All of the previous revolutions in tech, internet, right, like social media, like whatever, have been totally meaningless to biotechnology and biopharma. Like, yeah, it's nice. We communicate slightly better or whatever. It's just some like back-office IT crap, right? Like, not this. Like, this, this is actually going to change the fundamentals of how we do science, and our big science industries like biopharma are going to get disrupted. I really believe that, and that's not been true for the last 30 years of tech.
We are thrilled to have Jason Kelly, founder and CEO of Ginkgo Bioworks, with us today. Thank you for joining us.
So you started Ginkgo Bioworks in 2008 with the goal of making biology programmable, and 'programmable' has taken on a completely different meaning in the era of AI. So I'm very excited for the conversation today. Maybe tell us about the journey so far.
I mean, I'll do the Ginkgo journey in short, right? So yeah, we started in 2008, but we didn't actually raise any capital until 2014. So we bootstrapped for four or five years, which, like, if you're not a bio person, this doesn't make sense. But in biotech VC, they really don't like young people, for example. So we had started the company like straight out of grad school. It was 2008. We weren't trying to make a drug. So we were like totally uninvestable.
Were you full-time focused on the company for those first six years?
Oh yeah. Okay. Oh yeah. Yeah. We were basically like applying. We did government grants and service business. It was like pretty brutal start. And then summer '14, Sam Altman, now Mr. Famous, writes this blog post because he just took over YC and he's like, 'Hey, I think the Silicon Valley model can work for like deep tech, you know, nuclear fission, biotech, material science.' And so I wrote him an email. I was like, 'Oh man, like, thank you for... I mean, we're like five years old. I got 15 people in a lab in Boston. We don't make any sense for YC, but this is like an oasis in the desert, you know, like nobody will invest in weird companies like this.' And he's like, 'No, you got to meet me.' So I flew out to San Francisco, met him, and he's like, 'You should do YC.' I was like, 'I should do YC.' Yeah. So then we did YC. So we kind of, that was sort of when, you know, if you really want to mark Ginkgo from like having capital, it was sort of in 2014.
And how has the product changed since 2014?
Oh well, so the mission hasn't changed, but the product has gone all over the place, different roads. So we've always wanted to make biology easier to engineer. That was the idea. And so if you're, you know, this is very much...
Hang on, I remember 'make biology programmable.' Have the words always been 'make it easier to engineer'? Because I feel like that's a slight...
I was always, when I was talking to Sequoia, it was always like this, the computer science wrapper on 'make biology easier to engineer.' But yeah, that was always our mission.
But the analogy is solid, right? So you know, DNA is code, right? It's A's, T's, C's and G's, not zeros and ones. It's really our only other like coded product other than computers is really biotechnology. And so the core idea of Ginkgo was, well, if you could design DNA code, you could program cells to do things. And cells are, you know, they're programmable like computers, right? But unlike computers which just move information around, cells move atoms around. So if you can build whatever you want, that's we think ultimately going to be a huge market, a huge opportunity. But the challenge is our ability to program cells today is really bad. And so how could you fix that? That was the core idea behind Ginkgo.
And how's the product itself changed over time?
Yeah. So the way we went to market first was we're going to try to build what we called foundries, which was sort of a centralized laboratory that would kind of automate the lab work associated with doing biotech. And the reason, again, if you're a computer scientist, the way to think about this is if you wanted to compile and debug DNA code, that's a physical process.
Right. So you're like A, T, C, G, like we have to do phosphoramidite chemistry. You got to build the piece of DNA you want and then put it into a cell, grow the cell and test the cell, and that's your kind of compile-debug cycle. Does that make sense? And so one half of what we worked on technologically was how do we make that cheaper, right? Because if you want to get better at doing this, you got to do more of it faster and for less expense. And then the second thing we worked on was how do you get better at your programming? In other words, like that design you choose to test in the lab, how do you improve the odds that it does what you want? So sort of like get better at designing the biology and make it cheaper to try things, where like basically for the last 15 years, the twin activities. And in an era of AI, I see opportunities on both sides for that today, and that's, we've shifted a little bit over the years in terms of how we do it. But does that make sense, the two kind of halves?
And roughly speaking, the design piece sounds like it's a bit more software, the testing piece sounds like it's a bit more hardware controlled by software.
Very much. Yeah, that's exactly right. And today, the folks leading on the design side, you might see companies like Chai Bio, for example, like has these protein models, Bolt's. The folks at Arc Institute just came out today with a paper, we'll call it, about EVO 2, which is like a genomic model. There's a whole community of people now trying to solve the problem of designing biology with AI.
The big change at Ginkgo over the last two years is I've kind of stopped working on that problem. I'm like, we had our own approach to solving it. It's hard, right? Like designing cells is tricky. We're going to try to solve this half of the problem, which is how do you make it cheaper and faster to try things in the lab, and how can you... We can talk a little bit more. We just did a project with OpenAI. How could you have AI models help you do that?
Is anybody else focused on the backend, so to speak? I kind of think about design as the front end of the process and testing as the back end. And when I say anybody else, obviously people do this. Is anybody else with a similar approach focused on that part of the market?
There's some new companies, right? So there's companies like Medra out here is one that's doing it with like robotic arms trying to accelerate it. You have like the life science tools industry, but I would say it does not have a Silicon Valley attitude about things in the sense that they're not really trying to change the fundamentals of how you do it. They're sort of like just providing the next tool to people doing it the way they've always done it. And so we've always been this kind of unique force trying to say, hey, is there a new platform, right? Is there something like the jump to planar semiconductor manufacturing in electronics at the beginning of Intel? You know, like is there some way we should just do all this stuff differently that could make it way better in the future? And that's always been the Ginkgo, like what I think is unique about what we've been doing the last 10 years.
What catalyzed the focus on this part of the business, on this half of the house?
I think this is an engineering problem and I think this is a science problem. So, and I went after both initially and I kind of took my licks for that. And I think the good thing about an engineering problem is you can ultimately like render it to dust, right? A science problem I think is great if you hit it, but it's much more unpredictable. And so in this era of Ginkgo, I've got like the resources marshaled at this point to go after this and kind of see it through. And so that's why you see me pointing in that direction.
And with the efforts that we see with the Arc Institute and others of that ilk, what inning are we in, so to speak? Like, is the ecosystem around the science problem going to start producing meaningful results soon other than papers?
It's a good question. I think the hard part about designing bio... So biology is amazing, by the way, right? Like just as a substrate again, right? Like if you think about what's happening inside of a cell, it is producing, you know, Intel or now Nvidia, TSMC-level caliber atomic placement basically for free. Right. So it's able to do molecular assembly. It self-repairs. It self-replicates. Like as a physical substrate, it's insane. It is the product of four billion years of evolution.
So the complexity embedded in a cell is actually a lot bigger, I think, than people give it credit for. And so there's a march there. Now that said, you know, more than half of your drugs today are produced by biotechnology. We cure cancer, we do this. We have huge value coming out of even with the limited tools we have today. So you don't have to solve the whole problem over here to create a lot of value. You just have to be better than how we do it today. Does that make sense? And so I think they have really good opportunities already in the near term with all the protein models. You're seeing that, right? Like Chai just did a big deal with Lilly. Like there's real opportunities there, I think, right in the near term. Yeah. Does that make sense?
So speaking of OpenAI and Mr. Altman, you recently announced a partnership research result with them. Can you say more about that?
Yeah. Okay. So this is pretty exciting, I think, for the folks that are following AI, it's pretty neat. So basically what we did was we took our, we call it autonomous lab, right? And so I can talk more about this, but the short answer is if you really want to drive efficiency on the lab side, you need to get the human beings off of the lab bench, right? So the way we do, and this is true in biotechnology, it's kind of true for science broadly. The way we do science today, 95% of science, the stuff that's not theoretical. So not, you know, everybody's working on math, like let's work on Terence Tao, let's get a Terence Tao in a box or whatever, right? Like, you know, like the reason is you can just simulate all that stuff on a computer. Let's play chess, you know, right? Like no kidding, right? But if you look at the majority of like what we spend money on in the United States and just generally across the world in science, it's largely on experimental work.
And the reason is if you want to learn something new about the world, which is what science is fundamentally, you have to go out usually and like poke it. You have an opinion, a hypothesis, but you got to go test it to actually figure it out. So it's experimental science that moves the needle in my view. And so the question was, could a reasoning model do the work of experimental science if you gave it a robotic lab? That was the question. And the answer was, yeah, it's actually pretty damn good. So we did, basically the way the project worked was we had, there's a biochemistry problem called cell-free synthesis.
So you take a piece of DNA, A, T, C, G, right? If you were to put it in your cell right now, remember like central dogma in high school, right? Like it's like DNA makes RNA makes a protein, right? And so you put that DNA into a cell and it'll make a protein. Well, you can do a thing called cell-free where you pop a cell open, take the guts, put it in a test tube, and then add the DNA to that. And because the guts are still there, it makes the protein.
So this is kind of like, it's like the world's smallest 3D printer or something, right? Okay. And so scientists use this. They try to optimize. It's very expensive usually. And so there was a paper that came out of Stanford from Mike Jewett's lab in August that set the benchmark for like how cheap people had been able to do cell-free protein synthesis. And so we said, all right, let's try to optimize that with the model. And so we gave the model, we did each round, we would do 384-well plates. Okay. So each well in a plate is like a little kind of cup of liquid and you can do an experiment in there. And so we gave it, you know, 30,000 experiments to run. And after it would run those experiments, gets the data back and designs another set. So after four rounds of that, we beat state-of-the-art. And after six rounds, we beat it by 40%.
And so that was a, I think it's the most interesting sort of model doing experimental work result that's been shown to date by a lot.
And the 40% was a function of what? Just faster cycle time or more intelligent experiment design? How did it beat the state-of-the-art?
So this is my point. This is, I think, my larger point about science because I think we're going to do science differently in the future in my view based on what I'm starting to see here. So what does a scientist do when they're doing experimental work? They're coming up with an idea and then they're trying to design an experiment to ask a question about that idea. Then they're going to run the experiment, take the data back, interpret it, and then poke again based on what they learned. And they're going to go through that process a few times to like resolve something. Oh, this is how, you know, whatever, this cancer works. This is how this piece of materials, you know, works. This is this, this is that. And so that cycling is just logic.
Right. And so it doesn't require you to model biology or simulate anything. It's not that half of the house. It just requires you to be almost like a programmer, like you need to be logical, run through a set of things, do data analysis and draw conclusions. And so that's all it has to do. Does that make sense?
And so we didn't do anything other than that. Right. What really let it break through wasn't that it was so smart. It was that it could run experiments. And the question was just, could it design them like a scientist could? And the answer was, yeah, hell yeah, it could. And so now I think that opens a real interesting question about like how we do science in the United States.
Well, and it's easy to imagine a version of the future in which the scientific method, the design and the hypothesis testing and all that is done by reasoning models of some sort and the actual testing is done by, you know, autonomous labs. What's wrong with that vision of the future? And if that is the right vision of the future, how far out is it?
I mean, I think this is how it's going to happen. I really do. Like, I mean, I'm probably substantially more aggro on this than like the average scientist today. But like, it's so... I'll just explain like why, where I think if you like had a head-to-head competition, which I want to do this, right? So...
And then tell me how the average scientist would push back and say like, no, no, no, it's not going to happen that way because...
Right, so I think what the scientists will push back on is like, this thing can't just be as creative or as smart as me or something, right? Which I actually, I'm sympathetic to that.
I'm not saying it's going to be more creative. I'm saying...
It's going to be way more creative.
Yeah. No, not that. Silicon Valley. I live in Boston. I mean, I'm saying it can run a lab 24 hours a day.
I'll give you another example. Like the way science works today is you would have a lab, you'd have a lab, I'd have a lab. We're all working on the same area. Let's say we're working on like Alzheimer's. You have a hypothesis, you have a hypothesis, I have a hypothesis. We each kind of pursue it. We're collecting data over the course of a year or two and then based on what I see at the end, I write a paper. And when it comes out in the published literature, you get to read it and you get to read it and you're all doing the same thing. So we're kind of like exchanging information every year or two and I'm not getting to see every experiment you did, by the way. I'm getting like the distilled output of what you think you saw over two years. Does that make sense?
All right. So let's contrast that to like what I think should start to happen now based on what I saw with this OpenAI project. What I think should happen now is you should have a robotic lab that has every piece of equipment that we all have in our labs. So it can run any experiment you want. We can talk in a minute. That's actually a pretty technically difficult problem, but let's just wave that away. Okay, solved. All right, great. So then I'm going to put a hundred AI scientists on top of this thing. Each one is going to pursue a different hypothesis for Alzheimer's. All right, great. And they're going to run their experiments just like you would in your lab that day. But at the end of the day, they're going to pass the data on those experiments, like what experiment they ran and the raw data that came off it, to the other hundred AIs.
Daily. Every day. Okay. And so they're going to learn from each other. Like even though your hypothesis is different, we're working in the same area. So your failed result might, like for example, say your experiment went the wrong way from your hypothesis, that data might be relevant to my hypothesis and I would never see that normally. Does that make sense? And so that's all just chugging along and every week it dumps a lab notebook entry or like a mini paper, like a conclusion about what the hundred of them have figured out that week that we can all read and see and use, that we can direct, we can say, hey, cut that line of research or whatever. And so like that's number one, I think the information sharing and like the ability to handle like really broad context across a lot of projects for the AIs is just better than, it's just socially different even than how we do it today. Does that make sense? That's unfair advantage number one.
Okay. Unfair advantage number two. If you look at how we spend money in science, remember all this stuff like the NIH was like, 'What's up with the indirect rates at the academic universities and all this hullabaloo,' right? Well, what's an indirect rate? Well, it's basically paying for manual laboratories. That's what it pays for. Okay? You've got these labs and they're there 24/7, but they're used five days a week. Okay. They have equipment, but every lab, all three of our labs, we have same copies of the same equipment. Because we all got to do the same work. We don't share each other's. No, no, no. I have like a door in my lab. Only my lab gets to use it. Your lab uses yours. So we have all low utilization rate of our equipment. It's just how it works. Okay. Right. And so you have a very inefficient, like if you look at the spending on research, and this is true, the 60 to 80 billion a year that biopharma spends or the 40 billion that NIH spends, less than 5% is on the reagents. Everything is on overhead, is basically the overhead, the people, the regulatory, the lab space.
If we were running it efficiently, you would budget a research program at the NIH, not on indirect and heads and everything else, but just on the reagents because that's like the usage-based pricing of science. Because to actually do experimental work, I have to consume some chemicals. I have to consume a piece of plastic plateware. Like whatever the hell it is, like I'm actually doing atoms in the physical world. I got to burn some stuff up. That should be the dominant cost. It's the opposite right now. It's like less than 5%. So the other advantage those AIs will have is if they're able to run robotic labs, now they're running where 90% of the cost of a research project goes to the reagents.
Oh my God. Right. So that's like a 10x increase in the amount of data per dollar that you're getting compared to how we do it today. So I think you combine those two things, and but without the AIs even being smarter, right? They can even be dumber than the scientists. I think they win. I really think they win. And so I think we got to reevaluate like how we fund what we've done with the NIH. I think every biopharma head of R&D needs to care about this. And like, and I think there's a blind spot, by the way. Like we did YC, I know all the tech people and, you know, because I've always been adjacent to this stuff, right? All of the previous revolutions in tech, internet, right, like social media, like whatever, have been totally meaningless to biotechnology and biopharma. Like, yeah, it's nice, we communicate slightly better, whatever. It's just some like back-office IT crap, right? Like, not this. Like, this, this is actually going to change the fundamentals of how we do science, and our big science industries like biopharma are going to get disrupted. I really believe that, and that's not been true for the last 30 years of tech.
Our partner Constantine has a good framework for that. He talks about how there are revolutions in computation and revolutions in communication. Communication is about the distribution of information. Computation is about the processing of information. And what you're talking about here is just a different way to process the information. Like, and so the last several revolutions have been about the distribution side of the equation, which doesn't get to the core of what it is you're doing.
Completely agree. And that's, I think, fundamentally true. And so again, I think the leaders of biopharma companies and also the leaders of research universities and these people that are in the business of doing science to produce either products or for the government cannot ignore this. They cannot ignore AI. It is just different. And I'm telling you, I'm a person who has been adjacent to this crap for 15, 20 years now. And this is the first time I've, like, the tech guys finally did something cool.
Yeah.
And just so you understand, like, so you think AI is the catalyzing force behind, you know, cloud labs should be a thing. Nobody really ever moved to them, but like AI will be the reason they move.
Yeah. Well, we can talk about cloud labs. Yeah. So let's talk about autonomous labs and then I'll explain the cloud. All right. So why has it been hard? Right. Like the average tech person's look at how science is done where you have literally PhD-trained people. These are brilliant people. Paid a decent amount of money. Standing... I did a PhD at MIT in bioengineering. It's five years of moving liquids around a lab bench by hand. I swear to God, like, it is like, my friends from undergrad, just like, they would never, right? Like, you know, right? Like, it's ridiculous. You would do like manual labor, right? Like absurdity, right, in our crew. But like, that's what you have to, you have to do that if you want to play at the edge of science, you got to do physical work. And so that's what you learn to do. Okay. And so it's like, well, everyone in Silicon Valley is like, well, just automate it, bro. Like, you know, like, why don't we just do that? Okay. So why is it, why is it hard? All right. And the reason it's hard is it's like the technical, like automation-y term is like high-mix, low-volume work.
Okay. And this is true at like places like Hadrian today, for example, that are working on this on the manufacturing side, industrial high-mix, low-volume.
Historically. All right. And my like transportation analogy that I've been giving to people in bio is like, okay, so imagine on the y-axis you have like level of automation.
And then on the x-axis, okay, you have like flexibility, like that variability, the mix in what you're asking it to do. So in transportation, low-mix, high automation, that's like a subway. Sit down, takes you away, right? Like, you know, maybe you're like, I don't have to do anything, but you got to want to go to one of the stops on that subway line.
Low automation, high variability. That's a car.
Right. Hands on the wheel, foot on the pedal, take it right to your house or the grocery store. And that's what the transportation system looked like for the last hundred years until, thank you very much, you know, Google, we got Waymo up here in the corner where you get the flexibility, you know, the automation of a subway, but the flexibility of a car. Okay? And it's so surprising that we don't even call it automation anymore. We make up a new word. We call it autonomous. It's an autonomous car. Because the way I look at it is since the Industrial Revolution, we've basically been automating everything that is low-mix. That's like low-variable from the loom on. Okay. Right. And we just hit a wall, and AI, we hit a wall on flexibility of what you can do, and AI pushes us past it.
