DNA Is Code, but You Compile It with a Pipette

Sarah Guo and Elad Gil introduce Jason Kelly, co-founder and CEO of Ginkgo Bioworks, with a question the whole tech industry has been repeating: if DNA and proteins are code, what does synthetic biology actually mean, and what does Ginkgo do with that idea? Kelly grants the hook. Inside cells, A's, T's, C's, and G's sit on a tape in a way that is surprisingly analogous to zeros and ones in memory. Then the analogy dies. The code is a polymer, a physical thing. It makes proteins, little pieces of nanotechnology that bump into one another inside a bag that is not physically isolated like a semiconductor chip and was not designed by humans. Tech people show up for the metaphor and then say, what the hell.

He tells the origin story through Tom Knight, a co-founder who joined the MIT faculty in 1972, designed a refrigerator-sized Lisp machine, and appears in Steven Levy's *Hackers*. In the mid-1990s Knight decided DNA was code, opened a wet lab in the MIT computer science building, and started growing bacteria, freaking everyone out. He hung a flag for computer scientists: come try it. What they got was a pipette. Compile this physical code by doing molecular cloning at a bench. Day one, a result. Day two, the same steps, a different result. Tom, they said, I did the exact same thing. "You're never gonna know," Knight told them. Computer science is pure logic in clocks we designed; a bug can be run to ground. Biology is sometimes that, and sometimes a part we do not understand. Tough luck. You have to stomach it.

Kelly's favorite rhyme is neural nets. How does the network work? Nobody knows. Analyzing them, he says, is going to look like systems biology: reverse-engineering a thing you did not design. That will be many computer scientists' first taste of feeling like a biological engineer. Why bother with models you cannot debug, or cells that will not do what you want? Because they are powerful. The substrate is incredible. It self-replicates and self-assembles; we have nothing else like it in the physical world. Guo adds that once systems write their own code, neural nets will enter a non-designed evolutionary world, messy, redundant, and optimized for utility rather than form. Some minds want clocks. Some like the magic. Biology's unreadability is part of the charm.

The piece of computer science Ginkgo tried to import on day one, fifteen years ago, is abstraction. In Knight's era you could not be a computer scientist without being an electrical engineer. Then came assembly, operating systems, programming languages, graphical tools; an eight-year-old now programs an iPad by drawing boxes. Electrical engineering and computer science split into two long roads. Ginkgo forced that split immediately. Lab infrastructure is a foundry, a word stolen from semiconductor fabs. Its job is to automate and scale the work so that DNA programmers—typically biology PhDs—order experiments from a system instead of doing them by hand.

Culturally it is brutal. A scientist does not want someone else to run their experiments; that is part of being a good scientist. The first infrastructure also sucks. The payoff, once the foundry finds scale economics, is data: lots of genetic designs actually tested. That corpus is what you want for AI, and it is what you want anyway, because making a cell do a specified job means trying a lot of designs. Kelly calls driving that scale one of the biggest activities of Ginkgo's last decade.

Customers do not sit on the foundry themselves. Recent names include Merck and Biogen on the pharma side and Bayer, Syngenta, and Corteva among the largest agriculture companies, plus startups. The interface is a spec: here is what we want a cell to do, here is a development timeline. Ginkgo is, in Kelly's phrase, a custom software shop. It builds the program, licenses it to the customer to turn into a product, and takes development payments plus a royalty. Customer scientists still talk to Ginkgo scientists rather than hitting the infrastructure directly. That is mostly a technical limit. He would like every company's scientists on the system. It is too early.

A typical job is protein engineering. A bacterium has a three-million-letter genome. A customer wants a lot of a protein—those nanotech pieces bumping around in the bag of you—and they want it more active. Cold-water laundry detergent works without hot water because of enzymes; the spec is make a ton of catalyst and make the catalyst better. Ginkgo starts from a host strain already good at pumping out protein, a physical software library left over from an earlier project. Design tools propose a thousand DNA variants; the lab returns performance; the tools update; you iterate. Neural nets have been in that loop for a long time. The new bet, announced with Google a couple of weeks before the taping, is a foundation model that is not specific to catalysis, a model that "speaks protein" the way GPT-4 speaks English. AlphaFold was a first generation. They want bigger and better.

Fine-tuning on English, Kelly says, is a high bar. A legal model trained on LexisNexis still has to compete with a lawyer at Ropes & Gray: fifteen years of human apprenticeship, contracts designed for human brains, written in a language that co-evolved with those brains. You are asking a computer brain to play on our turf. For a while the applications replace the intern, not the partner. Biology also runs on sequential letters that feel like language. It is not our language. We did not invent it, do not understand it, cannot read or write it. Those same computer brains, he thinks, will beat us much faster here. English-side AI looks like Clayton Christensen low-end disruption. Biology looks like splitting the atom.

Gil asks why, if people have talked about AI drugs forever, we have not seen an AI-discovered one. Kelly's retort is to ask, three years earlier, why we had not seen good NLP: you needed the breakthrough. The bottleneck in biology is training data. Everyone trains on the same public sets. Ginkgo's edge is a 300,000-square-foot robotic lab built over a decade, generating private data in the course of customer projects. Nothing now stops you from standing up a giant neural net; the tech industry will commoditize that infrastructure. You might still not have enough data. Recursion is another company trying at scale. Gil notes that a new drug costs about $1.5 billion, mostly in trials, so little of the money sits in the molecule, whereas agriculture and industrial catalysts spend more on the molecule itself. Kelly's attitude is AWS-like: cell engineering for medicine or detergent, rock and roll. The wilder fact for tech people is the market structure. There are almost no platform services. Merck, Pfizer, Bayer, Syngenta—each is vertical from top to bottom. The closest thing to sharing is buying equipment from the same vendors. The industry's rationalization is that the work is too dissimilar across products. Kelly's whole business model is a bet that this is false for organism engineering, at least in degree. Plenty of people think he is wrong.

Covid's lesson, Kelly says, is not subtle. Modern health-care systems do not make developed countries immune to pandemic-scale infectious disease. This is not a fanciful threat; it happened two years ago. You do not get to leave society-killing, known risks undefended. One build is rapid vaccine response: Operation Warp Speed and mRNA showed that once you have a target you can put shots in a country in three months, though not every pathogen is vaccinable. The other build, the one Ginkgo has pushed, is monitoring that feels like radar.

Kelly grew up in Florida with hurricane warnings. Tom Knight, a generation older, remembers three hours of notice. We found out about "hurricane Covid" after it had already landed in New York a week earlier. Unacceptable. With the CDC, Ginkgo collects wastewater from inbound flights, sequences the DNA, and looks for pathogens and variants of flu and Covid, with room to add others. Similar programs run at Doha airport in Qatar and in Ukraine. Think of them as bio-radar stations: establish a baseline, then hunt anomalies. The mental model for what comes next is cybersecurity—persistent monitoring, rapid response, kill it, all in the background. Pathogens replicate, which is scary and beautiful: if you snuff the thing at the beginning, you win. Speed is everything. Kelly chairs a national-security commission on emerging biotech in Washington; the Defense Department had just put out a biodefense posture review. The DOD keeps millisecond preparedness. It sounds crazy. That is roughly how he thinks you ought to treat this.

Gil recalls that original SARS was snuffed, then leaked four times in the first two years after it was cultured, and that rapid shutdowns kept those outbreaks from becoming pandemics. Kelly agrees: if you get it early, you win, though Covid was harder than MERS and the world had nothing in place. On the AI-safety story of a lone actor designing a killer virus, he is unimpressed today. Exact design knowledge is low; you can try something; you cannot follow a recipe. Catalysis data is easy to accumulate. Case-fatality data is not. Gil needles the safety argument's odd needle: the lone actor is supposed to be someone not well versed in biology, because the people who are would not do it. Kelly's turn: whether or not lone actors are real, we are unacceptably exposed. We would never let our computers live as naked to viruses as we let our bodies. Computers have an edifice of detection and global notification running in the background. Human defense should feel like that. Nature will toss another pathogen at us regardless. Gil takes comfort from sloppy lab behavior that still rarely produces leaks; jumping to humans is hard. Kelly's reply is that we are the evolutionary product of that arms race. If it were easy, we would not have made it.

Ginkgo is a weird bug in Kelly's telling. The founders left graduate school in 2008, common for software and almost unheard of in biotech. They were unfundable: not a drug, so biotech investors passed; a wet lab, so tech investors laughed. Five years of DARPA, ARPA-E, and NSF SBIR grants kept them alive. They became the first biotech in Y Combinator after Sam Altman took over from Paul Graham and wanted nuclear and biotech in the batch. Kelly thinks entrepreneurial energy is common; hard-tech companies do not need a different training than the one YC perfected.

Platform power is the live question. Social networks got a pass because bits did not feel dangerous; a medicine goes into a child's body. Pharma and tech grate on each other culturally because tech sees therapeutics as slow, while those people run trials where patients die. If Ginkgo is right, a cell-programming platform will read on lives in a first-order way. Who should control it? Founders, the Facebook model, with Mark's super-voting shares passing to his kids. Or capital markets: BlackRock, Fidelity, an arm's-length majority that replaces the board and the CEO when returns disappoint. Kelly likes neither. Founder control hands a platform to one person. Capital's job, at the end of the day, is a high return.

Ginkgo took Silicon Valley founder super-voting and extended it to the whole employee base. Anyone at the company gets 10x votes on B shares versus 1x on A shares; the privilege disappears if you leave. The math: if employees collectively own more than 9.1 percent, times ten, they outvote the rest. Control, in the theory, sits with humans who go home at Thanksgiving and have to explain why they are proud to work there. Gil notes that the *New York Times* is still the *Times* because a family kept super-voting for a hundred years. Guo's objection is the hard one: a CEO sometimes has to run an unpopular RIF. Without founder authority, answering to employee-voters, do you slide into a popularity contest? Kelly does not dispute the risk. The answer is still share voting, not one person one vote. You accumulate shares by staying longer and by joining when the stock was cheaper and growing the value. The imagined scene is senior holders saying this is hard and right for the organization. He does not know. They are trying it. Worker ownership appeals to him. The real point, he says, is that someone has to have platform governance.

Guo closes on energy. The largest labs talk about a trillion dollars of compute over the next decade; if half is energy, you are in trillions of kilowatt-hours. A human might spend 5,000 kilowatt-hours before learning to read. Do we get more biological inspiration in AI architecture? Kelly does not have a sharp intuition, then offers a wilder option: stop limiting ourselves to a brain that fits in a skull. Grow a room-sized biological brain and throw it at GPT-4. Gil likes neural nets because they were computer scientists' excuse to build software they would not understand, an escape from logic. Open your minds, he says; the net is probably not the best architecture. Go in different directions. Do something crazier.

The moment code writes its own code, Gil argues, you leave hand-picked design for a Cambrian explosion of messy, duplicated, hyper-optimized systems, because energetics will sit in the utility function whether you put it there or not. Kelly recommends *The Plausibility of Life*. Learn from evolution, not only from brains, because evolution itself evolved things to become more evolvable. The book's example: a person born with a sixth finger does not grow a bare bone. Skin and nerves wrap it, because skin and nerves are adaptive to bone. That layering makes exploring body plans far more efficient. Four billion years is a long time and a lot of incident solar radiation. It was a random walk, and it figured out cool stuff we have barely tapped. If you are going to do a crazy mixture-of-experts routing scheme, Kelly's advice is simple. Wrap the skin around the bones.