Once a Scientist
Once a Scientist
93. Jason Kelly, CEO and co-founder of Ginkgo Bioworks, on the future of autonomous laboratories
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Episode 93. Jason Kelly is the co-founder and CEO of Ginkgo Bioworks. He has a Ph.D. in Biological Engineering and a B.S. in Chemical Engineering and Biology from the Massachusetts Institute of Technology.
References:
https://www.ginkgo.bio/autonomous-lab
OUTLINE:
00:00 – From Jurassic Park Pilled to MIT
03:00 – Founding Ginkgo During the 2008 Crash
07:00 – Tom Knight & Thinking About Biology Like Software
11:00 – Why Biology Lacked Abstraction (Until Now)
16:00 – The Foundry Model: Centralizing Biology
20:00 – The 5 Levels of Lab Automation
23:00 – Automation vs. Autonomy (Subway vs. Waymo)
27:00 – The Technical Bottleneck: Connecting 150+ Machines
34:30 – The Zymergen Acquisition & Strategic Pivot
38:00 – Level 4 & 5 Labs: The Future of Programmable Biology
The reason it works so well in computers, the reason computers have never ended in the way that like planes ended. Okay, is it is a coded discipline. So it. The. The like range of what you can do with computer code is like wild man. And cells are like that. Like biology is the only working nanotechnology. It's for billion years of evolutionary effort. It's totally programmable and we suck at programming it. If we got good at programming it and then it was accessible to everybody, it isn't. It is like computers, but in the physical world. It is unending applications.
Nick Edwards:This is the Once a Scientist podcast. I'm Nick Edwards. We're back with new episodes, so keep an eye out and subscribe. Subscribe to the podcast if you haven't already. This episode's brought to you by Potato. We're building AI agents for closed loop autonomous science. Check it out at Potato AI or email us at hello Potato AI if you want to figure out ways to collaborate. So super excited to have you on Jason. For listeners that haven't heard of you, this is the one and only Jason Kelly. You know Jason, do you want to give a quick background on yourself? Sure.
Jason Kelly:I grew up in Florida. I went to MIT for undergrad, did chemical engineering and biology. I always wanted to be a genetic engineer. Like Jurassic park pilled at a young age. And then there was no bioengineering major at mit by the way. Back then I was like, okay, I knew biology. And then I need to pick an engineering field because it's like genetic engineering. And so I picked chemical engineering because cells are made out of chemicals, Right. And about two years in, I realized that chemical engineering is designing oil and gas plants primarily. It's a good engineering background. But then I did my PhD in bioengineering with Drew Endy and kind of like with Tom Knight. We had like mutual lab meetings. And then after grad school, myself and four other folks I went to school with started Ginkgo along with Tom Knight. That was back in 2008. We were the first YC Biotech and we took the company public in 2021. I didn't realize that then.
Nick Edwards:I didn't realize you were the first YC biotech.
Jason Kelly:So Sam Ullman is her famous now. Had just taken over YC and he put this blog post up that was like. He was like, you know, they do those like call for startups at yc. Have you seen that? Yeah. And so it was like, I want the nuclear energy companies, material science, biotech. I think the Silicon Valley model could work. In these like hard tech areas. And Giga was like five years old at that point. I had been, we had been like bootstrapped on government grants. We were like 15 people in a lab here in South Boston area. And I emailed him and I was like, look, we don't make sense for yc. I got like a lab in Boston. But thanks for writing this blog post because it's like Oasis in the desert, you know, like if you were doing, if you weren't making it therapeutic and you weren't making software, there were no venture capitalists for you. Yeah. And so basically, and so I was like, you know, thanks. And he's like, oh, you gotta come meet me. So I flew out of San Francisco and I met him and was like, you know, love at first sight. It was awesome. I like why he's the best. And so we had just this amazing experience doing it. And by the way, everyone in Boston was like, why are you doing yC? Anything about biotech? And I.
Nick Edwards:But that's the argument. But like they've done a lot more like since then. I mean, especially now. I think recently there's tons.
Jason Kelly:Yeah, it's a dumb argument. I was like I wasn't getting any. You know, these guys weren't investing me anyway. So like, yeah, like, who cares? Terrible.
Nick Edwards:But you started ginkgo like straight out of your PhD, right? Because this is like the beginning of synthetic biology revolution. Is that the story? Like, and then five years government grants. And that's like after YC is when things really started to kind of take off.
Jason Kelly:Yeah. So basically what? Yeah, so a couple of things that come together. By then we had started to get commercial deals. So like, we started on government grants, but then in like 23rd, you know, 2013, about a year before we did YC, we started to get our first like commercial customers with our like solutions business. Which is, it's kind of like if you're in the pharma space, it's like adamab style business model. Like, we do research, partnership, we get some money and then we get royalties and milestones on the back end. And, and except we. Our model was like, we'll do that for any genetic engineering project. And back in 2013, we got those first projects in engineering microbes for like flavors and fragrances and nutritional ingredients, like weird biotech. And so we did YC and we had like commercial momentum, which really helped us, like, because then went to generalist tech like angels. But they were able to look and See, okay, well, you're getting business. I understand the gist of what you're talking about. And they invested and so then we basically got treated like a, like more or less like a tech growth company.
Nick Edwards:Okay.
Jason Kelly:Yeah, yeah, it was that era. It was like Uber, you know. Right. It was, it was like zer big private round investments.
Nick Edwards:We could do a whole podcast around this history because I'm sure it's like you've been through everything like the, you know, bring the company public. One of the reasons why I reached out recently and I was like, hey, we gotta. Because we've been talking about this for a while. Right?
Jason Kelly:Like, yeah.
Nick Edwards:But then you posted about the Genesis mission and you have like 97 racks deployed at PNNL and that's super exciting. So I'm excited to hear about that. But like, I'd love to kind of learn a little bit about what's the evolution of your guys's journey in the automation frontier. Because like, you know, some of the things you're talking about are like autonomy versus automation and lab automation. And so like how did you kind of start down that path? Like, because I know that there was the acquisition. Zimergen.
Jason Kelly:Yeah, love to. Yeah. And maybe the last point I'd make on just the general startup story. Cause I know you have a lot of folks listen in who are.
Nick Edwards:Yeah.
Jason Kelly:You know, grad school or students and things like that. So I mean, we started the company in 2008 which was like a little bit like the vibes of biotech today. Like horrible times, you know. Right. Like it was, you know, financial crisis all the company. Like my co founder Baron and I were basically backing u hauls up to shuttering startup biotech companies around Cambridge and getting all the equipment, like 10 cents on the dollar in auctions and stuff. Like for the first year I was basically like moving equipment. I actually think it's a good time. It's. It's a really good time to start a startup.
Nick Edwards:Yeah.
Jason Kelly:Because it's not as much of like a race, you know. And so you can particularly if you're like a founder led, younger person, startup entrepreneur in biotech that is a. That remain like I was hoping by now and we started to go in 2008, we're like dating myself here. But like, you know, I was hoping that by now it would be way more normalized to have like founders like were at the beginning.
Nick Edwards:Yeah, yeah.
Jason Kelly:You know, and some things have gotten better. Like, you know, we didn't have any rent a bench stuff like you know these things like QB3 use and whatever it's called now Nanolab central. All these different things like you know we, everything existing. So I you know like we built our own like lab in like an office building because it was like the only way to make it cheap enough. Like you know like it was just like nonsense and so some things have gotten better but like still like the venture community is not really on the hunt for young inexperienced founders in biotech.
Nick Edwards:It's brutal out there for a lot of people. Like so it sounds like it was similar times when you get.
Jason Kelly:Right. So my point is the good fact is you're going to have a hard time anyway at least if you want to do something like apply for grants or do it in some like oblique way in a down environment you have like more time like you know, like you're not like racing anybody for your idea as much. So I actually think it isn't a terrible time for like people who want to kind of bootstrap it and go for it. The battle is getting started and, and it applied for grants and that was a real boon for us early on. Just as another thing that people never do or get told not to do. It takes too long or whatever. Like just do it like you know it's non diluted money. It's great. You should do it. You know the like first 4 million in the get go or so was SBIRS and RV like you know, I don't know, probably 30 grants.
Nick Edwards:Yeah, yeah. You guys continue to go down the, I mean you continue to bring in revenue, obviously you have lots of gum stuff.
Jason Kelly:But like it was like basically like our initial translational venture money was USG back then, you know. And so the and that's partially like the other point of my background. I chaired a national security commission on emerging biotech over the. For like first two years of the commission and like you know, part of this like I feel like we should give back. Like, like we like really do exist thanks to that kind of like money that is there to translate tech out of the lab into the private sector on the US side. I would just still pitch to people to go for it right now especially so. But yeah I could talk about automation now. But anyway so you were asking like how do we get here? And maybe a little bit on like Autonomy and Genesis mission. So I'll try to break it up. So first off like how do we get here right so Ginkgo's focus was always on a lesson I learned from Tom Knight. And I don't know if you know Tom's background.
Nick Edwards:Not a ton. I know Drew, but.
Jason Kelly:Yeah, yeah. Okay, so Tom's an interesting guy, right? He's in 70s now. He started on the MIT faculty in 1972 in electron engineering, computer science. Wow. This is like old school computers, Nick. Okay, right. Like Tom programmed on like the timeshare mainframes. Like movie punch cards. Yeah, absolutely. Punch cards, yeah. And so he grew up on that and he watched the transition of that era of computing and he's a computer architect. So he designed like a pretty famous mini computer called the Lisp machine back in the day. And, and he watched the transition into planar semiconductor manufacturing and he became a chip architect and did that whole thing for like 25 years. And then in the 90s he switched his focus from ECS to biology basically, I think because he was getting bored and they wasn't exactly why, but you know, because as the other like programmatic discipline. Yeah, right. Like DNA is the other like coded engineering field basically along with computers. And so I get the cell, you got computers, they both run on code. That's true. All right. A lot of other differences, but that part is true. And so Tom, to his great credit, because he's like a hands on guy, opens up a wet lab in the MIT computer science building, sets up benches, starts freaking out his colleagues by growing bacteria and gets his hands dirty. And back then we tried a million different things from electrical engineering and computer science to bring them into biologic circuits and all the symbio shit that ended up being stupid. And like I'm overstating, like a little bit of logic is really valuable. Right? Like is my therapy next to the tumor? If yes, deliver it.
Nick Edwards:Yeah, right.
Jason Kelly:You see like really cool stuff. Companies like Strand, there's like cool RNA stuff, cool, blah, blah. There's like a lot of really cool, like simple amounts of logic.
Nick Edwards:Yeah. Biology.
Jason Kelly:You're trying to build like an 8 bit adder out of 37 genetic repressors. Like pointless levels of computation. Not with the right substrate. Right. Like that's actually not what biology is good at. All Right, Yeah, but Tom brought an idea that I think is like foundational to autonomy and automation and to ginkgo that he brought from electrical engineering, computer science that I think like everybody who's a biotechnologist should internalize. All right. And this is the idea of abstraction.
Nick Edwards:Okay.
Jason Kelly:All right. And the reason Tom understands this deeply is he uniquely lived this transition in electrical engineering, computer science. So when Tom started, in order to be a computer scientist, you had to be an electrical engineer. Okay, okay, Nick. Because, like, how could you program a computer?
Nick Edwards:Made sense.
Jason Kelly:You didn't know how it worked. Yeah, I mean, like, obviously there's no way in the world you're going to be able to program that thing if you.
Nick Edwards:Right, yeah.
Jason Kelly:Okay. You know, wires, you got all this shit. You got to know all that stuff. Like, of course.
Nick Edwards:Right, yeah, absolutely.
Jason Kelly:Okay, Right. But. But now somehow, right, my 10 year old can like drive boxes around on iPad or something. Program a computer today.
Nick Edwards:Yep. Yeah.
Jason Kelly:Okay. What happened?
Nick Edwards:So many layers of abstraction. Yeah, yeah, Right.
Jason Kelly:And, you know, it started at the low level. You had assembly language, and then you had operating systems, then you had programming languages, and then you had graphical programming languages. And now you have ChatGPT, and you can talk to it and it'll write code for you. And the consequence of that, like, why did that matter? Okay. And the reason it really matters is it let computer science and electrical engineering become two different fields. Two different disciplines could talk to each other through the abstraction layers and benefit from each other's innovations. Okay, you made a higher processing chip, and now suddenly you could do a new type of software. Oh, you created new software applications that create demand. Now we're buying more high processing. No, you invented 3D graphics. Guess what we need? We need GPUs. Okay, right. So you had this like tic tac between the. Between the computer science people and the electrical engineers that benefited each other because everyone's innovation could get passed back and forth. All right, that's point number one. Point number two. Let me tell you, the culture and community of computer science and electrical engineering are super different. Okay, so computer science, okay, this is like a theory of the mind, right? It's about math and algorithms and. Okay, right. Electrical engineering, this is like physics. It's not math. Okay, so it's about automation, it's about lithography, it's about miniaturization, it's about atoms. Very different people do that than do this. So why do I tell the story? So Tom gets into biology and he's like, all right, let's program some DNA, baby. Right? Like, that's why I'm here. I want to write some code in this new substrate and, okay, I want to program it. And the first thing I got to learn how to do if I want to program DNA is pick up a pipette. Right?
Nick Edwards:Okay.
Jason Kelly:Because how do you have the right to design a gene in E. Coli if you don't know what PCR is, If you don't know what a restriction enzyme is, if you don't know how to do a mini prep? You don't get the right to design DNA if you don't know how to work at the lab bench.
Nick Edwards:Nick. That feels right to me as a scientist at least.
Jason Kelly:That has meant that the only people who get to design organisms are the people who can tolerate like the minutiae of like working at the lab.
Nick Edwards:Yeah, yeah, that's right.
Jason Kelly:And it is very much like, if you think about the number of computer scientists that could also handle electrical engineering at the same time. Small.
Nick Edwards:Interesting.
Jason Kelly:I mean, it was like, that's why it was like, you know, the crew, like it was so narrow of a field back then because it required so much.
Nick Edwards:Interesting.
Jason Kelly:So, you know, Tom brought this concept that we got to abstract. The design of the organism, the design of the genome, which is fundamentally what bioengineering is about, is designing genomes.
Nick Edwards:Yeah.
Jason Kelly:From the lab work. And my argument is it's very analogous to computer science and electrical engineering, just a slightly different way. You're designing the genome. What you really are is a biologist.
Nick Edwards:Yeah.
Jason Kelly:You're a person who understands the biology.
Nick Edwards:Yeah.
Jason Kelly:You understand how the cell works and this, that blah all my God. It's like a turtles all the way down in complexity. It's crazy how much, you know, right. Like how much there is to understand about how a cell works. You got to have all that in your head to be good at designing it. Okay.
Nick Edwards:Yeah.
Jason Kelly:But over here, if you want to drop the cost of a mini prep by a factor of 100 or blah, blah, or like that is automation and miniaturization. It's being a hardware person and a software person. You also don't have to know anything about the biology.
Nick Edwards:Yeah.
Jason Kelly:Okay. And you could do some real damage there. And so those communities need to be split and talk to each other with abstraction. Interesting.
Nick Edwards:It does make sense. Yeah. So like, I mean, okay, do you see those lines starting to blur like in the future? And, and because like, I guess a big question that I have is in the next generation of science, I'm skipping too many levels ahead probably right now, but like in the next generation where, you know, we have more autonomy and people collaborating directly with AI scientists and autonomous systems. What's like the role of, of that experimental taste and like, and the role of the biologist, maybe we can come back to that question if, you know, if you want to kind of talk a little bit more about, like, how you guys are thinking about the autonomy and.
Jason Kelly:Yeah, so, yeah, maybe I'll take a little bit of a drive down autonomy lane. Automation lane, and then we come back to like, how will people use it? So.
Nick Edwards:Yeah.
Jason Kelly:Yeah. So the. The point I want to make about the Gingo journey is that whole concept of abstracting away the lab from the design of the DNA was in the founding of the company. So we have been working on that problem for 17 years now.
Nick Edwards:Yeah. Yeah.
Jason Kelly:Is not an easy problem. I'll tell you how we attack the problem. First, the. The big thing we tried was what we called a foundry. Right. And the idea was the foundry was like the electrical engineering half of the house.
Nick Edwards:Yep.
Jason Kelly:Okay. It was centralized, automated. It was filled with people that understood lab automation, processes, operation stuff. And then we had a separate team, our code base, our organism engineers, the people that were like the biology experts. And we did separate them. The idea was these people would order from the foundry. We built custom software to manage their requests. So you would request the work you wanted and it would get handed to the foundry to do your lab work. Okay. So we did structurally do that. And like, no one does this. Like. Like, if you walk into a research lab at Merck or Takeda or something, it's like everyone's on a bench, they're thinking about designs and they're doing work at the bench. It's like all like, it's, you know, it's exactly what it was when I was in grad school. Yeah. So.
Nick Edwards:So we.
Jason Kelly:Okay. Tried to do it this way. And. And I think importantly, we tried to do it this way across whatever kind of research were asked to do.
Nick Edwards:Okay. And the foundry model was like.
Jason Kelly:Was.
Nick Edwards:Was inspired by like semiconductor foundry.
Jason Kelly:Okay. That's where we sold the name. Yeah. Yeah. It's not exactly a perfect analogy because it's not really manufacturing.
Nick Edwards:Makes sense. Yeah.
Jason Kelly:The core idea behind it was separate the people doing the lab work from the people ordering the lab work.
Nick Edwards:Yeah. Makes sense.
Jason Kelly:That's the big idea. Like, separate the people writing the software from the people making the chips. Like, that's the loose analogy. All right. And then what I did in the foundry was I operated it with what I'll call like a services based architecture. Okay. All right. So we had a team that built DNA. We had a team that purified proteins. We had a team that ran genomics, we had a team that did metabolomics, we had A diva. And we had like, you know, 40 different foundry groups. And inside the groups or sub teams and like you had like specialists. And each team did see like, and. And the level of automation we had was a mix of walk up. Okay, so walk up to a Hamilton, walk up to a T can, whatever. And also work cell.
Nick Edwards:This is what you were calling like level 01 in.
Jason Kelly:Yeah, yeah. Did you read my.
Nick Edwards:Everybody should read your blog. It's. Yeah, it's on Ginkgo's site. We'll link to it in the show notes. So.
Jason Kelly:Yeah, so were sort of at level one.
Nick Edwards:Okay. Yeah.
Jason Kelly:What we call like narrow lab automation. Okay. So the lab automation was to do a certain operation at Hydro. Like, I'm the DNA build team. I have a thing that stitches two pieces of DNA together. I have a different operation that stitches four pieces of DNA together, whatever. Right. But if you want to do some random ass experiment, I can't do that. So I have like a suite of services and then I pass among those services to get your complete job done. And that was how the foundry worked. And what we found was you saw efficiency gains through the centralization, but then you had efficiency losses through the handoffs among the services that were run by these different groups of people. And like, I'd say were net more efficient, but weren't like a thousand times more efficient. We didn't see like crazy, like cleaner semiconductor gains on that type of setup. It was getting better every year, but, like, not at the end of the day. It didn't get good enough. All right. I guess it's maybe the short way to say it. And so, okay, fine. No one had tried this before, you know, and so what was our learning? So our learning was. The split, I think is still right. Okay, Like, Tom is fundamentally right about the abstraction. Makes no fucking sense to have the same people do these things. And this comes back to your taste thing. So what would be a new thing to try? A new thing to try is the same kind of like, I want to order, but there is no independent groups doing the services. There's a giant level two, general lab automation. And I'm saying level two. Just so you know what Nick was talking about. I wrote, we wrote the blog post that was like talking about levels of lab automation. Lola. Just like there are levels of autonomous driving, and so there's like 0 to 5 on like autonomy in Tesla. And so we have 0 to 5 for autonomy in the lab. And the first level is you're working at the bench by hand. That's level zero, level one. So it's all manual. Level one is like walk up and work cell. So I've got a system, but I walk up and I tell it to do one thing and it runs it. And then level two is general lab automation where I can submit lots of different requests into one big system. And that's what we have now at Ginkgo. And that's what the Genesis mission just bought for PNNL Pacific National Lab is a general purpose system. And so we sell our like unit of sale as these rack carts. So it's like a reconfigurable automation cell. And it's basically a robotic arm, a piece of magnum ocean track and an instrument inside it. And so videos of.
Nick Edwards:Watch videos of seeing these things run are so fun.
Jason Kelly:Yeah, yeah. We now have 36 of them in Boston. Yeah, it's going to be 50 soon. Like it's really neat setup to see.
Nick Edwards:All right, I'll be there for SLAs. Let's, let's.
Jason Kelly:You should come by at SLS. You'll like it. Anybody who's on SLA is welcome to come visit. Like SLS is like the automation meeting in February in Boston and anybody that's listening that just wants to come check it out, we're very happy to have people come see it. But the. In any case, what we've done is we basically can Lego block these carts where you connect as many pieces of lab equipment into one integrated setup. And you can send a sample to any piece of equipment by just like passing it along on the, on this magnetic rail and it gets delivered to an arm. The arm picks it up and puts it on the equipment. And we have like an arm for every piece of equipment. It's like overdo it on the arms. Like the arms are reliable and not a bottleneck. The transport system is reliable and not a bottleneck. And so you're kind of taking away a lot of the limitations from traditional work cells. Let me give like one more analogy on what I think is different about traditional lab automation. Work cell. Like you would get from like a high res or thermo and autonomous lab, what we're building. So let's still take transport. Again, transportation. Subway is automated transport. You're not driving it. Right. Like you get in the subway and it takes you somewhere. It is automated transport. We've had it for a hundred years, right?
Nick Edwards:Yeah.
Jason Kelly:Okay. A waymo is autonomous transport. Yes. And the difference is that automation is basically been applied like since the Industrial revolution to things that we are just repeating the same thing over and over again. Because that was like the level of computation that we had available to us from the, from, you know, the loom was just mechanical computation to the subway system, which is like mostly the rails, but there's also some little computing in there that knows where to stop and handles the brakes. Right. Like so. So there's sort of like low levels of smartness on the computing. But still, if it's the same operation and there's demand for a zillion of it, we can automate it. Yeah, but driving your car was not like that. Yeah, it was variable.
Nick Edwards:Yeah.
Jason Kelly:So you want to go different places every day. You don't want to go the same place every day. Different things are going to pop up in FR a car. There's a kid, there's a stop sign, whatever. Right. Like it, like it's raining. There's all this variability. And, and so autonomy is the waymo that can use smarts to manage that. That variability. Okay. And so why do I give this analogy? Because the work cells, all the automation we have built so far for biotechnology outside of what I have here in Boston and we're selling to P and L and other people is subways, run my high throughput screen 30,000 times a week.
Nick Edwards:Yeah.
Jason Kelly:Right. It's like the same thing over and over again. Same thing. I want to combi cam. Right. Bang, bang, bang. Right. Like the origin of the work cells was the human genome project. That was like when we first had integrated automation and labs because they were like sequencing. Same thing. Same thing, same thing. But the problem is it's limited to only going to the stops on the subway. Yeah.
Nick Edwards:It's not dynamic. Like what you're looking for is dynamic automation.
Jason Kelly:Right. Like flexible and dynamic. Yep. And so what has been the flexible and dynamic tool for genetic engineering over the last 30 years? The lab bench. So the lab bench still remains the only thing that is flexible enough to like. It's like driving the car. It does get you there and you can't automate it. And so we're stuck with it.
Nick Edwards:Yeah.
Jason Kelly:And most of the places you want to go are on the lab bench. So 95% of what people spend money on in biotech R and D is the lab bench.
Nick Edwards:So this is what you're saying, this is what. How like when you talk about in the blog, like an autonomous lab is like a new scientific instrument. This is kind of. This is what you mean. It's, it's an entire Lab, it's your new bench.
Jason Kelly:Like, it's your new bench and pipette like it King should go anywhere the bench can take you. Right. Except you just ask for it. As you go up the levels of autonomy, which we can talk about, you should eventually be able to talk to it roughly, like you would talk to a technician. Explain the protocol you want done, get some, you know, go back and forth, tighten some stuff up, specify, give them a protocol or a protocol and like they get the work done and you get to talk about results later. That's what it should feel like. That's climbing the levels of autonomy as opposed to higher throughput automation, which we've already beaten that dead horse pretty heavily over the last 30 years.
Nick Edwards:Yeah, that's a super helpful description of it. And I like that, I really like that analogy because I remember actually the first time I got into Waymo, it was in downtown San Francisco and were turning right and somebody was stopped at the corner and they were walking across. And the moment when I realized, oh, this is like, this is an autonomous system, was when it was like inching forward behind the person that was walking across the street, almost like nudging them to move.
Jason Kelly:And our.
Nick Edwards:So I'm curious because there's a level of agency that's important there. And so like you talk about like AI agents in this process. How are you thinking about that? How are you thinking about agents? And, and like the integration of lab automation?
Jason Kelly:There's a few different, There's a bunch of, there's like a million problems to work on here. Okay, so the first one is, can I just connect like all the equipment you need in a lab to go wherever you need to go? Because if you walk around a lab at Merck or at an academic lab or wherever, there's 150 different pieces of benchtop lab equipment around. Yep. There's center view, there's region, there's some of the basic stuff. But then there's like, hey, some weird esoteric assay machine and this other thing and like, whatever, and sometimes you need that. So one of the key things is the system at a hardware level needs to be able to connect a lot of things all at once.
Nick Edwards:Okay.
Jason Kelly:So we think we've solved that problem with our automation cards. Okay, racks. Next. It now needs to handle. This is level two automation, getting many orders submitted to it all at once. Okay. So again, traditionally the subway has a schedule, right? Like it moves and then the next one goes. And so when you have these traditional automation work cells. They're running bash. It's not like you come in halfway through and they're like, hey, run this for me. And just throw it in and run Arm grabs a different thing and works you into the system. No, so it's more, it's serial,
Nick Edwards:it's operating serial fashion.
Jason Kelly:And you want it to be like a parallel bus. Like you want to be able to like basically bang stuff into it. Yeah, for sure. And that sort of like scheduling problem is made much more complex if you're letting people just add stuff whenever you want versus if you have like a schedule like a subway.
Nick Edwards:Makes sense.
Jason Kelly:Okay, so that's the next software layer that you have to work on is like the scheduling. Okay, so we have that just sake of argument, that works. Okay, so then you're on to the next layer which is how does a person ask for something to be run on that system?
Nick Edwards:Yeah.
Jason Kelly:All right. Well, the straightforward way is write the code that the system accepts for knowing how to move a plate, you know, a sample through a various set of equipment, and importantly, what you need to tell each piece of equipment that you stop at. So there's like the parameterized control of the equipment too. But you need in your code to say, run the PCR machine for this long, spin the centrifuge at this many rpm. So the low level software, again, assume it can handle all that. You still have to actually write that code.
Nick Edwards:Yeah, which most biologists don't really know how to do.
Jason Kelly:Correct. Correct. Well, you don't certainly don't have to do that at the bench. Right. And so now I'm putting a burden on you to do that. Okay, so there's then the next level which is like, what are the tools I give you to be able to write that code? And up till now the primary tooling has been like graphical programming languages. Yep.
Nick Edwards:Okay.
Jason Kelly:And so, and that's fine. Like it is useful and you can kind of see it a lot of these pro. And by the way, like we're not talking about code like a website level of complexity coding. It's like Microsoft Basic coding. Right. Like it's like move down the line, you know, again, dating myself, but like go line by line through this code, you know? Right. Like do this step, then do that step, then do this step, then do that step for any individual protocol, you know, like it's not often that complex of like logical loops at the level of just do this thing. But you still have to write it. Okay. And so what's really exciting, and I think, like, things like potato are leading the way on this is could you interact as a scientist with human language? Specify, hey, there's a protocol I want done, and get it converted into code for the automation. And you just hit go, and then it just runs. And. And so that. That's sort of my level three. Like the conditional autonomy. That's half of it. Half of it is that you argue.
Nick Edwards:toward the closed loop systems.
Jason Kelly:Right. Leave the closed. I'm. Come back to the closed loop in a second. Yeah, yeah. Okay. Separate from that. Yeah. So. So this is. I am a scientist and I want to use an autonomous lab. I need to get my job order onto it.
Nick Edwards:Yep.
Jason Kelly:Okay. And the autonomy, because I was deliberate about this, I think the closed loop stuff is super cool. It's a slightly different. It's like an orthogonal technology branch, in my view, from the autonomous lab. It's like the AI scientist. The AI scientist can use an autonomous lab, but a scientist could use an autonomous lab. I can get in the back of a Waymo as a person and say, go there. Uber can have an AI agent order a lame O somewhere with no one in it and go get something picked up at the delivery store or whatever. Right. So if like Uber has an AI agent doing, you know, closed loop delivery pickup or something, that doesn't change the fact that the Waymo is autonomous, whether there's a person in the back or Uber is telling it where to go. Right?
Nick Edwards:Yep.
Jason Kelly:Okay. So that's my analogy. Like, my analogy on the autonomous lab is to focus on. Not actually the closed loop stuff is to focus on can the lab go the fuck where you tell it to go.
Nick Edwards:Yeah.
Jason Kelly:Can the Waymo get you there safely? I want to go there. Yeah.
Nick Edwards:Don't screw it up.
Jason Kelly:Okay. Not plan my trip for me or come up with where I want to go and, you know, for the next step, you know?
Nick Edwards:Yeah.
Jason Kelly:Okay. And that's what I want the autonomous lab to do. I want to say I want to go there. I want this experiment done. Can you get me there? And there's two parts to that one. I got to submit it. Now here's the problem. I wrote down a protocol I actually submitted into the automation. Just perfect. It's exactly the code. The coding agent writes the code just like what I asked for.
Nick Edwards:Yep.
Jason Kelly:And it goes into the automation and it encounters some very common physical automation, like error condition, like, oh, the plate size is incompatible. It's the wrong plate size for the echo, you know, acoustic liquid analytics by the way, it's the plate type. You said you wanted scientist, so I did a good job doing what you wanted. It's just not a good experiment. And you don't know that because you're not usually doing in an automated way. Something, something. Something about you doing at the bench, you usually don't catch it. So that's the second thing, which is the physical debugging of a protocol when you move it from the bench to automation.
Nick Edwards:Yeah.
Jason Kelly:And that is not trivial. Okay, Right. Like. Like, automation engineers are a whole, like, job title of people who talk to scientists to basically transfer their stuff onto work cells, onto the subways.
Nick Edwards:Okay.
Jason Kelly:Right. And it's just like, one protocol, and they spend a bunch of time and da, da. And so. So to get to level three, autonomy in my LOLA scheme, you have to solve both those problems. One, you got to be able to have a human language interface that just turns into code. And then two, that software needs to debug your code not for errors in compiling, but for physical errors, likely physical errors on the automation.
Nick Edwards:Yeah.
Jason Kelly:So that, by and large, you do not need to debug it yourself.
Nick Edwards:How much of that can happen through formulation versus.
Jason Kelly:Great question, great question. I think some of it. Yeah. And I think this is an open area. Like, we're like, this is not working yet.
Nick Edwards:Yeah, yeah.
Jason Kelly:Like, we're working our asses off at level two just to be able to handle, you know, 100 requests coming into a system. No one's ever done any of that stuff. Like, everyone's just been running subways. Right. So I'm trying to get, like, a solid level two right now.
Nick Edwards:Yeah.
Jason Kelly:And then once I have that, like, then, you know, then I'm like, okay, cool, maybe I can solve this problem next. But, like, the. It's an overall. It's a big one. It's a big one because scientists hate this shit. They're like. Like, when I do it at the Bench, it just works. But then I. I put it on these robots and they do stupid things, you know, like they, you know, and you're like, yeah, but, like, you know, but they could work all night, you know? Right. Like it's worth it. You'll be able to run an experiment and come in the morning, get a cup of coffee and look at your data. Like, wow. You know? Right. Like, what would that be? Like, so it's both of those things. It's. It's the physical debugger and the code writer. You got to solve both, and then you're at level three.
Nick Edwards:I think the level three automation makes sense and, you know, the agents kind of fit into in there because it's still a foundry model. Right. And that's interesting that it's like, remained consistent that whole time. This is like the next.
Jason Kelly:Still very much our attitude that like the human is ordering from a centralized thing. I've just got. I'm like changing my paradigm for the centralized thing based on my learnings over the last 10 years. Yeah. And then the other thing we did, just so you know, I can't go over that period of time, is we started to at the beginning, like you mentioned the acquisition of Zymergen. So I bought Zymogen 3 1/2 years ago for the automation because I believe that they had it right. And this stuff takes a lot. Oh, my God. It's like software on software and hardware and debug. I mean, we're like in the most esoteric corner of Automate. And I was like, put a humanoid in the lab. Like, it's literally the dumbest thing I've ever heard. Are you talking about. So, like, they had just worked out a lot of the bugs and I mean, these things start. Started being built at Zymergen 8, you know, eight years ago now. Right. And Zybergen was also well capitalized. So. So, like a lot went into this technology. So we bought it. But here's the rub, Nick, because I'm such a business genius, they were selling the. They'd started to sell these things, right? Because remember Zyworge like tumultuousness and so they, they're like, oh, shoot, we should start selling them. And so they started selling the. Selling the cards. They sold some. There's been. We've had customers that have been using it for like 30 years now. Really, like effectively super solid over a long period of time. When I bought it, I was like, stop selling them because I'm a genius, okay? And like, and because my business model was research partnerships where I would get royalties and like the Adama model. And so I wanted more proprietary technology to drive people to do research partnerships with me. Okay. I have found religion over the last like two and a half years now that was like, not worth it. And so what I'm doing now is I'm taking a lot. I mean, we spent, you know, if you'd done a couple billion dollars on technology development in these centralized platforms. And I'm now basically selling that, like, thermo would sell it.
Nick Edwards:Okay, that's it.
Jason Kelly:You. You want an autonomous Lab. You like what you see in my lab here in Boston. I'll. I'll just put one in your lab, no problem. Like you, we're selling reagents now. Like you like we have like the cell free kit. That's awesome. Like, like you know, it's way better than the ones that are in the market right now. Like you can just buy it. Yeah. That used to be a bunch of secret sauce for us in our protein expression pipeline. So we have like, I'm like exposing those things now as just tools, products with a no royalty, nothing. You just buy it and you know, I just got a check and so that's been great. Like, like I've really. It's been a good mission fit for us, Nick. Like so like that's given us energy over the last year in particular as those businesses have picked up. But like that's a change from the like 10 years after, you know, or whatever, the seven years after YC before went public where like were really only doing research partnerships. So we are like new as a vendor in this way. But we are way ahead of I'd say traditional vendors in the technology stack because we could have bought their stuff. Like I have five other work cells around here. The whole reason I like we built in house arbor is because they suck. Right.
Nick Edwards:Like there's so many questions I could ask you because like I think we really vibe on these types of questions for sake of time.
Jason Kelly:Yeah.
Nick Edwards:Let's say we get to level four automation. What are the implications of that on how we do science and on society?
Jason Kelly:So level four is at the point where the extra bit is still the same as level three. Right. It does like you know, it does the programming for you and it debugs it. The level four is now you don't have to specify what the protocol is. You specify like goal you're trying to achieve. I want to answer a question about X in the cell or Y. And the model builds the experimental plan and then submits the job order. Doesn't. You don't need to debug it and you get back data and it helps you interpret the data or you can interpret the data either way towards your experimental target. So it's now handling some of the experimental planning. But it's not just that you don't get to level four by just having the model that can do that for you. Because who gives a shit? You can't do lab work. Right. Like you're back to. Yeah, they go to the Lab bench. And people are like, I don't want to do the lab work of all model, you know. Right. Like. And so, like, you gotta. I believe it's like that whole stack has to exist to really get the most out of you guys and the Edison and stuff like that. I think it's all together, in my view, to like, really blow it open. And the last part, just so you know, Level four and five in cars are basically the same, except you take off the geofencing.
Nick Edwards:Okay.
Jason Kelly:Okay. So it's like not just wayos in San Francisco. It's wayos everywhere. That's when you get to level five. No, no one actually has a level five yet.
Nick Edwards:It's like a generalized lab intelligence then.
Jason Kelly:So my point is, level four is pretty much where you want to hit that. That is like the fucking thing. And all I'm saying is once you go to level five, it's any experiment you want. It's not like, limited in a certain window of like, well, we only have these things or it's only in this area. This stuff's hard to automate. It's like all the things. I mean, the thing that's really exciting to me is it's like with computers, as we did this, like, you start to get like, just consumer scientists, right? Like someone who says, hey, I want to understand this thing. Like, maybe. I mean, human biology, I think, is an obvious one, but maybe it's something about. Maybe they're into horticulture, plants, gardener, you know, like, maybe they want to ask that question there. Maybe they want to do this. Like, they suddenly can say, I want to do this thing in biotechnology. Make me an experimental plan to ask the question, to develop the product, right? Like, I want to make petunias that glow in the dark. Let's talk about it, right? Like, you know, like, yeah, that's a product on the market today. But that was like a lot of experts and, like, could people have the ability to just have these ideas using the substrate of biotechnology? Right? And. And I know that sounds crazy because biology seems complicated and everything else, but I will draw you back at the end here, Nick. To Tom Knight in 1972 with a fucking pocket protector and his fellow tribe of nerds who were the only people in the world who could program. And then today, and it would have seemed absurd that, like, average people and children would be able to program the timeshare mainframe at mit.
Nick Edwards:Yeah.
Jason Kelly:Like, are you fucking kidding me? How absurd that would have seemed, right? So if we do this, it will happen. Like, we will get biology weirdness out in the world because people will have crazy ideas and they'll have the tools to do it themselves. That's what we get. I'll take one last point on this. The reason it works so well in computers, the reason computers have never ended in the way that, like, planes ended, okay, is it is a coded discipline. So it. The, like, range of what you can do with computer code is like wild, man. And cells are like that. Like, biology is the only working nanotechnology. It's 4 billion years of evolutionary effort. It's fucking incredible. It's totally programmable, and we suck at programming it. If we got good at programming it and then it was accessible to everybody, it isn't. It is like computers, but in the physical world, it is unending applications. That train doesn't stop, you know, like, it really doesn't. And that's what we get if we get to level fucking four or five. Yeah. You know, like, we could do that.
Nick Edwards:Yeah, man. It's a compelling vision for the future. I love it. What would you say to a graduate student or postdoc that's listening to this and thinking about the future?
Jason Kelly:Buy potato. This is fun, man.
Nick Edwards:Yeah, thanks. Thanks, Jason.