Gregor Mendel

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI returns to Mendel as the patron saint of the premature truth. His story—correct, complete, available, invisible for thirty-four years because the conceptual equipment to recognize it did not yet exist—is the secret history of artificial intelligence, a field littered with ideas that were decades early. Neural networks were dismissed for lack of compute; probabilistic methods waited for data; architectures sketched and shelved until the hardware caught up. Mendel’s fate is a permanent rebuke to the confidence that the important ideas of any moment are the ones currently being noticed. Somewhere in the current literature, probably, sits the key idea of the next era—published and ignored, waiting for the conditions that will suddenly make it legible.

Mendel also illuminates the question of statistics and meaning that haunts every discussion of large language models. He had the complete mathematics of inheritance—the ratios, the factors, the rules of transmission—while remaining entirely ignorant of what was being transmitted. He did not know what a gene was made of, where it resided, or how it worked. He knew only how it behaved statistically. This is the question hanging over every language model: whether a system that captures the statistical structure of language with extraordinary fidelity—the patterns, the regularities, the probabilities of what follows what—understands anything of what the language means, or possesses only the form without the content, the statistics without the semantics. Mendel makes the question vivid because his case is undeniable: he genuinely had the statistics and genuinely lacked the meaning, and the two came apart completely.

The simultaneous rediscovery of his laws by three botanists in 1900 carries the darkest implication for AI governance. When the supporting conditions are in place, a discovery becomes available to the whole field at once, and which particular person makes it is almost a matter of accident. The frequency of independent parallel breakthroughs in AI—multiple groups reaching the same architecture or capability within months—suggests that the advances belong more to the moment than to the person, and that the development of powerful capabilities may be similarly “inevitable” once the enabling substrate exists. A capability that can be built will be built, by someone, once the compute and data are in place; no individual decision to refrain prevents it. Mendel’s three rediscoverers are a parable of why technological restraint at the level of individual actors is so structurally difficult.

Origin

Born in 1822 in Heinzendorf bei Odrau (now Hynčice, Czech Republic) to peasant farmers, Mendel entered the Augustinian monastery of Saint Thomas in Brünn (now Brno) in 1843 after struggling to support himself through university. The monastery proved a remarkable institution for science: it had a library, a meteorological station, and an abbot who encouraged research. Mendel studied physics under Christian Doppler in Vienna from 1851 to 1853, which gave him the physicist’s conviction that nature’s regularities are quantitative and emerge from aggregates—the precise attitude he brought to his garden.

He chose the garden pea for reasons of experimental control: it self-fertilized, establishing pure lines; it had sharply distinguishable traits with no intermediate forms; and it produced enough offspring to make counting meaningful. Eight years, twenty-eight thousand plants, and the careful use of crossing and self-fertilization produced the ratios that revealed the hidden units. He presented his results to the Natural History Society of Brünn in February and March of 1865, published the paper in the Society’s proceedings in 1866, and was elected abbot in 1868, after which administrative duties and a prolonged dispute with the government over monastery taxes consumed the remainder of his life. He died in January 1884 believing his work had failed.

Key Ideas

Inheritance is discrete. Traits do not blend; they sort. The hidden factors of inheritance behave like discrete tokens shuffled and dealt rather than like paint mixed to a single shade. This discovery placed biology, for the first time, on a combinatorial foundation—and opened the fault line between discrete and continuous representation that now runs through all of AI. Symbolic AI holds that intelligence requires discrete symbols combined by exact rules; the connectionist tradition holds that continuous statistical learning suffices; the question of which structures are crisply discrete and which genuinely continuous is still being contested.

The law lives in the ratio. The pattern of inheritance is invisible in any individual case and present everywhere in the aggregate. This is the structural foundation of machine learning: a model learns not from a single example but from the statistical structure of enormous datasets, regularities emergent only in the aggregate just as Mendel’s three-to-one emerged from thousands of peas. The hunger of modern AI for data is the direct consequence of this principle.

Hidden factors (latent variables). Behind the visible plant stood hidden factors that Mendel could not observe, could not isolate, and could prove existed only through the shadows they cast in the ratios of offspring. This is the genotype-phenotype distinction in modern language, and it is the governing structure of representation learning: observed data (tokens, pixels) are the phenotype; latent representations are the inferred genotype; and learning is the inference from one to the other. The same underdetermination Mendel faced—many hidden configurations can produce the same observable result—applies to the latent factors of any neural network.

Recombination generates novelty. Independent assortment produces endless new combinations from a fixed alphabet of discrete units. This is the mechanism of generative AI: a learned vocabulary of parts recombined into arrangements that did not previously exist. The combinatorial space available to a model with a large vocabulary and a long context is effectively infinite, explaining why generative systems seem inexhaustibly creative—and illuminating what they may not be able to do, since Mendelian inheritance cannot introduce a new unit through recombination alone. Genuine novelty, in biology, requires mutation: the introduction of something not already present in the alphabet.

Statistics without mechanism. Mendel’s knowledge was real, predictive, and lawful while remaining entirely ignorant of the physical substance whose behavior it described. This middle category—genuine structural knowledge that nonetheless lacks mechanistic comprehension—is precisely where a language model may reside: possessing the statistics of language as Mendel possessed the statistics of inheritance, really, powerfully, and incompletely.