Even a dumber AI can win science if it never leaves the lab

Jason Kelly started Ginkgo Bioworks in 2008, straight out of an MIT bioengineering Ph.D., with a mission that sounded like computer science: make biology easier to engineer. Biotech venture capital did not like young founders, and the company was not trying to make a drug, so for the first four or five years it raised no private capital at all. The team lived on government grants and a service business. Fifteen people sat in a Boston lab. Kelly calls it a pretty brutal start, and tells Sequoia partners Sonya Huang and Pat Grady that, by the standards of the industry, they were "totally uninvestable."

The turn came in the summer of 2014. Sam Altman had just taken over Y Combinator and wrote that the Silicon Valley model could work for deep tech — nuclear fission, biotech, materials science. Kelly emailed him: the company was five years old, made no sense for YC, and the post felt like an oasis in the desert. Altman told him to fly out. They met, Ginkgo joined YC, and that is when Kelly dates the company as actually having capital. The mission never changed. The product, he says, has gone all over the place.

The analogy he still uses with Sequoia is the original one. DNA is code, written in As, Ts, Cs, and Gs rather than zeros and ones, and it is the only other coded product besides software. If you can design that code, you can program cells. Unlike computers, which move information, cells move atoms — they build whatever you want, they self-repair, they self-replicate, and they do molecular assembly at a TSMC-like caliber for nearly free. That substrate is four billion years of evolution. The catch, which has been Ginkgo's problem for fifteen years, is that our ability to program cells is still really bad.

Kelly splits the work into two stacks. The first is compile-and-debug: DNA is physical, so you have to do phosphoramidite chemistry, build the strand, put it in a cell, grow the cell, and test it. That is a hardware problem dressed as software, and the way you get better is to do more of it, faster, for less money. The second is design: choose a sequence that is more likely to do what you want before you ever wet a plate. For fifteen years Ginkgo worked both sides. In the AI era he sees opportunity on both, and he has shifted hard toward one of them.

The design side is now a crowded software market. Chai Bio trains protein models and recently signed a large deal with Eli Lilly. Arc Institute just published Evo 2, a genomic model. Kelly says there is a whole community trying to design biology with AI, and that you do not have to solve the whole cell to create value — more than half of drugs are already produced by biotechnology. Two years ago he stopped working on that problem. Ginkgo had its own approach, designing cells is tricky, and he decided to spend the company's resources on making it cheaper and faster to try things in the lab, including with models in the loop.

Is anyone else on the testing side with a Silicon Valley attitude? There are new firms, including one called Madra that uses robotic arms, and there is the incumbent life-science tools industry. Kelly's complaint is that the tools vendors are not trying to change the fundamentals. They sell the next instrument to people who still work the way they always have. What he wants is closer to Intel's jump to planar semiconductor manufacturing: a different platform, not a nicer pipette. He went after both halves early, "took my licks," and concluded that testing is an engineering problem you can render to dust, while design is a science problem — great if you hit it, much more unpredictable if you do not.

The recent OpenAI collaboration is the exhibit. Kelly's question was whether a reasoning model could do experimental science if you gave it a robotic lab — not a simulation, an actual lab. Most of what the United States spends on science is experimental work, he notes, because if you want to learn something new you have to go poke the world. Math and chess can live on a computer. Cell-free protein synthesis cannot. The protocol is simple to describe and expensive to run: pop a cell open, put the guts in a test tube, add a DNA sequence, and the remaining machinery still makes the protein. Kelly calls it the world's smallest 3D printer.

A paper from Stanford's Mike Jewett lab had just set the benchmark for how cheaply anyone had done cell-free synthesis. Ginkgo gave the model a budget of 100 plates of 384 wells per round, about 30,000 experiments, then handed back the data and asked for the next design. After four rounds it beat the state of the art. After six, it beat that record by 40%. Kelly thinks it is the most interesting result yet of a model doing experimental work, "by a lot." The model did not simulate biology and did not need a private theory of the cell. It did the loop a scientist does: idea, experiment, data, interpretation, poke again.

That loop, he says, is just logic, closer to programming than to a world model. "What really let it break through wasn't that it was so smart, it was that it could run experiments." The open question was only whether it could design those experiments the way a scientist would. His answer is yes — "hell yeah, it could." Huang and Grady can already imagine a future in which reasoning models run the scientific method and autonomous labs run the tests. Kelly is more aggressive than the average scientist about that future. He wants a heads-up competition. The pushback he expects is about creativity. He is not claiming the model is more creative. He is claiming it can run a lab 24 hours a day.

Today three Alzheimer's labs each chase a hypothesis, collect data for a year or two, and publish a distilled paper. Nobody sees the other lab's failed wells. Kelly's alternative is a robotic facility that can run any experiment, with a hundred AI scientists on top of it, each pursuing a different hypothesis. At the end of every day they pass the protocol and the raw data to the other ninety-nine. A result that killed your hypothesis might be exactly the clue mine needed, and today that information goes in the trash. Once a week the hundred dump a notebook entry or a mini-paper that humans can read, redirect, or kill. Information sharing and broad context, he says, are an unfair advantage even before anyone claims the models are brilliant.

The second advantage is money. Academic "indirect rates" are, in his telling, the cost of manual laboratories: buildings that sit there seven days a week and get used five, plus three copies of the same underused instruments behind three locked doors. Biopharma spends $60 to $80 billion a year on research; NIH spends about $40 billion; less than 5% of that is reagents. The rest is people, regulatory overhead, and space. Efficient science would budget a program on chemicals and plastic plateware — usage-based pricing — because experimental work burns atoms. On a robotic lab, he wants 90% of the cost on reagents, which is a 10x increase in data per dollar.

Put those two together and the models do not even have to be smarter. "They can even be dumber than the scientists, I think they win. I really think they win." That is why he wants NIH to rethink what it funds, and why every biopharma head of R&D should care. He has been adjacent to tech CEOs since YC, and his line is blunt: previous revolutions — the internet, social media — were meaningless to biotechnology. Nice for email. "Just some back office IT crap." This one changes the fundamentals of how science is done. Grady cites a Sequoia partner's split between revolutions in communication, which distribute information, and revolutions in computation, which process it. Science lives on the processing side, which is why this wave is different.

Kelly's MIT Ph.D. was five years of moving liquids around a bench by hand. Silicon Valley's instinct is "just automate it, bro." The reason it has been hard is the automation term of art: high-mix, low-volume work. His transportation map puts automation on one axis and flexibility on the other. A subway is low mix and high automation. A car is high mix and low automation. Waymo sits in the corner that used to be empty, which is why we stopped saying automation and started saying autonomous. Since the Industrial Revolution, he argues, we automated everything that was low mix, from the loom on, and then hit a wall. AI pushes past it. Labs already have the subway version — Quest Diagnostics, pharma high-throughput screens — but 99% of lab work, like 99% of miles, is still the car.

The stack is not Waymo's stack. A car's hard problem is that the physical world changes every mile. A lab is the same room every day. The hard problem is the variability of what the scientist asks for. One mountain is integrating a long tail of a thousand third-party benchtop instruments, whose software Kelly says you cannot believe, into one controllable system. The other is liquid handling: a pipette is the world's fanciest straw, and when the liquid is viscous a human thumb and eye quietly adjust pressure. Nail those two, he says, and you are done. It is tractable. It is a lot of work. Ginkgo's unfair advantage is that it still employs scientists on research partnerships, so they can dogfood the robots the way Google engineers once sat next to a Waymo wheel in Palo Alto, ready to grab it.

The surprise third mountain is that scientists do not code. The industry's answer has been visual languages like LabVIEW, which Kelly calls complete trash; even flowcharts, they hate. A few weeks before the taping, someone sent a sealed plate to a pipetting robot, the tips stuck in the film, and Slack filled with "de-seal the plate, dum-dum." That is a humiliating UI for people who are experts at the bench. The new rule is that protocols go in as writing and Claude Code or Codex writes the robot code; if the model ships a sealed plate, they update the skills file so it never happens again. Boston now has about 50 robots heading to 100, and 50 scientists submitting jobs into one setup that Kelly says exists nowhere else on the planet. Remaining bugs — scheduling with tight timing, making equipment reliable when it runs all day — are engineering.

Humanoids in the lab? No. Samples should ride a maglev track and be delivered with micron specificity to a six-axis arm, not walk bipedally between benches. Biology is a microscopic discipline; nobody expects humanoids to etch chips at TSMC. Counterintuitively, autonomous labs get smaller, not larger. A Merck campus can hide a million square feet of underused benches because labs have to follow scientists, including after every acquisition. If experiments are ordered on a computer into a centralized facility, duplication disappears, utilization goes from under 20% to around 70%, and boxes pack tight around a track. Ginkgo's unit is a RAC, a reconfigurable automation cart: benchtop instrument, arm, and maglev segment, Lego-blocked together. They sell the box plus a software-and-service subscription, want to sell automation-friendly reagents as usage pricing, and also run the Boston lab as a cloud. The Department of Energy just bought 97 of them for the Genesis mission; Energy Secretary Wright and Kelly ribbon-cut the first 18.

Genesis came out of the White House Office of Science and Technology Policy and is run by the Department of Energy, the home of project-scale American science from the Manhattan Project through the Human Genome Project. The target is to double the pace of science in a few years, first by feeding existing national-lab data into models and then by using autonomous labs to generate new data in the direction the models point. Kelly chaired a National Security Commission on Emerging Biotechnology for two years — a sibling of the AI commission Eric Schmidt chaired seven years earlier — and the competitive picture is no longer the post-Soviet default in which America automatically led because nobody else could spend. China now publishes more papers. In his own market, startups discover drugs and sell them to Merck or Pfizer for $2 billion to $10 billion depending on the trial stage. Three years ago, less than 5% of those assets came from China. Last quarter it was more than 40%.

The explanation is not a mystery. Chinese scientists are as numerous and as smart, they are paid less, and science is driven by hands in the lab. More experimentalists per dollar means more research. DARPA exists because Sputnik was a technological surprise; Kelly thinks biotech is standing at a similar gate. He also wants industrial R&D to wake back up. OpenAI began as a pie-in-the-sky, Bell Labs-ish research project and is now worth on the order of half a trillion dollars, which he reads as a reminder that fundamental research inside industry can still be valuable — something the chemical industry, for one, has mostly forgotten. Grady's two-phase map is human-level AI that does today's work faster, then pockets of superintelligence, one category at a time. Sequoia recently backed the Google AlphaChip team, whose systems already design chips that beat human designers. Kelly's secret is that accelerating the experiment can look a lot like superintelligence, because much of the physical world is not simulatable.

Native ATCG models such as Arc's Evo, trained on a trillion bases of DNA, will be tools a reasoning model can call — along with AlphaFold — while the reasoning model itself does the job of the experimental scientist. "A thousand experimental scientists in a box" is, in his view, already true; he does not need a miracle. What he does need is to break a grim trend: it has gotten more expensive to develop drugs, year over year, for 25 years. Baumol's cost disease is the diagnosis. Scientists and rent get pricier; new tools make people slightly more productive; the majority of the cost is still manual work, and that does not get cheaper. China's Phase 1 trials can take six months against two and a half years in the United States; Australia is faster too, and Kelly expects FDA to match Australia. Even then, a drug is not a phone app. The application mix is itself embarrassing: 85% of biotechnology is therapeutics, 10% is agriculture, 5% is industrial enzymes such as the ones in cold-water laundry detergent. That is the entire app ecosystem for programmable matter. The computer industry would look absurd if its only working application were drug discovery.

The near-term application Kelly is actually excited about is not another oncology asset. GLP-1 drugs are worth close to a trillion dollars, and he takes them; he calls the category the best thing since the iPhone because you get your willpower back and stop spending the day thinking about food. The reason they are worth that much is that they are not really treating a disease. Biotech today is the disease industry, and most of a life is not spent being sick. People want to weigh fifteen pounds less, sleep better, have more muscle, feel better, add two years or five. He asks Sequoia what a product that adds five years of healthy life is worth — "$50 trillion? Like it's infinity." The catch is that FDA's pathway is built around treating disease, which is why he likes the part of the MAHA conversation that is about being healthy when you are not sick.

Ginkgo's first angel out of YC was Bryan Johnson, back when he was, in Kelly's phrase, "pudgy VC Bryan," not a longevity icon. Johnson's real contribution, Kelly thinks, is that he normalized continuous monitoring: two million dollars a year of diagnostics, tests every week, the most measured person in the room. Oura rings are fine — Kelly has worn one for ten years — and they are also, for this purpose, pathetic. Heart rate is not the story. Aging is molecular. The product he wants is a weekly blood draw and a longitudinal molecular readout so you can try interventions and see what they do inside you, not a once-a-year Quest panel of ten vials at Christmas. Function and similar services are early, and the experience is still miserable, which he reads as a market that has barely started.

Ginkgo just launched a cloud-lab service with experiments as cheap as $39: you do not get a sample back, you get the data. Science has a reputation as a precious genius activity. Kelly says it is formalized curiosity, a process for answering questions that everyone already has. What actually gate-keeps science is the lab, almost legally. If the cost of an experiment falls, and a model helps ordinary people design one, he thinks millions might want to be scientists. Tell someone in the 1960s, in the IBM mainframe era, that children would program computers, and they would have called you insane. Drop the cost of the physical loop, and original scientific questions stop belonging only to people with a keycard. Getting AI-for-science working is the prerequisite. After that, he thinks the market on the other side is much bigger than drugs.