State Space

State Space

Download on the App Store

State Space episodes

  • EP10: Making the Physical World Observable with Mike Borg (Index Bio/Authentica)

    Zhane Yamin: Welcome to State Space, where we illustrate some of most important and timely ideas in the science acceleration landscape. In this episode, we’re talking about a part of the conversation you don’t hear a lot in AI for science circles. Everyone talks about how the model isn’t the bottleneck, the data is, and we need more data, but sometimes the reason the data isn’t there is that no one has built the instrument yet. We’ll be talking to someone building not just one of those instruments, but something that could fundamentally change the game in domain spanning food safety and quality assurance and epidemic modeling.

    My guest today is Mike Borg. Mike is a software engineer by training who fell sideways into molecular biology, and he now runs two companies that came out of that collision. One of them, Index Bio, based in Toronto, invented something called the Biotag, a microscopic food-safe barcode made from inactive baker’s yeast that lets food safety teams simulate contamination without actually contaminating anything.

    One of the upshots of biotags is that they generate a lot of data, and as we know, data is important for AI use cases. I think Mike’s work is a really interesting case study in that. So I’m excited to have him on the show. And Mike, welcome to State Space.

    Mike Borg: Thanks for having me. Excited to be here.

    Zhane Yamin: Yeah, excited to have you. So to set the table a little bit, I just wanted to talk about how people with their AI and their data brains coming into this conversation can understand the collision of AI and data and food safety, quality assurance, that sort of stuff, because I think a lot of people that isn’t necessarily an intuitive combination to them. So if we could walk back, what is that problem that biotags were trying to solve? And is it related to AI, is it related to data, or is it something else entirely?

    Mike Borg: The food sector and food safety is our wedge commercially, but the same problem-solution pair applies to lots of other fields and domains: pharmaceuticals, agriculture, natural products, beauty, health, et cetera. Across all of these, these quality and food safety people care most about the thing that they least want to observe. And that tends to be contamination risk, things like pathogens or allergens. You’ve got your salmonellas, Cronobacter, Listeria, things you hear a lot in the news these days.

    And it begs the question, how do you learn about catastrophic events if you’re unable or you can’t afford to actually learn from catastrophic events? This is the exact problem that we’ve really focused on with Index and the biotag technology.

    Zhane Yamin: So just on to the AI point. How is AI currently being talked about in the industry? Is it necessarily being used? Is it being used usefully? And how do you see that as it differs from your vision on how it could be used if it was maximized its potential?

    Mike Borg: What you hear people saying right now is they want to move from reactive to predictive. This is very popular within the food sector. The challenge is, a better model does not fix the missing data problem. And you might have, for example, a food manufacturer who could go several years without a contamination event.

    So this is a very long-tail kind of data problem, and it’s also not something they want to routinely measure. So if you don’t have the data, you have to find a way to create it, and the traditional synthetic data approach doesn’t really work because you don’t have enough observations of the actual data. And every manufacturing facility is a snowflake. So you have to create a biological signal that you can leverage to prove your ability to control risk within all of these scenarios: food manufacturing, farming, supply chain, et cetera.

    Zhane Yamin: Yeah. So do you see that as something that generalizes to a lot of other different domains too, where people are bolting a model or something on very bad data instead of moving toward the instrumentation point as well? Do you see that as something that generalizes to other domains?

    Mike Borg: Yeah, absolutely. I think especially within any product category that touches a human, you can leverage something like the biotag to glean the sorts of observations and insights that you’d want in order to reduce risk around manufacturing. There’s interesting use cases beyond that, obviously. Right now, the big conversation within AI is AI safety and alignment.

    You could red team, right? And sometimes it works if they’re containerized correctly; other times it’s a little bit more challenging. You could think about what we do as a sort of penetration testing for food manufacturing. And that generalizes quite well to all manufacturers.

    Zhane Yamin: So to break that down a little bit, if I’m a Hill staffer with my background knowledge, what is a biotag? How does it work? I hear that it uses baker’s yeast. How does it use baker’s yeast? How exactly does it work?

    Mike Borg: So if you’ve ever purchased nutritional yeast from the grocery store, some people put it on their popcorn, that’s effectively the product. Except the difference between nutritional yeast and the biotag is that we bioengineer each batch of biotag that we create with a unique, harmless DNA sequence that allows it to be uniquely detected within any matrix, within any food product. And what you can then do with that is build models, understand the movement of materials, correlate with pathogens.

    What would a pathogen do in this scenario? You’re processing grain, let’s say, and you’re turning it into flour. If a pathogen were there, could you use the biotag to function as a sort of proxy for that pathogen? And the answer is yes. So you could actually leverage the biotag as a proxy for these contamination risks.

    Zhane Yamin: So a question on the proxy point: how much can it generalize? Is there not some diffusion dynamics? Is it that the particles are different? Does that make a difference in how well you can proxy your model these things? Or is it the yeast generalizes very well? Or is that part of the engineering is where you’re engineering it to be a good proxy for X, Y, and Z?

    Mike Borg: So the yeast, the biotag, is inactive. It’s not living. So it’s non-viable, whereas most pathogens are. So the challenge becomes, as you’re posing, can it actually function as a good proxy? And the answer is yes.

    Most of these manufacturing facilities are closed systems. And to a point, the biotag just goes along for the ride with the pathogen. One of the interesting scenarios that you would wanna consider is, do you have a biofilm somewhere that tends to persist over long periods of time, not just that run? The biotag has been shown to actually stay within those biofilms and be released over time.

    So you can effectively correlate pretty well. It’s not always linear, but it’s predictable and consistent. And that’s all that matters. This is the common misconception: it doesn’t need to be exactly like the pathogen. It just needs to be predictable and consistent relative to the pathogen.

    Zhane Yamin: And forgive me if this is a very simple question, but are you spraying it on there? And then also, how are you reading out the DNA sequences or the yeast particles or whatever they may be?

    Mike Borg: Yeah, so it could be mixed in dry, can mix it in with another ingredient, you can spray it on. Often we’ll mix it in with water and we’ll include it alongside some pre-existing application of an ingredient or solution. So it tends to just be piggybacking on with something else, right? An input, or in some cases, a solution for cleaning and sanitation, et cetera.

    Zhane Yamin: So once you had all of that data being generated, was that the only bottleneck to unlocking AI capabilities in this sort of domain, or modeling in this sort of way? Or were there other bottlenecks that you yourself and your team had to try to overcome in order to really maximize the AI use case?

    Mike Borg: So once we’ve mixed it in with whatever and we’…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    23 min
  • EP9: How Project ProtoStar Puts Public Research to Work with Emma Scharfmann (Hugging Face)

    Zhane Yamin: Every year, the federal government spends almost $200 billion on research and development. It used to be that that data wasn’t open, and now, more and more, it is. But open is not the same as usable. A data set buried in a supplementary file, in a format nobody’s used for decades, with no documentation on where it came from and how it was made, is technically public and practically lost.

    That’s a problem in our new AI native paradigm, where the downstream positive effects of research depend more and more on whether a machine can find it, read it, and combine it with everything else. Just a few days ago, Hugging Face launched Project Protostar, a bet that the open source AI community, which has spent years learning how to share models and data at scale, can help science do the same. Emma Scharfmann is a machine learning research engineer at Hugging Face, working directly in their AI for science portfolio.

    She’s helped to stand up the project as well as set up its first challenge, which launched on Monday. Emma, welcome to the show.

    Emma Scharfmann: Thank you. Really glad to be here.

    Zhane Yamin: Of course. So data is a bit of a recurring theme on the show. We’ve talked a lot about open science as well. A few episodes back, we had Praty Avasthi, who’s leading open science at Astera Institute, and we often talk about machine readable data.

    But I’m interested to hear your perspective as someone who has kind of more experience on the software side of things. I guess the first question would be: what is the problem that Protostar is trying to solve?

    Emma Scharfmann: Sure. So in general, I mean, especially these days, there’s more and more data which are available depending on the projects. Even in science, the journals, the funding institutes and the universities, they try to encourage people to release their data. But very often, as you said, these data are not really usable.

    It could be for different reasons. The first reasons that come to my mind are that there’s some technical friction. Formats might be hard to use, hard to read, could depend on some different software. Also, especially these days, more and more huge data sets require specific resources, which is also a big reason why, even if data sets are available and accessible, people can’t really use it.

    There’s also the whole problem of lacking documentation. When you have a huge data set but you don’t know what’s the number in it, if you don’t have the units, if you don’t have the metadata, you don’t know what the columns correspond to, of course it’s still open source, but it’s not really usable. And finally, especially in research papers that could be hard to read for people, hard to understand all the details, you often lack the context of the data.

    And if you don’t have a proper documentation which explain the limits or the bias of the data and the scope of the study, it makes it really hard to use for people. And I think that’s the main reason why there might be a lot of friction when using such data sets.

    Zhane Yamin: So I’m interested just to kind of hone in on a specific example. So what is the sort of failure mode that you would say that on the software side of things you guys see the most? And what is the sort of negative outcome of that failure mode that’s most typical?

    Emma Scharfmann: So I don’t know if it’s the one that comes the most often, but the one that I see more and more is that people use agents, more and more AI, to build their research projects or build on top of releases which are open source. When you do that, there’s one part of the process that people were doing before, such as reading and understanding what was missing. And if you see something missing, reaching out to the authors that now disappeared.

    When you ask an agent to work on something, if there’s a piece missing, often the AI may just assume and just fill the gap that would have been solved otherwise before. And so when these gaps are missing from the documentation before, even if it could make you lose time, it was not such a problem for the outcome of the project. Well now it can really be problematic. So you might have wrong result or result that only holds in one situation but not others. So yeah, that’s something I see these days.

    Zhane Yamin: Got it. And so I guess another question, maybe as an aside, because I assume we’ll get into this a little bit more later, but where does institutional policy fit in this? ‘Cause it seems like, at least from where I’m standing, a lot of people are very pro keeping things open and everybody always talks about making things machine readable.

    But is the problem that everybody’s encouraging it, but nobody’s really required to have it open and machine readable? And that’s what we’re trying to address here. Where does institutional policy in your guys’ mind fit into this conversation?

    Emma Scharfmann: So I think ideally, everybody would like to have everything easy to use and open source. But at the end, while researchers are encouraged to do it, they don’t really get recognized for it, they don’t really get credit out of it. Which I think personally is one of the reasons why it’s hard to do it, it’s time consuming and you don’t get anything in exchange.

    And that’s also where these projects stand. So we are working with different institutions, different universities, to start by taking all these data sets that were publicly funded and try to do something with it. What we try to do first is to make it just easier to use and then to create situations.

    So these challenges that we just launched this last week, a few days ago, to create use cases where people can use these data sets. So it’s really the two parts of the project: helping making it easier to use and just encouraging people to use it.

    Zhane Yamin: I guess if you could talk a little bit more about Project Protostar and where the idea came from. I know we’ve already talked about the problem a little bit and you just gave a little bit of an intro to it, but what is it in a nutshell and how is it addressing this problem?

    Emma Scharfmann: Yeah, so the idea of Project Protostar is to bring together different institutions and, with the help of everybody, try to make data that were publicly funded and that already exists more accessible, easier to use and to encourage applications on top from these data sets. So the idea is really to bring together a lot of people, so institutions and industry and also the open source community and some infrastructure, and from this collaboration try to make something out of these data sets. So we have different directions. Hugging Face is helping especially in the infrastructure, also that on the hub we have infrastructure that’s quite ready for that.

    So we’re trying to take these data sets and change the data set into a format that’s easier to read with proper documentation, proper README cards, and we also have the dataset viewer where we can see what’s the contents of the data set just very, very easily. So yeah, the idea is really to fill the gap between being open source and being usable. And the second part of it is to create situations where people can just build on top of these data sets.

    So that’s why we’re launching different challenges. We launched our first challenge last Thursday, so less than a week ago. And the idea is to take one of these data sets that come from a research paper and to create a challenge out of it to see what we can do and how we can just create the best out of it.

    Zhane Yamin: So I’m interested to hear also your perspective on how this is sort of meeting the AI moment as well, and why does it matter or why does this sort of project come up now rather than five or ten years ago? Is AI sort of raising the stakes? Or could you have seen this project also being sort of important, even if we weren’t in our AI mindsets of “we need data sets” and that sort of stuff?

    And I also wonder what more could we gain by having these data sets open and machine readable? And …



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    23 min
  • EP8: Writing a Biotech Action Plan Congress Can Act On with Caitlin Frazer (NSCEB)

    Zhane Yamin: I am fascinated by the phase shift that happens when a scientific field stops being something that we study and starts becoming something that we build. In the late 19th century, parts of chemistry made that leap, and we got plastics, rubber, and synthetic fibers, and parts of physics made it in the 20th century, and we got transistors, computers, and wireless communication. My guest today is working at the forefront of the next leap. Thanks to AI, gene editing, and automation, and more tools, we’re learning to engineer living systems, to engineer biology the way we once learned to engineer materials and machines.

    That’s a huge opportunity, and it’s also a major front of strategic competition. The National Security Commission on Emerging Biotechnology was charged with finding and recommending how America could keep that lead, and it has been remarkably effective. Twenty-five of its recommendations, in full or in part, have made it into law, and more are seeing major action from the federal government. So in this episode, I wanted to answer two questions.

    What is this biotech moment? And how does it relate to science acceleration? And how is the commission meeting it so effectively?

    There’s no better person, in my opinion, to ask than Caitlin Frazer, the executive director of the National Security Commission on Emerging Biotechnology. She leads the staff that supports the commissioners, and she spent much of her career on the inside of the legislative machine. Caitlin, welcome to the show.

    Caitlin Frazer: Thank you so much for having me.

    Zhane Yamin: Of course. So just to start off with, I was wondering if we could just go in your definition of biotechnology. Because I think for some people it’s vague, for some people they might have different definitions. So when your commission says emerging biotechnology, what are you guys actually talking about?

    Caitlin Frazer: What do we mean? Fair question. It’s actually something that we spent a number of weeks when we first started this project making sure that we made the right scoping and definitional decisions because this project was huge at the outset, and we knew that to get it right, we needed to be very clear with our audience and with ourselves about what we meant. So biotechnology, of course, is a millennia’s old technology.

    Humans have been fermenting beer and wine and cheese and bread for millennia. We’ve been using biotech for food and for medicine for a long time. But emerging biotechnology or modern biotechnology is really about the newer ability to engineer living systems and start to treat biology as a field of design and engineering. What has unlocked that over the last two decades or so is that researchers have figured out how to read, write, and edit DNA.

    And in doing so, we can now program, the living code in the way that we might program a computer. When you apply AI and automation, on top of that, you get an accelerant effect. And so we’re just seeing this enormous potential of biotechnology, which humans have worked with and lived alongside forever, explode into — now we think about biotech as a general purpose technology because it has applications across economic sectors. And we think that that is even poised to grow even more, if we get the policy foundation right.

    Zhane Yamin: So one thing I wanted to dig into a little bit too is why specifically a national security commission is a good fit for this, right? ‘Cause you know, you think about maybe health or science commissions, but why national security? What does that framing sort of mean for you guys?

    Caitlin Frazer: Yeah, absolutely. So the Congress created our commission in an act of law about four years, four and a half years ago now. And we’re actually the third in a series of national security technology commissions. And so there was a history to Congress reaching for bipartisan outside experts to help them think through the convergence of national security concerns with emerging technology issues.

    So the first of our tech commissions was on cybersecurity, then there was one on AI, and now emerging biotech. And I could imagine that there’ll be appetite for more coming. We’ve seen proposals for national security commissions on robotics and fusion and all kinds of other quantum and other technology disciplines. But I think what Congress was really asking for in creating this commission was for our group of outside experts, our commissioners, and our staff to think about the way that US leadership in biotechnology — which has historically been quite strong in many ways, modern or emerging biotechnology is an American invention — to think about that with respect to our national security, our supply chain resilience, our economic prosperity.

    And so one of the other scoping decisions that we made very early was to define national security. Right, so we didn’t have to just define emerging biotechnology, but we had to think about what national security meant to us. Because you could define it fairly narrowly and say, Okay, what are the conventional military applications of biotechnology?

    Like, how should we be thinking about these tools to advance military medicine or trauma care or combat trauma care? How should we be thinking about the ability to biomanufacture reshaping military logistical supply chains. And we did that.

    There’s a whole chapter of our action plan that is explicitly about biotechnology for conventional defense capabilities. But our commissioners wanted to think much more broadly and to define national security to include, like I said, supply chain resilience, food security and agriculture security, energy resilience and energy supply, biotech for critical minerals, and of course, the advancement and reliability of our medical ecosystem as well. So we bit off quite a bit when we defined national security that way, but I think it was an important decision because what it enabled us to do was to share with the Congress all of these myriad applications of biotech that are real and available today, but also where they’re going in the direction of travel over the next five, ten, fifteen years.

    Zhane Yamin: So I’m interested to hear what your perspective is on that national security framing for sort of like the normal person. Your report also cites some amount of companies relying on other international competitors for parts of their manufacturing processes. How does this broader national security framing relate to how people themselves should feel about or might feel it in their lives?

    Caitlin Frazer: Yeah, absolutely. Well, candidly, when we started, one of the first things that my team and I did was to just immerse ourselves into the existing emerging biotechnology, synthetic biology community. We wanted to listen before we started talking. And we spent about a year doing that.

    So a lot of conferences, a lot of stakeholder calls. And very honestly, the reaction to us showing up and saying, ‘Hey, I’m Caitlin, I’m with the National Security Commission on Emerging Biotech, we’re a government entity, we’re a legislative branch advisory group,’ was mixed. And so about half of the people that we would engage with would say, ‘Cool, I want to learn more. I’m interested in things that the federal government can do to accelerate biotechnology R&D, or maybe streamline product regulations, or maybe, you know, fund some of the work that we’re doing.’ But about half the people that we worked with in that first year were resistant to that framing.

    And in many cases, the life sciences industry and the biotechnology industry has not thought of itself historically as part of the national security industrial base of the United States. And so we would meet with people and they would say, ‘Well, I’m never gonna work with the Department of Defense or Department of War. I’m never gonna work with the intelligence community. I don’t know why I should be a part of this project.’ And so it was kind of a two-way conversation, us learning about the sh…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    29 min
  • EP7: Why AI in Science Needs Informative Data, Not Just More Data with Peter Koo (CSHL)

    Zhane Yamin: People are always going on about data these days, especially with AI. Get the data, create the data, protect the data, manufacture the data, don’t steal the data, you get the idea. People especially talk about data sets when they’re trying to train AI models on them because their predictive ability improves with more data.

    So everyone agrees, yes, we need more data, but more data is pretty much where the agreement ends and the trouble starts. My guest today has thought harder than most about why that happens.

    Peter Koo is an associate professor at Cold Spring Harbor Laboratory, where he runs a lab at the Simon Center for Quantitative Biology, building machine learning methods to decode the regulatory genome. I would say that’s a pretty relevant background for this conversation, but he also wrote one of the winning pieces on Hesero’s MetaScience Essay Competition, a piece called “Before the Next Protein Data Bank,” where he provides an argument about how a field decided what to build before it spends a fortune building the wrong thing.

    Excited to dive into it right from Peter’s lab in Cold Spring Harbor Laboratory. Peter, welcome to the show.

    Peter Koo: Thanks for having me.

    Zhane Yamin: It’s great to have you. So, softball, I guess, to start off with, depending on how you see it, but maybe also a very big question: why do we need data sets, and why are they helpful for machine learning or AI or these models that we like to build for science?

    Peter Koo: Yeah, in the sciences, and in machine learning generally, we’re trying to understand the world. We’re trying to build a predictive understanding of the world. And in some fields, like in theoretical physics, we can write down equations that describe how dynamical processes will evolve over time, and that in itself is a predictive model of the world.

    But we can’t build those kinds of models for everything. In other areas we need data to build statistical models, to learn patterns from the data in order to extract the rules of how the world works. So it’s a combination of the data itself, which encodes information about some aspect of the world that we’re trying to understand, and the statistical model that’s going to extract that information out from these models so that we can end up using it to predict future events or phenomena.

    Zhane Yamin: So in your work specifically, what does that look like? What is that data? What are you trying to build? What are you trying to predict?

    Peter Koo: So we are focused in the life sciences, and specifically we cover a lot of different scales of biology. The majority of my lab, we want to build models that understand how the DNA code functions. Within every cell in our body, they share the same copy of DNA, yet every cell in our body functions so differently.

    So we want to understand what parts of the DNA encode and control different kinds of cell types all throughout our body. And just to step back a second, the genome is about three billion bases, meaning As, Cs, Gs, and Ts, and only 1% of those code for proteins. The vast majority, it’s not really known exactly what it does.

    Sometimes it acts structurally to compact the genome into the nucleus, but there are large portions of it that control which genes to turn on and off. For instance, my heart doesn’t need a neurotransmitter, so I could just turn that off. Turn that gene off. The gene regulation program controls how much each protein should be expressed, and we want to understand how that code works.

    Zhane Yamin: So then when we’re building these models, and maybe your work as an example, why is a model only as good as the data? What role does the quality of the data play into the quality of the model or the predictions?

    Peter Koo: That’s a good question. The quality of the data is a very challenging question, because what we want to measure is at a very small molecular scale. If we could just see everything that’s going on, just visualize it, and see all the interactions taking place, and just take a movie of that, then we understand how the process works. But we can’t do that at this molecular scale.

    All we have tools to probe different aspects of this process. Maybe we can see which proteins are binding to DNA, but we can’t look at all the proteins binding to DNA. We can only look at the specific protein we’re interested in and see where that protein is bound to DNA all throughout the genome. These all provide little noisy snapshots.

    Also, even when we look at a protein binding to DNA, we’re not actually visualizing that. We’re indirectly measuring that, because we have to figure out biochemical processes, steps to interlock the protein with the DNA. Then we have to get rid of the cell walls and extract out the DNA bound to the protein, and we have to somehow sequence where that protein was bound.

    So there are a lot of processes happening. Biochemical steps where you cut up the DNA, where you remove the protein from the DNA, and then you have to sequence it. And even then you have lots of processes happening. So what you end up getting at the end, the actual raw data measurements, is going to be a very indirect and very noisy measurement of the underlying biological process that we were truly wanting to capture.

    We need models to be able to take in that raw data and try to go back and prune out all that noise and extract out the biological signals from this. Different data types have di fferent noise levels. They have different ways of indirectly monitoring this, and so the quality is of course different across the board.

    Zhane Yamin: Is it more so the job of the model to take out the patterns in that data? Or how much responsibility do we have as researchers, or the creators of these data sets, to make sure that those patterns are easily as noticeable?

    I guess what I’m saying is, in terms of cleaning the data, what does that mean?

    Peter Koo: So cleaning the data means processing it and getting it ready for machine learning models to analyze it. If you give a machine learning model too much noisy data, maybe it could learn the underlying biology on its own, but the noise level is so large that it’s often hard for it to find those patterns. And maybe in this sort of framework you would need lots more data, because the noise level is so high that you need to be able to see so many different samples of noise that the shared underlying biological signal, you can identify that.

    But by using human priors, humans figuring out how to clean up the noise of the data, it ends up making the data quality a lot higher, and so you don’t need as much data for a machine learning model than to learn the underlying patterns. But you still need lots of data. It’s just you don’t need as much data, and it’s more data efficient in that sense.

    Zhane Yamin: Would it be fair to say then that there’s active trade-off between the amount of data we need and the quality of that data? Like if you need less data... Could you theoretically just brute force it with a bunch of data that maybe is very noisy? Obviously that’s suboptimal, but is that theoretically possible, or is there sort of a roof or a ceiling or something?

    Peter Koo: I think when we think of data as high quality versus noisy, there are other axes that are important at play here. The task that we want these models, what patterns should they be learning from the data? If it’s a relatively simple set of rules or patterns that the models need to infer, like a protein binding to DNA might, if the binding is very, very strong for a specific subset of nucleotides, then learning that rule can be more straightforward, versus a protein that is very promiscuous and just binds loosely across the genome.

    And so the underlying patterns aren’t as strong, so the complexity of the data can also matter. When the underlying patterns aren’t that complex and the underlying rules aren’t so complex, then you don’t need as much data to learn those rules. As the complexity of the underlying mechanisms …



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    16 min
  • EP6: Building a Time Machine for Science with Chris Fussell (Lila Sciences/McChrystal Group)

    Zhane Yamin: My guest today isn’t a scientist, though I think you could call him an engineer of sorts. Chris Fussell is president of business operations at Lila Sciences and vice chairman of McChrystal Group. Before that, he spent 15 years as an officer on Navy SEAL teams. And suffice it to say, Chris has got some knowledge in trying to get large siloed organizations to move like fast connected networks.

    Luckily, or perhaps unluckily for us, the same thing could be said about science. That’s why I think it’s such an amazing opportunity to talk to him right now, because the question of whether America meets the AI for science moment might not be entirely a question of science. It might be a question of whether a sprawling enterprise, whether the agencies, labs, universities, and companies can act like they’re on the same team. And that’s the one problem Chris has a lot of experience solving. Chris, welcome to the show.

    Chris Fussell: Thanks, Zhane. Yeah, and I think the scientists that I work with at Lila would definitely agree that I’m not a scientist. But I appreciate it. Yeah.

    Zhane Yamin: No problem. So I guess I just wanted to start with an honest assessment because I think when we talk about scientific competitiveness, there are a lot of alarm bells that are rung, and they’re rung pretty frequently. And I don’t want to just ring them like everyone else. So I want to start off with one question: when you look at where the U.S. actually stands scientifically right now, how competitive are we?

    Chris Fussell: The entire field is at a transition point, right? So there’s a lot of unknowns in that question, right? Go back two years, where are we competitively? We’ve got, you know, but a 60-year advantage, post-World War II, right? But now as AI starts to intersect with science, which is a totally unique space inside of the AI zip code, you know, AI and science is a very specific neighborhood inside of that much broader world. Where are we competitively in that neighborhood? I feel very good about it.

    But it’s not one—it needs a special type of attention, both from the U.S. government and allies and from companies like Lila. And there are not a lot of companies in our space. I would say we are uniquely positioned globally. But where are we in that competition globally? I think we have an advantage and I think we are ahead of others, but I think it requires a certain level of attention and understanding from government authorities, private industry, et cetera. And if we don’t give it the proper attention, there’s a world where you get passed. I’m confident we won’t, and I’m confident we have the right pole position right now, but we can’t take our eye off of it.

    Zhane Yamin: So one thing I kind of want to track with that is, honestly, this idea of watching, and trying to know what we need to watch. Because I think when we’re talking about scientific competitiveness, there’s always a lot of metrics that are thrown around, like China’s taking us in this, but we’ve got this in the U.S. And so I want to get your thoughts on what sort of things we should be watching and metrics we should be taking account of when we’re talking about AI for science and AI competitiveness globally.

    Chris Fussell: Yeah, on the AI for science question, and again, caveating with your opening, I am not a scientist, but I spend my days with them. And so there’s, again, in this neighborhood inside of the AI discussion, once you’ve seen this, you can’t unsee it, which is: the winners of AI and science, I believe, will have to do what we are doing. And this gets to your point of measurability and how do we know how far along we are?

    Lila’s thesis and mission and what we are building is an AI model that’s able to connect with physical labs. We’re building many tens of thousands of square feet of labs that will be controlled by AI models, with autonomous systems, robotics, humans very much involved in the labs, which allows incredible throughput, right? And so I like to think of it as a time machine for science. We’ll be able to reach years, decades into the future and pull knowledge into today. So what will be measurable there, rather than just the speed of your model or a lot of stuff that the average person just can’t really get their heads around?

    We’re right on the cusp of being able to look at a lab and say—sort of like what happened in manufacturing, if you go back to, I don’t know, the post-industrial revolution all the way through the early 20th century—we changed the way we looked at output. Like, how much can this factory produce per square foot of X, and how much energy does it need? It wasn’t just a human question or a knowledge question or an artisan question. You could look at the size of a factory and say it’s going to be able to produce this much X.

    We’re going to be able to do that with science. Like, this lab, with this configuration and the power of this model, can produce this much new scientific knowledge. And we’ll eventually be able to measure that, I believe, down to sort of the square foot. And that’s where the measurement question, I think, becomes really interesting and different than it’s ever been in science. We’ve seen that evolution in other industries, but we’re on the brink of doing that in the scientific space.

    Zhane Yamin: So on that point, I want to backtrack into your background and the perspective it gives you. Because again, you’re saying you’re not a scientist, but I think that actually offers you a very unique perspective on this entire landscape. How does that inform your work with Lila Sciences and this idea of autonomous science and building out these labs of the future? What does that strategic perspective offer you in that sense?

    Chris Fussell: Yeah, just quickly on my background. My first career, for about 15 years from the late nineties until 2012, I guess, I was an officer in the SEAL teams. And so for your listeners, the U.S. special operations community is now, I don’t know, twice as big as it was 25 years ago, but say 70,000 people. Each branch has its own component of special operations. The SEAL teams sit inside the Navy. And there’s a long history to that that goes back to 1980 when those communities were really created. For the first 20 years or so of their history—and this is the force that I joined in the late ’90s—they were built for siloed excellence. The SEALs had to be really good at X. Army units had to be really good at Y, et cetera. And there wasn’t a world where we’d have to integrate together.

    Well, post-9/11, now 25 years ago, we were put into a little bit of a time machine ourselves, because we jumped onto a battlefield that was networked and digitally enabled, and our much less capable, much less well-funded adversaries were connected globally through a digital network, right? That’s obvious now, right? It’s at our fingertips every day. Twenty-five years ago that was new and interesting, and it broke the way that we had to operate. And we had to rebuild and say, okay, how do we create a truly networked set of units that weren’t built for that culturally, operationally—none of our systems integrated, from the human level to the technical level? But we have to be able to do that to defeat this adversary.

    And so we went through this whole rebuilding process, and then I spent years post-military thinking about that, writing about that. I teach on it, and then obviously built a consulting firm based on it with retired General Stan McChrystal, who was sort of the father of this inside of special operations. So that’s a quick background on what I think is hyper-applicable here, because if you go back just a few minutes, what we were talking about is building these highly integrated lab systems.

    AI doesn’t care if you’re a chemist or a life scientist, et cetera. Eventually it will just understand, like we see in large language models, except there are way more combinations in science. It’s just going to see atoms and molecules and the w…



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    28 min
  • EP5: The Making of an AI Scientist with Sam Rodriques (Edison Scientific/Future House)

    In a world where scientific discovery is often hampered by the limited number of exceptional scientists, Sam Rodriques, co-founder and CEO of Edison Scientific, sheds light on how AI can bridge this talent gap. With a background in physics and a history of pioneering work in AI for scientific research, Sam is at the forefront of a movement to augment human talent with intelligent systems.

    Defining AI Scientists

    Sam begins by addressing the common misconceptions surrounding AI scientists. He notes that while the term is frequently used, its meaning can often be unclear. For Sam, the purpose of an AI scientist is to **augment human scientific talent**. By providing access to exceptional scientific intelligence, AI scientists aim to accelerate the pace of research and discovery.

    “Removing talent as the bottleneck in science is going to be one of the fastest ways to accelerate progress in science today.” — Sam Rodriques

    The Role of AI Agents

    Sam explains that the AI agents developed at Edison Scientific are designed to operate within the scientific process, facilitating everything from hypothesis generation to laboratory experimentation and data analysis. By collaborating with pharmaceutical companies, these agents are already making novel discoveries that can be validated in real-world settings.

    Key Components of Scientific Progress

    According to Sam, successful scientific inquiry requires three essential components: intelligence, knowledge, and experimentation. He emphasizes that while intelligence in AI is becoming increasingly commoditized, the focus will be on ensuring that models have the knowledge and experimental capabilities necessary for groundbreaking research.

    Addressing the Knowledge Explosion

    With the vast amount of scientific literature published today, it’s becoming impractical for any single individual to keep up. Sam argues that AI scientists are uniquely positioned to manage this overwhelming influx of information. Their ability to aggregate knowledge and facilitate new discoveries can lead to significant advancements in various fields.

    “The proliferation of knowledge means that there’s now too much for any one person to hold in their head. AI can help manage and leverage that information effectively.” — Sam Rodriques

    Infrastructure and Sovereignty in AI

    As AI tools become more integral to scientific research, the infrastructure supporting these technologies is evolving. Sam highlights the importance of **sovereignty in AI**, where companies need to run models independently to protect their proprietary data. This shift is crucial for companies that want to leverage AI without relinquishing control over their most valuable asset—their data.

    The Importance of Benchmarks

    To evaluate the effectiveness of AI agents, Sam and his team have developed benchmarks like Labbench and Bixbench. These tools assess the performance of AI in scientific tasks, ensuring that as capabilities advance, the benchmarks do too.

    “You can’t tell whether or not the thing that you’re doing is correct if you don’t have a benchmark.” — Sam Rodriques

    Conclusion

    Sam’s insights into the role of AI in science reveal a future where human and artificial intelligence work together to push the boundaries of research. As the landscape of scientific inquiry evolves, the integration of AI promises to unlock new possibilities for discovery and innovation.

    For those interested in the intersection of technology and science, this conversation offers a compelling glimpse into what the future may hold.



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    31 min
  • EP4: How AI Learned to Speak Science with Rafael Gomez-Bombarelli (MIT/Lila Sciences)

    Rafael Gomez-Bombarelli sits down with Zhane to discuss the intersection of AI and science, exploring how AI models are transforming research in chemistry, materials science, and beyond. He shares insights on the evolution of AI for science, the role of scalable architectures, and the future of autonomous laboratories.

    The Evolution of AI in Science

    Rafa’s career illustrates a pivotal moment in science when computational methods began merging with experimental research. He shared,

    “Back around 2014, 2015, the simulation community realized that what they were doing was highly scalable.”

    This scalability allowed researchers to harness vast amounts of simulation data, coinciding with the rise of deep learning techniques. As Rafa explained, this period marked a significant shift:

    “We had the ability to make large data sets and deep learning architectures that performed well on these data sets.”

    He emphasized how the emergence of specialized neural networks transformed the way scientists approach data and simulations. The integration of machine learning into scientific research has sparked what Rafa terms the first incarnation of AI for science.

    The Bitter Lesson

    Rafa introduced the concept of the “bitter lesson,” which posits that the most effective AI models are those that leverage large datasets rather than relying on human intuition or domain-specific knowledge. He noted,

    “The models that ultimately succeed are the ones that can just take the most available data into the most expressive general models.”

    This understanding has profound implications for how scientists approach problem-solving in their fields, suggesting a need to adapt to more scalable methods of inquiry.

    A New Era for Scientific Discovery

    So why is there such a surge of interest in AI for science right now? According to Rafa, the urgency of scientific advancement is a major driver. He articulated that

    “Science is a very noble, important, and sort of unambiguously good for humanity.”

    The need to innovate, especially in fields like medicine and technology, creates a fertile ground for the application of AI. As Rafa pointed out, science inherently generates new knowledge, making it a compelling domain for AI applications.

    The Role of Lila Sciences

    At Lila Sciences, Rafa is pushing the boundaries of what’s possible with AI in the lab. He describes a vision for creating programmable laboratories or “AI Science Factories” that integrate AI’s capabilities with experimental verification. He explained that,

    “The only way to have scalable verification for AI-driven science is to give the whole wheel of science to AI.”

    This revolutionary approach aims to automate not only data analysis but also the execution of experiments, allowing AI to verify its hypotheses in real-time.

    Rafa’s work serves as a bridge between language-based AI models and the nuanced, complex world of scientific inquiry. His role involves translating scientific concepts into formats that AI can understand, ensuring that the rich data of the scientific domain can be effectively utilized by machine learning models.

    In this exciting intersection of AI and science, we are witnessing a transformation that promises to accelerate discovery and innovation like never before. To explore more about these insights and the future of AI in scientific research, listen to the full episode.

    Chapters

    00:00 Introduction to Rafa Gomez-Bombarelli and his background in AI and science

    01:24 The merging of experimental science and simulation in Rafa’s career

    02:24 The rise of data science and deep learning in materials and chemistry

    03:20 The impact of neural networks and large data sets in AI for science

    04:20 The difference between physics-based simulation and machine learning models

    05:43 Generative models for chemistry and materials design

    06:37 How language models relate to scientific data and models

    09:03 Transformers and the scaling of AI architectures

    10:38 Why now is a pivotal moment for AI in science

    Lila Sciences and the role of AI in scientific discovery

    The concept of AI Science Factories and autonomous labs

    Concrete examples of AI enabling scientific breakthroughs

    The potential and limitations of AI in generating new knowledge

    Open questions and the future role of human scientists



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    36 min
  • EP3: The Case for Open Science with Prachee Avasthi (Astera Institute/Arcadia Science)

    In this episode, Dr. Prachee Avasthi discusses the transformative potential of open science, challenging traditional publishing and funding models, and exploring how technology and policy can accelerate scientific progress.

    Chapters

    00:00 Introduction to open science and its importance

    01:51 What is open science and how is it defined?

    04:00 The role of open access and open data in research

    06:04 Challenges of traditional publishing and peer review

    07:57 Funding models and their influence on research direction

    10:10 The impact of AI and new technologies on science dissemination

    12:09 Strategies for implementing open science practices

    14:02 The distinction and relationship between for-profit and nonprofit research

    15:52 Policy initiatives and organizational efforts to promote open science

    18:12 The cultural shift needed for open science adoption

    20:08 The future of scientific collaboration and data sharing

    22:06 The societal stakes and benefits of open science

    23:54 Overcoming inertia and resistance in the research community

    26:01 The role of grassroots efforts and individual action

    28:06 The importance of infrastructure and resources for open science

    30:07 The potential of AI and machine learning to transform research

    32:04 How to navigate the evolving landscape of scientific publishing

    33:49 The impact of open science on innovation and societal progress

    36:07 Final thoughts: embracing change for a better scientific future



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    39 min
  • EP2: The Challenge of Focus & Messaging in AI for Science with Dan Turner-Evans (Institute for Progress)

    In this episode, we explore the Genesis mission’s focus, funding challenges, and the importance of clear communication in AI for science. Dan Turner-Evans of the Institute for Progress shares insights from his diverse background in physics, neuroscience, and policy, emphasizing the need for strategic focus and effective messaging to advance AI initiatives in government and research.

    Chapters

    00:00 Introduction to the Genesis mission and guest background

    02:01 Dan’s diverse scientific background and policy experience

    07:08 Overview of the Genesis mission and its recent expansion

    11:10 The importance of focus and funding for the mission

    15:11 Challenges of broad AI initiatives and strategic focus

    19:52 Communication strategies for AI in science to the public and Congress

    25:08 Platforms versus challenges in AI research and development

    30:00 Balancing basic research with targeted AI applications

    34:53 Concluding thoughts on focus, communication, and the future of AI in science



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    35 min
  • EP1: How AI Can Accelerate Scientific Discovery with Josh New (SeedAI)

    In this episode, Josh New, director of policy at SeedAI, shares insights on accelerating science through AI, policy challenges, and innovative models like autonomous labs and FROs. We explore how to unlock AI’s potential for scientific breakthroughs and the role of policy in fostering a more dynamic research ecosystem.

    Chapters

    00:00 Introduction and episode overview

    01:00 Guest introduction: Josh New and SeedAI

    02:05 Josh’s path from tech policy to AI for science

    03:23 The rise of AI in scientific research

    05:14 Changing conversations in AI policy

    07:43 Rapid evolution of AI and policy challenges

    08:35 The impact of AI saturation on scientific workflows

    12:00 Unlocking AI’s potential versus steering its direction

    14:02 Decentralized science and policy intervention

    17:28 Reimagining research labs with AI and automation

    20:55 Funding models and data infrastructure for AI-driven science

    23:07 Lessons from the Human Genome Project and big science

    29:45 The ASN model as connective tissue in AI and science

    34:30 Closing thoughts and call to action

    Episode Resources

    SeedAI - https://seedai.org

    AlphaFold - Protein structure prediction - https://deepmind.com/research/highlighted-research/alphafold

    Human Genome Project - https://www.genome.gov/human-genome-project

    ASN Coalition - Accelerate Science Now - https://acceleratesciencenow.com/



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit acceleratesciencenow.substack.com
    34 min

About State Space

From the publisher's feed

The scientific research ecosystem is its own kind of state space — full of decades-old ideas, signal mixed with noise, and breakthroughs increasingly shaped by AI. State Space sits down with leading…