The AI race to save American science

Note: Transcripts are automatically generated from episode audio, and are not fully corrected for spelling, grammar, and formatting.

RENEE WEGRZYN: Nobody who’s practicing science today has practiced science in the United States at a time when we weren’t, quote-unquote, winning from a competitiveness perspective. We were at the top. And I genuinely believe that we may lose that competitiveness, that edge, that incumbency, and I’m not sure I see us fighting for it in the way that we should be.

BOB SAFIAN: That’s Renee Wegrzyn, founding director of ARPA-H, the federal agency created to speed up health science breakthroughs. Renee is now chief innovation officer of Transfyr, which she co-founded after leaving government. I wanted to talk with Renee to better understand how U.S. federal research cuts are affecting America’s scientific progress at this moment, and whether the private sector is filling the gap. We also talk about her company’s efforts to accelerate the next big breakthroughs, how AI and physical lab work combine, and her optimism about what she calls the next Sputnik moment. I’ll admit lab science isn’t my natural habitat, but Renee has a gift for making complex things accessible. So let’s get to it. I’m Bob Safian, and this is Rapid Response. I

‘m here with Dr. Renee Wegrzyn, co-founder of Transfyr. Renee, welcome to the show.

WEGRZYN: Hey, Bob. Thanks for having me.

Copy LinkFilling the research gap left by federal funding cuts

SAFIAN: You and I first connected when you were at ARPA-H, the inaugural director of a White House effort to accelerate health research. You were amped up about the possibilities, but with the Trump administration, you had to pivot. What did that feel like? Was it hard not to take it personally that they wanted different leadership?

WEGRZYN: I’ll tell you, I’m still amped up about it. It’s still an amazing idea. Obviously, DARPA’s been around for almost 80 years. I hope ARPA-H will have the same longevity and continue to live on today. I’m a builder. I love to take things forward and really show what’s possible. So was it hard when I left? Of course it was hard. I had a really unusual appointment: I was a presidential appointee, but I was term-limited, so I served through the transition and had a few weeks in the new administration. I served in Trump one, by the way, as well. So it was a great opportunity to make sure they really understood why we built the agency. We needed an ARPA-H because there was something missing in the way we were doing science broadly, which was: How do you move really quickly? How do you get focused on a problem versus building a general knowledge set? I was able to show them what we were building and convince them that it was worth keeping around. I couldn’t be more proud of the agency and its perseverance.

SAFIAN: Your optimism about that is in contrast, in some ways, to some of the stories about what’s happened during Trump, too. The New York Times did a story about the CDC being hobbled by cuts and controversies over scientific advisory boards being dismissed and skepticism toward federally funded research. How different are things really? Most things in government go through cycles, and you’ve obviously experienced a few of those cycles.

WEGRZYN: Yeah, absolutely. A lot of tragedy and chaos happened. There’s no doubt about it. I don’t want to minimize that. But where I find optimism and inspiration is that sometimes, through hardship, one of the realizations I hope we’ve all had over the last two years is that the way we fund science is not sustainable, right? My own retirement account is not 90 percent based on one stock or fund. It is really diversified, and a lot of the scientific enterprise in the United States right now is really focused on federal funding. And that’s a challenge. How do we have better connections to the private sector? How do we capitalize on that investment to incentivize it in new and different ways? That has been a forcing function that I’m really optimistic about.

Having universities think about all of the great IP that comes out of those universities: How can that be more revenue-generating to return to their universities so they can become less dependent on some of those external sources of funding? How does this look different to make sure that it is robust and resilient going forward?

SAFIAN: There is more pressure on the private sector to step up and fill that research gap, right? You worked in the private sector even before your current thing, at Ginkgo Bioworks, where Transfyr co-founder Anna Marie Wagner was head of AI. Are companies up to the task of spurring science in this way? We’ve certainly seen the spurring of technology with AI labs.

WEGRZYN: I think so. The intersection of AI and bio in particular, which is my domain, is moving at such a pace that it’s hard to keep up if you’re not in the private sector. It seems like the smaller your organization is, the better, because you can make quick decisions, fail fast, and move quickly. That is hard for governments to do. In fact, that was one of the mandates of ARPA-H: to be able to identify failures and not pursue work streams that were no longer really delivering on the science.

Copy LinkThe problem Transfyr is trying to address

SAFIAN: So your new business, Transfyr — The New York Times wrote about you, and the article kept referring to magic hands. I’m curious what magic hands are. What problem is Transfyr trying to solve or address?

WEGRZYN: Have you spent time in a research lab before?

SAFIAN: Not a ton of time. I’m not that kind of researcher.

WEGRZYN: Okay, yeah. But certainly you’ve spent time in the kitchen, and you probably have your favorite dish, or your parents’ or grandparents’ favorite dish. Sometimes they just have the magic hands. Nobody can cook it like they do. Nobody can make it taste the special way they can. And that happens in science as well. There are some scientists who just have the knack and can deliver really great science over and over. The challenge is that a lot of those scientists are such experts they can hardly describe what they’re doing. It’s an art in some ways. And so the magic hands idea was really a testament to what those scientists can do in the lab.

The pipeline we’re creating at Transfyr is: Can we create computer vision and perception tools to capture the work of science as it’s being done so that we understand, if the protocol says, “Mix the cells,” whether someone is inverting the tube in a certain gentle way, while another researcher may be rigorously shaking or vortexing it, which could destroy the sample? Those little things, compounded over the many decisions a scientist may make over the course of several hours of experimentation, lead to what is called magic hands.

SAFIAN: So it’s all this kind of tacit knowledge that that expert isn’t even really thinking about. Was this an issue you saw at ARPA-H, the sort of breakthroughs that were trapped in a single lab in some ways?

WEGRZYN: I think tacit knowledge is part of it. I would call it more embodied knowledge. It’s what a scientist does that is not written down. It’s not even just the way they’re holding something in a special way; it’s the fact that they pre-labeled all of their tubes before they started the experiment, so they saved 10 minutes between steps. That embodied knowledge, in all of those little decisions, is really what we’re capturing.

Where that manifested at DARPA and at ARPA-H was in a realization, through my career, of how fragile breakthroughs are. Usually, the way a breakthrough is first communicated to the world is through a research publication, a Science or Nature paper. That paper is not a product. It’s not commercializable. It’s not ready for a patient that day. It’s not the period at the end of the sentence. That’s the starting gun for what is going to happen. The ARPAs of this world are really fantastic at getting to that initial breakthrough and demonstrating what is possible, but following through and delivering is not a 100 percent guaranteed activity. When I started ARPA-H, we wanted customers, who would be the final users of that technology, to give us feedback on what we were developing so that, when we got to the last day of the experiment, they were demanding that product, versus having that be the first contact with a customer, which unfortunately is pretty common around traditional grant-making organizations.

It’s rethinking how you make sure you can transfer that knowledge off the back end of a program.

SAFIAN: Because the folks at the lab aren’t necessarily worrying about, or prioritizing, how to make it scalable. They’re just trying to see if it works. But of course, eventually we want it to be scalable.

WEGRZYN: Sometimes the folks in the lab don’t even have the expertise or understanding of who they need to talk to next to do that, right? We think about science and those transfers as handoffs. We deploy our technology to an originating lab to understand the science and what was important about that discovery, but we can also deploy it in a receiving lab so that we can see how they’re reproducing it. And instead of the way we transfer science today, which is looking over each other’s shoulders and saying, “Oh, you’re doing it a little wrong,” we have this video-connectedness that lets us help accelerate those iterations of transferring the science.

Copy LinkUsing AI to cut lab waste

SAFIAN: The procedural parts of Transfyr could sound less sexy than claims about how AI can design new drugs, but you’re doing some cool stuff. You’re putting cameras on scientists’ headbands in the lab, and you’re capturing video and audio and equipment settings in real time. Then you’re using AI to mine that data to find out what needs to be consistent?

WEGRZYN: Our system of record for science today is a lab notebook or a LIMS system, which, again, is sort of an outline of what happened. It’s not the actual lossless representation of the work. And so for a scientist at the bench, that’s really important because you may spend six hours preparing a sample, you run your gel to do the analysis, you look at the results, and you think you might have skipped a well. Having the confidence to look back at the tape and say, “Yep, there it is. I skipped a well. I know exactly what went wrong. I can move forward with the next step of my experiment,” versus, “I don’t know what went wrong. Maybe that sample didn’t work. I have to repeat the whole experiment” — that’s huge.

The loss of resources across sciences is vast because we need to repeat things simply due to a lack of understanding of what happened. Sometimes you don’t ever find out what went wrong. It’s disappointing how many discoveries we probably lose to some sort of procedural variation that we haven’t picked up on.

SAFIAN: So some of this is about saving costs, and some of it is about accelerating the discovery process by pointing you in the right direction.

WEGRZYN: You talked about how AI is transforming drug design. We think of that as top of the funnel. Now you can test billions of designs in silico. Some subset of those you can actually test in the lab, but we haven’t seen billions of new clinical trials. It’s because the rest of that pipeline is still bottlenecked. We’re not transferring the assays between labs faster than we were before. We’re not scaling up manufacturing faster than we did before, or better and with more choices than we did before. This physical AI infrastructure that we’re creating is intended to break open some of those bottlenecks so that we can do one-shot tech transfer of manufacturing to scale up quickly, which is important to get cures to patients faster, candidly.

Copy LinkWhy scientists need control over lab recording

SAFIAN: Now I could see some people looking at your data capture process as a kind of surveillance within labs. Do you get that pushback from scientists who don’t want to be watched?

WEGRZYN: Yes, we get that a lot, and I think it’s something that my co-founder and I care really deeply about: we’re doing this for the scientists. We want scientists to be top of license. If a scientist is successful, we all win, right? Humanity wins if science is successful. So how can we empower that scientist to make better decisions based on real evidence? We’re not just engaging with the chief operating officer who sees this opportunity to make sure everything’s more reproducible and that we don’t lose a batch. We’re working with the scientist to understand how this can be frictionless for them. A GoPro is heavy to wear on your head, so our headsets are much lighter and have longer recording capability, based on direct feedback from those scientists. It’s important that they have ways to do their work. We capture voice as well, so it’s a way to capture their lab notebook in real time without having to take their gloves off.

That is really our goal: to empower the scientists. The interfaces that we have are also controlled by the scientists. They turn on the recording, they turn off the recording, so they are in control of that.

SAFIAN: The bottleneck that you refer to makes me wonder whether science is less rigorous than maybe we’ve always been led to believe. I know there’s been rising skepticism in certain places, including within the Trump White House, about assertions of expertise and even published research. Is that skepticism well-founded?

WEGRZYN: Science is just incredibly hard, and I think because we’re not capturing the real work of what’s happening, there’s a misunderstanding of why science goes wrong. A lot of it is that we’re just not providing a high-enough-fidelity account of the work that’s being done. If you asked an AI agent today, “How do humans do science?” it would say, “Oh, well, someone comes up with a hypothesis, and then they publish a materials and methods section, and then there’s the data showing how it works, of course.” My postdoc was three years long, and I had a three-page paper summary at the end. It was the highlights reel. It was not all the hard work that failed. There were a lot of successes, too, that were probably different paths that I decided not to pursue because they were less interesting, but that could be incredibly useful for the next set of scientists. What’s different now, I think, is AI allows us to, in the same way it allows us to read a journal article cover to cover, which I couldn’t possibly do for all the journals in my lifetime, kind of watch an experiment from beginning to end in many, many contexts to start to understand emergent patterns, different types of contexts like temperature and humidity. You even see tribal behaviors in labs, which is kind of interesting.

SAFIAN: Tribal behavior?

WEGRZYN: Yeah. What I mean by that is we’ll sometimes deploy to a lab and, when a step says something like, “Mix the tube,” they’ll flick the tube. Everybody in that lab flicks the tube, and you’re like, “Somebody trained the people in that lab to flick the tube,” because if you see that same protocol from a vendor kit implemented anywhere else, everybody does it differently. Why do you do it that way? Something I tried to teach my team at ARPA-H was just to ask why constantly and not assume what’s presented in front of you.

SAFIAN: Let’s take a beat here to absorb what Renee is saying. Each lab, each company, each of us may have our own way of doing something, and that variation is exactly why we can come up with creative, game-changing results. So can any business experiment at the speed of AI frontier models? And is America doing enough to protect its scientific edge? We’ll talk about that and more after the break. Stay with us.

[AD BREAK]

Before the break, ARPA-H Founding Director Renee Wegrzyn talked about why magic hands inside a lab are so valuable and so hard to pass on. Now she talks about autonomous labs, how to adjust at the speed of AI frontier models, and searching for the next Sputnik moment.

Let’s jump back in.

Copy LinkHow automation can assist research labs

SAFIAN: The very idea of experimentation is, by definition, messy, because you’re not really sure where you’re going. I guess sometimes you might find things that you didn’t realize, and sometimes you don’t really know how you got there.

WEGRZYN: We were talking about this yesterday, actually. There are moments in a scientist’s life where they’re tinkering over the course of several weeks trying to get something to work, and the moment it works, you’re like, “What did I do? What did I do to get it to work?” So you’re not following a protocol, but this is where I think robotics and humans collaborating can become really interesting. I might recognize that I have a hunch that it’s the magnesium concentration of this buffer that really matters. It’s not the best use of my time. I’m not top of license if I test 50 magnesium concentrations, but I can outsource that to my robotic partner arm that’s here at the lab with me. That is where I see automation, at least in the near term, really empowering the scientist before we get to a world, and I don’t even know when it would be, where it’s full closed-loop automation. You hear about these autonomous labs, but I’ll tell you from my time, even during the pandemic, supporting testing nationwide through the work we did at Ginkgo, the number of labs across the country where the operator loads the 96-well plate in a flipped direction is in the double-digit percentages. The robot didn’t do anything wrong. It’s the human in the loop that is still part of autonomous science that we’re not observing. In those labs, you have many cameras pointing at the robots because we all assume the robot’s going to do something silly, but regular human error is not captured.

SAFIAN: I can see, as you give that example, why some people resist it, because they don’t want to be caught having put something in the wrong way. None of us want to be called out for having done something the wrong way.

WEGRZYN: But sometimes the stakes are really high, though, right? Some of our diagnostics customers are delivering a genetic result to a patient. Most scientists, I think, want to do the right thing and also want to have the information so they can be coached by understanding where they made that error.

Copy LinkBuilding model-agnostic AI for lab work

SAFIAN: You and your co-founder, Anna Marie, both worked at Ginkgo Bioworks, which focuses on genetics. Why not work on applying AI in that area? Biotech has so much runway ahead of it. There are so many different areas.

WEGRZYN: I think for us, our technology is agnostic to the science that is being performed. In some ways, Bob, our narrowing down is just focusing on life sciences to start. Chemistry and material science are other areas where we get a lot of incoming.

So maybe that’s a future path for us. But life sciences writ large is where we’re focused. If you think about it, it makes sense because when you break down molecular biology, genetics broadly, at the atomic action level, it’s like transferring liquids between vessels.

SAFIAN: You make it sound like I’m just a bartender making drinks.

WEGRZYN: But it’s a level the models can understand a lot about. They understand what it looks like when there’s a bubble in your pipette tip, and that is a deviation. It was a choice for us to focus on that deep foundational level because it is general-purpose for the rest of life sciences. That was the bet we took rather than saying, “We’re going to crush biologics manufacturing.” That is not generalizable to a broad area of life sciences.

SAFIAN: And the modeling that you’re doing, is it a different model for each lab? Is there one model that you’re training across all of them? Is there an underlying LLM that you use, or do you create your own?

WEGRZYN: It honestly depends on the day, because we are hardware- and model-agnostic, right? We know that hardware will commoditize, and we use whatever’s best in class today. When there are new things that are lighter, cheaper, longer-lasting, with better batteries, we will integrate that. It’s the same with the models. There’s a new model published every week from all of the frontier AI companies. We evaluate and test those, and whatever is performing best on the task at hand, we will adopt. It allows us to keep pace with the development of that entire ecosystem. It’s also how we can work really closely with the frontier AI companies to say, “Hey, we really appreciate this new model drop, but did you know that when it comes to recognizing X, you’re not as good as this other model?” It’s a helpful loop because we want those models to get better, so it’s a chance to provide some of that feedback.

SAFIAN: And moving to a potentially different underlying model every two weeks, of course that’s got to keep you on your toes. But does it mean you have to retrain your model with the new engine, essentially?

WEGRZYN: We’re constantly evaluating the models. We retrain as we get new data. We also use data to perturb experiments and then see if the model can detect deviations. So we tell the model we’re doing protocol X, but then we have operators come through and change the protocol slightly.

We’ll intentionally skip a step or add a step or give the wrong reagent.

SAFIAN: To test.

WEGRZYN: To test the models. We have the answer key in that case, so it gives us a good sense of how they’re performing.

SAFIAN: Yeah, because, as we know, these generative models do hallucinate or make things up, and you don’t want that happening in the lab.

WEGRZYN: Yes, absolutely they do, Bob. We also have human adjudication for our models. Sometimes it hallucinates. Sometimes it gets things a little wrong. It recognizes that it’s a pipette, but it’s actually a hundred-microliter pipette, not a two-hundred-microliter pipette. So we make those corrections along the way, and that data can be used to train and improve the models iteratively.

SAFIAN: It’s so much detail work, isn’t it? Each lab and each experiment and each question is different, and yet if the model’s going to work the right way, it needs all that detail.

WEGRZYN: Determining what level of detail it needs is really important. For a lot of scientific work, the most detailed level is just too much information, and it’s just going to waste compute if you’re always labeling at that effort. So it’s important for us internally to understand how we can make our models and training more efficient based on the level of information that’s needed. That’s where we experiment quite a bit in the lab.

SAFIAN: And there’s also cost for you every time you’re running compute.

WEGRZYN: Absolutely.

SAFIAN: Right?

WEGRZYN: Yes.

SAFIAN: Yeah.

WEGRZYN: And then storing data, that’s very expensive too.

Copy LinkWhy America risks losing its scientific edge

SAFIAN: Do you ever get frustrated about how resources are allocated in our world? You’ve put your life toward this work, and yet so much money is devoted to, I don’t know, space exploration or the next hot social app, the Meta Muse craze. How do you deal with even government and how it prioritizes things?

WEGRZYN: Yeah, it’s maddening. Do I get frustrated? How do I not pull my hair out? That is part of the motivation to do ARPA-H in the first place. We have to use these incredible resources, which are taxpayer dollars, in a really purposeful way to solve real problems for Americans. Full stop, that is what we’re doing, and at pace. My program managers had three-year terms based on statute from Congress. I would say to them, “How are things going?” And they’d say, “Well, maybe next week I’m going to do X.” And I’m like, “Next week? Every two weeks that go by, 1 percent of your time is up. Come on, what about this week?” How do you really do the work and keep pace? That type of resource allocation, especially now, is so important. But being able to do it smartly and not get lost in AI slop and things like this is becoming hard in this day and age.

One of the things that has me nervous right now is that I look back on the last 80 years following the Second World War. We’ve led the globe in science and engineering and innovation. Nobody who’s practicing science today has practiced science in the United States at a time when we weren’t, quote-unquote, winning from a competitiveness perspective, when we weren’t at the top. I genuinely believe we may lose that competitiveness, that edge, that incumbency, and I’m not sure I see us fighting for it in the way that we should be.

When somebody is facing the potential of a loss, every two weeks that go by, 1 percent of your time is up. We have to deliver on what we’ve promised to do, and that’s a commitment not just to my country, but to my kids and my family. So if anybody wants to talk about that, tell me your lab problems. We’ll try to fix those, and then we can get on to the bigger problems of competitiveness and making sure that we stay on top.

SAFIAN: Well, Renee, this was great. I always like to ask, is there anything that I haven’t asked you about that I should have?

WEGRZYN: What I think a lot about is Sputnik moments, given my ARPA trajectory. We’re going to look back at this moment in time like there’s something happening, and it’s going to impact whatever’s happening 20 to 50 years from now. You can feel it in the moment. Where that happened before, of course, was Sputnik. The United States didn’t know the Soviet Union was developing that capability, and we didn’t have a counter-capability. Eisenhower launched ARPA, which became DARPA, to create and prevent technological surprise.

When he started ARPA-H, the timing was around COVID, but what I realized is that each individual person has a Sputnik moment in their life that is usually health-related. My Sputnik moment was when my mother was diagnosed with cancer. That stopped everything in her life, in my life, in my dad’s life, in my sister’s life. Everybody has a story, whether it happened to them or their family, or if you talk to a member of Congress, something’s happening in their state that they can’t deal with.

So I’m like, “What is the Sputnik moment now?” If ARPA-H was bringing people from an unhealthy state to a healthy state, what’s happening now is taking us beyond that. If we can figure out and solve health and wellness, a lot of people talk about abundance. What is the next trajectory? What is the Sputnik that we’re working toward? We should talk again in a year, Bob, and maybe I’ve figured it out, because I don’t know that I have the answer to this. It’s augmenting, going from protecting a nation to enabling and healing an individual to now allowing that individual to rise and elevate, be top of license, be their best self. That’s why I’m optimistic. I can taste it. I don’t know what it is yet, but it’s here.”

SAFIAN: Well, Renee, this was great. Thanks so much for doing it.

WEGRZYN: Thanks, Bob.