intuition leaves a genetic trace. At the same physical distance, were older ancient populations more genetically differentiated than more recent ones?

After controlling for geographic distance and broad region-pair, older group pairs have higher Hudson FST. Put plainly: the same number of kilometers corresponded to more genetic distance in older periods.

First, distance matters today

Before asking whether distance mattered more in the past, we need a baseline.

Among present-day populations in the AADR dataset, populations that live farther apart tend to be more genetically different. This is hardly surprising, but it is important to establish before moving on to the ancient data.

To show this, I matched the modern AADR Hudson FST table to the geographic coordinates of each population and calculated the distance between population centers. I restricted the analysis to groups with at least five individuals and known coordinates.

The relationship is far from perfect, but it is clear: greater geographic distance generally corresponds to greater genetic distance.

Figure 1. Modern FST by geographic distance

The real question is whether a given distance meant the same thing throughout history. Did 1,000 kilometers separate populations more strongly 8,000 years ago than it did 2,000 years ago? Ancient DNA allows us to test exactly that.

The continuous trend

Even after accounting for how far apart two populations were and which broad regions they belonged to, older population pairs were consistently more genetically different from one another.

The effect is large enough to matter. For every additional 1,000 years into the past, Hudson FST increases by about 0.01.

In practical terms, this means that a journey of 1,000 kilometers connected people much less effectively in the distant past than it does in more recent periods. Populations separated by the same physical distance were genetically more distinct, suggesting that migration, trade, communication, and marriage networks became progressively better at bridging geographic space over time.

Table 1. Core model coefficients

One way to read the result is to fix the distance. At 1,000 km, the predicted FST rises steadily as pair age gets older.

Table 2. Predicted FST at 1,000 km by mean pair age

Because this is a linear model, the smooth spacing between rows is the fitted time slope. It should be read as a compact model summary, not as a claim that genetic distance changed mechanically by the same amount every millennium.

Figure 1. Predicted FST by age at fixed geographic distances

Figure 2 isolates the 1,000 km case. This is the most direct visualization of the question: hold distance fixed and move through time.

In this model, 1,000 km around 1,000 BP corresponds to predicted FST of about 0.064. The same 1,000 km around 9,000 BP corresponds to about 0.142. The exact numbers depend on the model and the AADR groups that survive the sample-size filter, but the direction is robust.

Subscribe now

Some concrete same-country-sized examples

The abstract version of the result is easy to miss. A more intuitive way to read it is to ask what distances inside modern countries would have implied in the Neolithic.

Modern countries did not exist in the Neolithic, of course. London and Edinburgh were not endpoints in a British population system, and Rome and Milan were not Italian cities. The point is only to give the distance scale a familiar shape. These are present-day city pairs separated by roughly 500-660 km.

Table 3. Model-predicted FST for same-country-sized European distances

Those values are large compared with modern within-country or near-within-country population proxies in the same AADR-derived modern FST benchmark.

Table 4. Modern within-country or regional proxy FST values

A 500 km separation inside present-day Europe often corresponds to tiny modern FST values, on the order of a few thousandths in these proxies. In the ancient model, a similar geographic separation around 6000-8000 BP corresponds to predicted FST around 0.11-0.13, similar to the distance between East Asians and Europeans. That is not a small difference in scale.

What this means historically

Older ancient populations were more genetically patchy over space. The same physical distance separated more differentiated populations in the deeper past than in later periods.

That does not require a single mechanism. Several processes can produce the pattern. Small effective population sizes can make local drift stronger. Sparse contact networks can allow neighboring regions to diverge. Later expansions and mixtures can smooth older discontinuities. Farming, pastoralism, metallurgy, trade, ships, horses, roads, empires, and marriage networks can all make distance less absolute, but FST does not tell us which one is responsible in any particular case.

Methods

I use FST as the genetic-distance measure. An FST of zero would mean two groups are genetically indistinguishable at the measured markers; higher values mean greater allele-frequency differentiation. Geographic distance is the great-circle distance between group centroids.

The unit of analysis is AADR Group ID pairs, not individual pairs. I first KING-pruned individuals, retained ancient groups with at least 25 post-pruning individuals, computed Hudson FST between retained groups, and then modeled group-pair FST as a function of geography and time.

The retained dataset contains 141 ancient groups, 7,494 individuals, and 9,870 group-pair comparisons. For the region-controlled continuous model, pairs with unclassified broad-region labels were dropped, leaving 9,045 group pairs.

The core model is simple:

Hudson FST ~ geographic distance + mean pair age + broad region-pair

Mean pair age is the mean of the two groups’ median Years BP. Broad region-pair is a dummy for combinations such as Europe-Europe, Europe-West Asia, West Asia-West Asia, and so on.

References

Bhatia, G., Patterson, N., Sankararaman, S. & Price, A.L. (2013). Estimating and interpreting FST: The impact of rare variants. *Genome Research*, 23(9), 1514-1521. https://doi.org/10.1101/gr.154831.113

Chang, C.C., Chow, C.C., Tellier, L.C.A.M., Vattikuti, S., Purcell, S.M. & Lee, J.J. (2015). Second-generation PLINK: rising to the challenge of larger and richer datasets. *GigaScience*, 4, 7. https://doi.org/10.1186/s13742-015-0047-8

PifferPilfer is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.Mallick, S., Micco, A., Mah, M., Ringbauer, H., Lazaridis, I., Olalde, I., Patterson, N. & Reich, D. (2024). The Allen Ancient DNA Resource (AADR) a curated compendium of ancient human genomes. *Scientific Data*, 11, 182. https://doi.org/10.1038/s41597-024-03031-7

Manichaikul, A., Mychaleckyj, J.C., Rich, S.S., Daly, K., Sale, M. & Chen, W.-M. (2010). Robust relationship inference in genome-wide association studies. *Bioinformatics*, 26(22), 2867-2873. https://doi.org/10.1093/bioinformatics/btq559

Educational attainment (EA) was the next obvious trait to examine.

Ancient people did not attend universities, take standardized tests, or spend years in formal schooling. Yet the modern educational attainment polygenic score captures much more than education itself. It reflects thousands of genetic variants associated with cognition, persistence, long-term planning, and the ability to navigate complex social systems.

That raises an intriguing question:

When a region became colder than it had been a millennium earlier, did the genetic profile associated with educational attainment also change?

The answer depends on how the question is asked.

Using my primary climate-history model, the effect is essentially zero. But when cooling and warming are allowed to operate differently, a more interesting pattern appears: cooling predicts higher EA polygenic scores, while warming shows little evidence of producing the reverse effect.

Unlike height, where the signal is straightforward, EA turns out to be a more complicated and perhaps more revealing case.

I will not repeat the full methodological details here because they are identical to those used in the previous height and personality analyses. This post uses the same AADR sample, climate reconstructions, Mundlak within/between-location framework, ancestry controls, and robustness procedures described earlier.

The only thing that changes is the outcome variable: educational attainment polygenic score (EA PGS).

The first question is simple: if a location became colder or warmer than it had been a millennium earlier, did its EA polygenic scores change as well?

To test that, I added 1,000-year temperature change to the baseline model and asked whether it improved prediction.

Table 1. Linear 1,000-Year Temperature Change and EA PGS

The result is straightforward: there is no detectable linear relationship.

Adding 1,000-year temperature change does not improve prediction of EA polygenic scores in either the pooled or Mundlak specifications.

If this were the only model examined, the conclusion would be simple:

There is no evidence that millennium-scale climate change influenced ancient EA polygenic scores.

Goodbye cold-winters theory!

The Split-Slope Diagnostic Changes the Picture

The problem with the linear model is that it assumes cooling and warming are mirror images of one another.

Nature does not necessarily behave that way.

To test this possibility, I replaced the single temperature-change term with separate cooling and warming variables:

• Cooling = max(−Δ1000, 0)

• Warming = max(Δ1000, 0)

This allows cooling and warming to have independent effects.

Read more

To be absolutely clear: I do not mean personality in the loose, everyday sense. I mean modern Genome-Wide Association Study (GWAS) weighted scores projected into ancient genomes. I am looking at Polygenic Scores (PGS) for the Big Five personality traits, plus two sparse psychiatric comparison scores for autism and schizophrenia as a social-behavioral sanity check.

The question I am asking is simple: When a specific geography became warmer or colder than it had been 1,000 years earlier, did later ancient samples from that exact place carry different psychological PGS?

To find out, I pushed the data through the exact same rigorous controls I used for the height analysis. Here is how the experiment was structured, where the signals emerged, and what happened when I forced the model to look closer.

Read more

When applied to human population history, this framework generates a straightforward prediction: populations in colder regions should carry a higher density of height-increasing alleles. If you look at a static modern genetic map, you’ll see that higher height polygenic scores (PGS) heavily cluster in colder northern latitudes, seemingly validating the model.

But this static observation is deeply flawed because it conflates two entirely different tests:

• The Geographic Correlation: Are colder places, on average, cross-sectionally associated with higher polygenic scores for height?

• The Historical Dynamics: When a specific location experiences a temporal shift in temperature, do subsequent generations in that exact place undergo a parallel genetic shift in height?

Relying on the geographic cross-section is highly problematic. Because human groups migrate, a static map cannot determine whether a genetic pattern is a direct adaptation to local climate or simply a reflection of old population structure and ancestral movements.

To run a direct, unconfounded test, we need a kind of natural experiment. In a true experiment, an experimenter manipulates an independent variable to observe its isolated effect on a dependent variable. In paleogenomics, we obviously cannot manipulate the climate of the past. However, by looking within the exact same location over time, we can let history run its own natural experiment. When a specific region experiences a sudden cooling or warming episode while its physical geography remains entirely constant, we can observe the direct effect of that climate “manipulation” on the subsequent gene pool.

To execute this, I reran an ancient-DNA climate model using a statistical framework that explicitly decouples stable differences between regions from these historical climate shifts occurring within those exact same regions.

The signal isn’t merely that cold places have higher allele scores. Rather, a 1,000-year local temperature drop adds predictive information that geographic variance alone cannot account for.

Controlling for Population Structure

The primary vulnerability of simpler climate-genetic models is that they risk merely rediscovering geography. Northern and eastern ancient populations differ genetically from southern ones due to deep ancestral population structure, and they happen to have lived in colder climates. If a model links baseline cold to height PGS without accounting for this, it is likely tracking old migrations rather than localized climate response.

To set a more rigorous baseline, I isolated the temporal component by asking: Given the average long-run temperature of a specific location, does the localized change in temperature over the previous 1,000 years add any predictive value?

Using data from ancient individuals in the Allen Ancient DNA Resource (AADR v66) matched with paleoclimate reconstructions from CHELSA-TraCE21k, I calculated a localized historical lag:

Delta1000 = Current Window Temperature - Temperature 1,000 Years Prior

• Negative values indicate local cooling over the preceding millennium.

• Positive values indicate local warming.

Figure 1. Local 1,000-Year Temperature Change Across Ancient Sample Locations

While a map shows us where these climate shifts occurred, it doesn't give us a clear sense of the raw data's balance. To see exactly how many individual samples lived through historical anomalies versus stable baselines, we can strip away the map coordinates and plot each individual's current-window winter temperature directly against its own 1,000-year historical lag.

Figure 2. Current vs. 1,000-Year-Lagged Winter Temperature

To prevent spatial population structure from driving the results, I utilized a Mundlak specification (a correlated-random-effects approach). This framework formally decomposes the climate variables into two distinct components:

• The between-location component: Captures whether a location is generally colder or warmer than other regions globally (susceptible to migration confounding).

• The within-location component: Captures whether a specific sample’s local climate was colder or warmer than that exact location’s own long-run historical baseline (the direct test).

Note on the genetic metric: This analysis tracks the frequency of modern GWAS height-associated alleles moving through ancient genomes. It measures shifts in weighted polygenic scores, not direct osteological measurements.

What the Data Says

When you force the model to look strictly at historical temperature change within a location rather than the latitude of the sample, the model’s predictive power improves significantly.

Table 1. Does 1,000-Year Temperature Change Add Predictive Value?

The key metric here is Delta adj. R2, which shows whether adding the 1,000-year temperature change improves the model after baseline temperature, geography, and ancestry controls are already factored in.

The cleanest, most robust result comes from winter temperatures (coldest-quarter). Adding the previous millennium’s temperature trend significantly improves the Mundlak model (p = 0.0015).

Figure 3. Local Temperature Change and Ancient Height PGS

In a standard linear regression, the resulting coefficient (beta = -0.0291, p = 0.0186) implies a clean, symmetrical relationship: cooling predicts higher height polygenic scores, and warming predicts lower ones.

We can see this general direction when we split the raw data by the direction of the climate trend:

Table 2. Height PGS by Direction of 1,000-Year Winter Temperature Change

The Symmetrical Trap: It’s Only the Cold That Matters

Linear models force data to fit a perfectly symmetrical narrative: if cold does X, warmth must do the exact opposite of X. But evolutionary history and climate shifts are rarely so neat.

To test if this relationship was actually symmetrical, I ran a split-slope diagnostic. This allowed the model to calculate entirely separate slopes for cooling trends and warming trends.

Table 3. Asymmetric Cooling and Warming Slopes

(Note: Both cooling and warming are calculated here as positive magnitudes in degrees Celsius to compare their direct strengths.)

This sensitivity check shatters the clean linear assumption. For the height score, the winter cooling slope carries almost the entire signal. It is strong, pronounced, and statistically significant (p = 0.019). Meanwhile, the warming slope—though pointing in the expected negative direction—is weak, flat, and statistical noise (p = 0.430).

Furthermore, the between-location version of this change variable shows absolutely no effect (beta = -0.008, p = 0.725). If this entire signal were just a proxy for “traditionally cold regions have taller people,” that between-location term would be dominant. Instead, the genetic action is happening strictly within populations experiencing a historical downswing in winter temperatures.

Why Climate History Beats Geography

Height is a highly heritable physical trait with well-documented links to thermoregulation, resource availability, and metabolic constraints. If an arbitrary, non-physical trait showed this exact same specific pattern, it could easily be dismissed as an artifact of statistical noise.

What this framework changes is how we evaluate climate-genetic interactions. By shifting the focus to within-location historical deviations, we move past the uninterpretable cross-sectional map. We are no longer just looking at a continent and observing that the colder north happens to have a higher genetic predisposition for height.

Instead, we are isolating specific historical points. We are showing that when a region grew colder over a millennium, later individuals from that exact location tended to display higher height polygenic scores.

While this signal cannot yet disentangle the precise underlying mechanism, which could involve direct selection, local climate-driven population replacements, or complex migratory mixtures, it demonstrates that a population’s localized climate history carries predictive information that baseline climate levels completely miss.

The bottom line: Ancient height polygenic scores are not simply elevated because a population resides in a cold environment; they are specifically elevated in populations whose local environments recently became colder.

This is not a claim that can be settled by simply eyeballing birth-rate charts. Fertility was already falling in many countries before COVID-19. If we compare raw fertility in 2024 across nations, we mostly just rediscover that older, richer, highly vaccinated countries already had lower fertility before the vaccine campaigns ever began. That is not a useful causal comparison.

But the question is still worth testing carefully because the hypothesis has a real, documented public history.

Where the Hypothesis Came From

The vaccine-fertility concern did not originate as an academic demographic hypothesis. It was an early social-media claim. One influential version appeared in December 2020 and argued that antibodies generated against the SARS-CoV-2 spike protein might also attack syncytin-1, a protein involved in placental development. Fact-checkers and immunologists pushed back quickly, noting that the proposed mechanism was biologically implausible, but the claim spread anyway.

FactCheck.org described the rumor as starting with an article headlined “Head of Pfizer Research: Covid Vaccine is Female Sterilization.” PolitiFact covered a related December 2020 claim that a former Pfizer employee had said the vaccine would make women infertile. AFP documented online claims that COVID vaccines would cause infertility and even global population decline. Later versions involved male sterility, menstrual changes, miscarriage fears, and vaccine “shedding.” By 2021 and 2022, medical journals and mainstream outlets were discussing vaccine-infertility misinformation as a real driver of hesitancy among people of reproductive age.

Most of those mechanisms were weak, false, or speculative. But a weak mechanism does not by itself answer the empirical question. If countries with higher vaccine uptake later fell below their expected fertility trajectories, that would still be worth noticing. The right standard here is not “prove causality from country-level data.” Instead, we can ask: “do the ecological data show any pattern consistent with the claim, after using reasonable controls and avoiding obviously bad comparisons?”

There is also now a directly relevant published paper, although it is not the same design as the country-level analysis here. Manniche, Fürst, Schmeling, Gilthorpe, and Riis Hansen (2025) examined Czech Republic data on successful conceptions by COVID-19 vaccination status among women aged 18-39. Because that paper studies individual vaccination status within one country rather than cross-country ecological variation, it is much closer to the ideal exposure definition than the global country-level proxies used here. I do not treat it as settled evidence, but it belongs in the motivation: the fertility question is no longer only a social-media rumor; there are now published analyses directly asking whether conceptions differed by vaccination status.

The Raw Comparison (And Why It Fails)

If you take a deliberately crude approach, you can plot the share of the total population fully vaccinated by the end of 2021 against each country's total fertility rate (TFR) in 2024.

The raw association is very strong and negative: Countries with higher vaccination coverage in 2021 tend to have lower TFR in 2024.

But this is completely unsurprising, and it is not evidence of a vaccine effect. High-vaccination countries were disproportionately rich, older, urban, administratively capable, and already experiencing low fertility. Switzerland and Japan did not become low-fertility societies because of a 2021 vaccination campaign. A raw TFR comparison confounds vaccine uptake with the entire global demographic transition.

Designing a Better Test: The Residualized Model

To fix this, we need a model that doesn’t ask whether vaccinated countries have low fertility, but whether vaccinated countries fell further below their own pre-existing trends.

For each country, I fit a pre-pandemic fertility trend using data from 2015–2019:

log(TFR) ~ year

Then, I predicted what each country’s TFR should have been in 2024 if that pre-pandemic trend had simply continued. The primary dependent variable is the residual:

observed log(TFR_2024) - predicted log(TFR_2024)

In plain English: How far fertility in 2024 was above or below the country’s own expected trajectory. Negative residuals mean fertility came in below expectations.

The final adjusted model controls for baseline TFR in 2019, the 2015–2019 fertility trend, median age in 2020, and log population density in 2020. These address the most obvious ecological confounding: demographic age structure, baseline fertility regime, pre-existing decline, and urbanization.

The Results

In the adjusted 2024-only residual model, being fully vaccinated by the end of 2021 is associated with a -0.686% lower-than-expected 2024 fertility per +10 percentage points of vaccination coverage (p = 0.0410).

That is not a large effect, and it is not definitive, but it is in the hypothesized direction and nominally significant.

Testing Different Metrics and Timings

The timing matters. Births in 2024 are cleanly aligned with mature vaccination exposure prior to conception. When we look at various vaccine exposure definitions, the results are mixed but trend in a specific direction:

The strongest adjusted signals come from boosters per 100 by end-2022 and the fully vaccinated share by end-2021.

However, this table also highlights why we must remain cautious. If the effect were a simple cumulative-dose effect at the population level, total vaccinations per 100 might be expected to be among the stronger predictors. It is not. That suggests the signal, if real, is timing- or subgroup-specific rather than a clean aggregate dose-response pattern.

The global result is suggestive, but it still uses a blunt exposure measure. What happens if we look more directly at vaccination among people of reproductive age?

I will also investigate possible effecs of vaccine product class (mRNA, viral vector, and non-mRNA other).

Read more

No academic setup. No complex categories. Just a few geographic labels and a standard metric used by geneticists to measure genetic distance, called FST.

How it works: This metric measures genetic differentiation between two groups. A Fst value means the two groups are genetically identical. The higher the number, the more genetically distinct (”distant”) the groups are from each other.

The labels below are intentionally partial, just familiar enough to give you context, but vague enough to keep the game honest.

What are you looking at? Modern human populations? Isolated island communities? Old-fashioned continental categories dressed up in modern genetics?

Make your guess before reading the next line.

The Labels Come Off

Those are bears. Specifically, these are brown bear (Ursus arctos) populations pulled from a genomic dataset. “European,” “North American,” and “West Coast” aren’t human groups here, they are bear populations with the word “bear” scrubbed out.

Now, let’s look at actual human data using the exact same metric, matched for comparable, small-to-moderate sample sizes so the math stays fair.

These are modern human population comparisons from theHudson FST table. I shortened some labels so the table stays readable, but these are human groups.

People expect a clean ladder of nature: humans at the bottom (low genetic diversity), animal subspecies in the middle, and distinct animal species at the top.

But look at the data. While the first table has a massive outlier (Hokkaido vs. West Coast at 0.41), the human table easily hits values between 0.22 and 0.34. These aren’t different human species; they are modern human populations.

The lazy argument, that animal populations are inherently far more genetically distinct than human populations, fails. It is true when comparing different species (like a brown bear vs. a polar bear). It is not true for standard regional populations within the same species.

A heatmap makes the same point more directly. Among these brown-bear populations, Europe and Mainland Alaska are relatively close, Europe and Hokkaido are farther apart, and Hokkaido versus West Coast is the most separated pair in this small set.

The same view for the matched human examples

For comparison, here is the FST matrix for the human populations used in the matched example table. The color scale is the same as the bear heatmap, so the two figures can be read against each other directly.

But this raises a massive problem. If the numbers overlap this perfectly, why does every textbook claim that human genetic diversity is uniquely low compared to animals?

The answer comes down to a glaring methodological flaw in how we sample animal populations, and a statistical illusion that has distorted the literature for decades.

Read more

Again and again, a field advances due to an unusually talented individual. Years later, when that creation becomes famous, commercialized, or weaponized, that same pioneer transforms into its dissident—not necessarily an enemy, but a disappointed parent or a heretic trying to recover the revolution’s original meaning.

Call it the Yann LeCun Effect.

Artificial Intelligence: The Contemporary Frontier

As a central figure in modern deep learning and a 2018 Turing Award winner, Yann LeCun helped construct the intellectual world that made today’s AI boom possible. Yet, in today’s discourse, he often sounds like a dissenter:

• The Paradigm Critique: He argues that Large Language Models (LLMs) are an impressive engineering feat, but they are not a true theory of mind or the path to human-level intelligence.

• The Reality Check: He maintains that current systems lack even animal-level intelligence and that existential risk rhetoric is wildly overblown.

LeCun’s skepticism, however, faces a massive empirical challenge. Critics argue that his rigid boundaries are increasingly disproven by the sheer capability of modern AI.

First, there is the staggering velocity of the technology: the sheer volume of advanced, multi-step tasks that AI can execute at blinding speed goes far beyond simple text prediction. Second, and more profoundly, AI has broken into the realm of elite human intellect by solving open, notoriously difficult mathematical puzzles such as Erdős problems.

When a machine achieves heuristic breakthroughs in pure mathematics that baffled humans for generations, it suggests that the line between “spectacular engineering” and genuine higher-order problem solving is blurring much faster than the pioneers care to admit.

LeCun is not alone. AI is uniquely fertile ground for this effect because the line between engineering and metaphysics is constantly blurred:

Physics & The Great Metaphysical Disputes

This phenomenon is not new. The classic historical example is Albert Einstein and quantum mechanics.

Popular memory labels Einstein as the old physicist who couldn’t accept the future. In reality, his 1905 paper on light quanta helped birth quantum physics. His famous objection—“God does not play dice”—was not a denial that the equations worked, but a dispute over meaning. He believed the field had mistaken predictive power for actual understanding.

The Moral Shock: Weapons & Biotechnology

Sometimes the pioneer’s turn is driven by a profound moral realization rather than philosophy. The technology escapes the lab, and the creator is left grappling with its political destination.

• The Nuclear Architects: Leo Szilard conceived the nuclear chain reaction and initiated the Manhattan Project, only to later aggressively lobby for international weapons control. J. Robert Oppenheimer and Andrei Sakharov followed similar trajectories, their later warnings carrying immense weight precisely because of their insider complicity.

• The Genetic Gatekeepers: Paul Berg (recombinant DNA pioneer) led the call for a voluntary moratorium on risky gene experiments in the 1970s. Decades later, Jennifer Doudna (CRISPR co-inventor) became a leading voice calling for strict global governance after the premature birth of gene-edited babies in China.

• The Industrial Legacy: Alfred Nobel amassed a fortune inventing dynamite for construction and mining. Witnessing its devastating use in warfare, he repurposed his wealth into the Nobel Prizes to rewrite his legacy toward peace and human advancement.

The Disappointed Internet Fathers

In the digital age, pioneers frequently find themselves loving the original research vision while loathing the implemented civilization.

1. Tim Berners-Lee invented the World Wide Web as an open, decentralized information-sharing system. He spends his later years criticizing what it became: an ecosystem of corporate surveillance, addictive algorithms, and commercial enclosure.

• The Architects: Jaron Lanier (VR pioneer) and Ted Nelson (hypertext creator) both became sharp critics of Web 2.0, arguing that ad-driven platforms and crude implementations eroded true human personhood and literary potential.

• The Feature Designers: On a product scale, Ethan Zuckerman (the pop-up ad), Aza Raskin (infinite scroll), and Justin Rosenstein (the Facebook Like button) have all expressed regret over how their local, clever mechanisms were weaponized by the attention economy.

The Ambiguous Cases: When Authority Meets Infallibility

The “pioneer turned critic” possesses massive rhetorical force because an insider cannot be easily dismissed as ignorant. However, this is precisely where the Yann LeCun effect becomes dangerous: insider status does not equal infallibility, and a pioneer is not automatically right.

Two specific cases illustrate the deep ambiguity and danger of abusing this trope.

The Case of Kary Mullis

Kary Mullis won the 1993 Nobel Prize in Chemistry for inventing PCR, which became a foundational technique of modern molecular biology. Later in life, however, he became a prominent dissenter from the scientific consensus on HIV and AIDS, famously arguing that HIV had never been proven to cause AIDS. Mullis also fiercely objected to certain applications of his own technology in that debate, specifically criticizing the idea that PCR could straightforwardly quantify viral load without making major interpretive assumptions.

While this makes him a powerful example of the pioneer-turned-critic trope, a crucial caveat is required: Mullis’s status as the brilliant inventor of PCR did not make his medical views on HIV correct. The scientific consensus that HIV causes AIDS remains overwhelming. Furthermore, public misquotations have often weaponized his legacy, exaggerating his actual position into the entirely false claim that PCR cannot detect viruses or infectious diseases at all.

There is also a less flattering possibility at play here: envy, or status displacement. A pioneer may create a tool or idea, only to watch later figures build larger careers, institutions, fortunes, or moral authority on top of it. The founder then becomes not merely a critic of misuse, but a resentful critic of successors.

Mullis is the delicate, textbook example of this social dynamic. PCR became central to molecular medicine, including HIV detection and viral-load monitoring, while Mullis became hostile to the HIV/AIDS consensus and to public health figures such as Robert Gallo, Luc Montagnier, and Anthony Fauci. While it is impossible to prove envy from the outside (and Mullis always gave intellectual reasons for his dissent) the social situation is recognizable: the inventor of a powerful method watches other people become authorities through applications of that method, and subsequently attacks the authority structure that grew around it.

The Case of Robert Malone

Robert Malone represents a more delicate and highly visible modern example. In the late 1980s, Malone made landmark early contributions to RNA delivery and vaccine-related mRNA technology. Decades later, during the COVID-19 pandemic, he emerged as one of the most prominent critics of mRNA vaccines and public vaccine mandates.

This case demands careful handling because it perfectly exposes the limits of the pioneer myth:

• In public discourse, Malone is frequently labeled as “the sole inventor of mRNA vaccines.” This claim is historically inaccurate.

• mRNA vaccines were actually the culmination of decades of collaborative work by a vast web of researchers, including Katalin Karikó, Drew Weissman, lipid nanoparticle experts, immunologists, and major industry teams.

Malone belongs in the history books as an important early contributor, not as a lone creator. He serves as a vital warning for this essay: the “pioneer turns critic” narrative is so rhetorically seductive that it tempts people to overstate a single person’s historical role, or to treat past insider status as a blank check for modern scientific correctness.

Why This Happens

There are several structural reasons why pioneers so regularly become dissenters:

• They remember the contingency: While the public sees a finished product as inevitable destiny, the pioneer remembers the duct tape, the toy assumptions, and the arbitrary choices that built it.

• They separate the idea from the implementation: It is entirely logical to love the concept of an open web or genetic editing while despising platform monopolies or reckless embryo manipulation.

• They are immune to category errors: The pioneer knows the machine too well to buy into the myth. Weizenbaum knew ELIZA didn’t care; LeCun knows prediction is not the same as embodied intelligence.

• Status Displacement and Resentment: When a quiet research field mutates into a multi-billion-dollar industry driven by capital, marketers, and politicians, the pioneer loses sovereignty over their life’s work. Dissent then becomes a complex psychological mix: partly intellectual, partly moral, and sometimes a deeply human reaction to watching new successors build massive fortunes, institutional authority, and global fame on top of their original foundations.

Conclusion

Einstein was not infallible about quantum mechanics; his intuition failed him on the metaphysics of reality. Weizenbaum was not entirely correct about the absolute limits of AI.

Similarly, LeCun’s skepticism does not automatically settle the future of LLMs. While he dismisses current architectures as limited engineering paradigms, the empirical evidence of AI executing advanced tasks at superhuman speed and solving open Erdős problems proves that the machine is already outgrowing the imagination of its creators.

The takeaway of the Yann LeCun effect is not to blindly trust the apostate founder, but to listen carefully when they choose to become an apostate. They are valuable because they interrupt the smooth, corporate myth at the exact moment it is becoming socially convenient.

Share

A Chihuahua and a Labrador Retriever differ dramatically in size, anatomy, behavior, and function. One is a tiny companion dog; the other is a large working retriever bred for swimming and carrying game.

Yet their genetic differentiation (FST = 0.116) is almost identical to several ordinary human intercontinental comparisons: northern Italians vs Japanese (0.116), southern Han Chinese vs northern Italians (0.116), Beijing Han Chinese vs Mozabites (0.116), and French vs Japanese (0.113).

This apparent paradox illustrates a central theme of this article: genetic distance and phenotypic difference are not the same thing. As we will see, FST is often treated as a universal ruler of biological difference, but that interpretation can be highly misleading.

There are countless dog and cat breeds, and they differ a lot in hair color, size, and even intelligence and personality. Consider the stark behavioral differences we take for granted: a Border Collie lives to herd livestock, a Siberian Husky is driven to pull sleds for miles, and a Cavalier King Charles Spaniel is content simply being a lapdog. Meanwhile, in the feline world, a Siamese is famously vocal and demanding of attention, while a British Shorthair remains placid, reserved, and fiercely independent.

Looking at these dramatic differences in appearance and behavior, an intriguing question naturally arises: how much do they differ genetically? And more provocatively, are the genetic differences between a Bulldog and a Chihuahua similar to those between different human populations?

When researchers want to measure this kind of genetic distance, they often reach for a mathematical metric called FST. But as it turns out, relying purely on this number to map out biological reality can lead us straight into a trap.

The Illusion of the Absolute Ruler

FST is often treated as if it were a simple, objective genetic distance ruler. The logic seems straightforward: put two populations into the formula, get a number, and then rank the distances. While that is undeniably useful, it is also dangerous.

The truth is that an FST value is not just pure, unadulterated biology. It is highly sensitive to the mechanics of the study itself, including:

• Sample size (N)

• Total marker count

• Missing data (”missingness”)

• Ascertainment bias

• The specific way the population labels were built in the first place

To demonstrate how easily these numbers can be misinterpreted, I compared pairwise Hudson FST values across three distinct within-species datasets:

1. Humans: Modern human groups from the Allen Ancient DNA Resource (AADR) v66 2M panel (179 population labels, yielding 15,931 pairwise values).

2. Dogs: Purebred dog SNP data from the Parker 2017 dataset after quality control, minority allele frequency (MAF) filtering, and linkage disequilibrium (LD) pruning (177 breed labels, yielding 15,576 pairwise values).

3. Cats: Cat breed and random-bred groups from the Matsumoto SNP dataset after QC and LD pruning (14 group labels, yielding 91 pairwise values).

The concrete result of this comparison is surprising: human, cat, and dog FST values live on a remarkably similar numerical scale.

What Exactly Is FST Measuring?

At its core, FST measures allele-frequency differentiation. If two groups share very similar allele frequencies across many single-nucleotide polymorphisms (SNPs), the FST is low. If those frequencies differ significantly, the FSTis high.

However, because we must estimate allele frequencies from finite samples, small sample sizes introduce a massive amount of statistical noise. Because FST behaves like a distance metric, random noise inherently pushes groups apart, creating the illusion of large genetic separations where none exist.

The Data: Uncomfortably Close Distributions

When we look at the raw distributions, the overlap between species is striking. Cats are especially close to the human median, and while dogs are shifted higher, as you would expect from strictly closed, inbred breed structures, the vast majority of dog breed-pair values still sit firmly within the central human range.

Table 1: Pairwise Hudson FST Summary Statistics

Using the human 5th-to-95th percentile range (0.017 to 0.468) as our baseline reference interval, we get a clear picture of just how much these distributions overlap:

Table 2: Overlap with the Human F_ST Distribution

This means that 98.9% of cat breed comparisons and 97.3% of dog breed comparisons fall squarely inside the normal genetic distance range found between human populations.

The visual density plot below illustrates this dramatic overlap, highlighting how the mathematical distributions occupy the exact same territory:

A Chihuahua and a Labrador can be separated by roughly the same FST as an Italian and a Japanese person.

So why don’t they look anything alike?

To answer that question, I match dog breeds to human population pairs with nearly identical genetic distances, reveal the complete cat FST matrix and an expanded matrix of major dog breeds, test the sample-size trap that inflates genetic distances, and explain the quantitative genetic logic behind Lewontin’s Fallacy.

Upgrade to a paid subscription to continue reading and access the full archive.

Your support directly funds the independent data pipelines, quality control, and computation required to turn raw genomic data into analyses like this one.

Read more

That was an unexpected combination. Northern Europeans and Inuit are not especially close genetically, linguistically, or culturally. What they do share is a harsh ecological reality: long winters, strong seasonality, and environments in which planning failures can be costly.

This observation led me to a hypothesis. Cold, seasonal environments should reward planning. They make food availability less continuous. They raise the value of storage, reliable tools, delayed returns, long-range coordination, and technologies that cannot afford to fail. If the Met(A) allele pushes cognition toward greater stability, working memory, and maintenance of task goals, then colder and more seasonal ecologies might have favored it.

There was also a second hypothesis. Farming is a delayed-return system. It requires storing seed, planning across seasons, defending land, coordinating labor, and accepting more formal rules than mobile foraging usually requires. If the same cognitive style helped with farming, then farmers should have higher Met(A) frequencies than hunter-gatherers.

COMT matters because it helps regulate catecholamines, including dopamine, and it is especially relevant in the prefrontal cortex, where dopamine transporters are relatively sparse. The Val158Met variant is functional: Val is associated with higher COMT activity, while Met(A) produces a lower-activity enzyme (Lachman et al. 1996; Chen et al. 2004). The usual cognitive story is that Met(A) can favor working memory and stable cognitive control, while Val(G) can favor flexibility under some conditions (Egan et al. 2001). But the cognitive literature is mixed, and meta-analytic work cautions against treating rs4680 as a simple intelligence switch (Barnett et al. 2008).

At the time, the available evidence was thin. There were no large ancient-DNA panels to lean on. The argument had to be made from modern population frequencies, ecological proxies, and a theoretical model about climate, cognition, and culture.

Thirteen years later, we can revisit those ideas with thousands of ancient genomes, paleoclimate reconstructions, and genome-wide ancestry controls.

The results are not what I expected.

What Changed

I used AADR v66 2M and extracted rs4680, treating the A allele as the Met allele. For ancient samples, I merged the allele counts with AADR metadata and with CHELSA/TraCE21k climate reconstructions. The main climate variables are coldest-quarter temperature, converted where needed into coldness, and quarter temperature seasonality, the contrast between the warmest and coldest quarters.

For global ancient climate models I used projected genome-wide PCs, PC1 through PC10, plus sample date and 2M coverage. For modern samples I re-geocoded obvious diaspora labels where possible, and also ran a strict version excluding admixed populations without a single homeland coordinate.

The Hypotheses

Table 1. The old hypotheses and the updated ancient-DNA result.

Highest Ancient Groups

Before collapsing samples into hunter-gatherer, farmer, and steppe categories, it is useful to look at the highest-frequency ancient labels directly. The plot below uses N >= 15. Small labels are pooled only when they share a clear geographic/cultural label; here that adds an Old Bering Sea / Chukotka bin.

Figure 1. Highest ancient Met(A) frequencies. Points are sized by sample count; the pooled Arctic/Beringian bin is marked separately.

Two of the most conspicuous high-frequency labels need a little interpretation. Serbia Iron Gates is not a farmer signal: it is a Danube Gorge hunter-gatherer context with eastern hunter-gatherer and steppe-adjacent affinities, sitting on the corridor between the Balkans and the Pontic-Caspian world. Russia_Tatarstan_Eneolithic is also not Yamnaya proper. It is a Volga-Kama Eneolithic group from Murzikha-2, chronologically earlier than Yamnaya but geographically and genetically close to the northern forest-steppe/EHG sphere. In both cases, the high Met(A) frequency points back toward forager and steppe-adjacent population history, not toward the arrival of farming. Curiously, Roman IA/Republican samples also have quite high A frequency.

Hunter-Gatherers Versus Farmers

The first result is the most damaging to the farming hypothesis. Ancient hunter-gatherer-rich groups are not low in Met(A). Several are very high. Iron Gates hunter-gatherers, Baltic hunter-gatherers, core WHG groups, and EHG or steppe-forager-like groups all sit above early Anatolian and European farmer groups. Yamnaya is elevated, but not uniquely so. The signal is older and broader than Yamnaya.

That reverses the simple modern farmer-vs-hunter-gatherer contrast from the earlier paper. With ancient DNA, Met(A) does not look like a farming allele. It looks like an allele that was already common in parts of the northern and western Eurasian forager world, then carried forward through later mixtures (Fig. 2).

Figure 2. Met(A) frequency in hunter-gatherer-rich, steppe, and farmer-related ancient groups.

Table 2. Category-level allele frequencies. These labels are interpretive bins built from AADR Group_ID strings, so they should be read as descriptive rather than formal qpAdm classifications.

Upgrade to a paid subscription to unlock the full climate analysis, see the exact data models, and find out what really drove the evolution of the COMT Met(A) allele across deep history.

Read more

That is the starting point of a hypothesis suggested to me by Dr. József Kukolya. The idea is that the Neolithic shift to large-scale grain cultivation and storage created better conditions for *Aspergillus flavus*, raising chronic exposure to aflatoxin B1. Aflatoxin is famous in cancer genetics because it leaves a recognizable scar: G>T-like mutations, especially in the SBS24/AFB1 mutational signature.

The speculative evolutionary version is this: if chronic aflatoxin exposure also increased germline mutation, then early agricultural populations may have received extra mutational input biased toward G>T changes. If some of that input was useful, then derived alleles that rose under selection after the Neolithic should show a slight excess of G>T-like changes, especially among variants with stronger inferred positive selection.

This is not a claim that aflatoxin explains human evolution. It is a narrower molecular test: when we look at selected derived alleles, is there a detectable tilt toward G>T?

I used the selection summary statistics from Akbari et al.’s 2026 paper. In that file, `S` is the estimated selection coefficient for the ALT allele: positive `S` means ALT is inferred to have risen in frequency, while negative `S` means the other allele is favoured by the model.

There is one small complication. The hypothesis is about derived alleles, but Akbari’s `S` is written for ALT, and ALT is not always the derived allele. So I first identified which allele was ancestral and which was derived. Then I flipped the sign of `S` whenever necessary so that `S_derived` always means the inferred selection coefficient for the derived allele. Larger positive `S_derived` means a stronger inferred rise of the derived allele.

The first version of this analysis used all filtered variants and produced a strong-looking trend. But that was too optimistic, because millions of nearby variants are not independent observations. For this revision, I LD-pruned the Akbari variant list using 1000 Genomes Phase 3 EUR genotypes and PLINK2:

After LD pruning, 118,567 variants remained; 78,635 had the favoured allele as derived.

I tested two mutation classes: strict G>T, and strand-collapsed G>T/C>A.

Before looking at selection strength, I first checked the simple average: do all favoured-derived variants have more G>T than the non-rising variants? Table 1 reports that comparison. It is not significant. In the LD-pruned data, the favoured-derived variants are actually a little lower than background.

Table 1. Overall G>T-like enrichment after LD pruning

So the significant result in this analysis is not a blanket excess of G>T among all favoured-derived variants. The signal, if there is one, is in the gradient: whether G>T-like changes become more common as `S_derived` gets larger.

The original hypothesis predicts more than a global average. It predicts that the G>T-like signal should become stronger as the inferred selection coefficient gets larger. To test that, I split the favoured-derived variants into twenty equal-count bins by `S_derived` and compared each bin to the same non-rising LD-pruned background. Table 2 summarizes the correlation tests across those bins.

Table 2. Correlation between bin mean S_derived and G>T-like enrichment

Figure 1 shows the main positive result. Each point is one of the twenty `S_derived` bins. The y-axis is enrichment relative to the non-rising background, so values above zero mean that a bin has more of that mutation class than expected from the comparison set. Under the 1000kb/r2=0.1 LD-pruned analysis, strict G>T enrichment correlates with mean `S_derived` at r = 0.472 (p = 0.035), and strand-collapsed G>T/C>A shows almost the same pattern at r = 0.470 (p = 0.037). In other words, the bins with stronger inferred derived-allele selection tend to be more G>T-like.

Figure 1. LD-pruned bin-wise enrichment

The plot also shows the main caveat. This is not a smooth staircase. The points are noisy, and the signal is clearest as a Pearson trend across bins rather than as a monotone increase in every adjacent bin.

Figure 2 puts the same result on the raw-proportion scale. This is useful because odds ratios can make a small compositional shift look visually larger than it is. In absolute terms, the effect is modest: for G>T/C>A, the first bin is 8.67% and the last bin is 9.10%. For strict G>T, the endpoints are nearly flat. The positive Pearson result is coming from the shape across all bins, not from a simple first-bin-versus-last-bin contrast.

Figure 2. Raw bin proportions

To make the bin pattern easier to inspect numerically, Table 3 pulls out five points from the twenty-bin series: the first bin, three middle positions, and the highest-`S_derived` bin.

Table 3. Selected bins across the S_derived range

Discussion

The unpruned result was too strong. LD pruning removes the global average excess and forces the conclusion to become narrower. What remains is the binned result: under the 1000kb/r2=0.1 pruning rule, G>T-like enrichment is positively correlated with `S_derived` across bins.

That is not proof of an aflatoxin mechanism. It is a surviving hint in the right direction. The fair interpretation is: no overall enrichment after LD pruning, but a significant Pearson bin-wise association between inferred selection strength and G>T-like mutation class, with weaker support from rank and logistic tests.

The next test should be more specific than the one in this post. Here I treated all G>T and G>T/C>A changes as if they were equally aflatoxin-like. That is only a rough proxy. Aflatoxin does not just produce generic G>T mutations; it tends to produce G>T changes in particular local sequence contexts.

In other words, the neighboring bases matter. A G>T change inside AGA is not molecularly identical to a G>T change inside CGC or TGG. Cancer geneticists describe this as trinucleotide context: the changed base plus the base immediately before it and the base immediately after it. SBS24, the aflatoxin-like signature, is defined at that more detailed level.

So the more precise version of the test is to take each selected SNP, recover the two neighboring bases from the reference genome, orient the mutation from ancestral allele to derived allele, and then ask whether the high-`S_derived` variants are enriched for the specific trinucleotide contexts expected under SBS24. A generic G>T trend is suggestive. A trend concentrated in SBS24-like contexts would be much stronger evidence for the proposed aflatoxin mechanism.

References

Kukolya, József. Personal communication; source of the aflatoxin-agriculture hypothesis tested here.

Akbari, A., Perry, A., Barton, A.R. et al. (2026). *Ancient DNA reveals pervasive directional selection across West Eurasia*. *Nature*. https://doi.org/10.1038/s41586-026-10358-1

Alexandrov, L.B. et al. (2020). *The repertoire of mutational signatures in human cancer*. *Nature*, 578, 94-101. https://doi.org/10.1038/s41586-020-1943-3

Huang, M.N. et al. (2017). *Genome-scale mutational signatures of aflatoxin in cells, mice, and human tumors*. *Genome Research*, 27, 1475-1486. https://doi.org/10.1101/gr.220038.116

COSMIC Mutational Signatures. SBS24/AFB1-like signature. https://cancer.sanger.ac.uk/signatures/sbs/sbs24/

The journal’s conclusion was straightforward: no correction or addendum would be pursued. That outcome itself did not surprise me. Journals are generally reluctant to intervene in citation disputes, especially when a big Harvard lab is involved, and I never assumed that Nature would compel the authors to amend their paper.

What surprised me was the reasoning.

After discussing the matter internally, Nature wrote:

“authors have a reasonable degree of flexibility in choosing to cite papers they feel are the most relevant to their study.”

The editor also stated:

“it is not required that all papers related to a topic are cited to support every point in a paper.”

When I read those sentences, my immediate reaction was confusion.

Who exactly is making that argument?

Certainly not me.

At no point did I claim that Akbari et al. were required to cite every paper related to polygenic adaptation, ancient DNA, or human evolution. Such a standard would be absurd. Every scientific paper omits relevant references. Every bibliography is selective. Nobody disputes this.

My complaint was about something entirely different.

The Akbari et al. paper presents evidence of directional selection on complex traits using ancient DNA time series. My concern was that my previous papers had already addressed essentially the same question and reported evidence consistent with directional selection on polygenic traits through time, using overlapping datasets and similar methods.

One can certainly argue that those papers were imperfect. One can argue that the datasets were smaller, the methods less sophisticated, or the conclusions less robust. Science is full of early studies that are later refined by better data and better methods.

But those would at least be arguments about the substance of the issue.

Instead, Nature’s response seems to treat my complaint as though I had simply requested a longer bibliography. That is what I find difficult to understand.

There is an obvious difference between saying:

“You failed to cite every relevant paper.”

and saying:

“You failed to acknowledge prior work directly related to your paper’s central contribution.”

The first complaint is unreasonable but the second is a question of scientific attribution.

Perhaps Nature genuinely sees no meaningful distinction between the two. If so, I would have preferred them to say so explicitly. At least then we would be debating the same issue.

Instead, I was left with the peculiar feeling that the argument I had spent weeks making had been replaced with a much weaker one and then rejected.

What Nature expressed was that authors cannot be expected to cite all papers related to a topic.

And that is a point with which I completely agree but it is also a point that has very little to do with the complaint I actually made.

The entire experience has made me reflect on how difficult it is to discuss scientific attribution in the first place. Nobody likes talking about priority. The moment a researcher raises the issue, there is an assumption that they are simply upset about not being cited.

Yet attribution matters. Science is cumulative. The historical record of a field is built from thousands of decisions about what gets acknowledged, what gets remembered, and what gets forgotten. Those decisions affect reputations, careers, and our understanding of how ideas developed.

For that reason, I think there is a meaningful difference between an incomplete bibliography and the omission of prior work on a study’s central question.

Nature apparently disagrees, or perhaps never reached that question at all.

There was one final part of the email that I found particularly puzzling.

After informing me that no further action would be taken, the editor wrote:

“Please note that discussions regarding potential amendments to our published papers are confidential, and should be treated as such, including email conversations such as this one.”

I confess I do not understand this request.

I understand why peer review is confidential. I understand why anonymous referees need protection. I understand why editors require private deliberations while decisions are being made.

But this was not an unpublished manuscript under review.

This was a discussion about whether a published paper accurately acknowledged prior work and whether the published scientific record should be amended.

Why should such discussions be confidential?

Indeed, one could argue the opposite. If journals are making decisions about attribution, priority, and corrections to the literature, there is a public interest in understanding how those decisions are reached.

Perhaps Nature has good reasons for preferring these discussions to remain private. If so, those reasons are not obvious to me.

The request struck me as particularly strange because it appeared at the end of a process that had already concluded. The journal had reviewed the complaint, consulted internally, reached a decision, and informed me that the matter was closed. Having done so, it then requested that the discussion itself be treated as confidential.

I genuinely wonder what purpose that serves.

In any case, the dispute is now over as far as Nature is concerned. The journal has reviewed the complaint, consulted internally, and decided that no further action will be taken. What remains, however, is the question that prompted the complaint in the first place: where exactly is the line between an omitted reference and an omitted contribution?

That was always the issue I was trying to raise. Not whether every relevant paper should be cited, but whether prior work directly addressing the same scientific question deserves acknowledgment when later studies achieve greater visibility. After reading Nature’s response several times, I am still left with the impression that this distinction was never really addressed.

It is easy to view the contrast between northern and southern Italy as a permanent feature of the landscape, a timeless divide separating the Alps and the Po Valley from the Mezzogiorno and the Mediterranean. Many of these physical differences are well known, having been famously mapped by the 19th-century anthropologist Renato Biasutti. His documentation of regional variations, such as the percentage of blonde hair in each province shown in the map below, highlights just how deeply ingrained these geographic contrasts are in our historical understanding.

In economic history, there is a fierce debate over how old the regional gaps in wealth and human capital really are (Pescosolido, Rotondi, and Scoppa 2025; Barbero, Deandrea, and Negri 2021). One influential view sees the North-South divide as having deep pre-unification roots, especially in literacy and human development (Ciccarelli and Weisdorf 2019; Bozzano 2023). A revisionist view, by contrast, argues that the economic gap at unification was small or even absent, and that later state policies helped widen it (Daniele and Malanima 2007, 2011; Federico, Nuvolari, and Vasta 2019).

But this entire debate quietly assumes a baseline that needs to be tested. Was the underlying genetic divergence between the North and South always present in its current form, or did it sharpen, weaken, or shift as Italy transitioned through the Bronze Age, Iron Age, Roman period, and later historic eras?

Ancient DNA (aDNA) allows us to address this question directly, at least from a purely genetic angle. Rather than projecting modern geographic differences backward, we can map ancient individuals, categorize them by region, and track whether polygenic score (PGS) variations show up consistently across different eras.

Specifically, we can look at the very traits that form the core of regional stereotypes, testing whether the data backs up the familiar imagery of shorter, darker southerners compared to taller, more blonde, and more educated northerners. By looking at traits like height, pigmentation, Educational Attainment (EA), and BMI, we can finally ask a concrete question: does the genetic map of Italy look like a fixed, permanent background fact, or did these specific differences emerge, fade, or flip over time?

Pooled North-South Differences

Before investigating how things changed over time, we need a baseline comparison. By collapsing all historical periods together, we can see whether coverage-adjusted, standardized PGS values differ fundamentally between the North and South.

The table reports the South relative to the North using Cohen’s d, meaning negative values indicate that the South exhibits a lower average score than the North.

Table 1. Pooled north-south differences

The largest pooled differences are quite straightforward: the South scores lower than the North for height, blonde hair, and light skin. Educational Attainment (EA) is also lower in the South. The proxy for blue eyes leans in the same direction but is only borderline in this pooled layout. Meanwhile, traits like delay discounting and the two BMI variations show no meaningful regional separation when time is ignored

Figure 1. Overall north-south distributions, coverage-adjusted

Tracking the Contrast Through Time

While a pooled comparison gives us a useful overview, it can easily mask historical shifts. A regional contrast might be ancient and stubborn, or it might have started weak and intensified later.

To keep the sample sizes viable for analysis, the data is divided into two broad chronological bins:

1. Pre-Iron Age and older

2. A merged Iron Age / Roman / Medieval-to-recent bin

Merging the later periods is a practical necessity; the southern Medieval/recent sample size on its own is simply too small to yield a reliable independent estimate.

Figure 2. Coverage-adjusted period means

Think of this period plot as a descriptive bridge between the pooled boxplots and a formal interaction model. It maps the mean and baseline uncertainty across broad eras rather than forcing a smooth, continuous time trend. Because the older bin remains relatively sparse, we shouldn’t over-interpret those early-period means. The real takeaway is seeing whether the direction of the contrast remains stable or visibly shifts.

The data reveals that the gaps in Educational Attainment (EA) is remarkably ancient, maintaining a relatively stable trajectory over time. Conversely, the temporal trajectories of Height, Light Skin and Blonde Hair diverged, increasing the gap between North and South in more recent times.

The broad period means in the plot above suggest a shifting landscape, but eye-balling a couple of historical bins isn’t enough. Ancient DNA is notoriously messy, sample sizes fluctuate, and if you want to know whether Italy’s genetic geography was truly fluid, you have to put it to a formal regression test.

When you run the formal interaction model, the results are striking.

To unlock the formal temporal test, and the robustness checks proving this isn’t just data noise, consider becoming a paid subscriber.

Your support directly funds the independent data pipelines and computation time required to pull these historical insights out of raw ancient genomes. Join the community to read the rest of this post and access the full archive.

Read more

The specific hypothesis tested here is not that the plague directly selected for modern educational attainment (EA). Rather, the pandemic may have selected on a partly heritable bundle of cognitive, personality, and status-linked traits for which an educational attainment polygenic score (EA PGS) serves as an imperfect modern proxy.

The idea that the plague acted as a major selection event for cognitive and personality traits has been proposed before (Dutton & Woodley of Menie, 2018), given the substantial heritability reported in twin studies for intelligence and personality (Plomin & Deary, 2015; Nettle, 2007). In practical terms, the hypothesized mechanism is that higher-status, more future-oriented, more conscientious, and cognitively stronger lineages may have been better equipped to secure food, move away from danger, bargain, delay exposure, or simply survive the chaotic shock-and-recovery period that followed the plague.

That mechanism is also historically plausible in a narrower social sense. The plague struck populations already weakened by famine, meaning lineages with better access to resources may have held a vital survival advantage. Differential mortality estimates are consistent with strong class gradients in at least some settings: while overall mortality is often placed around 40%, much higher rates have been reported for serfs and laborers in some regions (Dodds, 2008).

None of that proves genetic selection for intelligence, but it does provide a concrete model for why an EA-linked signal might shift after 1350. Body Mass Index (BMI) and Height are included as secondary controls because famine should imply directional expectations for them. If repeated food scarcity selected for a thriftier metabolism, one might expect a higher post-crisis BMI-linked signal, especially in harder-hit areas. If famine and chronic scarcity favored lower maintenance costs, one might expect a lower post-crisis Height-linked signal, or a more negative post-1350 trajectory, again especially where mortality and associated disruption were most severe. Those control traits therefore help distinguish a general famine-selection story from an EA-specific one.

I look at the question two ways. First, I test whether post-1350 Europeans differ from pre-1350 Europeans after adjusting for Dataset, broad region, coverage, PCs, and imputation uncertainty. Second, I fit a hinge model with a knot at 1350 CE and then ask whether higher Black Death mortality is associated with a steeper post-1350 slope.

The Sample: Tracking 12,000 Years of European History

To capture this moment in time, the analysis uses a few different sample slices depending on the model:

• The Broad Window Sample: 754 individuals in total. For the covariate-adjusted means, this includes 508 individuals living before the plague (1000–1349 CE) and 248 individuals living after (1350–1599 CE) across 19 distinct populations.

• The Hinge Model Sample: To see the long-run evolutionary trajectory, I use a broader Europe-only sample of 5,183 individuals spanning the last 12,000 years.

• The Severity Subsets: To test whether harder-hit regions saw bigger changes, I use a subset of individuals (643 for the broad window, 2,342 for the hinge model) mapped to Christakos locality-level plague mortality data.

  ---

Before vs After

The raw means are useful as a visual check, but the adjusted means are the relevant comparison because the samples are not evenly distributed across datasets, regions, coverage levels, principal components, or imputation uncertainty. The adjustment does not make the data representative, but it reduces the most obvious technical and sampling differences between the pre- and post-1350 groups.

In the broad 1000-1599 CE window, the post-1350 samples sit higher on EA after adjustment. This is the simple version of the Black Death hypothesis: if the mortality shock changed who survived or reproduced, the later medieval sample should differ from the earlier medieval sample. The important caveat is that the broad-window design deliberately sacrifices temporal detail. It treats 1000 CE and 1349 CE as one pre period, and 1350 CE and 1599 CE as one post period, so it is a blunt test of a level shift rather than a model of historical dynamics.

The adjusted pre/post regression estimates a post-1350 gap of 0.142 SD with cluster p=1.99e-04. Read literally, this says that individuals sampled after 1350 have higher EA PGS than individuals sampled before 1350, after adjusting for Dataset, broad region, coverage, PC1-PC5, and log imputation variance. The sign is therefore consistent with the hypothesis, but the design is still vulnerable to composition: if the post-1350 sample is concentrated in different places, cemeteries, social strata, or source studies than the pre-1350 sample, the coefficient can partly reflect those differences despite the covariates.

This regression coefficient is the inferential result, not just the difference between the two adjusted means in the plot. In the current data, the raw post-minus-pre mean difference is 0.142 SD; residualizing EA on the covariates and then comparing period means gives a slightly smaller descriptive gap of 0.120 SD. The full regression, EA_z ~ PostBD + covariates, estimates 0.142 SD because it adjusts the period indicator and the outcome jointly. In other words, the adjusted-means plot is a readable visualization, while the regression coefficient is the formal covariate-adjusted pre/post test.

That is why I treat the pre/post model as supportive rather than decisive. It is easy to interpret, but historically coarse. It answers the question, ‘Are the later medieval individuals higher than the earlier medieval individuals in this window?’ It does not by itself tell us whether 1350 is a real turning point in the longer series.

The Hinge Model: A Change in Trajectory

But a simple before/after comparison assumes that genetic change happens like a clean step-stool. What if the Black Death didn’t cause a sudden leap, but instead acted as a turning point, a “hinge” that accelerated an existing trend?

To test this, I fit a hinge model across 12,000 years of European history, setting a “knot” exactly at the year 1350 CE. This allows us to see if the slope of the trajectory changed after the pandemic.

When we analyze the slopes, a much more compelling story emerges.

Table 1. Long-Run EA Hinge Model Coefficients

The results here are statistically robust. For the millennia leading up to the Black Death, the polygenic score for educational attainment was creeping upward at a very slow, steady pace (slope = 0.003, p = 0.005).

After 1350 CE, that trajectory dramatically steepened. The slope experienced a highly significant upward shift of +0.054 (p = 0.034), resulting in a post-plague slope of 0.057. Crucially, comparing the models shows that the hinge model is a statistically superior fit over a basic linear model (p = 0.01). The Black Death didn’t just push the numbers up; it accelerated the entire genetic timeline.

The long-run hinge model proves that Europe’s genetic landscape didn’t just experience a temporary bump—the entire trajectory accelerated after 1350 CE.

But this leaves us with a massive causal riddle: Was this dramatic acceleration actually driven by the pandemic, or are we just looking at a massive historical coincidence?

To prove a true evolutionary signature, we have to look for direct geographical alignment. If the Black Death caused this shift, we should see an undeniable pattern: the regions hit hardest by the plague must show a significantly steeper trajectory than the regions that escaped relatively unscathed.

When we map country-level mortality data directly to these ancient skeletons, the data reveals a striking, highly significant interaction.

Coming up in the premium section:

• The Severity Models: The exact regression outputs and p-values proving how plague mortality alters the post-1350 hinge.

• The Famine Control Test: I run the identical workflow on Height and BMI to see if this was just a general story of human bodies surviving starvation, or if something completely unique happened to educational attainment proxies.

• The Boundary Lines: Why one specific scoring method completely changes the BMI results, and what that tells us about the limits of ancient DNA data.

Upgrade to a paying subscription to unlock the complete data tables, severity models, and the final conclusion.

Read more

And emotionally? Maybe they are.

But when Italians actually pack their bags, when they change their legal residence and enter the dry, unromantic ledger of municipal registers, the data points to a reality that is far more decisive: work.

Not work in the abstract, mind you. What actually moves Italians is a very practical, localized promise: that a specific province can turn their education into a salary, their youth into stable employment, and their household into something sustainable. A province doesn’t need to be postcard-beautiful to attract people; it just needs to make the next step of adult life look structurally possible.

I wanted to dig past the sentimental cultural clichés, so I looked at the hardest evidence available: ISTAT records of where people actually registered their residence in 2024. I combined these internal migration tables with ISTAT’s 2025 Bes dei territori indicators—a massive dataset that measures provincial quality of life across indicators like jobs, innovation, civic participation, and safety. To ensure massive cities like Rome and Milan didn’t skew the reality simply by being huge, I converted every flow into rates per 1,000 residents.

Then I ran a regression model to ask a simple question: What actually predicts where people move?

The results don’t look like a tourism brochure. In fact, they break the usual North-South clichés entirely.

1. The Real Migration Scoreboard

If you look at net migration (arrivals minus departures per 1,000 residents), you find an unexpected leaderboard.

In 2024, the biggest net winners weren’t glamour destinations like Florence or Milan proper. They were places like Pavia, Alessandria, Vercelli, Imperia, Ferrara, Piacenza, Genova, and Cremona. On the flip side, the steepest net losses were in Caltanissetta, Napoli, Enna, Nuoro, Agrigento, Barletta-Andria-Trani, Reggio di Calabria, and Palermo.

Figure 1. Net migration per 1,000 residents by province, 2024.

As you can clearly see on the map in Figure 1, this isn’t just a basic Milan or Rome growth story. Take Pavia: it wins heavily because it is tightly plugged into the Milanese labour and university ecosystem without forcing residents to pay Milanese rents.

Piacenza and Cremona thrive because they sit in the hyper-productive geographical belly of the Po Valley. Alessandria and Vercelli aren’t industrial chic or major tourist hubs, but they sit within a northern industrial belt where job-hopping and economic mobility are seamlessly legible.

Crucially, none of these high-performing provinces boast the majestic, snow-capped scenery of the Alps, or the iconic beach coastlines of Liguria, Amalfi, or Sardinia. Yet, they pull people in anyway. A province doesn’t need an alpine view or a beach resort to be a demographic magnet; it just needs a functioning local economy.

Meanwhile, the losing side of Figure 1 is brutal. Interior Sicily, Calabria, and parts of Puglia aren’t just graying due to a low birth rate—they are actively losing the competition for working-age adults.

2. High-Mobility vs. Being “Stuck”

A common trap in migration analysis is treating arrivals as the ultimate sign of attraction. But that’s only half the equation. To understand what’s really happening, we have to look at how people arrive and how they leave across the country.

Figure 2. Total arrivals per 1,000 residents by province, 2024.

Figure 3. Total departures per 1,000 residents by province, 2024.

If you compare the arrival rates in Figure 2 with the departure rates in Figure 3, you notice something fascinating about Northern Italy: it operates on intense economic churn. The provinces with the highest raw arrival rates in 2024 were places like Pavia, Biella, Valle d’Aosta, and Imperia, but many of these exact same regions also suffered very high departure rates.

People move in, people move out, and they constantly trade one province for another as jobs, rents, relationships, and childcare needs evolve. This churn is actually a sign of economic health. A high-mobility province is a functioning labor market where people feel secure enough to relocate for a job, knowing they can also easily leave if a better alternative pops up nearby.

The opposite of mobility isn’t cozy regional stability; it’s being stuck.

When you look at what drives these patterns, the data is crystalline: provinces with higher overall employment, vibrant youth job markets, robust patenting (innovation), and high voter turnout see people flood in. Conversely, areas heavily defined by economic fragility and low-income pensioners keep outsiders from ever showing up.

3. The Graduate Filter or The Internal Brain Drain

The cleanest, most telling statistical signal comes from internal migration between Italian provinces.

Figure 4. Internal net migration per 1,000 residents by province, 2024.

Figure 5. ISTAT’s graduate-mobility indicator, meaning whether a province gains or loses Italian graduates aged 25-39, and internal net migration.

As shown in the internal migration map (Figure 4), the flight from the South to the North becomes even starker when you isolate domestic moves. And if you look at the regression plot in Figure 5, the single strongest predictor of whether an Italian province will grow internally is what ISTAT calls the “mobility of Italian graduates aged 25-39”.

This tracks whether a region is gaining or losing young adults with university degrees. This is the exact demographic for whom moving matters most: they are young enough to change paths, educated enough to rigorously compare opportunities, and old enough that their choice dictates where they will lay down long-term family roots.

Provinces that can attract or hold onto young educated adults win the broader migration game across the board. It is the ultimate statistical translation of the everyday phrase “there is work there”.

Conversely, areas plagued by low pension incomes, weak labor market participation, and poor school competencies suffer devastating population drains. This tells us that people aren’t just moving for a paycheck today; they are reacting to the local stock of human capital and mapping out the perceived future of the place.

This leads to a harsh truth: The South isn’t just being abandoned. It is facing a highly selective demographic and brain drain filter.

Migration doesn’t hit every group equally; it actively sorts the population. The filter allows those with the highest resources, university degrees, and economic runway to exit toward northern labor markets. Meanwhile, it effectively traps those who lack the liquid capital to relocate or who are bound by local family obligations.

Over time, this brain drain filter creates a compounding downward spiral. It hollows out the origin provinces by removing the very human capital—young, educated professionals and tax-payers—needed to revitalize the local economy. This leaves the remaining community older, poorer, and more reliant on low-income pensions, performing even worse on the exact indicators that attract future investment. Migration isn’t just an after-effect of geographic inequality; it is the exact mechanism by which geographic inequality reproduces itself.

4. Foreign Inflows Follow a Different Logic

When we shift the lens from internal moves to foreign net migration (international arrivals minus international departures), the map transforms completely.

Figure 6. Foreign net migration per 1,000 residents by province, 2024.

Figure 7. Youth employment and foreign arrivals.

Suddenly, as displayed in Figure 6, positive inflows light up in regions that are completely bypassed by Italian graduates. In 2024, provinces like Ragusa, Vibo Valentia, Imperia, Campobasso, Cosenza, Pavia, Genova, and Rieti saw major foreign net gains per resident.

Why the divergence? Statistically, the standard quality-of-life indicators explain very little here. While the scatter plot in Figure 7 shows that higher youth employment offers a moderate pull, international migration is dictated by entirely different forces: agricultural labor networks, local care economies, the hospitality industry, existing migrant communities, and transnational family ties.

A province can look structurally weak or unattractive to a young Italian professional, yet remain a vital landing pad for foreign workers responding to highly specific local labor shortages. When it comes to foreign departures, the statistical models find virtually no robust signal at all. Leaving Italy as an international resident has less to do with a province’s baseline economy and far more to do with transnational routes, family networks, and the specific passports individuals hold.

5. Leaving Italy Altogether

What about the over 103,000 young adults (aged 18-39) who left Italy entirely in 2024? (Crucially, 85% of them were Italian citizens ).

Figure 8. Young adult emigration abroad and youth labour-market distress. Province-level emigration is age 18-39; labour-market distress is the BES youth non-participation rate.

If you look at absolute numbers, they pour out of giant reservoirs like Milan, Rome, Turin, and Naples. But if you look at the rates per 1,000 residents featured in Figure 8, the highest young adult emigration is found in an incredibly diverse mix of provinces: Bolzano, Campobasso, Vibo Valentia, Rieti, and Sondrio.

This tells us that leaving Italy isn’t a single phenomenon. It’s a mix of three completely different dynamics:

• Rich-border mobility: Highly productive alpine regions (like Bolzano) where popping over the border for work is a normal, integrated economic choice.

• Southern exit: Young people fleeing structurally weak labor markets.

• Metropolitan professional mobility: Corporate and academic high-flyers chasing global hubs.

Is leaving the country acting as a “safety valve” that at least lowers local youth unemployment? The scatter plot in Figure 8 gives an absolute answer: No.

The correlation between young adult emigration abroad and youth labor market distress is practically zero. As the data points show, rich Bolzano has massive youth emigration and a booming youth labor market; struggling Vibo Valentia and Enna have massive youth emigration alongside completely stalled youth labor markets.

The complete data breakdown, regression plots, and supplementary indicator maps (Employment Rate, Graduate Mobility, and Low Pension Income by province) are available below for paid subscribers. Upgrade your subscription to unlock the full model and support deep-dive, independent data journalism.

Read more

To test this model, I examine whether a specific cluster of genetic traits, including lighter skin, blonde hair, taller height, and higher educational attainment, co-occurs in the ancient genomic data. I then track how this trait cluster behaves across different archaeological periods, climate variations, group structures, and modern replication samples to see if the pattern holds under scrutiny.

When analyzing ancient European genomic data, a strange statistical pattern emerges. It isn’t just about skin pigmentation: a full phenotype cluster, combining lighter skin, blonde hair, and taller height, correlates significantly with a higher polygenic score for educational attainment (EA PGS).

The real issue isn’t whether ancient Europeans were magically programmed for formal schooling. It is whether this genetic proxy aligns with visible physical traits in ways that track archaeological and social structures.

The Literary Foil: Phenotype in the Cultural Imagination

There is also a deeply rooted cultural reason this question is so tempting to ask. European folk traditions have long imagined a “noble” body through a distinct aesthetic package: height, fairness, bright hair, and light eyes. To be clear, classical literature does not prove that this phenotype was biologically elite, and ancient color words are notorious for not mapping cleanly onto modern racial or genetic categories.

Yet, physical appearance was already intensely legible in both social and erotic spheres, not as a simple, monolithic blond aristocratic ideal, but as a site of intense cultural friction. The Roman upper classes certainly prized an indoor, fair-skinned aesthetic as a status symbol, but Latin poetry consistently complicates and wrestles with this very code.

We see this exact tension play out in Virgil’s Bucolics. In Eclogue 2, the shepherd Corydon is spurned by the fair, elite city-boy Alexis, prompting Corydon to famously warn him: “O formose puer, nimium ne crede colori” (”O beautiful boy, trust not too much to your complexion”). He reminds Alexis that in nature, white flowers are often left to wither while dark hyacinths are eagerly gathered. Virgil explicitly returns to this defense of the darker, sun-baked complexion in Eclogue 10. There, the heartbroken Gallus fantasizes about a rustic lover, Amyntas, and immediately cuts off the predictable high-status objection to his skin tone: “quid tum, si fuscus Amyntas? et nigrae violae sunt” (”What of it if Amyntas is dark? Violets, too, are black”).

The point, then, is historical: ancient societies were hyper-aware of these physical variations and actively argued over how they mapped onto desirability, rank, and beauty. If classical literature shows that phenotype and status were already entangled in a complicated cultural debate, the genomic data allows us to test whether they were structured on the ground.

Unpacking the Mechanisms: How Do These Relationships Emerge?

Before diving into the numbers, we need to clarify how these relationships could actually materialize in the ancient world. It is easy to misinterpret a correlation between physical phenotypes and a modern behavioral proxy like an EA PGS. This isn’t an exercise in abstract genetic determinism; there are distinct historical, environmental, and social pathways that could bundle these traits together.

First, consider height. In ancient societies, a positive relationship between height and social status (or its genetic proxies) likely emerged through two clear, reinforcing pathways:

• The Nutritional Pathway: Higher social classes consistently had better, more secure access to high-protein diets, specifically meat and dairy. Over generations, access to superior nutrition correlates heavily with both increased physical stature and the overall consolidation of wealth, privilege, and social power.

• The Physical Force Pathway: In volatile, early stratified societies, physical dominance and martial capability directly translated to high status. Taller, physically imposing lineages were structurally more likely to secure elite, protective, or ruling roles, binding height alleles to the upper strata of the social structure over time.

Second, consider pigmentation. A link between lighter skin and higher educational or status proxies can be explained by a very practical social shift:

• The Occupational Pathway: As societies grew more complex and stratified, elites, scholars, and administrative rulers transitioned away from grueling, dawn-to-dusk agricultural labor to indoor, sheltered roles. Individuals from lineages that spent less time outdoors facing intense UV exposure faced different selective pressures, or actively selected partners within an insulated, non-agrarian social class that prized unexposed skin.

When these dynamics pair with assortative mating (high-status families marrying strictly within restricted elite networks) or migrant-local stratification (incoming groups arriving with distinct physical traits and immediately occupying a specific social tier), a correlated phenotypic package is born. Over time, selection and social sorting do not act on each trait in isolation. Height, pigmentation, and behavioral proxies move together because they are attached to the exact same socio-demographic package.

Coming up next: Is this hypothesis just a compelling narrative, or does it hold up under rigorous statistical scrutiny?

To find out, we have to look at the actual numbers. Below, I unpack the results of a series of strict regressions across thousands of ancient genomes, tracing how this phenotype cluster behaves across different historical eras—from egalitarian hunter-gatherers to the complex hierarchies of the classical world.

We will also dive into:

• The “Universal Slope” Illusion: A look at how an individual’s position in ancient genetic space drastically moderates this correlation, proving that the signal is conditioned by historical migration and ancestry rather than a uniform biological rule.

• The Environmental Reality Check: We examine the role of latitude, UV exposure, and winter temperatures to see if the signal can simply be explained away by climate geography.

• The Within-Group Proof: A decomposition of the data to prove that this isn’t a statistical illusion driven by broad “group selection” or shifting population means.

• The Control Settings: A direct comparison with modern Europeans and ancient East Asians, establishing exactly why this is a unique, highly localized archaeological phenomenon.

To unlock the full regression models, the historical timelines, and the statistical proof of ancient social stratification, consider becoming a paid subscriber. Your support directly funds this deep-dive research into ancient genomic data.

Read more

But from a genetic standpoint, the Jomon are an anomaly. They aren’t just “ancient Japanese” in a modern sense. Instead, modern mainland Japanese ancestry is a tripartite (three-part) cocktail (Okada et al. 2024).:

• ~12.4% Jomon (the original foragers)

• ~21.2% Northeast Asian (associated with the influx of Yayoi period farmers)

• ~66.4% East Asian (linked to Kofun period state formation)

Because of this distinct genetic layering, Japan serves as a fascinating evolutionary test case. If the ancient Jomon possessed a unique polygenic score (PGS) profile—the collective genetic variants that influence complex traits—does a shadow of that profile still linger in modern Japanese populations today?

To find out, I looked at three specific polygenic scores: educational attainment (EA), standing height, and a light-skin pigmentation score based on an East Asian GWAS (Kim et al., 2024).

The Ancient Baseline: How the Jomon Stood Out

To set the baseline, I pulled imputed ancient DNA from a broad East Asian dataset, grouping 17 ancient samples explicitly identified as Jomon. I then stacked them up against other ancient East Asian populations.

Because these ancient samples span a wide range of eras, a raw comparison could easily mask or mimic evolutionary shifts. To fix this, I adjusted the model for sample age. The numbers below reflect the standardized PGS effects driven strictly by Jomon group identity, independent of time.

Table 1. Time-adjusted Jomon PGS effects in the ancient dataset.

The Jomon are low for EA, extremely low for height, and high for the Kim light-skin score.

Figure 1. Time-adjusted ancient Jomon PGS compared with other metadata population groups.

The ancient Jomon pattern is clear: low EA, very low height, and high light-skin PGS. But the more interesting question is whether any of this survives in modern Japanese, who are mostly descended from later mainland-derived Yayoi and Kofun ancestry. That is where the result gets more complicated.

Read more

Paid newsletters are fundamentally subscription businesses. Their growth relies on a simple truth: new paying readers must arrive faster than existing ones leave.

While this sounds obvious, it introduces a critical mathematical reality: even if a publication consistently acquires new subscribers every month, net growth will flatten if cancellations rise alongside the growing subscriber base.

This model is generalizable to any subscription product where:

• New paid subscribers arrive at an average daily rate (x).

• Existing paid subscribers cancel at an average per-subscriber rate (c).

By using my own Substack export as a case study, we can test these principles against a concrete subscriber curve.

Case Study: My Substack Data

When applied to my actual data, the model reveals that the current growth curve is not exponential.

The Baseline Forecast

If recent acquisition and finalized-cancellation rates hold steady, the paid subscriber list is projected to plateau around 173 paid subscribers. Given a current count of 123, the publication has roughly 50 more paid subscribers to gain before net growth grinds to a near-halt.

Note: This estimate is a baseline, not a prophecy. Depending on the data window analyzed, the plateau estimate fluctuates between 171 and 176. While the precise number moves, the core conclusion does not: under recent conditions, the curve bends.

This slowing growth is not due to an audience ceiling or Substack limitations. It is basic subscription arithmetic: if daily acquisitions (x) remain flat while total cancellations scale naturally with a larger subscriber base, net growth must slow down.

Figure 1. Paid subscribers rose quickly, but the curve is already bending

The Mathematical Model

The model is deliberately simple: tomorrow’s paid subscribers equal today’s paid subscribers, plus new paid subscribers, minus churn.

N[t+1] = N[t] + x - cN[t]

Here N is the active paid subscriber count, x is the average number of paid acquisitions per day, and c is the daily cancellation rate per active paid subscriber. When x equals cN, growth stops. The implied plateau is N* = x/c.

Data Definitions: Acquisition vs. Churn

• Churn: Measured by finalized cancellations (when a subscriber is officially removed from the paid count) rather than cancellation initiations (intent to cancel).

• Acquisition: Analyzed through two lenses:

  1. Visible Event Count: New paid sign-ups plus tier upgrades.

  2. Accounting-Consistent Acquisition: The net change in paid subscribers plus finalized cancellations. This version better matches the actual paid-count series, accounting for minor ledger inconsistencies in the event export.

Figure 2. Acquisition is the unstable part of the model

Has churn stayed stable?

The simple model assumes a stable per-subscriber churn rate. If I use the whole history, that assumption looks too strong: a formal test rejects one constant churn rate over the whole observed period. But the publication started on July 1, 2025, and the first few months are a cold-start period. When the paid base is tiny, one or two cancellations can produce a huge apparent churn-rate swing.

Excluding July through September changes the result. From October 1, 2025 onward, the estimated daily churn rate is 0.42%. Adding a time trend to daily churn gives p = 0.61; the weekly version gives p = 0.868; and monthly churn differences give p = 0.311. In other words, after the launch period there is not strong evidence that churn kept trending up or down.

Recent daily churn estimates also cluster around 0.38% to 0.46% of the active paid base per day. In plain English, that is about 3.3 finalized cancellations per week in the recent March-to-May window.

I tested this more formally with Poisson churn models that treat cancellations as proportional to paid-subscriber exposure. Adding a linear time trend to daily churn gives p = 0.0025. The weekly version gives p = 0.00261. Monthly churn differences give p = 0.00613. Those full-history p-values are evidence against a single stable churn rate across the entire history, mostly because they include the launch period.

So the model should be read as a post-launch/recent-conditions model, not as a universal law. After the first three months, the stable-churn approximation looks much more defensible. The bigger practical uncertainty is still acquisition: x moves around a lot.

Figure 3. Recent churn has hovered around 0.4% per day

Figure 4. Weekly losses are roughly proportional to the paid base

The plateau estimate

Using the March 1 to May 18 period, the accounting-consistent acquisition rate is 0.722 subscribers per day and daily churn is 0.416%. That gives N* = x/c = 173.4. Rounded, the model says about 173 paid subscribers.

The approach to a plateau is gradual. Under this model, I reach 90 percent of the remaining distance to the plateau around November 22, 2027, 95 percent around May 06, 2028, and 99 percent around May 27, 2029. That last date should not be read too literally; it mostly illustrates that plateaus are asymptotic.

Figure 5. Under recent conditions, the model plateaus near 170 paid subscribers

Window sensitivity

Figure 6. December was not normal growth

What would change the answer?

Left unchanged, the current growth engine has a visible structural ceiling. To push past 170 subscribers, the publication needs either a higher steady-state acquisition rate or lower churn.

Which means, of course, that there is a remarkably simple way for you to help me test (and completely break) this model in real time: subscribe. If enough of you convert to paid, the acquisition rate spikes, the formula resets, and this entire forecast becomes delightfully, beautifully wrong. You know, for science.

Methods note

The paid subscriber count comes from the daily Substack subscriber export through May 18, 2026. I use Paid only, excluding comps, gifts, and free trials. The growth-event file does not perfectly reconcile with the paid-count file, so the Fortunately, there is a remarkably simple way to test—and break—this model in real-time: subscribe. If enough readers convert to paid, the acquisition rate rises, the formula resets, and this entire forecast becomes safely obsolete.paid-count series is treated as ground truth and finalized cancellations are treated as the best observed loss series.

The fascinating question is what happens to this identity when we transition from chronicles and archaeology into the genome.

We know the early medieval Slavic expansion happened. But did it leave a distinct genetic signature, or did it dissolve into the existing northern and eastern European landscape? The data suggests a complex mix of both. While West Slavic ancient samples align well with modern populations, formal tests show they sit in a tight genetic cluster alongside medieval Germanic, North Sea, and Iron Age Celtic references.

The reality is not a unified Slavic genetic block. Instead, the data reveals a set of regional populations sitting on overlapping European clines. These groups were shaped as much by local pre-Slavic substrate populations as they were by the medieval spread of the language itself.

Past the paywall, I break down the hard data:

• A comprehensive Europe/Levant PCA mapping Slavic, Baltic, Finnic, Germanic, Balkan, and West Asian groups.

• An analysis of ancient Slavic proxies and their modern descendants.

• Whether Scythian and Sarmatian groups are better sources for Slavic ancestry than Baltic, West Slavic, or Corded Ware-related references.

• The methodology explaining why the West Slavic signal remains the clearest.

• Hudson Fst distances comparing modern Slavs to neighboring populations.

• The deep history of haplogroup R1a, and why its connection to the Slavic expansion is frequently misunderstood.

Read more

Even commercial DNA services like NexoGENO lean into this, framing the North as a blend of Continental, Celtic, and Germanic influences, while the South is defined by Greek, Near Eastern, and Mediterranean roots, like in the picture below.

This is the same intuition in a more commercial form. Italy is presented as two visible types, one northern and one southern. The question is whether that visual shorthand corresponds to anything measurable in the genome, once we stop staring at faces and actually test allele sharing.

The Genetic Question

Once we move past the “Barbarian Invasion” myths—the Goths, Lombards, and Franks—we have to ask:

1. Did the North actually absorb measurable ancestry from these northern migrants?

2. Is that migration what creates the North-South divide, or is the “pull” of the Eastern Mediterranean on the South the more significant factor?

Moving Beyond Fragile Models

Previously, I used Genoplot and G25 models to test these theories. While they confirmed a general North-South shift, these models are notoriously prone to “overfitting.” By swapping a Danish Iron Age proxy for a Hungarian Lombard one, you can make the percentages jump around until they fit a specific narrative.

To get a more robust answer, I’ve shifted my methodology for this follow-up. Using AADR Human Origins data and formal statistics, I’ve moved away from arguing over whether a population is “12% or 25% Germanic” and started asking a more precise question:

Do Northern Italians share more genetic drift with Migration Period northern Europeans than Southerners do? Conversely, do Southern Italians share more drift with Greek, Levantine, and Anatolian populations?

The Reality

The short answer is yes.

The long answer? The direction of the shift is undeniable, but the actual genetic distance is relatively small. When you strip away the percentage models and look at the raw data (the PCA plots and the f4-statistics) the pattern remains. The North pulls toward the Vikings, Saxons, and Lombards; the South pulls toward Cyprus, Anatolia, and the Levant.

The history is messier than a meme, but the genetic gradient is very real.

Read more