Once a Scientist
Once a Scientist
96. Chris Anderson, Advisor to Pillar VC and former Wired Editor, on How Adjacent Technologies Will Transform Lab Automation
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Chris Anderson currently serves as an advisor at Pillar VC and Renaissance Philanthropy. He was editor of WIRED and spent seven years at The Economist, where he worked as editor of both the technology and business sections. Anderson holds a degree in physics from UC Berkeley, has conducted research at the Los Alamos National Laboratory, and has done stints at the leading journals Nature and Science.
I think, you know, the most important one is being sort of polymathic. It's an extension of being curious. But basically being interested in everything means that you're able to spot connections across disciplines. And often some of the most transformative moments have been interdisciplinary.
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 to the podcast if you haven't already. Chris Anderson is an advisor at Pillar VC and a senior advisor at Renaissance Philanthropy. He's got a really cool background, and I think many people that are listening are probably familiar with your work because you've been kind of in this space for a little while. Chris.
Chris Anderson:You've.
Nick Edwards:You've got a really interesting career. I. I'd love for you to piece it together a little bit, because I know you started out as a scientist and working as a physicist, but then spent some time as an editor both at Nature and Science. I think you're what, the senior editor at Wired.
Chris Anderson:I was the editor in chief.
Nick Edwards:Okay, all right. And editor in chief, and then wrote a couple books, one of which was New York Times bestseller. You spent some time in the maker community. We're CEO of a robotics startup. And now you've been interested in DIY Labs. So super excited to have you on. And thanks for, you know, coming on the podcast, first of all, thanks.
Chris Anderson:It's hard to make my career make sense, although every step of the way, you know, it did. But basically, scientists sort of, you know, not a good enough physicist to be a physicist, and so transitioned into media science media in particular, and then tech media, and then got back into tech through drones. So that's the robotics we did were flying robotics and sort of it helped start the, you know, the modern drone industry that was, you know, kind of cheap and low cost. And. And my. My single stupid trick is I put the letters DIY in front of industries to see what happened. So DIY drones became a community, including technology that. That we built in open source.
Nick Edwards:You started out as a. As a physicist. How did you end up moving from that into editing? And, you know, tell me a little bit more about kind of some of these transitions, because it's always good for, I think, junior people to hear the different ways that people move through science in these different careers.
Chris Anderson:So physics in the middle of the 20th century was incredibly exciting. As a teenager, I dropped out of high school and was a truant, but for the best possible reason, which is I was reading the Feynman lectures in the Library. Oh yeah, rather than going to classes. So, you know, the 20th century, between the Manhattan Project and Feynman and quantum mechanics, you know, the physics was the kind of the golden scientific domain. And, and so, you know, I aspired to, to do that. But, but at the time I like, you know, was like 27 years old and going back to college and doing physics. You know, two things had changed. It was now, you know, the world's best physicists were coming to, I was at, there was at Berkeley and you know, the world's best physicists, especially in China were just now showing up and they were just unbelievably good. Second is that physics kind of fizzled out for a kind of interesting reason, which is that any scientific domain as you know, the sort of, the gap between theory and experiment can only be so big. In physics if you have a theory about, you know, some particle and you can test it within three to five years, that's great. And you know, it's either falsifies or doesn't. If the experiment falsifies theory, then you come up with a new theory. If it confirms theory, then that becomes canon. That sort of finite gap between theory, experiment is super important. The problem with physics, at least the kind of physics I was doing, which is particle physics, is that you're basically trying to get closer and closer to the Big Bang, which has more and more exotic particles that only exist at higher and higher energies. And the problem is that the cost of those particle accelerators we're talking about like, you know, 10 mile diameter, you know, underground rings and then, you know, 20 mile and then 30 mile, you know, they're just.
Nick Edwards:How many of those can you make in the world? Right?
Chris Anderson:How many can you make in the world? So basically physics got harder. And I, you know, I was not, you know, among the best. And so, you know, the funnel shrank. And so my whole generation of physicists came to the same conclusion I did. And almost everybody at that time went to Wall street to become a quantum. They dealt with high performance computing, they were good at math, they were good at programming. And the next adjacent space that had those things was Wall Street. They got paid a lot better there. So almost everybody went to Wall street to become quant. But the other thing that happened in the physics labs at that time was the Internet. That the Internet was designed to connect physics laboratories to share things like supercomputers. And the web was created at CERN in Switzerland, that physics facility. So physicists are like, you know, Crap, I can't do physics. What can I do? And the answer is you can do Wall street or you can do the Internet. And I chose path B. And that took me into, well, I mean, then I went into the journals in Nature and Science and then the Economist, where I was for seven years. I increasingly started realizing I was doing less science and more sort of Internet coverage. And that took me into tech.
Nick Edwards:Interesting. Do you feel like you're kind of coming back home a little bit in the last little while you've been looking at DIY labs or.
Chris Anderson:Exactly. I mean, so after I did sort of robotics and AI and you know that both drones, and then I did full scale aircraft developing air taxis, and then went from there to a kind of a spinoff that was working on AI advanced manufacturing, basically. Still sort of thinking about how would we, if we, you know, we're going to redo the industrial revolution in 2026 or now, how would we do it differently, you know, both on the design side and the manufacturing side. So that gave me some real sort of familiarity with like, it's a good time to like rethink any industry from first principles. And I thought, what is like, adjacent to that? I realized that, you know, science is kind of a form of manufacturing. You're generating, I mean, you're working with atoms. Those atoms could be molecules or cells and things like that, but you're basically working with the physical world. Although much of manufacturing is now already automated, as I discovered, you know, not enough of science is automated. That was fine for decades. You just throw grad students, you know, at it. But that when AI kind of enters the picture, and this is, this plays entirely to your strengths.
Nick Edwards: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.
Chris Anderson:When AI enters the picture and AI is generating candidates, you have a hypothesis. AI suggests some paths to pursue. AI can generate candidates faster than you can do them.
Nick Edwards:Yeah.
Chris Anderson:And so you basically need automation and.
Nick Edwards:Better one and better ones.
Chris Anderson:But the reality is that any AI needs to be validated. It needs to be ground truthed because it'll start hallucinating. And so like I was saying before, if the gap between theory and experiment gets too big, then theory becomes not just hallucination, but almost like metaphysics. It's like philosophy, it's poetry, you know, it just sounds good. And so you got to keep that kind of theory validation loop pretty tight. And because the AIs are generating hypotheses so quickly, or can mandate so quickly that the only way to validate some significant fraction of them, so to keep them honest is to automate the process. So here we are.
Nick Edwards:That's something I think about a lot, is just like, how do we close that gap? How do we close the loop faster? I mean, you will never get to a commensurate degree of the amount of hypotheses you could test. But at least we can widen the funnel, right?
Chris Anderson:Absolutely. And you don't need to test everything, but you need to test representative samples so that you can, you know, establish that the gradient is accurate. I had a really interesting conversation with Andrew White from Feature House now.
Nick Edwards:Yeah, he's great.
Chris Anderson:I was saying, you know, because I'd worked in the publishing industry and I knew the publishers, like Elsevier, I was like, if we could write them a check so that they would open up all their paywalled papers so that the AIs could learn from it, what's that worth? Well, he thought a little bit and he sort of said, you know, it's probably worth, you know, it's a decreasing amount over time because, you know, we're generating more data faster than we're, than the old did. And then he pondered for a bit and then he came back and said nothing. I don't think it's worth anything. I was like, wow. And he says, we can now generate new data so much faster and higher quality data. And data has all the metadata that you describe, not just like, you know, a one paragraph obfuscated method section, but like literally the code that ran the machines that generated the data. And this replication crisis we have, I mean, how much of that is due to slop or error or fraud and how much of it is just due to the fact that they. People are not precise enough in what their methods are and so it's impossible to replicate them.
Nick Edwards:Yeah, a hundred percent agree with that, I think. Yeah, obviously you want to stand on the shoulders of giants and that's part of what we do when we go and really deeply understand an area and a body of literature, but it's really just like vectors for thought and you know, the ability to give you, like to help you develop that taste. Right. So that you can then use the tools to allow for the throughput and to test, you know, like a number, any number of hypotheses within, you know, kind of like that given vector, the.
Chris Anderson:Genome project was what, $3 billion for the first genome? I can't remember was it, remember how much it was?
Nick Edwards:I don't remember. I was in the billions though, for sure.
Chris Anderson:Let's say it was 3 billion and now it's like 3,000. Right. So if you could sort of rerun the 20th century's, you know, key science at 0.1% of the cost of doing it the first time, and this time it's going to be super reliable. I mean, you wouldn't redo everything. It's certainly asking a question, you know, so in physics, for example, what you do is you have what's called a cross section of particles, which is basically how big are they when they, you know, how big do they appear in collisions? And the first time you do it, your air bars are like this. And the second time you do it, your air bars are shrink over time. And there's a lot of value in just getting more decimals of precision for any measurement. And so, you know, getting, you know, decimal like 27, you know, adding decimal 27 to the previous 26 is not Nobel prize winning work. It's important for a sector to kind of like refine its measurement accuracy. And I think the same thing could apply to biology. This is open question. I have nothing, I'm not a biologist, but let's say like take the PDB for example, which I guess is mostly done with crystallography. Yep. Is that, is that right?
Nick Edwards:Yeah. And it's like pretty established methods that people are kind of running similar methods over and over the same method basically to generate these questions.
Chris Anderson:What if you, what if you redid all of them? How different do you think doing it with modern methods would be from previous. I mean, in other words, I guess, how much error do you think there is there in the pdb?
Nick Edwards:I don't know. I'm not a crystallographer, like so it's hard for me to have a real accurate description. But like, I think that the crystallography and the reason why we have the PDB is because people contributed obviously to this open source movement. And also there was an established standard for how to generate that data. And so I think with new methods you could probably get to a better precision. You've already overcome a bunch of the standard deviations. Right. Because people standardize these methods and like that's something that we've been thinking about at potato a lot is because we built this tool for optimization of experiments and it touches on the point that you're talking about Is you need to have really good assays that, you know are working and have been optimized and can give you a strong signal to noise ratio. Because, like, the challenge is a lot of data just is super noisy because we don't have good standard working assays for them. So I really think that's the bottleneck is getting really good standardized assays that allow you to generate the kinds of data that you could build, you know, an alphafold with or whatever the next modality that you really want to go after is.
Chris Anderson:I think the answer is it's now possible to gather data that's higher quality, more precise, more repeatable. Absolutely. Way deeper than it was in the past. Obviously, going forward, we'll do that and we'll do. We'll share it and we'll have like, the proper repos and we'll have. The message will be much more detailed. But I'm just wondering whether there's a case for going back and repeating experiments on the grounds that. Not that they're wrong, but there's probably the error bars could be shrunk if you did it with modern methods.
Nick Edwards:I don't know for sure, but I think probably, I mean, it seems rational to me because, like, that's the thing about science is, like, it's continual progression. We move forward with new tools and collaboration. You know, the fact that we have tools to be able to better design experiments and those. I would imagine those crystallography methods have improved. So. Yeah, I think it's. It's an interesting question.
Chris Anderson:I learned, by the way, how to crystallize proteins. Have you done crystallography before? Have you. Have you crystallized?
Nick Edwards:Nope.
Chris Anderson:It's. It's so cool. I mean, first of all, it's crazy hard. And, you know, and there's.
Nick Edwards:I hear it's an art form. Yeah, it's an art.
Chris Anderson:I'm gonna get this a little bit wrong. So, you know, the scientists and the listeners will. Will correct me. But, you know, what they were explained to me is that. Is that basically, you know that you've got these like flash freezers and you know that there's. I think it was a Nobel Prize for, like, if you combine. So it's not ethanol, maybe it's methanol or something like that, you know. You know, at the right temperature, the right pressure, it sort of instantly freezes, you know, and so the water molecules don't destroy the protein, but there's lots of knobs and dials in sort of like the right. The recipe that does that and every protein has a different recipe. And you know, God willing, one of them works and it's like, boom, that's the recipe, you know, but imagine doing that by hand, which they did for decades.
Nick Edwards:I think that's pretty representative of basically all of science. My expertise was in electrophysiology and neurons and looking at circuitry and very much an art form, like, dependent on individual hands and very specific methods. Yeah, I mean, this is the classic.
Chris Anderson:Thing in biology is like, it worked on Tuesday, but it didn't work on Thursday.
Nick Edwards:Yeah, well, that's, I guess that's one of the big questions I have about, like, when we think about, you know, robotics in the lab and if you make the comparison to manufacturing, which is a world that you've lived. I mean, by the way, for listeners, Chris wrote a book called Makers, which has great reviews. Check it out. But talking about the kind of the future of manufacturing back in what, 2012 maybe.
Chris Anderson:Yeah, I think that was, it was probably about then. It was kind of like inspired a little bit. My experience in sort of becoming a, you know, a drone manufacturer. I had no business. I mean, as I was a magazine editor and I was messing around with Lego with my kids on the dining room table, and it's like, whoa, that looks like it could be a robot and then it could be a flying robot and oh my goodness, that was a flying robot. Anyway, going from that to being America's biggest, you know, drone manufacturer was a crash course in manufacturing. And what I realized about it is that anybody can do it. We were just buying used machines on ebay and figuring out how to use them. And it's just like the Chinese turned out to better at it than were, but. But still, it was, there was definitely a democratization moment. And 3D printing was part of that. And open source software.
Nick Edwards:That's one of the questions I have is like, is science as modular as manufacturing? So I'm curious what your perspective is, because last time we talked, I think you said you visited around like a hundred labs. What have you learned across that process about, you know.
Chris Anderson:Yeah, right now there's superficial similarities. You're definitely taking bits and turning them into atoms. There's definitely room for automation in both sectors. So let's take the three sort of major domains of science that at least that we're talking about, which is there's chemistry, there's material science, and then there's. And there's biology. Biology has the advantage of being basically room temperature, basically water based so you can do it pretty safely with regular folks, but we really don't have any. We don't have a fundamental model of what's going on there. It's incredibly complex. It's complex. It's probably chaotic and emergent and we're kind of groping in the dark. So the good news is that we can do experiments quite easily with automation. The bad news is that we kind of have to do experiments because we don't have a theory that will really guide us. I was just reading a book called How Life Works by Phil Ball, a former editor I worked with at Nature. And, you know, one of his points is that the central dogma has been basically just doesn't apply anymore. We thought, we thought it was like, yeah, we got a theory. It's the central dogma. DNA, rna, which creates proteins. Done. It's like, yeah, the fallacy of the genome project, it was like, DNA is destiny. We're done. Central dogma. That's just not true. It's just way more complex at every layer on. And all immersion. So the good news about biology is that you can test it, you know, pretty safely in a normal environment. The bad news is that you basically have to, because, you know, you can't just write a whiteboard theory and expect it to work. Chemistry, periodic tables, pretty well understood. Chemistry is better. And we, and we do. We can't just sort of apply Schrodinger's wave equation at scale. And of course that's computationally irreducible and impossible to do. But at least theoretically, we could sort of predict things. Computationally, it's too hard. We can't do it. You know, at least theory is pretty solid. Material science, you know, kind of likewise. It's basically atoms interacting, there's certain wave functions, et cetera. We start with the solid theory, but we just don't have an ability to extend it to the macro phase of actual material science. The problem, of course, is with chemistry, you're dealing with high temperatures and caustic materials. With material science, we're dealing with powders or dealing with kilns. If you look at sort of the chaos of room temp, the search for room temperature superconductivity, which typically a ceramic. What'll happen is you remember this. What happened is someone says, hey, I mixed these weird powders under these weird conditions and stuck it in the oven and it came out and I hit it with a hammer. And this one little flip over there kind of levitated. Boom. And it's like replicate that and which they can't because it's just so freaking, like unique and one off and, and not properly documented. And it's just freaking chaos. And so, you know, all of science can benefit from better structured, more automated, better instrumented experiments, but all in different ways.
Nick Edwards:That's interesting. It reminds me, actually, my grandpa was. He had a friend that in the, I think the 80s, he had made a discovery that was when went big and everybody kind of knew about it and it was cold fusion.
Chris Anderson:Oh, wait, who was your uncle?
Nick Edwards:So. Oh, this is my. This is my grandpa's friend that went to. I would go to lunch with him when he was older one time, but it ended up. It ended up not being replicable. And he had a lab in his basement, still in his old age, in his 80s, where he was trying to reproduce this experiment at home after he had retired as a professor. And I was just like, that's the, that's the challenge is like. I mean, I don't know what the story is behind it, but I remember hearing the story when I was young. This is before I even went into science, before I even. It was even a twinkling in my eye. That story's always stuck with me. And you know, I. Because I've had things that I remember seeing in the lab and like, it was reproducible and it worked for me over and over for a while and it just went away. And like, I never published it because, like, I couldn't find it again. It's an interesting, like, tidbit of science that like, is somewhat representative of the challenge of doing research. Right. You've seen a lot of things. Like you were a big part of like, the Internet moment when you have been part of this manufacturing push. Like, how do you. How are you spotting all these important waves? How are you seeing through the trees?
Chris Anderson:Yeah, you know, I mean, a lot of it's luck. Just right place, right time.
Nick Edwards:Okay.
Chris Anderson:That sort of thing. I didn't spot the Internet as being. Being a big deal. I just happened to be in the physics world at the time when the Internet was being used to connect physics labs. You know, in that particular case, I didn't even realize what I was sitting on until this magazine called Wired shows up in 1993. And there's just like all day Glow and Super Sci Fi and. And it was just like, the Internet's gonna. Actually initially, they, it was even pre Internet and they were like, the information Superhighway is going to change everything. And, you know, it's going to be the end of the nation state. It was going to be a cultural revolution. I was just like, I mean, that thing I, that nerdy thing I use every day in the lab is going to like change the world. So I would at least, you know, in that case I credit, you know, Wired for illuminating the potential. But, you know, once I saw it, I realized that they were right. It was like, well, I got to do it. And so I think that, you know, being in the right place at the right time is one important thing. Being kind of optimistic and sort of curious and open to new ideas and sort of, and being willing to sort of like, you know, get sort of excessively enthusiastic about them is also key. And then I think, you know, the most important one is being sort of polymathic, sort of, you know, it's an extension of being curious. But basically sort of having, you know, sort of being interested in everything means that you're able to spot connections across disciplines. And often some of the most transformative moments have been interdisciplinary moments, like what I'm doing right now with like the AI scientist moment. You've got scientists, you've got AI people, and then you've got the lab automation people who are validating it. It's very hard, very rare to find a lab which is strong in all three. Yeah. And so it's a great moment for people who may feel that they're not, you know, I'm not of enough of a scientist to be useful in a lab, or I'm not enough of an engineer to be useful. It's one of these moments where it's just like, it's okay to bring your own skill set to the mix because we need all of them. It's quite liberating in many ways. And it's a good time to have like, you have permission to enter areas that normally would have required a PhD in something you have permission to enter because you're adding value in your own way. And this is why I can. I call myself a value added tourist because I visit labs and they're kind enough to show me how they do their science. But, you know, what I can bring to bear is like, well, you know, the engineering mindset. And I was like, oh, here's how you could do the automation, or here's how you could do the cameras better. Here's how you could sort of handle your data, you know, better not, because I know a lot about that particular science. But because you take a tool from an adjacent industry and you import it. And like, one of the things I did recently is I was helping out with a lab that said DNA barcoding insects, where we have a huge backlog of bags of bugs in California in particular, that it's millions and millions of bugs that were sort of bagged, you know, collected, stuck in ethanol and bagged and stuck in the refrigerator. They don't really know what they, you know, what they got. And there's no real census of biodiversity of California. So the best way to do it is DNA barcoding. But, you know, it's a real, I mean, these are bags of bugs and there's like hundreds in each bag. And so you have to take, you know, take them out. You have to put them under a microscope with an old tweezer. So these are parasitic wasps that it's like the size like a period of the tiny things under a microscope. You take them out with tweezers. You take them out and you stick them in a well. You then put in the enzymes to digest the root of soft tissues. And then ultimately. And they tried to automate him with like robot tweezers. I walked in and I was like. Of my industry, electronics, you know how we pick up like tiny little things and move them. We have this thing called pick and place machine. And it's electronics you take up. You have spools and tiny little electronic components. You pick them up with a little vacuum, with a little nozzle, picks them up, looks at it with the camera and sticks it very precisely on a prison circuit board. Yeah. And I was like, I think we could just pick up bugs instead of like electronics. And by the way, I think that a standard office, you know, open source pick and place machines cost less than $2,000. They are open source, so you can modify them. And I was like, I got one. Just so happened I had one that I could borrow, brought it in, literally. Codecs, Claude, code, whatever. Like, like two hours of like write a script that will not do electronics, but instead pick up bugs, look at them under a camera and stick them in a. Well, two hours of codecs and the machine basically works. And we've now we could potentially take an order of magnitude out of the cost of DNA barcoding insects by simply utilizing a off the shelf industrial machine from an adjacent industry that no biologist would have thought of. But any electronics guy automatically knows it was just the ability to sort of like import best practice from Adjacent industry and sort of say, look, this is. You don't want a robot tweezer. What you want is a little vacuum with a nozzle and those things you can buy them. It's like, who knew?
Nick Edwards:That's. I love that. This is Biokia lab.
Chris Anderson:Yeah, yeah. Out of Berkeley. Out of Berkeley. And it was a classic case where they. I said, can I visit your lab? I visited their lab. They showed me refrigerators full of bugs, and they showed me the problem tweezing them into wells. And I was like, looks like a pick and place machine to me.
Nick Edwards:And it works very cool. So if you're thinking about then, like, all these kind of experiences that you've had in these different. Well, first of all, why. Why DIY labs? Like, what. What is it about AI science, this moment in AI scientists like science that's driving you toward this?
Chris Anderson:Well, I mean, the DIY is simply. It's just code for democratizing, you know, just. Just make it. Make it. Yeah.
Nick Edwards:Lower the barrier. Answer.
Chris Anderson:So more people can participate. And we've had DIY bio for a long time, and it's. It. It's struggled for reasons. Because biology is hard. I think that, you know, in general right now, because it is so interdisciplinary and because you people have. Who have skill sets in AI, but not necessarily, you know, biology and automation, but not necessarily. It's the right time to just make it easier to participate, build community, open source. The tools, make them cheap. Your company's entirely about making them easier to use. And anytime you can lower the barrier, you expand the participation. And when you expand the participation, you get not only just get more labor, but also new ideas. And this is like the history of, like, at least the last 20 or 30 years. Certainly the history of the Internet is about democratizing. Just make it easy to make video, make it easy to do media, make it easy to, you know, share code. Just make it easier. And I think it's science's moment.
Nick Edwards:Yeah, I love it. Tell me what you're seeing in terms of, like, where the future is going with lab automation. What does the lab of the future look like to you? How are you thinking about it?
Chris Anderson:Yeah, once I started to realize that lab automation involved automation, which is something I'm good at, but I needed to kind of baseline myself on kind of where the. What the gold standard is in the moment. I basically visited a lot of labs and I went to SLAs, the, you know, the big conference. I think you and I, you know, I met at Slash.
Nick Edwards:Yes.
Chris Anderson:And I want to just kind of like say, where is the industry today? And the industry today struck me as being mature. It sort of had a kind of a 20th century conceit, which is largely about sort of the lab bench and the human scale stuff by and large. And it was very mature. The enclosures were nice, the touch panels, you know, were nice. They were sold, the pricing was like in you know, 50 to 100K and machines were made in units of like hundreds. Maybe the number one liquid handler sells a couple thousand per year. In other words, it looks like a mature industry. However, none of it looked like it had been truly transformed by AI in a couple ways. I mean, first of all, a lot of them had proprietary interfaces, so they weren't easy to automate. This is obviously the core of your business. They were also human scale, which is to say they were kind of boxes that would sit on a bench. And they were designed, they had doors that need to be opened by humans and sort of buttons that need to be pushed by humans. And the notion is that these machines, automated machines, work in proximity with humans. And so it's kind of a hybrid lab where you have humans doing some jobs and machines doing others. Imagine if a data center worked that way, where you basically sort of said, well you know, we have all these laptops and they're on benches and you know, and then when technicians come in and sort of start programs on laptops. You imagine if a data center worked that way as opposed to the way a data center really works, which is like it's a lights out, everything is racked super high, the board, you know, everything. All the computers are stripped down to their bare motherboard essentials. It's just optimized for scale. And a data center will have hundreds of thousands or millions of, you know, CPUs in there. What would you know, a self driving lab or an automated lab look like if you really put on data center, a hyperscaler kind of conceit and you know, it's clear that you have a human scale science and then you have machine scale science. And machine scale science wouldn't have enclosures and wouldn't have touch screens and they'd be designed to be racked really high. I think we're seeing glimpses of it like the ginkgo racks for example, still sort of human scale, but at least there's no humans involved as a robots and take, move things around from machines to machines and open doors when necessary and stick things on little trolleys that go onto the next one, et cetera. But even that's too expensive. What if like every experiment had a bespoke? Not all experiments suit, you know, microfluidics, but let's imagine they did and that, you know, you said, I want to do this experiment. The experiment's gonna involve certain inputs and certain outputs. There are certain reagents that go in, there's certain transformations that happen, and we're just gonna create a bespoke chip or, you know, it could be 3D printed, it could be, you know, PCBFAB or something like that. But as they got a single purpose experiment that does like one thing properly. If it's a cartridge and it fits into, you know, into a standard slot and the standard tubes come in, come out and it's like this big, not that big. I realize that doesn't suit everything and obviously microfluidics and dili fluidics are a thing right now. But you can sort of see that to keep up with the AI opportunity, we're going to have to sort of rethink, you know, rethink the kind of human centric approach we right now have to lab automation and go for things that are truly designed to be done, lights out, you know, no human involvement. And that's going to involve shrinking things to the bare minimum. It's also going to involve bringing in sort of gold standard automation from elsewhere. Robot arms, humanoids, dogs, you name it. I mean, there's, you know, obviously the robotics industry is on fire right now and there's a ton of iteration happening there. Companies like Medra are focusing on robot arms on the grounds that they have the ability to automate potentially every step of the life cycle, starting from the lab bench in basic biology, all the way to, like, animals and plants and then even, maybe even to the clinic and beyond. Because A, they're incredibly flexible and they have lots of degrees of freedom and they can pick up things and all this stuff. And B, that you're surfing on the wave of all the investment that's going into general purpose robotics. Right now. Right now I'm slightly on Team Robot army only because I think, you know, to the pattern matching point you made earlier, you want to surf waves. And it feels to me that's the biggest wave right now. And although the robot arms are not as good right now in terms of this throughput, as a bespoke, as a single purpose liquid handler, they are getting better, faster. And you have the ability to sort of solve the inter device problem, moving things from one place to the other without having to invest in Expensive trolleys and, you know, maglev tracks and things.
Nick Edwards:Like that in some of the labs that you've been visiting. I mean, do you see people starting to implement robotic arms? I mean, I guess, like, Medra is a great example. They've got, I think, a hundred workstations or work cells that they've. They've stood up in a really short amount of time. So it's pretty cool to see. I think Ginkgo is a really interesting example from a commercial perspective. But are you seeing people use, like, starting to build these things in like a DIY setting?
Chris Anderson:Like smaller ones? Just a bit. I mean, the learning curve is still higher than it. Than it should be. I mean, I think at the clear, at the DIY setting. And I think you and I were talking about this really inspiring thread on X the other day. This guy who kind of refined an Alzheimer's drug using an Opentrons in his garage. So I think for the DIY stuff, just like, there's a lot of complexity. Let's make it simple. Just like use a liquid handler. It could be like either an off the shelf one, like an Opentrons, or you could repurpose the 3D printer. Then you need to close the loop with computer vision. And like, one of the things I'm working on with the lab right now is just like, in my world of robotics and manufacturing, everything has fiducials. Fiducials are just like April tags or little sort of machine readable, you know, like a QR code. The nice thing about a fiducial is very easy for computers to read and it tells you the orientation of the plate. Everything should have fiducials. Everything in the lab should have fiducials. And once you have fiducials, then the robots are like, oh, I can see okay. You know, and I can see where the plate is. I can see what orientation is. I can see his offset. I know how to pick it up. You can have stickers. You can like, literally, I don't know why labs don't have like sheets and sheets of stickers with these little QR tags or QR codes or April tags and just fiducial everything. Put fiducials everywhere.
Nick Edwards:Yeah.
Chris Anderson:And what you've done is you. Basically, it's training wheels for the robots. Now. They look around and it's just like, whoa, what is that? I've never seen this sort of, oh, that's an April tag. Not only do I. Do I know the tag, do I recognize it as a tag. I recognize its orientation. And because the tag has a unique id, I know what it represents. It's like, oh, that's a test tube. Oh, and that this kind of test tube is most easily sort of unscrewed with this technique that I've already learned. So basically there's going to be an initial element where we kind of have to. If you want robotics to come in, we have to kind of prep the lab to make it easier for the robots to kind of find the way around. But over time, they're going to learn and there's things called visual language, action models. And over time, just like if you drive a Tesla, all of the weird edge cases that you see all become training data for Tesla, the company. And this and the next model will incorporate that. And they just get smarter over time. You want a robot to be able to do one thing well and that gives it, like, license to get into the lab. You have permission to come into the lab because you can do this one job well. But once it's in the lab, then, you know, like, okay, you know, foot in the door now you start to say, well, let's just like, what else could you do? And when you start putting traditionals around and sort of prepping the lab for, you know, for robotics, then suddenly it gets a lot easier to do it.
Nick Edwards:Yeah, I think high res is doing some really interesting things around that, where they're tagging all the instruments in the, you know, the broader work cells and basically it generates kind of a, a digital twin of the lab and then their agents can reason across that for experimental planning. Very exciting, really cool stuff.
Chris Anderson:And it's just so easy. It's like literally sheets of stickers.
Nick Edwards:That's the exciting part. That's the challenge also. Right. Is like, we're serving an industry where the workflows are so variable. It's like, you know, R and D by definition is like you're changing things constantly. And so how do you allow for that flexibility? And I think it's exciting times. So you're team Humanoid. It sounds like team robot arm.
Chris Anderson:Well, I'm probably team robot army. I played with all of them. The humanoids seem like overkill for the lab. I mean, you know, great humanoid has got legs and arms. If all you need to do is move the arms from place A, you know, place A to place B and stick them on wheels also, I mean, in some situations, the dogs are actually, robot dogs are a better choice than the humanoids because there's stable platforms. I got four Points of contact. If you ask me, like what My sort of MVP of automation is. It's a robot arm. A 7 DOF, 7 degrees of freedom robot arm on a rail. So that adds, let's say it's 6 degrees plus the rail, or maybe it's 7 degrees plus the rail. So it's 8. That is the sweet spot right now. It's super easy to program. It has the ability as long as the rail is. So it could be, you know, 10, 12ft, you know, moving it, moving up and down. It has the reach on both sides of the rail. The robot, honestly, the ones I write I like a lot right now are the seed Studio Rebot, which is about $1,500, can lift 5 kg. Wow. Incredible repeatability. Comes like ready to go out of the box. Actually out of the box one's $2,000, but you know, close enough. They program with the hugging face, you know, stack, which is all imitation learning, which is really easy to use, just really solidly made. And just to think that, you know, for like 1500 bucks you can basically get a pro quality arm that is designed for AI that was not the case 10 years ago.
Nick Edwards:That's crazy. It's totally crazy. In this future, like, how does the future look bright? How does it look bright for the individual scientists? If we're automating a lot of work, like, what is the role of researchers? What do you think? Yeah, I mean, that's a really good question.
Chris Anderson:Let's divide science. And again, this is kind of, maybe I should ask you this question. So I'm going to give you kind of a half baked answer, but I really want to hear your fully baked answer.
Nick Edwards:Sure.
Chris Anderson:Let's divide science into sort of two categories. There's kind of novel discovery and then there's sort of like optimization, you know, so novel discovery would be like, I've got some crazy hypothesis and let's see if it works. It works, right? And you know, in that case, I think automation is probably a nice to have, but because it's by definition novel, these are protocols that have not been done before, maybe involving certain recipes or ingredients and things that have not been put together. Again, that probably for the novel side, that's probably pretty good on, I think AI is going to be quite useful for lit search. So like Edison and the Spin out of Future House is really good at sort of, you know, crisping up your hypothesis. It's like, okay, I've done the lit search, here's what I think about it. So just kind of like sanity Checking your hypothesis or maybe helping you generate better hypotheses. But that's kind of purely on the AI side. It doesn't involve a lot of automation once you get then to the sort of like the process improvement. So it's like here I've got some enzyme that has a certain yield or actually what's the, how do we measure the success? The utility enzyme. Yeah, activity. Yeah, an enzyme has a certain activity and if we can improve the activity, that has a huge industrial potential. That's an example. And you see this a lot in chemistry and material sciences and a lot of industrial processes where you have Pareto frontier, which basically we found a recipe that works and you know, you turn the knobs and dials and this thing is like 40% yield. There's probably some recipe that works better, but this is like good enough. And who has the time to explore all the various other permutations of knobs and dials? Well, AI has AI and automation has time.
Nick Edwards:Yeah.
Chris Anderson:And so we saw this. The ginkgo experiment with OpenAI was a, I think it was a protein synthesis, you know, experiment. And they basically just sort of like, you know what, we're just going to sweep across the domain range that's like 10,000 different variations of knobs and dials, of settings, parameters, hyperparameters, whatever. Let's just run through them all, or if not all of them, let's at least sort of like here's the Pareto frontier and let's do a couple samples here and there. Ooh, yeah, this one seems like a little bit promising there. Maybe there's a gradient. Let's explore in that direction. And that sort of classic exploration exploitation that AI is so good at is the way that we're going to sort of get better recipes with higher efficiency processes without a huge amount of labor. And that's your classic. That's groping the dark. That doesn't sound quite right. But let Ali sort of wander through the forest in the, in the dark on our behalf, such that you can sort of find a better recipe and that's that idea. So that you may not call that science, that's maybe closer to engineering, but there's a lot of industrial utility there.
Nick Edwards:Yeah, I think that's a really good point. That's where the strengths are, as in processes and optimization. And I personally think that like human intelligence and machine intelligence are fundamentally different. We're creating a world of collaboration where each individual person has, you know, an area of expertise that they've got a synthesis of a certain body of information, but they don't have much expanse outside of that. And it's really hard to understand adjacent areas. And so the agents help us to better be informed of like different areas that are close to that niche, you know, even much further out, generate better hypotheses. But there's always going to be a taste component, I think, of people, especially a new discovery like you're talking about, where it's like getting to a new novel concept. I think we're getting into a world where every individual scientist has like so much more leverage to explore and to like find new threads to pull on and then the AI will help pull on that thread and unravel kind of the tapestry in a way. I got, I don't know if that's a great analogy, but have you used.
Chris Anderson:Any of the kind of AI lit search tools, whether it's, you know, Edison or Cosmos or the. What do you think?
Nick Edwards:I think it's, I think they're interesting. I think like perplexity. I've, I've used a decent amount and we built some. And I think that what happens is they come to interesting ideas, they compete and they, you know, the agents are kind of like using this competitive process to compare ideas against each other. And in the end it comes out a lot of times with interesting hypotheses, interesting concepts, but it's really hard to be able to like analyze all the exhaust that comes off of that. How do you know like that it's making the right logical conclusions. And so that's where I think the challenge comes. It's still a little bit hard to kind of fully trust the system that it's getting the right conclusions and you.
Chris Anderson:Can't just ensemble them. So basically have one AI fact check the other AI.
Nick Edwards:You can, you can, and it does a pretty good job. And so I got to want to downplay, I think that like super valuable and exciting. I think just also fundamentally one of the challenges is that the literature is incredibly flawed. You know, there's a lot of assumptions that are coming out. And so that's one of the reasons why we as a company focus a lot more on kind of methods because they're pretty verifiable. Interesting.
Chris Anderson:You know, computer science doesn't have this problem because it's by definition verifiable. You just, you know, run the code.
Nick Edwards:Yeah, yeah, exactly.
Chris Anderson:And the more biology can start to look like computer science and that is replicatable. And, you know, the methods are fully described in that. Like, it compiles.
Nick Edwards:Yeah.
Chris Anderson:For lack of a better word, the happier we'll all be.
Nick Edwards:For sure. I want to be mindful of your time. This has been really fun. Is there anything that we didn't talk about that you wanted to mention? Anything that you want to.
Chris Anderson:No, no, no. This has been a lot of fun.
Nick Edwards:Thank you. I really appreciate it. It's been a fun conversation. Well, I think it'll be really interesting to have this conversation and to. And to further it in a year from now. Like, where's the world going to go? There's so many interesting options. Let's see. I'm excited to see the future, man. Thanks so much for coming on the podcast. And, you know, it's really. It was really fun to kind of shoot around some ideas with you. I think this is probably the most exploratory episode I've done in a little while.
Chris Anderson:Wide ranging.
Nick Edwards:Yes.
Chris Anderson:You know, it's a mile wide inch deep, but no, it's perfect.
Nick Edwards:It's perfect. It's like, this is exactly why you. Why it's, like, good to have different people come in with. With area, different areas of expertise. Right. Because, like, we're all exploring this future together. We're trying to figure it out. And, like, I don't know. I would say I. I know as much as many people, but, like, we have to. We have to, like, build this conversation and we have to.
Chris Anderson:Absolutely.
Nick Edwards:And we have to build it together. No shame in.
Chris Anderson:In confessing to ignorance in some part of the domain because there's just so many of them.
Nick Edwards:Exactly. All right, thanks, Chris. This is the Once the Scientist podcast. I'm Nick Edwards. We're back with new episodes, so keep an eye out and subscribe to the podcast if you haven't already.