Richard Socher argues that AI's history is the steady replacement of human expertise with scale, and that the last step is automating the scientific method and pointing it at AI itself. He says Recursive, his startup, could reach strong recursive self-improvement within a year or two, that this will not mean an overnight takeoff, and that the payoff is a century of scientific progress in a decade, with some software and jobs enduring however capable AI becomes.
From features to the scientific method. NLP moved from expert-designed features to architecture engineering, then to one model prompted with any question, his rejected but GPT-cited decaNLP paper. The field is now stuck one level up, with humans generating ideas and AI mostly implementing them. He calls labs' use of coding tools a weak form of self-improvement; in the strong form humans set goals, rewards and environments, while AI proposes, implements and validates. He dismisses doom fears as imagination of the kind that held back nuclear power, which he says kills fewer people than coal each year, and says nearly all data centers use closed water systems.
The Eureka machine. His bestseller argues AI can link hypothesis, simulation, experiment and validation across physics, chemistry, biology, neuroscience, economics and astrophysics. Discoveries should pass four pillars, world knowledge, measurements, simulations and real experiments, orchestrated by an agent swarm. Hallucination aids exploration of a vast combinatorial space, but facts need grounded, cited search, which is why he still runs you.com as search for agents. Some explanations may exceed human intuition; he notes even GLP-1 drugs turned out to work differently than first thought, and says neural networks are far easier to probe than brains.
Simulation and economics. Asked how to stop imperfect simulations breeding false confidence, he concedes humanity lacks full answers but says fidelity keeps rising: neural approximations made weather modelling cheaper, and persona-prompted LLMs increasingly match real survey outcomes. His Salesforce "AI economist" paper used two-level reinforcement learning, with a planner setting taxes and subsidies while self-interested agents dodge them, to iterate policy far beyond what history allows. He says economics has not had even a machine-learning moment, being ideological and reliant on one-step, linear models.
Biology becomes engineering. He predicts AI will make biology programmable, as large models displaced hand-crafted linguistics. Organoids predict drug effects better than mice, and one company has FDA approval to skip animal trials for immunotherapy toxicity testing, saving animals and patent time; brain organoids are kept smaller than a bee's brain. He sees longevity as a major application.
Why takeoff is not instant. He now says strong recursive self-improvement may take only one year, but bigger ideas need bigger models and more compute, physical constraints rule out instant robot armies, and experiments still take time, though software experiments run far faster than biology. Recursive uses its system internally and plans to open-source some artifacts, not the machine.
What survives. Coding went superhuman because it is simulatable and verifiable, the test he applies to any field, but critical code should stay readable by a few humans, and coding remains a "gym for the mind". He argues most innovation recombines past ideas, so even single-inference novelty is plausible. Researchers will act like grant program directors steering AI, and one weekend now lets him run 10,000 experiments alone. Durable software has unique data, network effects or a brand; durable jobs lie in entertainment, health, research and AI; fame and attention stay scarce amid abundance. He closes on "constructive optimism": the future is something to build, not predict.