Genesis Machines

Chapter 1: The New Collaborator

On a gray November morning in 2019, a researcher named Jürgen Schmidhuber stood in front of a room of chemists at the Royal Society in London and asked them to consider a molecule. It wasn’t a particularly impressive molecule—a small, unassuming arrangement of atoms that, on paper, looked like the kind of thing a graduate student might sketch during a particularly dull seminar. It had a core of carbon and nitrogen, a few dangling hydrogen atoms, and a ring structure that suggested it might do something interesting, though nobody could say exactly what. Schmidhuber’s team at the Swiss AI lab IDSIA had generated this molecule using a neural network—a computational system loosely inspired by the brain that had been trained on millions of known chemical structures. The network wasn’t following a recipe. It wasn’t searching a database for similar compounds. It was doing something that looked, from the outside, uncannily like imagination: it was producing a novel chemical formula that no human had ever proposed. The chemists in the room were polite. They nodded. A few of them scribbled notes. But the mood was one of quiet skepticism, the kind that settles over a scientific audience when someone from outside the field presents a result that seems too easy. Schmidhuber, after all, was a computer scientist, not a chemist. He had never synthesized a molecule in his life. He couldn’t even pronounce half the compounds his network had generated. But he stood there, wire-rimmed glasses