On the evening of 8 February 1865, in the meeting room of the Natural History Society of Brünn, a forty-two-year-old Augustinian friar named Gregor Mendel stood up and read a paper. The room was half-full. The audience had come expecting plant hybridisation. What they heard was arithmetic. Mendel described seven traits in pea plants, how they segregated across generations in ratios of exactly three to one, and what this implied about the invisible "factors" of inheritance. He read for an hour. When he finished, nobody asked a question. The society printed the paper the following year in volume four of their Verhandlungen.

Mendel was not a hobbyist. He had grown, cross-pollinated, and counted twenty-eight thousand pea plants in the two-hectare garden of St Thomas's Abbey between 1856 and 1863. He had chosen peas because they were cheap, because they self-fertilised cleanly if left alone and cross-pollinated reliably when opened with tweezers, and because the traits he selected (seed shape, seed colour, pod colour, pod shape, flower colour, flower position, stem length) were each cleanly binary. He had done the mathematics himself. The central numbers of his result, the three-to-one ratio, the one-to-two-to-one when recombined, the independent assortment of unlinked traits, were stable across tens of thousands of offspring.

He was also a beekeeper with fifty hives, a founder of the Austrian Meteorological Society, and, from 1868, the abbot of his monastery. After the promotion his research effectively stopped. He spent the last fifteen years of his life in a tax dispute with the Habsburg state, which was trying to levy the monasteries. He died in 1884, and the monks burned most of his papers.

The paper he had read in Brünn was published in a journal with limited distribution and no real presence outside central Europe. In thirty-five years it was cited approximately three times. The men who might have been expected to read it, Charles Darwin among them, either did not see it or saw it and did not recognise what it was. Mendel had sent a reprint to Carl Nägeli, the most eminent botanist in Munich, who wrote back encouragingly and suggested, patiently, that Mendel should try the experiments again using hawkweed instead. Hawkweed, it turned out, reproduces apomictically, meaning without meaningful genetic mixing, so Mendel's subsequent hawkweed experiments produced no clean ratios and quietly buried his confidence in his own results.

And then, in the spring of 1900, three European botanists working independently published essentially the same rediscovery within two months of each other. Hugo de Vries in Amsterdam, Carl Correns in Tübingen, Erich von Tschermak in Vienna. Each had been crossing plants and arriving at three-to-one ratios and asking what could possibly produce them. Each, in the late stages of preparing a paper, found the Brünn Verhandlungen article of 1866. Correns had in fact cited it the year before. The paper had been sitting in libraries for a third of a century, entirely correct, waiting for a generation with the right questions.

Within a decade, the word "gene" existed. Half a century later, the structure of DNA was solved. The biology of the twentieth century rested on a pile of unread reprints from Brünn.

The Echo

This month, a team led by the Tufts roboticist Matthias Scheutz reported a neuro-symbolic artificial intelligence that uses one-hundredth of the energy of a comparable neural system and performs better on structured reasoning tasks. On the Tower of Hanoi, their robot succeeded ninety-five per cent of the time against the neural-network baseline's thirty-four per cent. Training consumed one per cent of the energy a conventional model would need. At inference, five per cent. The work will be presented at the International Conference of Robotics and Automation in Vienna in May.

The key word is "neuro-symbolic". The "symbolic" half is old. Symbolic artificial intelligence, meaning logic rules, structured knowledge, hand-coded reasoning, was the founding project of the field in the 1950s and 1960s. It was the approach behind the expert systems of the 1980s, which promised to capture specialist human knowledge in if-then rules. When expert systems failed to scale, symbolic AI fell out of fashion. By the 2010s the consensus was total. The field had moved on to neural networks, which learned statistical patterns from data rather than reasoning from rules. The biggest neural networks, trained on the biggest data, won every benchmark. Symbolic reasoning was a fossil.

The fossil has just walked out of the museum. The current frontier problem in neural AI is energy: the data centres powering large models consume gigawatts, and every scaling step doubles the bill. The current frontier limitation is reliability: neural systems hallucinate, fail at multi-step logic, and cannot reliably explain themselves. A symbolic layer addresses both. It imposes structure. It can be inspected. It runs on a laptop.

Scheutz's paper is not the first neuro-symbolic result, and it will not be the last. DeepMind's AlphaGeometry in 2024 used a similar architecture to solve International Mathematical Olympiad problems. Half the field is quietly reading papers from the 1980s that had been filed under "obsolete" for thirty years. The obsolete papers are turning out to contain the missing piece.

Mendel's paper sat in the library for thirty-five years because nobody who read it had the framework to notice what it said. The symbolic-AI papers sat on arXiv for a shorter interval and a different reason: everybody had the framework, and everybody had decided the framework was wrong.

The fastest way to discover something is sometimes to read what was already discovered, by someone you were told not to read, in a journal you were told not to bother with.