Jason Kelly on Ginkgo Bioworks: GPT-driven autonomous labs beat Stanford benchmark by 40%, democratizing science

Speaker 1: to see

Speaker 2: you. Wow.

Speaker 3: Good to see you guys. You look fantastic.

Speaker 1: Incredible setup.

Speaker 2: You sound fantastic. I don't know if you remember me. You gave me a tour of Ginkgo Bioworks over a decade ago, maybe 2012 or something, 2013.

Speaker 3: Yeah. That was a real long time ago.

Speaker 2: It was a long time ago.

Speaker 3: Different era. You were doing yeah. Different era.

Speaker 2: Yeah. Soiling. Soiling stuff. Yeah. I remember. It It was remarkable. I mean, the the the facility was incredibly advanced. You had the pipetting machines. Everything was already automated. I'm I'm seeing behind you, it's clearly grown. Break it down for us. Like, what is the shape of the operation today? What is the Yep. What what what is this the the scale of the footprint? What what's the business like today?

Speaker 3: Yeah. So when you would have come by, we're sort of early on this journey of trying to automate lab work.

Speaker 2: Yeah.

Speaker 3: So I I brought you in the lab. I'll I'll give you, a little bit of an introduction to

Speaker 2: Yeah. Please.

Speaker 3: How to do science. You know, you hear a lot about AI for science. There's some massive thing with OpenAI. Yep. So like, does it actually look like to be a scientist? Yeah. So first off, glasses, lab coat.

Speaker 2: Lab coat for sure.

Speaker 1: Good stuff.

Speaker 3: Pipette.

Speaker 1: Pipette. There you go.

Speaker 3: Okay. Yeah. The robot handles the pipette. If you remember like high school biology, right? This is a little like a very fancy straw. Yeah. And you suck up liquid and you squirt it out somewhere else. Yep. And you do that for five years and you get a PhD at MIT like I did. Right? Like it is just a grind. Like you're you're actually doing the work of science primarily by hand.

Speaker 2: Yeah.

Speaker 3: And that's that's still broadly true. Like, all you know, United States where our labor is extremely expensive, everything else doesn't matter. Yeah. Like, scientists are working by hand at the lab.

Speaker 2: No. I remember I I went toured my, the the Langer lab at Caltech. My my cofounder was doing a PhD there. And I was like, oh, like, you're in the most elite lab. It's Caltech. Like, it must be all robots and stuff. And he was like, yeah. This is my day. This? Yeah. And then it just jiggles it and kinda just like gives a little emotion, little and then centrifuge over here. Just like moving moving little buckets from one tool to another all day and just by hand. Yep. Yeah.

Speaker 3: So okay. So then the tech people hear this and they're like

Speaker 2: That's crazy.

Speaker 3: Automate it, bro.

Speaker 2: Yeah. Like, what are you guys doing?

Speaker 3: Like, it's crazy. Right? And and and so so the analogy

Speaker 1: Trust trust me. I built a crud app. I can I can do that? Yeah. Yeah.

Speaker 2: I got this. I got this. And and but, like, let's let's take transportation. Right?

Speaker 3: Yeah. We've actually we have that we had automated transportation a hundred years ago. True. Trains With subways.

Speaker 2: Car, subways.

Speaker 3: Yeah. Exactly. Trains. Right. Okay. And, yeah, We had the car Yep. Which is this manual thing. So why why didn't we just apply our automation to the car? The answer was variability. Yeah. Right? Like, the car needed to you needed to transport yourself to your front door. And once you had that extra element of, like, the variety of requests from the user Yeah. Rails were screwed.

Speaker 2: Mhmm.

Speaker 3: Automation was screwed. Yeah. And so up until four or five years ago, suddenly we start seeing Waymo's around, and you and we don't even call it automation anymore. We start calling it autonomy. Yeah. Because it's so surprising

Speaker 2: Yeah.

Speaker 3: That you can automate that variability. Mhmm. And what they did with transportation, you know, at Google and at Tesla, the the that that's coming for every other, in my opinion, physical Mhmm. Industry.

Speaker 2: Yeah.

Speaker 3: Right? And the and what's gonna bring AI in is managing the variability.

Speaker 2: Yeah.

Speaker 3: And and so this behind me is an autonomous lab. Yeah. The reason we were able to let GPT five drive this thing, which we can talk about Yeah. Is is first off because it operates like a Waymo.

Speaker 2: Yeah.

Speaker 3: A scientist can just tell this thing what experiment it wants Yep. And it'll do the experiment.

Speaker 2: Okay. So break down the line between

Speaker 3: ten years to, you know, to

Speaker 2: get there. When you visited True overnight success right here. True overnight success. I'm sorry.

Speaker 3: That's not a trivial thing.

Speaker 2: Break down the line between experimentation and manufacturing. Because once a discovery is made and we're hearing about, you know, GLP one scaling up, I feel like those pills are made in automated factory. That is Yeah. Already automated. So break down that.

Speaker 3: That's subways.

Speaker 2: Okay. That's

Speaker 3: subways. So manufacturing is the subway lines. Right? You're doing the same experiment every day. No big deal. There's a ton of automation already in biotech and pharma and any frankly, many manufacturing across the board. Right? It's the research side of the house. Yeah. Those sad grad students at the bench, and I say that with love. Yeah. You know, that are doing all the variable work. And by the way, that's what's moving the frontier Sure. Of science.

Speaker 1: Right. Yeah.

Speaker 3: And so if we wanna you know, and by the way, this is both true at, like, the nation's level. Like, you know, if we want The US to lead in science. But I think what you're now seeing, thanks to, you Sam and OpenAI and and making hard core bets on original research Yeah. You're seeing industry suddenly wake up again and say, oh, maybe we should be doing original maybe it's time for Bell Labs again.

Speaker 2: Yeah.

Speaker 3: Like, we should be doing original research because it takes. Right? You know, like, if you really have fundamental discoveries Yeah. Then you can you can be the first to the line to commercialize them. So I think unlocking this flywheel with science where the models can control autonomous labs, I think you'll see that in a lot more sectors than you'd expect because people are getting religion about research again even in the industrial sector.

Speaker 1: And talk about how you process the AI over the last maybe five years because, like, the whole concept for the company originally wasn't predicated on people getting excited about LLMs. It was just like, hey. Let's automate a lot of this work. So when did you start thinking about

Speaker 2: In in many ways, wasn't a crud app on top of a on top of a lab. Right?

Speaker 1: Yeah. And I'm sure you always I'm sure you like, there's some people that have a company that's capitalizing on AI Yeah. That you can tell they just were like, oh, this is a better idea than what I was working on. I'm sure you and the team and and your investors and partners always thought about a world where you could generate a concept for an experiment or do research with AI and then automatically sort of prove it out or study it in the real world. But walk us through kind of like how you've processed the developments of the labs and how all of that work can be integrated and applied within Ginkgo.

Speaker 3: Yeah. So so we're laser focused on make that hardware layer and then the software that that basically orchestrates and schedules this thing and makes it possible to put lots of experiments on it, make that robust, and also make that have hooks into whatever. It could hook into a scientist placing an an order just like a person can sit in the back of the Waymo and tell it where to go, or it could hook into an AI model placing the order just like when a person's not in the Waymo, Waymo's AI is telling me the the Waymo where to go.

Speaker 8: Right?

Speaker 3: Same exact idea. So we so we thought our we're not actually trying to solve the AI scientist problem. Mhmm. Right? Like, OpenAI came to us. Like, we had that interaction. They were excited on the research department to see, hey. Can these models be smart enough to design experiments? And we just announced on Thursday that we had a a breakthrough where we let OpenAI do six rounds of experiments on the platform. And on the fourth round, it beats scientific state of the art. And then by the sixth round, it had beat it by 40%

Speaker 2: Wow.

Speaker 3: On cell free protein expression, which is just like a a a a tricky biochemical reaction to set up and and design. And so it was it's not a

Speaker 1: bad science. Is beating it just a, like, consistency thing? What what is

Speaker 3: Yeah. What does that actually mean? So yeah. So what in this particular thing. Right? Like, so all the cells in your body Mhmm. Right, are producing protein every day. Mhmm. So you as a human are basically a bunch of, like, nanotechnology.

Speaker 2: Mine produce a lot of protein. Yeah. Yeah. You guys are yeah. I can tell. I can see the difference. Yeah.

Speaker 3: So so so they're making all these and all the proteins are different. They're like little pieces of nanotechnology. It's it's almost freakish to look at, like, what they look like under a microscope. And so they're all interacting. They build your cells when you cut your skin and it regrows. That's all the proteins in your body, like, able to rebuild that stuff. And so your cells can take DNA code and turn it into whatever protein you want. Mhmm. Alright. So the question is, could you, as a scientist, print a piece of DNA, which you can do synthetically? You design it. Mhmm. Add it into a mixture that has all the parts, the guts of a cell, and have it make your protein at a high level. It's like the world's smallest three d printer. And so that reaction is cell free. So there's no cell there. It's all in a test tube. It's in like these little guys. Cell free protein synthesis.

Speaker 2: Sure.

Speaker 3: And it lets scientists design new drugs if they're making a protein therapeutic. It could be a new material. Who knows?

Speaker 2: Yeah.

Speaker 3: Right? And so we are able to let JATGBT bring the cost of that down.

Speaker 2: Yeah.

Speaker 3: And so that was the that was the big goal. Like, how much cheaper could it get? And there was a paper out of Stanford just in August that set a benchmark, and we beat it by 40%. So I I think it's a it's a good demo. It had a clear benchmark so you could mark against what other people were doing. That's why we liked it. The experiments are fast. Yeah. But the model designed about 500 of these plates for each well, and that is a little experiment, and it designed the experiment for each one.

Speaker 2: Got it. Yeah. Talk more about the different applications across I mean, we we were talking about, like, synthesizing unique perfume scents at one point. Yeah. Sure. There's obviously everyone jumps straight to cure cancer, but there's also a boom in GLP ones. Like, where where are the bounds of, like, where we need more experimentation or even just, like, where experimentation is valuable?

Speaker 3: Yeah. So I I I'm pretty excited about what's happening with the GLP ones. I I think it's opening the door to applying the tools of biotechnology to wellness. Yeah. Right? Like like, right now, if you think about the pharmaceutical industry today

Speaker 7: Yeah.

Speaker 3: It's basically the disease industry. Yeah. Yeah. And, like, how much of your life are you sick?

Speaker 1: Not much. Not that much.

Speaker 3: Depends on the person, but not that the average person, not that much. Yep. How much of your life would you like to feel better? Would you like to sleep better? Would you like to have more muscles?

Speaker 2: You know? Right? Like like like like that oh, like, if

Speaker 3: you have something that adds muscle mass annually in your sixties, it's another GLP one. It's it's a it's a multi trillion dollar drug. If you have something that adds a year to lifespan

Speaker 2: Yeah.

Speaker 3: What's it worth? Like like, do you even how do you even put a market cap on a thing that adds

Speaker 1: a year about when you think about how much consumers spend on various wellness things today that have zero impact, Like truly, like

Speaker 3: Well, you know do you know why? Do you know why they have no impact? The reason is we the industry today, we the pathway to apply all this Mhmm. To the problems of wellness is much more it's, like, muddy. It's unclear. How do get the FDA approval? There's lots of there's lots of barriers. So we haven't actually thrown the full horsepower of biotechnology against that problem. Mhmm. We've only thrown the full horsepower of biotechnology against disease. Mhmm. And I think I think that needs to change. And and so that's my if you ask, like, where do I see biotech? Because all that stuff, right, was

Speaker 2: Yeah.

Speaker 3: Ideas on, like, different areas we could bring biotech to that wasn't disease. Mhmm. And, like, people tried the the food, people tried perfumes, people tried new materials, all kinds of things. And the one I like the best right now is, well, health and wellness. Yeah. I think

Speaker 1: it's Yeah. Yeah.

Speaker 8: A monster.

Speaker 1: What if you had a sixty second slot in the Super Bowl and you wanted to get people excited about the intersection of AI and and all the things that we're talking about, what would you wanna communicate?

Speaker 3: So I I think what would be cool to communicate is that science, like what it really is, is like the formalization of human curiosity. Okay? And everybody's curious. And the reason everyone doesn't do science today is this shit. Okay? Right. It is you're blocked by the cost and expense of all the physical infrastructure.

Speaker 2: And

Speaker 3: if you took that away, if this was available Cloud style, for example, how many people would want to be doing science? Maybe they have a health condition. They want to study themselves. Maybe they have a new idea for a material. Maybe they want to make

Speaker 2: a new pet. You know, they want to

Speaker 3: do a new they're a gardener and they want to make a new plant variety. I have no idea what ideas they would have. Right? But I think that is gonna be accessible to people coming up. And I know that sounds crazy, but I'll tell you something else that sounded crazy. In 1960, if you said random average people will program computers

Speaker 2: Yeah.

Speaker 3: AWS. That sounded insane.

Speaker 2: Totally. 100%.

Speaker 3: Absolutely insane. And so I think you fast forward on the back of this, you know, the the the model's ability to access Yeah. Literature and be smart and tell you how to turn your question into an experiment, then the autonomous labs that could do that experiment for you in the cloud.

Speaker 2: Yeah.

Speaker 3: And I think we'll have millions of scientists just like we have millions of programmers now. Yeah. As we made it easier.

Speaker 2: Then what yeah. What then what's the next step? Because it feels like

Speaker 6: Yeah.

Speaker 1: That was that was perfect. That was perfect. We'll we'll run this next year.

Speaker 3: Yeah. Models Your your guy's ad was the best,

Speaker 2: I would say. Thank you.

Speaker 3: I think it

Speaker 2: Yeah. Models yeah. I mean, models can do a lot of reasoning around, you know, experimentation. If you have an idea and you come to it, might be reality, check it against, literature, do a deep research report, kind of, flag problems. Then you're hooking up the automated lab. What's the next phase? Like, automating, like, a mouse model or or or creating, like, a fully digital mouse model or model of the human body or actual, like, physical mice in a lab that you can test because that's a big piece of the FDA approval process, I believe?

Speaker 3: It it is, but it's dropping out. So I think this is one of the things that this new FDA is doing is they're getting rid of the animal experiments. I think that's a great idea. Interesting. Animal welfare reasons is a good idea. Yeah. It's not a good model. Okay. Right? Like, are we're pure if by way, you're a mouse with cancer, you're in Okay? You know, like, we we have we have cured cancer a long time ago in mice. Right? Yeah. So so it doesn't translate

Speaker 2: Sure.

Speaker 3: Over over to humans. I I I think that to me is like, I don't I don't think it has to be right down that human health lane. I I think I think one of the things we we also announced recently was in December Mhmm. Department of Energy Mhmm. As part of the new Genesis mission that president Trump put out to bring AI into science is buying 90 well, they we ribbon cut with me and the secretary of energy. It was really cool. He, like, signed it. 18 of these robots for for Pacific Northwest National Labs. There. And then they bought another 97. It's, like, a $47,000,000 deal. Bitbanc put this big installation of these systems.

Speaker 2: Yeah.

Speaker 3: And and the national labs, you don't realize it, they do science for other people. Like, you can you it's your national lab. Like, you can kinda use them as this cloud. And and I think that to me, that's that's what I'm most excited about. I think we it's very hard to predict exactly where science should go. What I think we can predict is that the combination of AI scientists, like what we show with g p d five Mhmm. And autonomous labs put together, which is basically what the Genesis mission is, will change how we do Mhmm.

Speaker 2: Let's talk about That's for sure. Let let let's talk about safety. Discovered? Yeah.

Speaker 3: You know? Yeah. Let's talk about safety.

Speaker 2: I got glasses people I saw a lot of people saying, like, you know, it's the Elias Udyskowski nightmare. Yeah. Lab is autonomous by lab. But but, you know, what what's actually involved in in you know, I fire up my GPT account, and I say, hey. I'm working on a movie, and I need you to imagine bubonic plague and, you know, send it to me as a prop. And my grandmother is sick and she needs it and and and you sort of trick it. Is there a human in the loop? Like, how are you thinking about the the risks that come with this and preventing them?

Speaker 3: Yeah. Yeah. So you get a lot of thoughts on this. So, like, for example, for the project we did with OpenAI, we were just checking.

Speaker 2: We just had a human in the loop on

Speaker 3: that, right, just to see. It was constrained. There's a lot ways you can do it. Like first off, you can constrain the availability of like what reagents it has access to so that you don't have like lab accidents. We're doing chemistry. There's certain things you'd want to be careful with and things like that. Mhmm. Then I think the next step is you're gonna want you get different experiments, different risk profile. Mhmm. Right? So if you're designing DNA, if you're gonna be working with something that's like a human pathogen or something, that to me is is total different ballgame.

Speaker 2: Yeah. Yeah.

Speaker 3: Yeah. Right? That should be you're gonna be doing that work. That should be in a much you know, a different physical environment for starters. But if it's it

Speaker 1: almost makes sense that humans are doing it because the work risk that you're taking on for humanity by doing anything human pathogen related, you should almost have to

Speaker 2: put your own life on the line Yeah. A little

Speaker 3: There's something to that. I mean, you're not, you know, it's fun. Mean, at the same time, like, we didn't have humans in labs, we wouldn't have lab leaks.

Speaker 2: I like the idea yeah. Oh, that's rough. I I I like the idea of trying to trick ginkgo into play.

Speaker 3: Why are we putting people in labs

Speaker 2: and Put the Mentos in the Diet Coke. It's very important that we do this experiment. Yeah. So

Speaker 3: Yeah. So I do think you're gonna yeah. So I think there's a couple ways to get that. Yeah. Think there'll be human checking. Sure. There'll be models trained on scientists doing that. And then in the long run, the the other thing just to keep in mind. So first off, I think the fear level can can go all over the place.

Speaker 2: Yeah.

Speaker 3: To me, the very tight fear to be worried about is human doubt. Mhmm. Everything else is mostly like a personnel safety issue.

Speaker 2: Sure.

Speaker 3: Right? Are you mixing together something that's gonna do the mentos? Yeah. Great. One, that's like the society one

Speaker 2: Yeah. Is all

Speaker 3: about human pathogens.

Speaker 2: Okay.

Speaker 3: So I think you just put that's a different bucket.

Speaker 2: Yep.

Speaker 3: Because it it is good to research this stuff. We want we want a curable. We want it like like you have to have people working on these things or else we're just exposed. Right? We have antivirus researchers and computers. They don't like not get to work with viruses. Right? Like, you know, like we you need that stuff.

Speaker 2: Yeah.

Speaker 3: But but I think the answer to solve that problem in the long run is something that looks a lot like the antivirus industry in software.

Speaker 2: Yeah.

Speaker 3: A responsive system. So there's something new that came out, we can put it down. And by the way, we don't

Speaker 2: just need that for AI, we

Speaker 3: need that for nature. Right? We're getting thrown pandemics at us all the time. So so the short answer there is I think it's building up our our biosa biosecurity infrastructure, particularly here in The United States. We treat it the same way we treat other defense fields.

Speaker 1: Yeah. It feels like one

Speaker 3: of those things.

Speaker 1: There's risk there's risk associated with creating an autonomous lab, but the risk of not innovating here and just not having it as a capability set as a as a country feels like way higher.

Speaker 2: Yeah.

Speaker 3: Oh, we're I mean, the thing that's happening in the biotech industry, like the ugly secret, is all the startups that used to be, like, the the innovation engine for discovering new drugs over the last two years have been moving to China. So when you see these acquisitions of new drug candidates by the large pharmas, it went from less than 5% from China to more than 40% over the last two years.

Speaker 8: Wow.

Speaker 3: So that and that's not manufacturing. That's innovation and discovery. And you know the reason why? You know what China has cheaper than The United States?

Speaker 1: Pipetting. Labor.

Speaker 3: Hands pipetting. Yeah. So if we're gonna keep up, we gotta move to a topic. And this is by the this also how we're gonna reindustrialize manufacturing. What do you think we're gonna do? We're gonna compete with on hands? No way.

Speaker 2: So advice for No.

Speaker 3: I look at hatering and everybody else. It's it's all it's all automation.

Speaker 2: Yeah. So so advice for young people, should you if you want to have an impact in science, should you learn to code? Should you learn to prompt? Should you learn to pipette? What should you learn?

Speaker 3: You should learn about the domain. Mhmm. So so you should learn if you wanted to have a breakthrough in biology, you should learn about biology. Sure. Right? So that you're the one who understands

Speaker 2: Yeah.

Speaker 3: The limits of that. And then second, I think

Speaker 2: you do need to like, don't

Speaker 3: think PhDs are not gonna work at the lab bench at the start. Yeah. Because you got to understand the limits of experimentation. Yep. So that it's just like today, who are the best users of the coding aid?

Speaker 13: Yeah.

Speaker 3: Coders.

Speaker 13: Yep.

Speaker 3: Yeah. You know? Right? So who are going be the best users of autonomous labs? Scientists at the lab

Speaker 2: bench. Yeah.

Speaker 3: Right? That that already know. And and I think the thing is I people get scared. Oh, is it gonna take scientist jobs? That's like, there's a great IBM ad from, the night 1951 IBM ad. It's like, the the automated calculator will do the work of a 150 engineers, and it shows engineers with slide rules. Yeah. I swear to God. Wow. Like a sea of engineers with slide rules. And and you might have said, oh, will take a 150

Speaker 2: engineer jobs. And, of course, total opposite. Yeah.

Speaker 3: Right? By increasing the return on investment of computation, what was in the minds of engineers became worth a 100 times more. Jobs increased by a 100 fold, all post IBM. Right? All post the automation of computation. So if we automate all these industries, whether it's science, manufacture blah, blah, blah, it only favors the people that have the know how in those industries. But we gotta automate it because otherwise, we're not competitive.

Speaker 2: Yeah.

Speaker 3: And and in the modern era, we already automated the easy stuff. We automated the subways. It's all the autonomy. It's the things that can handle the variability. That's what the hardware guys like us have to work on, and then let the AI models go go nuts.

Speaker 2: That's awesome.

Speaker 1: Jordan, this is fascinating.

Speaker 7: Is great.

Speaker 1: Really enjoyed it. You're our new you're our two new in house science yeah. Our our science corner, our our biotechs are.

Speaker 3: Yes. So CVPN science corner.

Speaker 2: It's happening. I'm here for it. I'm talking to Yeah.

Speaker 1: Incredible incredible setup again. So looking looking forward to the next one.