Past midnight in the Hearst Memorial Mining Building on the campus of the University of California, Berkeley, beyond a vaulted entrance and down a marble staircase, the experiments in the A-Lab are running without people. Powdered precursors and oxides twirl through the laboratory in crucibles shaped like sake cups, then are slurried and spun in centrifuges with zirconium beads, baked in industrial ovens, scanned using x-ray diffraction and, in battery tests, measured for ionic conductivity. Each result feeds the next experiment.

When something goes wrong—a jammed rack, a spilled sample, a precursor running out—the choreography halts. Minerva, Alfred, Prometheus, Jeeves, and a handful of other artificial intelligence–enabled robots that run the lab overnight can’t always reset it themselves. A sleeping graduate student gets an e-mail and a Slack alert, then can log in from bed, check the lab’s cameras and try to fix the problem.

“We want a better material,” says Gerbrand Ceder, the U.C. Berkeley materials scientist who runs the A-Lab. “But we’re also really interested in: Can you build an AI that acts like a scientist?”

On supporting science journalism

If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

The A-Lab is one of a few sites where a new research infrastructure is taking shape. Operated by U.C. Berkeley and Lawrence Berkeley National Laboratory (LBNL), it pairs robotics and lab automation with a custom AI agent that interprets results and proposes the next round of experiments, backed by LBNL’s computing resources. Researchers call it a lab in the loop: a system that can experiment, iterate and suggest the next step. “I don’t think people know what’s about to hit them,” Ceder says.

But speed is not the same as reliability. The A-Lab has also become a test case for the field’s current limits. One of its early high-profile papers, published in 2023 in Nature, reported that the lab had autonomously synthesized dozens of new materials in a matter of days. It was later corrected after outside researchers raised questions about whether the materials were genuinely new and whether the data supported the claims. The episode exposed a central tension in autonomous science: machines can run experiments faster than any human, but the results still have to be verified and interpreted by people.

The tension extends well beyond the A-Lab. Robots now gather data at a scale humans can’t match; machine learning finds patterns in the torrent; AI agents are starting to help researchers decide what to try next. Together, proponents say, the three tools could compress the timeline of scientific discovery—with stakes that run from the cost of new drugs to the global race for biotech leadership. The question is whether faster science will also be better science.

In the three years the A-Lab has been running, it has often outpaced the commercial tools available; early on, Ceder’s team had to rig a fake finger to a machine so the robots could start it. Today the lab iterates at roughly 100 times the speed of a human researcher, and the humans serve as architects of the process, refining the machines, setting the direction of inquiry and deciding what to test next.

That speed is possible because the A-Lab draws on LBNL’s data infrastructure, including the National Energy Research Scientific Computing Center (NERSC), whose supercomputers support large-scale scientific modeling and AI work. NERSC, a part of LBNL, is also building Doudna—named for Jennifer Doudna, a U.C. Berkeley biochemist who shared the 2020 Nobel Prize in Chemistry for CRISPR—a next-generation supercomputer, in partnership with Dell and Nvidia, designed to, among other things, link AI tools, scientific instruments and data across the Department of Energy. For autonomous labs, computing power is becoming part of the bench.

A new furnace system being installed in the A-Lab will track chemical synthesis moment by moment. Researchers currently know the before and after—bake these compounds at 1,000 degrees Fahrenheit for three hours, and you get this result—but not the intervening reactions. By capturing the entire sequence in real time and adapting experiments on the fly, the lab hopes to build the dataset needed for predictive synthesis.

“I’ve also learned that speed makes people think differently,” Ceder says. “If you have an idea and you need to wait three months for the answer, you intellectually don’t remain engaged.”

“When people get rapid answers, they stay engaged with things, and they tend to ask different questions,” Ceder adds.

A bigger hope is that scale plus speed will yield better models of the messy systems researchers are trying to understand—and questions more likely to lead to real answers. “It frustrates the hell out of me that we fail 90 percent of the time,” says Brad Ringeisen, executive director of the Innovative Genomics Institute (IGI), which focuses on genome editing. The biotech industry spends billions of dollars on drug development because so many experiments fail. Ringeisen proposes two solutions: (1) do things in a more automated way with the same failure rate but just do a whole “hell of a lot more of those experiments” or (2) take the IGI approach and try to build a better, more precise model of disease.

On a Sunday in early February, attendees of the annual Society for Laboratory Automation and Screening International Conference, this year held in Boston, could take a complimentary Uber over to Ginkgo Bioworks. The company was demonstrating its Reconfigurable Automation Carts—modular blocks on wheels with barcode scanners and robotic arms, arrayed like a souped-up bank of arcade claw games—which could be programmed on the fly to replicate any sequence of lab steps.

Earlier that day Ginkgo had asked several conference-goers to suggest experiments; they could type plain-language commands into the company’s interface. When visitors arrived that afternoon, the lab was conducting dozens of experiments at once. What stood out, says CEO and co-founder Jason Kelly, wasn’t speed or precision but experimental flexibility. “If you’re a scientist, you’re like, ‘Wait, I can run an experiment overnight? I can wake up in the morning to data with my coffee?’” Kelly says. “That’s a totally new experience.”

Falling robotics costs, better data pipelines and AI-powered natural-language control have together made the modern autonomous lab possible. Ginkgo, founded in 2008, shifted among different business models, including engineering yeast strains for fragrances and foods, before focusing on lab robotics. It’s spent years refining this approach; its aim, as it notes in a promotional video, is to make the lab bench extinct. The modular system can operate more than 100 pieces of equipment and pushes 384-sample plates through any configuration a scientist programs.

Kelly compares lab automation to self-driving cars, saying in his lab it is roughly where Waymo cars were five years ago. Ginkgo recently launched a cloud-lab service that lets scientists across the globe submit an experiment, receive a cost estimate, and, if they proceed, have the work run remotely. It’s getting a handful of new inquiries every day.

“Can you build an AI that acts like a scientist?” —Gerbrand Ceder U.C. Berkeley

Flexibility is one model; industrial scale is another. In Salt Lake City, on the floor of a former Dick’s Sporting Goods, the robotics system at Recursion—which takes images of cells and cultures—can run up to 2.2 million experiments a week. Index-card-size plates with 1,536 wells cycle through incubation, treatment and microscope imaging from multiple angles; the resulting data are analyzed by an AI system.

Recursion processes its more than 50 petabytes of proprietary data through BioHive-2, its in-house supercomputer, and uses those data to map biological processes and search for unexpected drug targets. Its platform has helped build large-scale cellular maps, using models of neurons and microglia, says Christopher Winrow, Recursion’s vice president of neuroscience. The company has advanced several drug candidates into clinical trials.

The rest of the pharmaceutical industry is moving in the same direction, although the payoff for most remains unproven. Lab buildings are being redesigned with more room for servers and heavier power supplies for robotic systems, biomanufacturing and data centers, says Matt Gardner, a biotechnology specialist at commercial real-estate firm CBRE. Swiss firm Roche said in March that AI had helped it develop an oncology drug candidate 25 percent faster than conventional methods; Nvidia and Eli Lilly recently announced a five-year AI drug-discovery lab worth up to $1 billion. “There’s a hope down the road that this leads to faster, better, cheaper drug discovery,” Gardner says. “We’re not there yet.”

This model of high-throughput discovery extends beyond human therapeutics. IGI is pointing the same approach at a planetary threat: methane. With support from Google and TED’s Audacious Project, the lab is sampling the gut microbiomes of a herd of cattle—even tracing how a mother’s microbiome shapes her calf’s—then replicating them and running them through an autonomous system with computer vision. Trained to recognize novel microbial formations, it is working to isolate the organisms that feed methanogens, which produce the potent greenhouse gas.

The project is generating terabytes of data and a working model of the cow microbiome. The hope is to find an intervention, potentially involving CRISPR, that could make the shift durable. The $1-million setup runs with one automation engineer and one microbiologist. “We see the robotics and automation as an assist,” IGI’s Ringeisen says.

Earlier this year Ginkgo partnered with OpenAI to test, in part, whether the robotic lab could operate as an experimental scientist, Kelly says. An OpenAI agent, trained on literature around cell-free protein synthesis—a method for generating proteins without growing living cells—was connected to Ginkgo’s setup. In a preprint, the team reported that over several rounds of experimentation and more than 36,000 unique reactions, the system beat a published benchmark with a 40 percent reduction in protein-production cost. Kelly felt the models proved themselves competent experimentalists. “The overwhelming majority of important stuff in science is happening in the world of atoms,” he says.

But competent at execution is not the same as competent at insight. Closing that gap has become the focus of a cluster of well-funded start-ups.

“Most Nobel Prize–level discoveries are not throughput-limited; they are intelligence-limited,” says Andrew Beam, chief technology officer at Lila Sciences, a start-up building AI designed for scientific reasoning. Most biology Nobels have been awarded for connecting different areas of the field that had been disconnected, he says. Brute force will get you a slightly better drug, he adds, “but it’s not going to get you the next breakthrough.”

Several start-ups and the large AI companies are racing to build that model, ingesting experimental data and partnering with research organizations to train and test them. Anthropic CEO Dario Amodei, a biophysicist by training, wrote “Machines of Loving Grace,” a 2024 essay arguing that AI could dramatically accelerate biological discovery.

In February, Anthropic announced partnerships with the Allen Institute and Howard Hughes Medical Institute intended not only to support lab operations but also to start formulating hypotheses and designing experiments, says Jonah Cool, the firm’s head of life sciences partnerships and deployment. A scientist’s career, Cool says, is often grindy observation or analysis—the kind of work Anthropic aims to accelerate.

The common thread is training on the right data with the right type of learning. Lila, for instance, uses bespoke data and a reasoning model trained with reinforcement learning. “If you think about the diet that ChatGPT and Claude have been fed, that diet comes from the Internet,” Beam says. “There’s some stuff that will give you reflux on the Internet if you eat too much of it.” Lila’s lab setup is built differently, too: not fixed tracks inside a closed loop but open-ended experimentation that, chief autonomous science officer John Gregoire says, is like jazz improvisation.

Beam thinks Lila’s model can outperform the frontier systems on scientific-reasoning tasks. The company claims it used the approach to produce and optimize in vivo chimeric antigen receptor T cell therapy, an experimental cancer treatment that aims to engineer a patient’s own immune cells to target tumors, at roughly 1 percent of the cost of the field-leading development effort. The claim remains a company case study, not an independently established benchmark.

Ringeisen is less sanguine. He worries that in the current austere funding environment, researchers will take the easy path and train AI on existing data—“scrub whatever’s out there; there’s a lot of snake oil that might be sold”—instead of taking the more expensive path of first improving those data. He points to AlphaFold, DeepMind’s protein-structure prediction model, as the template. “That was a highly curated, highly relevant, expansive dataset that allowed AlphaFold to work,” he says. “Let’s re-create that and make the right physiological choices about human disease to be able to better inform those AI models.”

Last December, U.S. Secretary of Energy Chris Wright cut the ribbon on a $47-million Ginkgo system at the Pacific Northwest National Laboratory in Richland, Wash. The installation is part of the Genesis Mission, a federal AI-for-science effort backed by hundreds of millions of dollars in awards, including funding that would connect frontier AI models, automated facilities and data troves from the national labs into a coordinated research network.

Federal investment in this kind of infrastructure is “low-hanging fruit,” says Erwin Gianchandani, the inaugural assistant director for technology, innovation and partnerships at the National Science Foundation, who is helping to deploy CHIPS Act funds to make scientists more productive.

Fully realizing the vision, proponents say, will require that the federal government do more than buy the equipment. A national system of cloud labs needs rules of the road—hardware and software standards, standards for data collection and sharing, conventions for cloud storage—all of which Gianchandani and his team are trying to set. It may also need federal money, as the government has provided many times before.

Even as the Trump administration has moved to cut or constrain many areas of scientific research, it has treated AI-enabled lab infrastructure as a strategic priority. Gains in Chinese biotech have alarmed industry leaders and members of Congress, who view autonomous labs as a critical dual-use capability. The National Security Commission on Emerging Biotechnology, along with congressional allies such as Senator Todd Young of Indiana, has called for sweeping regulatory changes and new federal funding for next-generation labs. China, the commission warns, will “weaponize biotechnology.”

The commission’s recommendations are moving through Congress in pieces, with some already reflected in legislation and others still awaiting hearings or committee action. Its roadmap calls for new federal investment in biomanufacturing and biodata infrastructure, expanded export controls on biotech equipment, and a coordinating body inside the executive branch. “We are now telling policymakers that the decisions they make right now, in this legislative year, are some of the most consequential that they could make with respect to whether the United States is positioned to lead the coming bioindustrial revolution,” says Caitlin Frazer, the commission’s executive director.

“Our job is to show that you can double the pace of science by using AI appropriately,” says Michael Witherell, director of LBNL. “We need fusion. We need better reprocessing of water. All these national challenges—and we need to go faster than China.”

Scientific investigation has often been compared to a streetlight in the dark: researchers cluster under the light of what’s already known, wary of the shadow. What happens when the entire sky is lit? Some of the possibilities, CBRE’s Gardner says, are staggering, such as being able to watch cellular action in real time.

But faster science will still have to prove itself based on old-fashioned scientific standards. The lab of the future may run through the night. The work for scientists will be to decide what should happen when morning comes.