Richard Socher thinks science slowed down because it split into too many pieces, not because it ran short of money. He is the founder and CEO of You.com and the co-founder and CEO of Recursive. In his new book, The Eureka Machine, he argues that AI can join those pieces back together.
Peter Diamandis, the founder of XPRIZE and Singularity University, summed up the book's main claim at the start of the conversation. He hosts the Moonshots podcast, which recorded this episode on October 2, 2026. The claim is that AI will deliver "a century of scientific breakthroughs in the next decade." It would do this by connecting every stage of research at once: forming a hypothesis, running the experiment, analyzing the data and building the theory. Socher calls this "full-stack AI" for science. The book's official site lists a September 22, 2026 release. It presents the compressed century as Socher's thesis, with examples from medicine, cell biology, economics, neuroscience and astronomy.
Socher's answers centered on two questions: why biology, more than physics, is where he expects AI to matter most, and why even fast discoveries will reach patients slowly.
Fragmented fields
Socher's diagnosis begins with specialization. AI researchers rarely work on "AI" as a whole, he said. They work on narrow topics such as optimization methods for neural networks. Biology is divided the same way: a researcher becomes a cell biologist, a molecular biologist or a medical researcher, and "there's so much separation that it's hard to weave it all together."
He said this is the moment for AI to "help us weave back together all these separate pieces." He compared it to language. People could never write down every rule of natural language and combine those rules into a system that holds a conversation. That took a large neural network and a great deal of data. Biology, he said, is about to produce that kind of data too. Fields that have mostly tried to understand nature could then become "programmable engineering sciences."
Socher listed what he sees as the ingredients now in place. Large language models hold broad world knowledge. More scientific data is being digitized. Simulations keep improving. Robotic lab automation is coming. On top of all of that would run an "agent swarm," many AI agents working together.
When Alexander Wissner-Gross, a computer scientist and the founder of Reified, pressed him for timelines, Socher answered with an argument about abstraction. Computer science, he said, has climbed so many levels above zeros and ones that it has met the rest of humanity in plain English. People can now program in ordinary language. He hopes other fields will climb the same way until they all meet in natural language.
"Flourishing, not getting cooked"
Diamandis and Wissner-Gross, who co-wrote a book called Solve Everything, have often said on the show that math is "cooked," meaning AI is taking over its hardest problems. Diamandis said "we're seeing millennium prizes all falling" and suggested physics would be next. The Clay Mathematics Institute's rules for the Millennium Prize Problems mean a claimed solution is not quickly accepted. It must be published in a qualifying outlet, wait at least two years, and win general acceptance from mathematicians worldwide.
Socher rejected the word. "I think they're flourishing, not getting cooked," he said. Some mathematicians have worried that AI could hurt their field because people still need to be trained. Socher compared that to a medical researcher calling it "really a bummer for the field that we cured all these diseases for people." That, he said, "would be insane." His advice to mathematicians is to "grab as many as you can and work with AI to solve them": to prove as many theorems as possible now. After that, he expects the work to shift toward inventing things. "What kind of new formalisms can you create," he asked, that would be interesting for AI and humans to solve together.
Wissner-Gross went along with it. Perhaps, he said, he should be "using flourishing as a transitive verb" and say AI is flourishing math.
Dave Blundin, founder and general partner of Link Ventures, brought a story from MIT. He had just spent an evening with Sertac Karaman, who runs the university's Laboratory for Information and Decision Systems. According to Blundin, Karaman said his mathematician friends know they are cooked, and "he literally said cooked." Blundin said those mathematicians are moving into directing AI systems and will end up ahead. Karaman's bigger worry, Blundin said, was biology professors, who are "in complete denial" and use almost no AI in their daily work.
Why biology before physics
Diamandis asked whether physics would be the next field to take off, since there has been relatively little fundamental progress since Einstein. Socher said the biggest field will be biology. Physics, in his view, has been "interestingly stuck." Its great ideas, such as E=mc², turned into engineering, which is where the real-world impact came from. He pointed to fusion. Keeping plasma stable inside a tokamak, a donut-shaped fusion reactor, is a very hard control problem where AI is already used. In a 2022 project with Switzerland's EPFL, DeepMind trained a controller in simulation. It then ran a real tokamak, coordinating 19 magnetic coils to shape the plasma.
Socher's main point was about data and fit. New physics data often needs huge particle colliders and billions of dollars. And he sees neural networks as better suited to biology's kind of problem. Calculus, he said, helped physics understand separate, individual phenomena very well. Neural networks "are great in combining all these little things." Scientists know what one neuron does, or what one bacterium in the gut does. But when many of them interact, he said, "we don't really know anymore how that works."
The book lays out four pillars, and Socher boiled the scientific method down to ideating, implementing and validating ideas. The faster that loop closes, the better, especially with an open-ended search process running on top. He said this idea also inspired Recursive.
Diamandis, who said he is an investor in Lila Sciences, described the company as building a "scientific superintelligence" that runs a million square feet of robotic lab space. Socher said Lila and Periodic Labs are among the few really going after this closed loop. Lila describes reasoning models connected to scientific tools and automated labs, with each experiment's results used to train the models further. Periodic Labs pairs AI scientists with autonomous labs, starting with physics and chemistry. In biology, Socher said, you still need real "wetware" experiments on living material.
Virtual cells and the bitter lesson
Wissner-Gross asked whether the path to solving biology runs through "digital twins" of cells: computer models that predict how a cell will respond to a change. Socher said he had written a chapter on the virtual cell when he started the book three years ago and was proud of it. Then he had to rewrite the whole thing, because in the meantime Mark Zuckerberg's foundation and many others had started building virtual cells. "Everyone has a virtual cell model," Wissner-Gross said. In October 2025, the Chan Zuckerberg Initiative announced a partnership with NVIDIA to build infrastructure for such models, handling billions of observations of individual cells.
Socher explained the idea through the "bitter lesson," a well-known essay by AI researcher Rich Sutton. In Socher's summary, experts keep coming up with clever ideas and elegant theories. But the approach that wins at scale is usually the simplest one: a large neural network trained end to end on lots of data and computing power. Socher said this happened in natural language processing. Experts once doubted that a single model could hold any conversation. He sees biology in the same place now: experts know so much that they can't imagine one model learning all of it. "But if you have enough data, you can," he said.
That data is starting to arrive. Socher pointed to Tahoe Therapeutics and its large "perturbation" studies, which record how cells change when something is added to them. The team's Tahoe-100M dataset includes more than 100 million profiles of individual cells from 50 cancer cell lines, exposed to 1,100 small-molecule perturbations. "Yes, all models are wrong," Socher said, "but more and more of them will be useful." Virtual cells are the book's third pillar. His rule of thumb is that AI can solve problems in a domain once you have "anything you can simulate, anything you can verify."
Salim Ismail, the founder of Open ExO, said science "has always been a coordination problem," organized into departments and journals. He described Socher's project as shrinking the time between a hypothesis and a result, which he called "domain collapse of the scientific method."
Organoids instead of mice
Socher's favorite example was a company the episode's transcript renders as "Peril Bio." It is most likely Parallel Bio, which describes its core platform as human immune organoids: tiny lab-grown tissues that copy part of an organ. Its biobank includes more than 170 donors of different ages, sexes, ethnicities and disease states. Socher said the company grows small lymph-node organoids from stem cells. Lymph nodes are part of the immune system, which matters for immunotherapy, the cancer treatment that turns a patient's own immune system against the disease.
"If you love animals, you too can love AI," Socher said. He said the company got FDA approval to skip animal trials by showing that how its organoids react to drugs and toxicity tests is more predictive of how those drugs interact in real human bodies. "Turns out we cured most diseases in mice. It's not that helpful," he said. The FDA's published material supports a policy direction but does not confirm a company-specific approval. In April 2025, the agency announced a roadmap to reduce animal testing, starting with antibody drugs. It encouraged organoids, human cell tests and computer models, and offered streamlined review to developers with strong non-animal safety data.
Diamandis took the idea further. Someone could take his skin cells, turn them into stem cells, grow his own organs and test how a drug would work for him in particular.
Why cures still take years
Socher refused to give a single date for when science is "solved." Instead he expects a steady stream of results. "Multiple different diseases will get cured in the next 12 months," he said, probably simpler ones where a single gene needs fixing. There are many such diseases. More complex diseases would follow, along with better battery materials. How fast this goes, he said, will depend on "how much money do you want to put into compute."
Human testing is the brake. Socher said biotech companies used to have one drug in development for 10 years. Sometimes they went public before knowing whether it worked, then died when late-stage trials failed. Now, he said, newer companies have five to 10 compounds in phase 3 trials after two or three years. Still, getting treatments to patients at scale in the United States with FDA approval involves "natural delays" that he put at half a decade to a decade. Chemistry and physics, where researchers can iterate without long human trials, will move faster.
The FDA's description of clinical research shows where that time goes. Phase 1 safety studies usually take several months. Phase 2 can take up to two years. Phase 3, typically with 300 to 3,000 patients, often runs one to four years. Later in the conversation, Socher argued that a future superintelligence could cure all diseases. Even then, he said, confirming that it had would take more than three years "just because of FDA things," even if every compound were ready to manufacture the next day.
Diamandis disagreed. He expects cell simulators to prove that a drug works "in this cell, your cell," and said "the idea of human trials is going to get incinerated, I think." Socher found himself on unfamiliar ground. "Usually on all podcasts, I'm the one who is the optimist," he said. "We're just disagreeing on timelines." His objection was about measurement. Today, "we cannot currently measure all the proteins that happen in one cell without destroying that cell." A perturbation study adds one molecule to one cell, watches how it changes and gets one data point. A perfect virtual cell would need huge numbers of those points, and then models of how many cells work together.
That is the company Socher said he would start if he had time: a system of organoids that models not just one lymph node but the whole lymphatic system. Diamandis said someone is already working on it and offered to introduce him.