The most consequential use of artificial intelligence will not be chatbots or workplace productivity tools but the automation of scientific discovery itself, according to pioneering AI researcher and entrepreneur Richard Socher. In his new book, «The Eureka Machine: Why AI Is the Key to Unlocking a New Era of Scientific Discoveries,» Socher argues that a system combining all human knowledge, simulating billions of experiments, and generating breakthroughs could help end disease, extend human life, and solve the hardest materials problems facing civilization.
Socher, who spent nearly two decades at the forefront of natural language processing, is the CEO and cofounder of Recursive, a company building self-improving superintelligence to automate knowledge discovery. He also founded the search infrastructure provider and the venture firm AIX Ventures, and previously served as chief scientist and executive vice president at Salesforce and as an adjunct professor of computer science at Stanford.
He points to October 2024, when two AI pioneers won Nobel Prizes, as a signal of where the field's significance lies. Demis Hassabis shared the chemistry prize with John Jumper and David Baker, while Geoffrey Hinton shared the physics prize with John Hopfield. Neither Hassabis nor Hinton was a trained chemist or physicist, yet the neural network work they and their teams did was making a profound impact in those sciences. Socher argues that doomsday scenarios dominate headlines while the positive effects of AI on science receive far less attention. Science leads to technology, he writes, and technology drives human flourishing.
A central part of his argument is that everything has a language, and AI can learn to speak it. A language model asked to predict the next word in a sentence about driving south from Berlin must acquire geographical knowledge to do so well. At sufficient scale, it builds something like a mental map of the world, including distances, city sizes, and relationships between regions, none of which was explicitly programmed. The same principle extends beyond human language. Socher describes ProGen, a project at Salesforce in which his colleagues and he trained a large language model on 280 million protein sequences across 19,000 protein families and then asked it to create new proteins. It generated millions of artificial sequences, including ones in the lysozyme family, enzymes that break down bacterial cell walls. Seventy-three percent folded correctly and showed the antibacterial properties the researchers sought. Socher frames the shift this way: what calculus did for physics, AI will do for biology.
He also argues that the next breakthrough may already be sitting in the scientific literature. In the early 1960s, Arno Penzias and Robert Wilson at Bell Labs could not eliminate a persistent hiss in their antenna, scrubbing pigeon droppings and taping every rivet without success. They were hearing the afterglow of the Big Bang, a phenomenon a Russian cosmologist had suggested two decades earlier and a Princeton team was chasing without knowing the Bell Labs researchers held the answer. A mutual acquaintance connected them, and the collaboration led to a Nobel Prize in 1978.
Today, 23 million researchers publish in more than 54,000 journals, and no human can see all the pieces at once. In the 1980s, information scientist Don Swanson linked separate bodies of literature on fish oil's effects on blood vessels and on Raynaud's syndrome, a vascular condition, pointing to what Socher calls undiscovered public knowledge: findings already published but connected to nothing. AI can digest the full literature and spot such connections, he argues. James Evans and colleagues at the University of Chicago used AI to map nearly 20 million biomedical papers and the nine million researchers who wrote them, modeling not just what science knows but where scientists are looking, then inverting the map to find unexplored regions where the next connection might wait. Evans calls these alien hypotheses.
Socher describes the Eureka machine as the culmination of several converging streams of AI development, potentially the last organic invention needed before a new era of automated discovery begins.