All episodes

60. Patrick Hsu - Speedrunning the Infinite Game

·Scientist, Founder, Researcher, Investor
Read transcript ↓Browse chapters ↓
Listen to the conversation
Watch the video

Related conversations

Description

Patrick Hsu (X, Website, Arc lab) is co-founder and core investigator of Arc Institute, assistant professor of pathology at Stanford, and venture partner at Thrive Capital.

Patrick has always been precocious: he began college as a young teenager and contributed to early CRISPR-Cas9 gene-editing breakthroughs as part of Feng Zhang’s lab at the Broad Institute of MIT and Harvard. Today, the Hsu Lab at Arc works where genome engineering meets AI, from bridge recombinases that can rearrange up to a million bases of DNA at once to Arc's Evo biological foundation models.

We discuss Patrick’s pace and ambition, infinite games, and why he was drawn to biology. Then, we traverse power laws in biotech and the reach of GLP-1s, research taste and success, Arc’s multidisciplinary model, data limits in AI-enabled bio, and cell and bio foundation models. We also discuss how AI and new science will help us solve diseases, design life, and push the boundaries of sci-fi when it comes to human health. I hope you are inspired to raise your bar, commit to reinvention, and hold a lightness that makes the long and challenging journey fun.

Dialectic is presented by Notion. Notion is a collaborative workspace for your life's work, and it works with whatever AI you use. New models are typically available in Notion the day they launch, you can tag agents from inside Notion, and you can tag Notion from your other AI tools. As the tools around you change, Notion stays the one place that holds your context. Notion is also hiring across product, engineering, design, storytelling, and more. Learn more at notion.com/dialectic.

This episode is supported by Google. Gemini 4 Argon recently launched and is built for long-horizon work. I used it inside Antigravity and gave it all 60 Dialectic transcripts: more than a million words. Instead of summarizing episodes one by one, Argon synthesized across them, tracing how ideas from early conversations connect to later ones. Together we made The Dialectic Compendium, an interactive book of the themes that run across the show, with a dossier on every episode and a 3D version you can page through on the web. Gemini 4 Argon is available in the Gemini app, Antigravity, Google AI Studio, and the Gemini API.

Compendium: https://jacksondahl.com/dialectic/compendium

Corrections:

  1. Periodic Labs (41:13): Patrick said “room-temperature semiconductors.” He meant superconductors.

  2. Yamanaka’s Nobel (1:03:44): Patrick said 2009. Yamanaka won the Nobel in 2012, for reprogramming factors discovered in 2006.

Timestamps

  • (0:00) Opening Highlights
  • (1:39) Intro: Patrick Hsu
  • (4:29) Start: Boldness, Infinite Games, & Ambition
  • (12:54) GLP-1s and Power Laws in Biotechnology
  • (20:17) Expelled at 13 and Falling for Biology
  • (25:51) What Makes a Good Biologist, Research Taste, and Biology’s Data Challenges
  • (37:58) The Arc Model: Full-Stack Science
  • (52:01) Betting on People, Style, & Learning to Simplify
  • (57:22) Perturbation, World Models, and Evo
  • (1:03:27) Virtual Cell Models
  • (1:09:56) Designing Life: From Stem Cells to Genome Writing
  • (1:18:39) Testing, Bro Science, and Human Data
  • (1:24:26) Can It Exist vs. Will It Exist: Science Meets Markets
  • (1:29:59) Biofuturism: Interoception, Peptides, Muscle Modulation, Sleep, and New Forms of Life
  • (1:43:47) Fun, Obsolescence, Going Slow, and the Future

Links & References

Show 49 more referencesShow fewer references

Transcript

4:29 Start: Boldness, Infinite Games, & Ambition

Jackson: (4:29) Patrick Hsu, thank you for having me. Thank you for joining me on the podcast.

Patrick: (4:32) It's a joy. Thanks, Jackson.

Jackson: (4:34) We are in the botanical garden that is also home. If people could see the amount of plants that are in my vision right now, it would be amazing.

Patrick: (4:43) And I water them myself.

Jackson: (4:44) You love life. What can we say? Impressive work. You have been prolific and you have also been prodigious. I think you went to college at 15, PhD was finished in your early 20s. Most of us are under-calibrated on risk and on boldness. I think particularly now, as technology just makes everything less capacity-gated and more like will-gated or even imagination-gated.

Patrick: (5:06) Mm.

Jackson: (5:13) You have a fervor and boldness that is really broadly impressive and especially suited for this time. I'm curious where you think that comes from.

Patrick: (5:23) You know, I grew up the dumb son. And so I always had to try really hard, right? I think the funny thing is I absolutely went to school young, had accelerated education, started working in a lab very early, started running one very early, and got very lucky to be part of important discoveries quite young. But this all came after an entire childhood of feeling like I had to really work for everything I've got. Things didn't come nearly as naturally to me, and I think that work ethic, combined with the sense that you have to really push for what you want, really shaped who I am today. That's the kind of thing that I try to teach my graduate students or postdocs when we're training superstar scientists. That's our goal here. As long as we're aligned on the big picture vision, how we get from A to B is going to be weird and murky. Research is 99% failure. It's a very weird profession because there are so many incredibly talented people. If you took my graduate class of 15 biology PhD students, none of them would have said that I'm the smartest. I'm certainly not, but I think I was one of the hardest working.

Jackson: (6:49) There are people who are incredibly hardworking who are maybe, to oversimplify it, sort of trapped in bounded games. They're climbing to the top of the corporate ladder or whatever. There are plenty of ladders to climb, and on some level, of course you are climbing ladders or you have climbed ladders. But there's also like an outside-of-the-box boldness here, whether it be on the explicit specific research side or even something like Arc. Clearly, yes, obviously you're very hardworking and very driven, but there's an element of, again, maybe—I don't know what the right word is—maybe boldness, maybe creativity. I'm curious where that part, and maybe when that part showed up.

Patrick: (7:32) I think I learned about the importance of infinite games pretty early on. The reason I went and did a PhD was because I wanted to do something really hard. I had no other good reason. I just felt I wasn't done with learning, I wasn't done with research. I wanted to go from cosplaying being a scientist and working on parts of a research problem and learning how to actually take an idea end-to-end to, at the time a paper, and I think now a drug. Your ambition increases and your vision increases over time, but back then I just wanted to learn how to do science. I think things like Alpha School are showing this really well, where for just two hours a day you can get done all of your academics. If you give students the right cognitive scaffolding, which is something that AI-assisted math training like Math Academy does really well, you can just speedrun a lot of things that normally take people many, many years into something that's 10% of the time or less. There's no real magic trick about it. This is something that anybody can do. I think I was able to, quite early on, speedrun a lot of things that have a lot of incumbency and a lot of "this is the way you're supposed to do things," or you're slowed down by the large batch size by the way that we do public education. I had something that was extremely personalized to me, being homeschooled until college. I think things like piano, things like science are examples of infinite games where, no matter how good you are, there's always another level. This is sort of the weird thing about research is all your very best ideas kind of suck, except the one or two that you're remembered for. By definition, almost everything you try will fail spectacularly. This breeds, by necessity, a humility and a criticality to take something, think about it super rigorously, and figure how to poke holes at every angle of your idea so you can succeed and get out of this web of no direction and actually find a way that you can actually ski downhill. Yet at the same time, that humility and wonder and doggedness needs to be paired with this supreme confidence that somehow you can do it. Otherwise, it just wouldn't be psychologically tenable to do this type of work in the first place. This is like deep insecurity plus sheer megalomania have to meld together in weird, complicated ways. You see this in ML research, you see this in biomedical research, and somehow we try to bring these two together at Arc. It's a fun experiment at the institutional scale.

Jackson: (10:39) You do that. You see that amongst entrepreneurs, of course, as well. And so I don't think it's that too much of a surprise that you've blended a bunch of those worlds. You said you're remembered for kind of like one idea. I don't know, it's funny to imagine—you've got a long career ahead of you, but it's funny to, if you were to try to predict what you would be remembered for, I don't know that it would be so obvious. There's probably a more interesting way of asking it beyond just like, how are you both prolific and really deep? Like how do you have breadth and depth at the same time? But like, I'm curious, and especially given the sort of like scope of like—I don't know, at some altitude you're thinking about broad, decade-level change and Arc, and change on the institutional level, and then you're also like, "We're actually going to do cell models in this very, very specific way." Is that just you're comfortable in both of those zones? Is that like a deliberate thing you've cultivated? How does that happen?

Patrick: (11:31) So in some sense I work three jobs.

Jackson: (11:33) Yeah.

Patrick: (11:34) right? And, you know, maybe part of the thing that I'm worried about is maybe I'm just going to be bad at all three.

Jackson: (11:39) Sure. Yeah, yeah, yeah.

Patrick: (11:40) But the way that I justify to myself both kind of intellectually and emotionally is I think these are all different facets of trying to solve the same fundamental problem, right? Like ultimately if I try to zoom out and think what do I care about doing, it's really like deep scientific innovation combined with capital and a sense of product in order to make something in the real world, right? And my own career and life has been a really kind of you know put my hand on the elephant attempt to try to discover what that process is, how it happens and how to make it more repeatable in our lifetime, right? And so in a way we're very lucky to live through some sort of really unprecedented technological and hopefully productivity revolution. And so when a rising tide lifts all boats, often times you just need to be in the right zip code, right? And I think I have been able to be in the right place at the right time at multiple points through the gene editing revolution, the AI revolution.

12:54 GLP-1s and Power Laws in Biotechnology

Patrick: (12:54) In many ways there's sort of a capital revolution that's happening as well and how you actually power a lot of this type of innovation. And so

Jackson: (13:01) You're saying specifically in regard to science?

Patrick: (13:03) Yeah, absolutely, right? So let's say in the 1940s, right, the vast majority of R&D spend was from industry. The, you know, NIH and a lot of our kind of post-war government funding apparatus became the dominant way by which we kind of funded and incentivized kind of biomedical innovation in this country. And then I think now today, we seem to be in a place where that kind of 50-year sort of default is now historically weak, and the sort of market share, right, of kind of corporate agenda and ideas and leadership of innovation is now ascendant, right? And that seems to be a trend that's likely to only continue, right? Eli Lilly, for example, is, you know, the largest biotech company in the world. It's valued about a trillion dollars. The, you know, GLP-1 revenue is going to be something like $90 billion this year, and they're going to scale their internal R&D budget linearly or along with their GLP-1 revenue, which they project over time will be half of the entire pharma industry. So half will be GLP-1, the other half will be everything else. Now this is interesting because Eli Lilly spends more money on R&D than the entire country of Germany—not just the government of Germany, but the aggregate national output of all German companies, right, in health and biopharma, right? And so by the early to mid-2030s, they'll surpass the United States if GLP-1 revenue continues to ramp the way that it has, right?

Jackson: (14:50) Sort of the power law we saw in software and technology happening in pharma on some basic level, or is that maybe that's an oversimplification?

Patrick: (14:59) The question is how repeatable this will be, right? GLP-1 is really special because it's so horizontal, right? Other medicines, while, you know, really important, generally work for a very specific disease, right? But targeting obesity seems to target so many other comorbidities and co-pathologies. Everyone's trying to figure out, what's the GLP-1 for sleep, right? For vitality, for healthspan. You know, that's I think ultimately why everyone's so intrigued and interested and skeptical about peptides. Like, why are peptides such a thing? Why does everyone ask me about them at dinner, right? It's sort of this idea that, you know, modern medicine has failed us. The drug industry has failed us. It's given all these really expensive, really targeted drugs that work only for very specific things with crazy side effects. And for the first time, we found that there's something that feels magical, right? Just the product feels magical. And we can see it, we can feel it. It's very tangible and visceral. It's not, you know, these weird pharma ads with happy people now able to stand up again and walk around, but you see all of your own friends, your entire friend group—or mine—is being shaped by this, right? And the idea that this could extend to, you know, other parts of things that plague humans or could enhance humans, I think is incredibly tantalizing. So this sort of proof of existence of biomedical product that can be for everybody, combined with this revolution in all of the AI models for biology, I think is creating today a massive wave of excitement for what geniuses in a data center can do when they break outside of the data center, right?

Jackson: (16:52) Is the implication there that you think there could be more GLP-1s? Not in the sense of more peptides for X, but more— That category goes from one to N in terms of these sort of like interventions that are broadly useful across a whole suite of categories.

Patrick: (17:09) One of the funny things about innovation is it's very obvious post hoc, right? And for example, you know, I was recently thinking about this, you know, with the four-minute mile, right? As soon as, you know, Roger Bannister ran faster than four minutes, so many people, like 10 people broke it within some very short period of time, right? As soon as, you know, we knew that, you know, you were able to make a nuclear bomb, right, other people were able to replicate it. And so both in scientific discovery and in quant trading, right, oftentimes just knowing that there's alpha and that someone is making progress in this general area, you don't need to know a lot of specific details. That's sufficient, right? And this has come up recently in the discussion between the frontier AI labs about who's actually solving the Millennium Prize problems and, you know, kind of how do they learn about specific mathematical approaches. I'm not well-versed in the merits of, you know, that kind of academic credit discussion, but I think the, you know, the fact that multiple people and multiple models are able to kind of converge on the same problem at the same time is something very kind of deep in how discovery happens. And I think that's, you know, that's something that's really interesting: that just knowing that this is possible creates the existence proof that gives people and markets somehow the courage and changes something about the plumbing of the universe to make it possible, right? And so I just think it is going to happen, right, because now people think it's possible. The prize is so large that it incentivizes a completely different way of thinking about how you create new types of medicines, how you design the experiments in the first place, how you run the clinical trials, right? Like a lot of companies are explicitly trying to attack this playbook of GLP-1 being tested first in diabetes, then obesity, then everybody. So now there's this blueprint of small disease where you can have very clear mechanism, bigger disease, then everybody. So you could go after narcolepsy, right, then you can go after sleep, then you can go after wakefulness and attention and nootropics and enhancements, right? And, you know, how do you kind of chart that path across multiple other things?

Jackson: (19:42) Yeah, it's funny. There's whether you—there's this view that it's sort of like there's a time for an idea and then it shows up, or it's as you say, maybe the right word is just courage or belief. But it's just—it's amazing. One of my favorite—it's not exactly the same idea, but one of my favorite versions of this is, I think it's like the fastest 100-meter dash. It's like obviously Usain Bolt has the record, and then like the next two guys, I think, were racing against him.

Patrick: (20:08) Yeah, yeah.

Jackson: (20:09) There's something—I mean, that's pretty inspiring in terms of what could happen over the next 25 years.

Patrick: (20:15) Absolutely, yeah.

20:17 Expelled at 13 and Falling for Biology

Jackson: (20:17) What—a very open-ended question, but what makes a good biologist?

Patrick: (20:25) I think I have a huge gripe with the way that biology is taught. It's incredibly boring.

Jackson: (20:34) You're saying taught in high school.

Patrick: (20:35) Yeah, taught in high school and college. The pedagogy of our field is terrible, right? And this is—it's just not compelling information. You're constantly—you're memorizing all kinds of weird esoteric stuff, and a lot of it just doesn't feel tangible the way that you take it in class, right?

Jackson: (20:47) Right.

Patrick: (20:58) And I got very lucky because I started working in a research lab in high school at a university. And I learned by doing, right?

Jackson: (21:07) So back up—that wasn't totally luck. How did that happen?

Patrick: (21:11) I emailed. So, when I was 13, I got expelled from college.

Jackson: (21:16) Expelled from college.

Patrick: (21:17) We moved down to Palo Alto.

Jackson: (21:19) So you were already in college?

Patrick: (21:20) Well, not anymore.

Jackson: (21:23) Wait, wait, I thought you started college at 15.

Patrick: (21:27) That was round two. Yeah, yeah, of course.

Jackson: (21:32) I was going to say, because I think I read somewhere your brother graduated at 12, and so you were—maybe that's incorrect.

Patrick: (21:39) He graduated at 15, and then he started his PhD at Stanford when he was 15.

Jackson: (21:44) But point being, you actually were on a similar track, and then you got expelled.

Patrick: (21:49) I was homeschooled on a mountain with internet, you know? I was just—I didn't have a calibrated sense of what normal people do.

Jackson: (21:58) Why did you get expelled?

Patrick: (22:00) I was hacking my professor's email.

Jackson: (22:01) Okay.

Patrick: (22:02) Yeah.

Jackson: (22:03) Okay, so you got expelled and then you were looking for—

Patrick: (22:05) I thought the world was ending at the time. Today I'm kind of proud.

Jackson: (22:10) Different. All kinds of 13-year-olds have different problems, I suppose.

Patrick: (22:14) Totally, totally.

Jackson: (22:15) Okay, so you got expelled and you're like, "I need something else to do. I'm going to email a bio..." Like, maybe ask it in a different way.

Patrick: (22:18) Yeah.

Jackson: (22:21) If we are bad at teaching biology, what clicked for you? What or what picked up for you that was like, "Biology is actually interesting. I do want to go work in a lab"?

Patrick: (22:31) So I don't have any doctors or professors in my family, and I didn't, you know, know that much about biology. But when I was 11, my grandfather got Alzheimer's and he was living with us at the time. And that was deeply moving for me, and it created—you know, the idea of decline and death became extremely visible and tangible for me in a way that, you know, normally you live hopefully a happy and healthy life surrounded by other happy and healthy people. There's sort of this saying, you know, that a healthy man has 99 problems and an unhealthy one only has one, right? Sort of kind of becomes like that. And that's sort of what made the problem feel relatable to me as an 11, 12, 13-year-old, and kind of, you know, got me into into the lab. Yeah. And I've never stopped. I've never stopped since. Yeah.

Jackson: (23:23) Okay, so we aren't great at teaching biology. You get into a lab and you get a relationship to it that makes it way more exciting.

Patrick: (23:32) Yeah, yeah. It's project-based learning where you're actually solving a real problem. That's really what was intoxicating for me. It wasn't just following some syllabus of book learning, but it was real research, right? So I did high school science fair, Intel ISEF—now Regeneron sponsors it, Regeneron ISEF—and I got selected at Intel ISEF for this experience where you got to meet Nobel laureates, right? And it was like, you know, Dudley Herschbach, who won the Nobel in Chemistry, and, you know, kind of Richard Roberts, and, you know, a whole bunch of other epic legends, right? And so I was like, "Wow, I'm going to meet these science gods," right? And I remember telling them about my project very enthusiastically and wet behind the ears, and telling him, you know, about one of the research problems I was trying to solve and trying to get his advice. And he kind of looked at me and asked me a question. He was like, "What do you think?" And that was crazy. That was kind of deeply moving for me because it kind of made me realize intuitively in that moment he had no idea, right? That this science god knew as much or, well, probably in retrospect much less than I did about this problem. And I was like, research is an even playing field kind of by definition. No matter how great or accomplished you are, we all have the same shot. And that made me feel so motivated. And yeah.

Jackson: (25:14) There's sort of open territory on the map.

Patrick: (25:17) Exactly. And actually most things in the world are like this. Like the reason why you're driven to found something, right? It's fundamentally an act of desperation of why has the universe, you know, not created this thing that is so obvious to me that this should happen. But that's just actually how combinatorics work, right? Most things in the world haven't happened, and so the kind of act of research creation or the act of founding something tap into this same kind of energy, this sort of will of existence.

25:51 What Makes a Good Biologist, Research Taste, and Biology's Data Challenges

Jackson: (25:51) Yes, it's very easy for us to underrate them when they're in unbounded space because we can't really imagine. I guess that speaks to like the interest in biology and in the opportunity. What—I'll ask again, though—what makes a good or what has made you a good biologist?

Patrick: (26:06) So sorry, I'll actually answer the question. I think there's a few primitives. The first is you need analytical depth. I think that's maybe the most important one is you need to be able to think very clearly and logically and objectively, and harness your mind and your attention to detail and your sharpness around pattern matching and knowing when something looks off pattern, right? I think it becomes something that you train to the point it becomes deeply intuitive, right? So like doing biology research is very different from, you know, math or physics.

Jackson: (26:49) I was going to say, is—not knowing much and certainly not being a scientist—that sort of sounds like something that would be useful across science. Do you think this is very distinct, or not only a biology lab?

Patrick: (26:59) I think the way that you deploy it in biology is different because it's fundamentally a non-quantifiable or not well-quantified type of sort of sport, right?

Jackson: (27:05) I see.

Patrick: (27:11) There are very, like, biotechnologies have been invented ground-up to try to make sense of these weird, miniature, kind of atomic-scale structures and dynamics and interactions that create life. And, you know, back to Robert Hooke pulling glass, right, or folks trying to figure out how to image these microscopic microorganisms, how to extract genetic information from them and amplify them exponentially, you know, to create these programmable molecules that can go inside of complex cell environments and turn it into a digital code so we can manipulate them in living cells. Like, biotechnology is weird and crazy. And so we kind of try to compress deep complexity and dynamics into something we can measure. And that's like very lossy, very low-dimensional characterization of something that's very complex. And I think one thing that I'm super bullish about is just the increasing scale of what we're able to measure and have really kind of high-content understanding and multi-parameter measurements of what biology is, to start to look at it more holistically rather than compressing it into kind of individual vectors, so to speak, which is what we've been doing for most of the last 50 years of biology. And, you know, so I think you need that taste that comes from being able to have this intuitive model of what the data is telling you about a system and have the analytical clarity to be able to navigate in and around and out of this maze. And you need to be extremely resilient. You have to be really psychologically, it's like irrational to, in the face of just constant failure, still think you can make it work. So you need persistence, you need analytical depth, you need conceptual creativity, and I think some sense of style.

Jackson: (29:25) Yeah. Yeah. A bunch of questions actually on that. One is in this world where sort of one of the meta-skills here is identifying, like, breaks in the pattern. In the world we're going in biology, in a large part thanks to work you have been a part of, we're getting much more holistic models, way more data, way broader tooling. Do you think that skill becomes less important in that world?

Patrick: (29:53) Which skill?

Jackson: (29:54) This sort of, like, intuition, anti-pattern.

Patrick: (29:58) I really struggle between the modes of being long-human and long-AGI. And so I feel kind of insecure whatever I say now is going to become wildly dated.

Jackson: (30:12) Maybe a better way of asking would just be, like, what have you seen in the last two years? Like, is that skill persisting? Is it moving to a different part of the maze? Like, is it becoming more important? Is it becoming like, "Wow, actually a lot of this, a lot of the things that maybe previously would have required a certain research taste or intuition are more evident in the model now"?

Patrick: (30:31) I would say research taste and intuition is only increasing in importance, in relative importance. Like, I would buy call options on it if I could, right, in terms of how important it's going to become. It's just biology is so high-dimensional and the models today need so much inductive bias. And so we just, we need a lot more kind of creative architecture and data. So if you take something like math, right, where we're able to solve Millennium Prize problems today, right, we have this verifiable problem and such to the point where we can unleash, like, 10,000 agents to simultaneously try to solve a problem from every possible angle. This is not possible in biology today. First of all, we can't token-max in this way because the agents fundamentally don't have enough context. They don't have the right evals, they don't know enough. So I think biology over the next decade is going to be a discipline of context-maxing.

Jackson: (31:37) Is that just a scale problem, or do you think that doesn't go away if you had 10x the scale?

Patrick: (31:42) I think scaling naively will be really inefficient, but it will work.

Jackson: (31:46) Okay. Which has, I guess, broadly been true in language models too.

Patrick: (31:50) Yeah, definitely. Yeah. I mean, there's so much work on data-efficient learning and, you know, continual learning and all of that that may pay dividends over time, but we just have such crazy, unprecedented scale of data in text and image modeling that we just don't have in biology. And so there's a huge amount of interest in trying to figure out what is the demand for biological data, what data types should we be creating, what does the Scale or Mercor or Surge for biodata actually look like? And, you know, wearing my investor hat, we think a lot about the demand versus supply side of that equation, but, you know, fundamentally I think it's obvious that we all just need more and that we kind of need to build complexity up, right? So there's a lot of interest in talking about digital twins, right, and predicting drug toxicity, right? But for me, in many ways I worry, you know, that starts to only work when you have something that's a whole, complete world model of biology, whatever that means, right? And if you can't even model what happens in a single cell, how can you model multicellular systems and then kind of organismal-level things like whether or not this drug is going to be toxic, right? You need to be able to model every organ and how they respond and interact across them. And so it will... I'm very AGI-pilled on one hand, but I think we need to figure out how to create the right evals and the right context for these models to be able to kind of work on these kind of, you know, semi-verifiable problems.

Jackson: (33:24) It sounds like you're saying there's... while there's certainly a data problem—like if we had whatever, 10, 100, 1,000x the current datasets, which I know you guys are also working on, it would be a lot better—and maybe there's something else here that is true in biology that doesn't perfectly analogize from language or math around evals. I almost wonder if on some level—this is probably a poor analogy, but I keep talking with friends about it's just amazing, like, I don't know, Astra can do Blender and all these things and yet it still can't write well. Part of that might just be like there's not a huge economic incentive to write well, but maybe there's something else about the verifiability of good writing. And maybe there's something else here, or put a different way, do you think if you just had 1,000x the datasets, we would make the same progress we've seen in AI broadly over the last few years?

Patrick: (34:18) I would say yes and no, depending on the type of modeling objective that you have. So for example, if you ask a machine learning person what type of biological data do you desperately want, one of the major categories would be time-series data,

Jackson: (34:34) Okay.

Patrick: (34:34) right? And biology does not tend to have great time-series data, right? We'll have time points, right? We can take a time point at 24 hours, at 48 hours, or if it's a patient, right, you can kind of look before and after, but, you know, you're not able to... Wearables are one example of being able to track really low-content things with really high resolution, like heart rate or skin conductance or glucose or whatever, but it's sort of obvious in some future we want really multiplexed, multi-analyte, continuous measurement: hormones, dopamine, serotonin, cortisol, testosterone, progesterone—all of those things you would want to continuously monitor. Similarly in genomics, you want to be able to measure gene expression profiles or protein expression profiles at the single-cell resolution across time. Our current technologies don't do that. We're looking at cell sets and averages.

Jackson: (35:37) There's so much lower fidelity.

Patrick: (35:39) Exactly, exactly. And so there's like a fundamental need for biotechnology invention that creates that shape of data to train that type of model.

Jackson: (35:50) Right, right, right. And is that a hardware problem? Is that a, like, are there like biological understanding? Like what is the sort of shape of that type of technology? I'm sure it's a super, super wide set, but—

Patrick: (36:04) So we have a number of lines in the water at Arc to try to invent those capabilities, right? There are a lot of other brilliant labs and investigators who are trying to do some similar, some distinct approaches technologically. So, for example, one of our brilliant faculty, Felix Horns at Arc, who's also faculty at Stanford, has invented these nanoparticles that can be budded out of cells, right? And you can engineer them to basically take a hand, if you will, and grab RNAs from inside that cell and then bud them out of the cell.

Jackson: (36:44) These are artificial and these are biological?

Patrick: (36:48) They're engineered proteins, right, that create a nanocage and sort of export them from the cell. So it's sort of like sending out a FedEx package, you know, every few hours, right?

Jackson: (37:00) Just sorry, a random aside, I suspect you and biologists broadly have to get really good at analogies.

Patrick: (37:06) Yeah, no, teaching definitely involves trying to make things visceral and relatable. And so, yeah, you have to create this FedEx sort of center where, you know, the house becomes the place where the things that you want to measure, you ship them out. Current technologies involve destroying the cell, harvesting its contents, and then that cell's dead afterwards. So you don't know what happens next because it's dead. You can only know what happens next to other cells that were similar, right? But if you want to know what happened in that exact cell, right? You need something that's non-invasive, right? And doesn't kill it. And so this type of exporter technology can allow you to take non-invasive multiple time points with deterministic single cell resolution, right?

37:58 The Arc Model: Full-Stack Science

Patrick: (37:58) And so that's really why we felt something like Arc needed to exist. No individual research lab or company really has both the ability and the time scale to work across the invention bucket, the engineering, and then the scaling bucket. So on one hand we're scaling an internet scale data set of perturbed single cells and are building the largest sort of repository of this in the entire industry, right? Yet on the other hand, we also need to invent the technologies to create a fundamentally new way of measuring this because we understand the fundamental limitations of what you can do with scaling perturb-seq today, right? And so on top of all that we want to train the models that allow us to capitalize on that data. So you just need something deeply multidisciplinary across, you know, biology, machine learning, the engineering that you need to bring all these things together. That's quite rare. It's sort of funny because I think today it feels like an obvious model and, you know,

Jackson: (39:00) The Arc model.

Patrick: (39:01) Yeah. But, you know, when we started Arc, there was so much deep skepticism that it could possibly work. Yeah.

Jackson: (39:11) How much of that last statement was on some level being right around the timing or fortunate with the timing of AI broadly and ML broadly?

Patrick: (39:21) You know, part of our initial thesis of Arc was that we needed much more computation and much more scale in biology. And we also had sort of an institutional bet that we needed to have our own institute that was still deeply partnered with universities, because they have sort of the lifeblood of innovation in biomedical research, but these organizations are large, sprawling bureaucracies that and the current grant funding system doesn't really unlock our best investigators to do their best work. We put those things together. And, you know, figuring out the right cultural ingredients and talking about it and writing blog posts about it is very different from actually operating in real life to actually do the thing, right? And we're still very Day 1 at Arc, but we always said that the long-term success of Arc is that there are many more things like it. And I think we are seeing that across both our earlier Fast Grants work on rapid science funding and the Arc model. There's a lot of interest in propagating this model forward, which I think is kind of the ultimate compliment.

Jackson: (40:38) And the simplest articulation of that being—and by that I mean more Arcs being this sort of multidisciplinary angle on it, the combination of scale and computation and engineering—what is the Arc-iness that, specifically if you were, I don't know, and maybe there have been over the last five years new things that are cropping up and you're like, "Okay, we have evidence"—what is the sort of atomic unit of that?

Patrick: (41:06) I think this idea that you need something that's kind of full stack in being multidisciplinary. Yeah.

Jackson: (41:10) That makes sense.

Patrick: (41:13) And you see other efforts that have very similar ingredients that, you know, like Periodic Labs, for example, is trying to really close the loop between machine learning and materials science—material design and innovation, making things like room-temperature semiconductors or new semi materials. And, you know, I think that's an example of something that we have the right ingredients to make many more Periodics or Arcs today. And I'm so deeply excited for, you know, seeing more of those, yeah.

Jackson: (41:53) It doesn't seem super obvious. I mean, I can think of a bunch of reasons why academia would struggle on various dimensions of this, but the multidisciplinary part of it doesn't seem so obvious to me. In some sense, it seems like, I don't know, Stanford or whatever would be like a great place for multidisciplinary things to happen. What is happening structurally that is causing these things to be so siloed?

Patrick: (42:13) Top down, Stanford in principle has all of these departments, right? They have a computer science department, they have a biochemistry department, et cetera, et cetera. They have all the capital equipment, all the human talent, right? But it's first of all spread across an entire campus, in multiple parts of a campus. And the sort of individual atomic unit of research, which are individual research groups, are incentivized to generally do their own thing, right? And of course there are interdisciplinary labs, but they'll usually be able to work across one domain boundary, right? But if you have to do three or four, that kind of exponentially declines. So I'll give you an example, right? We spent the last five years at Arc building out this really scalable platform for making brain organoids where you can take human embryonic stem cells and turn them into the three fundamental cell types of the brain: neurons, astrocytes, and microglia. We can grow them in 2D and in 3D to make these mini-brain models. There are entire labs dedicated just to making these mini-brain models, which is sort of like cell biology and neuroscience and developmental biology. So that's sort of like three different fields that come together to make this possible. Now we kind of add on top of this the sort of ability to do functional genomics, right? Where we kind of have state-of-the-art methods for doing genetic perturbations in these, gene editing them, being able to put in different genetic risk mutations from Alzheimer's because we're trying to model Alzheimer's and then do this at unprecedented scale. Where people generally make a few organoids at a time, we're able to make thousands, right? But then just doing that alone is very slow. And so we're now bringing in kind of like frontier lab automation, right? Where you take humanoid arms, right? That are able to orchestrate the workflows. And so there's sort of two different types of lab robots today. They're like box robots, similar to CLIA COVID diagnostics, right? They're really brittle and very scalable for one specific workflow. But most biology research you have to chain together like 20 different primitives in order to do an end-to-end complicated project. And the lab automation has sucked for decades. People have wanted this to create these cloud labs or whatever, but they just don't work. This is starting to work now with kind of agentic physical intelligence.

Jackson: (44:48) You have that in the lab now?

Patrick: (44:50) We're deploying these in the lab to automate all aspects of this work: the cell culture, the genomics, the genetic perturbations, the profiling, the sequencing, really making these end-to-end data pipelines. And even if you're a top lab that does the functional genomics and the cell biology, you won't have the machine learning part. And if you have the machine learning research part, you won't have the engineering and infrastructure ability to do distributed node pre-training, right? The ability to do RL. All of those things have to come together, right?

Jackson: (45:31) It sounds like you're describing some of the ways that Arc works as being more like a company. And again, maybe that's just you're more flexible, more dynamic, whatever. I'm curious if that's broadly true. And then second, I guess, why are you partnered with—I think it's UCSF, Berkeley, and Stanford? What is the reason for the academic collaboration, and how do those two things work together?

Patrick: (46:00) One of the really interesting things about companies and biotech, in my experience anyway, is they tend to get frozen in time relative to the time of their founding, right? So they'll often have the right to be founded because they come on the back of some technological innovation or inflection point where you've invented some new way of making molecular glue or whatever and you start the company. And then five, six, seven years passes and that's your platform. But other people have invented a whole bunch of other things that can be faster, more scalable.

Jackson: (46:39) By the way, this is true for all companies.

Patrick: (46:41) Some reinvent themselves.

Jackson: (46:42) Right, right.

Patrick: (46:43) Yeah, but yeah, and I think, at least in biology, right, the way that the system is set up is the kind of basic science labs have to, are sort of at the frontier of pushing for innovation, trying really crazy new stuff that are in the invention bucket. They're not in the engineering or scaling bucket, which is what industry really needs to specialize at, right? And so we're in many ways sort of in the first third of the equation of trying to really be at the fountain of discovery and innovation and then bring them from the invention to the scaling bucket, right?

Jackson: (47:25) So it's sort of that if academia only has the first thing, you've sort of like clipped off a little bit of the next phase and integrated it at a kind of like foundational level that makes it easier to go into. Is that a—

Patrick: (47:37) That's our hope. That's our hope, right? Yeah. And then our hope and belief is that this will be an Edison lab for many companies over time, right, on the basis of technology or insights that are much more well-validated and ready to spin out at the right time, yeah.

Jackson: (47:57) And then the integration with the universities, is it just sort of like you still want to be adjacent to either the pedigree or the talent or the inventiveness or—

Patrick: (48:08) In many ways it's talent capture, right? I think in our industry, there is this—it's sort of this annually renewing resource for some of the best global talent, and it has insane pull. You sort of zoom out and say, "What makes the Bay Area great, right? Why is it the birthplace of the AI revolution, right?" It's because all the AI PhDs are here, right? And they trained here, and you can combine the AI with the engineering to actually make it click, right? And I think for biomedicine it's also very similar, and we're combining both of those. So then that's why we can also only be here. And so it's the innovation and the kind of interdisciplinary reach across all of these domains that the universities have had historically, and then putting all of those ingredients into this sort of new organization that's able to really kind of operationally unlock all these people to work on integrative problems together. Like we ultimately have two flagship problems. We're not trying to solve 50 problems, right? We're trying to find drug targets for Alzheimer's; we're trying to make these virtual cell models.

Jackson: (49:31) Has that been the case since the very beginning?

Patrick: (49:34) We started Arc—the labs came online in July '22, right? We were working on it for a year and a half, two years before that. So this is pre-ChatGPT, certainly. We certainly believed computation would happen, but we didn't predict the—I mean, I don't think anyone really predicted the AI revolution in quite this way. And so I think we were again at the right place at the right time with the right intuitions and the right ingredients, but then you have to try to meet the moment.

52:01 Betting on People, Style, & Learning to Simplify

Jackson: (52:01) You mentioned talent. There's a couple of excerpts from, I think it's the very first Arc blog post. First you say nearly 80% of Fast Grants—which is what you guys had been working on in COVID to kind of bootstrap some quick, not solutions, but quick progress, and be a little less bureaucratic, let's say—nearly 80% of Fast Grants recipients said they would change their research program, quote, "a lot," if all of their existing funding could be spent in an unconstrained fashion. And then later, "The funding for these labs, which we envision will grow to 10 to 20 people each, is completely flexible because we are betting on people rather than specific project ideas. By pre-funding our scientists, we hope to avoid the Gordian knot of incentives that currently entangle grant funding success, single author credit, high-impact publications." Obviously this is a really interesting frame. What makes for a good, sort of like entrepreneurial scientist? It seems to me that that is like one of the most—for the sort of people certainly leading Arc labs, but maybe everyone there, there's like an element of this entrepreneurialism that might not be obviously there in the broader academia. And I'm curious how you guys have sort of selected for that, especially if you're going to trust somebody for an eight-year term to work on whatever they want, I suppose.

Patrick: (53:17) I think there's an aspect of your question, which is: how do you select a good founder?

Jackson: (53:21) Yes.

Patrick: (53:22) Right? And in a way you expect these top scientists to be founders of what hopefully becomes a field, right, or a fundamentally new way of thinking, right? And so you need multiple qualities beyond—like everyone is very technical, right, and spectacularly technical, but you need so many other qualities beyond that, right? You need to be able to convince and coax and give good talks and recruit people. You need to be not just the CEO or the Chief Innovation Officer, but you also have to be the Chief Scientific Officer, the Chief Marketing Officer. You know, this is also a business in a way. You need to be able to, against your top-line revenue, decide how you're investing your resources into different areas.

Jackson: (54:14) Right. You guys solve some of that, what you're saying, "We're going to solve the fundraising for you for a while," but even so you still have to allocate.

Patrick: (54:21) You still have to do BD. You still have to figure out how to bring your research program together with others to make something that's a much bigger collaboration than what you can do alone. I think you kind of try to distill those very similar primitives into investigator selection, right? I think at the end of the day, you also just—there needs to be a sense of style in the work, right, that the world with the innovations that these people create will, I think, just fundamentally enrich science in a different way. And I don't know if we've done a perfect job at this, but that's at least been our hope, yeah.

Jackson: (55:02) You mentioned the word "style" just now, and also earlier there was a comment you made somewhere on some podcast about grad school being one of the few places where you still have this kind of traditional master and apprentice model, as maybe the rest of education has gotten very scaled. I'm curious how specifically you benefited from that under Feng or otherwise, and how you guys try to implement that at Arc, maybe specifically in this direction of like style, or research taste, or creativity, or whatever you want to call it.

Patrick: (55:35) I think one of the most important things I learned from my PhD advisor was how to simplify. Biology is really complicated and most projects and teams get mired in endless complexity, right? There are just too many variables. It's hard to decide how to prioritize. You can talk yourself out of any idea, right? It's very easy to be too smart, right?

Jackson: (56:05) This is Feng or this is—

Patrick: (56:07) Yeah, Feng has such a unique talent in distilling, right? And I got to learn and see and do over that period of time, and I think that was something that I feel really grateful to have studied under him and learned, and try to implement in our own group and at the Arc-wide level as well, right? The concept of finding drug targets for Alzheimer's sounds, in a way, enormously complicated, right? And maybe it is, right? But we have a very specific scientific strategy by which we're trying to causally map all of the human genetic risk mutations and map all of those mutations to pathways to try to mechanistically understand the initiating steps of the disease to then find those drug targets, right? As opposed to the traditional way to do it, which has been: record what's going wrong in these old patient brains—they have high amyloid, they have high tau, let's drug those things. We're trying to go back to the mechanism. I think mechanistic understanding of disease is really going to be our most fundamental way out of this crazy maze, you know.

57:22 Perturbation, World Models, and Evo

Jackson: (57:22) It's a little more maybe like top-down, like coming over the top with a sort of a hypothesis or a theory of things rather than just being purely observational? Or is that—am I misunderstanding?

Patrick: (57:37) Yeah, we're big believers in perturbation, right? Most biotechnologies are observational. You image something, you sequence something, you're just trying to figure out what's happening, right? You're just kind of taking pictures, right? So there's a very interesting company called General Intuition that's building world models, right? And it's trained on a supreme amount of data.

Jackson: (57:59) Is this Pim's company? Okay. Funnily enough they spun out of Medal, this gaming—

Patrick: (58:04) Exactly. Exactly. It was gaming videos where they would do screen recording of kind of pro gamers, and they would have a keylogger where they could record every sort of keypress. So they would be able to map the action space into the frames, right? So they can go frames in, actions out; actions in, frames out, right? The really amazing thing is they can take real videos, just normal videos like in real life, and then they can pass it through this model that's frames in, actions out and then create an action space that would have created this video—

Jackson: (58:40) Oh my gosh.

Patrick: (58:40) —if you had been playing it as a video game even though it's a real-life video, right? And you can use this to control a drone, a robot, you know, something dextrous. I think it actually starts to get really trippy when you have that ability to have those labels, which you wouldn't if it—it would be much more inefficient to train such a model if you didn't have the labeled actions, right? And so in many ways, perturbational data in biology is having the labeled actions for what happens before and after, as opposed to the majority of observational data that we have today. That's been in many ways our fundamental bet across our institute initiatives, yeah.

Jackson: (59:18) I see. Yeah, I see. How does that frame connect to sort of—I mean, the two, at least as I understand it, a lot of the work is either models or interventions. Can we get way better data and thus build way better models? And then what can we do on the intervention side? On the model side, the two big things that at least you guys have talked a lot publicly about are the Evo models and the cell models. Is that perturbation sort of frame inside of both of those? I think I've heard you specifically talk about it on the cell model side, but is that the sort of working frame for how you do all of this work?

Patrick: (1:00:00) In a way, yes. I mean, obviously Evo's trained—or the current publicly released Evo models are trained—on evolutionary distributions of DNA.

Jackson: (1:00:11) And maybe it's worth saying Evo 2 is the biggest bio model basically ever created.

Patrick: (1:00:16) For now. For now, they're only getting bigger.

Jackson: (1:00:18) Bio foundation model, maybe, is the right frame for it, right?

Patrick: (1:00:21) Yes, and we're training bigger ones, as are others. And the goal is not to just have DNA language models alone, right? They need to be combined with these cell models because, you know, you learn in high school, genotype and environment create phenotype, right? And so models can't be of cells alone or cell states alone without understanding their genotype. For example, we know different cancer driver mutations will create cancers in different organs, right? The mutations that create pancreatic cancer are different from the ones that create breast cancer, even though both cell types have the same genome, right? And so cell state is mediated, in part, by inputs from the environment, but then also inputs from the genotype. And so these model paradigms need to be brought together in order to create—people talk a lot about using these AI analogies in biology. There's lots of talk about world models of biology, right? What the hell does that mean? I think for me, at the end of the day, if you think about it from a molecular or genomics point of view, you need something that goes from sequence to function to state, right, in order to predict how cells will respond. Sequence means genomic—genomic sequence, right? To how the sequence gets interpreted epigenetically to create RNA and protein dynamics, right? Which then creates some sort of functional consequence, right? Like ultimately you're not just trying to predict a genomic profile of a cell, but also if this cell is going to be more effective at killing this tumor, right? If it's going to become—

Jackson: (1:02:01) You need a testing environment.

Patrick: (1:02:02) Exactly, exactly. So the lab, you know, is the place where you can do active learning, right? Where the model will make predictions zero-shot. You can test those in the lab, feed them back into the model, and then improve the model's predictions in multi-round loops, right? And so really—

Jackson: (1:02:21) Sorry, just to forgive my ignorance here: this is the Evo-style genome models, or cell models?

Patrick: (1:02:28) So I'm talking about this sort of general paradigm of lab-in-the-loop, right? That will work both for genome models and for cell models and then for the jointly modeled space, right? And we need to do both, right? And this extends all the way to clinical perturbational information, right? Like people are training data on electronic health records, right? Where, you know, someone took a drug and then you have their EHR before and after, right? And then people try to use this to create like patient embeddings and patient clinical trajectories and try to use this to predict, you know, like what you should be doing for certain kind of patient classes. This is like very, very coarse, right? And I think gathering the right type of clinical perturbations with kind of genomic resolution information is going to be one of the most important outcomes for really like a new type of medicine or health-maxing, right?

1:03:27 Virtual Cell Models

Jackson: (1:03:27) Is there a threshold you sort of have in mind for cell models that is sort of—that either has been crossed or that is in the future—that allows them to be a good enough proxy for some level of like very material progress?

Patrick: (1:03:44) Yeah, yeah. So in 2009, I think Shinya Yamanaka won the Nobel Prize for discovering the reprogramming factors, which have been very popular lately for epigenetic reprogramming in longevity, right? And he basically discovered when you turn on four transcription factors, you can take any cell type in the body and turn it into a stem cell, right? And that stem cell is universal and can be turned into any other cell type in the body, right? And that's, you know, traversing this sort of Waddington landscape was so powerful because you can do like epigenetic resets or a cell identity reset, right? And folks have been trying to figure out how you can use this in aged cells to make them, you know, kind of transcriptionally or functionally younger, right? And you know, that would be like an eval, a Nobel Prize-winning eval from, you know, 15 years ago, right? Can a virtual cell model rediscover Yamanaka's reprogramming factors, right?

Jackson: (1:04:48) I see.

Patrick: (1:04:49) That would be one really tangible example of that's evaled on ground truth, right?

Jackson: (1:04:52) Yep.

Patrick: (1:04:56) And the question is if you can find such an important discovery from, you know, a decade and a half ago, then what are the prospective ones that you can make moving forward? But you need to be able to recapitulate like the most important discoveries of human-labeled biology in order to trust the machine ones, you know.

Jackson: (1:05:14) Yeah, again, probably oversimplifying, but that sounds actually a lot like a lot of the eval stuff in AI broadly. What is the current state of the sort of cell models now and maybe what is the rate limiter on progress there? Is that also largely a data problem? Is that just this kind of broad complexity of scale and compute and a million other things? Is it—I mean, I know you guys have this, I think it's like a billion or at least on track to a billion perturbed cells. But like what is the right way to think about progress there?

Patrick: (1:05:43) It's a good question and I think the short answer is it's all of the above, right? It's data, it's model architecture, and it's evals, right? And all three need to kind of co-evolve together to bring us to really pragmatic benchmarks like being able to simulate in a computer all Perturb-seq data, right? That's sort of our explicit goal for the models that Arc is building, at least in this current phase of our kind of virtual cell modeling work, right? And we've created this Virtual Cell Challenge that's sort of like a CASP or Kaggle-like competition where I think this is the second year we've done it. Over a thousand teams have submitted their own virtual cell models, right? And this is a way bigger community than just this very small and narrow bioML community, right? There are all these blogs that have tried to translate the bio stuff into something that engineers can understand. So there are just sort of engineers at tech companies all over the world who are trying to train and test their own virtual cell models against kind of internal eval data that we've generated in-house in our labs, so there's like no hacking of it, for example.

Jackson: (1:07:00) And sorry, how do you win that competition? What is it—it's just about the overall quality of the model?

Patrick: (1:07:07) Yeah, so we have sort of a variety of different evals that try to score the model's ability to make predictions for how a cell will respond after perturbation, right, in quantitative metrics. And you know, it's a public leaderboard. It's at virtualcellchallenge.org, but our goal was, you know, to ask the question of how well do virtual cell models work? Where, how quickly are they progressing? Not just in our labs, but this is an industry-wide problem and kind of a holy grail in the field that we should have this public, open, and kind of transparent evaluation of all of the models that everyone's trying. So it's not just overfitted on the two or three specific evals that they're testing in their paper, but it's, you know, across a variety of metrics and it's all like, you know, out in the open.

Jackson: (1:07:57) What's the right way to think about how you guys use and benefit from sort of state-of-the-art general AI models in this process? And what the lines are—a super unsophisticated view might say learning from the Bitter Lesson and a bunch of the scaling stuff is actually just like whatever GPT-9 is just going to be the best at everything including bio. From everything I can sort of loosely understand that seems like decidedly not the case. But what is the role beyond just speeding up workflow and everything else?

Patrick: (1:08:26) So Periodic recently released their own kind of open-source mid-train models on their own internally generated lab data for material science stuff, right? And then they compared it to Astra, right? And they were basically able to show that, you know, their own custom model is state-of-the-art and it's half the inference cost of Astra. But Astra sort of kind of max thinking is just below their, you know, kind of custom data-specific model, right? I think that's itself very interesting, right? And so the short answer is I don't know. Is—are all kind of domain-specific or data-specific models going to just become tool calls for these sort of, you know, kind of super agents? Or do you have to take that data and add it into training these big models across a variety of different domains? That's playing out in real time, right? But I think we're kind of agnostic to that future since end of the day, I care a lot about the capability existing so that we can make real progress—

Jackson: (1:09:29) Yep. Yep. Yep.

Patrick: (1:09:42) Yeah, make real progress on science, but then also on the products that will become unlocked in the tech tree of creating that scientific kind of product or discovery. Yeah.

1:09:56 Designing Life: From Stem Cells to Genome Writing

Jackson: (1:09:56) You have this comment somewhere where you said—I think it was maybe with the Sequoia folks on a podcast—getting to the point where a model makes a cell biologist feel emotion.

Patrick: (1:10:01) Hm.

Jackson: (1:10:06) Can you say more about what that means?

Patrick: (1:10:08) Yeah, I think take these Yamanaka factors, for example. You can physically—and like I have done these experiments myself with my own hands, right? You can take cells from an animal or from yourself, right? You can take your own like fibroblasts and turn them into induced pluripotent stem cells in a dish. It is one of the most kind of profound and magical and semi-religious feelings I've ever had, right? And the idea that a model could rediscover this and then find things beyond it, right? I think it is sort of when you start to, you know, in the same reason where like when you're watching progressions of image or video gen models from Runway or Genie or World Labs or whatever, they're just making these—every generation, the outputs of these things gets crazier and crazier, and you have all these holy shit moments. How do we have those holy shit moments in biology?

Jackson: (1:11:07) Do you have many of those currently?

Patrick: (1:11:10) Do we have many of those currently? Well, we've created them. We've created some of them. I think we recently had a paper from one of our kind of Arc and Stanford colleagues designing entire organisms, right? And sort of the first time that AI's been used to create entire viable organisms, right? And I think the idea of being able to do this not just for really small genomes, but for, you know, things like mammalian or eukaryotic genomes over time is like a really fundamentally new paradigm for designing biology rather than just observing it. Yeah.

Jackson: (1:11:55) One of the sort of tenets at least of your lab at Arc, and obviously so much of the work you did very early in your career, is on what you would call programmable biology. In that vein, I remember first hearing about CRISPR, whatever it was, about a decade ago. There's been a lot of progress in the broader programmable biology side of things. And then there's also maybe part of what you were I think maybe just gesturing at is also like synthetic human biology. I'd love just like an update on kind of your thinking on programmable biology in terms of where we sit now in late 2026, and then how the sort of like recent progress on the technology side has ramped that, whether it's informed by the model stuff or otherwise, like—

Patrick: (1:12:43) Yeah, yeah.

Jackson: (1:12:44) ’Cause I think that when people think about sci-fi and biology, like this feels like the kind of the big thing. And obviously there's been a tremendous amount of progress in the last decade on the CRISPR side and RNA and otherwise, but like it doesn't necessarily—at the very least doesn't seem to be coming to me as a layperson in this like very vivid way.

Patrick: (1:13:06) Yeah. I think first of all taking these machines and putting them in a human body has been really hard. This is sort of the delivery problem in gene and cell therapy.

Jackson: (1:13:13) Got it.

Patrick: (1:13:21) But there's also been a design problem fundamentally which is these tools or the kind of first generation of these kind of the kind of CRISPR revolution has been around fixing individual broken genes, right? That lead to a monogenic disease like sickle cell anemia, cystic fibrosis, muscular dystrophy, and so on. And so that's sort of like one problem where you've had to create a different CRISPR product for every different genetic mutation, right? And so you're making very small precise targeted changes, but end of the day like you couldn't program things—

Jackson: (1:14:02) Doesn't feel like GLP-1s, maybe, right now.

Patrick: (1:14:04) Right. Right. You want to make much larger-scale changes. And so a lot of the work that we've been doing in my lab lately has been developing large scale genome editors beyond CRISPR known as bridge RNAs and bridge recombinases that allow us to change up to a million bases in the genome at a time. The problem is once you start to be able to—

Jackson: (1:14:22) What is—a million bases is what kind of scale?

Patrick: (1:14:25) Uh, megabase, right? So, you know—

Jackson: (1:14:26) But like as a percent, I don't know, like—

Patrick: (1:14:28) Well, the human genome is extremely large. But a bacterial genome can be, you know, three megabases, right? So it can be a third of an entire bacterial genome, for example, or, you know, kind of like two, three, four percent of an entire human chromosome.

Jackson: (1:14:45) Wow.

Patrick: (1:14:46) And you know, the problem is once you're making changes of this scale, you quickly run out of ideas about what to do because we only know how to change things at the individual gene level. When you have a hundred genes, how do I design how a hundred genes should be interacting in order to create an entire biological program or higher-order logic?

Jackson: (1:15:06) Oh, oh, I see. Got it.

Patrick: (1:15:07) And so I started my research career in how you make programmable tools for manipulating the biology and the living molecules inside of cells. Our more recent work has really been around how do you model, right? So it's like writing and then thinking, right? How do you think about all of these different genetic components, how they interact, how they create higher-order logic that allows us to do top-down design?

Jackson: (1:15:34) I see. Right. We're currently rate-limited by our understanding of how these things will combine, basically.

Patrick: (1:15:38) Exactly. We barely speak the foreign language of the genome or the cell. And by understanding and predicting that better, we can use these programmable kind of manipulation tools to make those changes in a much kind of higher success and more targeted way, right?

Jackson: (1:15:56) Got it. What is, how would you think about roughly like the sort of near-term progress in this domain? Like let's say on a—I don't know, three-year time horizon, do you think we're going to make material—largely due to the model side or otherwise, we make material progress on the understanding side? Like what would progress actually look like if some of this goes right? What is the implication on a drug or a pharma or otherwise level in the next few years? I'm not saying—maybe it's worth talking about where things could go in 20 years, too, but I'm curious on the sort of on-deck side of things.

Patrick: (1:16:30) Yeah. So take RNA interference. It was sort of the original programmable genetic manipulation technology before CRISPR, right? Kind of came out in the early 2000s, before my time, right? And over the last 20, 25 years, it has really matured in terms of sort of the chemical modifications you need to make on these RNAs so they can last a really long time in the body, and also how you deliver them really efficiently to the liver and to other organs in the body, right? And so folks have been using this to create medicines for, you know, for high LDL, right? And so, you know, today you take a statin—small molecule, right? Works okay. But then now folks are targeting, you know, like PCSK9, right? They're also targeting high triglycerides, right? And they're creating—Arrowhead has been creating multiplexed RNAi that can target two things at the same time. So both high LDL and high triglycerides, right? And I think, you know, heart disease is still the number one killer in the world, right? And if we can actually—we could significantly reduce the incidence of coronary heart disease by just targeting these two well-established biomarkers in a programmable, multiplex fashion directly in the human patient liver. And this is going from these daily pills that people forget to take to a quarterly injection, right? Because we've dialed in the chemistry so much, you just shoot yourself four times a year, right? And then you can just wipe this out, right? I think that is so tremendous to the degree that 100 years ago, infectious disease was maybe the number one killer, right? Today that's heart disease. We may see both of those go away, right, in terms of their relative disease share. And, you know, that sort of leaves us with cancer, with Alzheimer's disease, and then, you know, there's going to be this really long tail of chronic autoimmune disorders, right?

Jackson: (1:18:35) Right, right. Yeah, and then wherever else we find once we no longer deal with heart disease.

1:18:39 Testing, Bro Science, and Human Data

Jackson: (1:18:39) You had a comment about ML for bio is not just drug design. Obviously, on some level, a huge part of that is all the progress on the modeling side, but I'm curious more broadly what's kind of inside that statement.

Patrick: (1:18:52) I think we need to be able to test things in people, right? So there are different types of tokens, if you will, for the models. There's like a protein structure token, there's maybe a DNA sequence token, there's maybe a cell perturbation token. We also need human tokens, right? Like the most valuable thing in the industry is data on how a human body will respond, right? And so I think that's going to be really important—is going to be innovations around our clinical trial sort of regulations and enrollment and recruitment and endpoints in order to be able to test things faster in people, right? That's going to be an incredibly important trend. I'm actually also just personally very excited about agricultural innovation as well. So the bridge recombinases that we've been developing have been getting deployed in plant gene insertion, which has been historically a huge problem. And being able to kind of genetically endow them with the ability to be more nutritious, to resist cold snaps, to be larger, and so on. I think many of the University of California's most valuable patents are for engineered strawberries, right? I think plant genome editing has been something that people have talked about for a very long time, but just hasn't worked really well. There's also a big real-world deployment gap. I think seeing more of that is going to be really exciting.

Jackson: (1:20:22) Broadly when it comes to I think a lot of people's response to genetic engineering in general, whether it be in plants or certainly in humans, is like—or frankly even stuff like peptides. I think a lot of people's response is like, "We don't have longitudinal data. How do we know these things are good?" I think some of that could be ignorance, some of that could actually be we don't have these answers. I'm curious how you think about maybe how so much of the stuff we've talked about, and maybe particularly the modeling side, might allow us to have more trust, data, whatever it is, to allow us to say, "Hey, yeah, we..." If drug discovery or any of these implementations stop being—like we can do them much faster, the bottleneck is going to go somewhere else. And it might be in testing, it might actually just be in people's belief that they should be eating whatever, your super corn.

Patrick: (1:21:12) I think the idea that you can scrape Reddit data. For example, there are all kinds of subreddits for these niche patient communities who are trying all kinds of things for their chronic, underserved, multisystem disease. Long COVID is Exhibit A for this. I mean, I think something like 10, 20 million Americans have had long COVID. I think something like four to five million still have deficits. And these are like these very chronic multisystem disorders that we don't do a really good job of diagnosing and treating. Medicine is really structured around treating single injuries, like acute care, or single-organ diseases, right? That's also how the residency system is sort of set up. If you have this autoimmune disorder, you have to see five different specialists for your liver, for your GI, etc., etc. And I think the idea that you can kind of train models on this kind of anecdotal, observational data could be a much more effective way to design very clear, interventional, smaller-N trials. Maybe this is a bit of a tongue-in-cheek example, but there was this huge wave of bro science from the 1980s to the early 2000s. A lot of the things that the bros tried for gym-maxing is standard today. And we sort of, in this kind of community population scale, figured out a bunch of interventions that work. And we didn't do RCTs for all of these things in the traditional clinical setting. So I think there should be ways, with the right oversight, regulation, informed consent, safety, to create additional mechanisms for human testing. There are so many people trying peptides today that if we had the ability to record what these people are doing, actually track and understand what's happening, we could deliver tremendous value in terms of knowledge and whether or not these things actually work.

Jackson: (1:23:43) Do you think that's mainly a crowdsourcing, like a coordination problem, or like an actually collecting the data, data technology problem?

Patrick: (1:23:50) Both. Both, yeah. I mean, people are just YOLOing it today, right?

Jackson: (1:23:54) Well sorry, I mean in terms of capturing it—like if I could put you on a text thread with every single person trialing peptides in the Bay Area and they were down to fill out a 10-minute form once a month, would that solve the problem? Or are there a whole bunch of other complexities around this in terms of getting the right data from them?

Patrick: (1:24:11) Oh, I think it depends on the peptide and its putative mechanism of action. But if you could do deep blood profiling of all of those patients before and after over time, I think that would be a tremendously interesting data set.

1:24:26 Can It Exist vs. Will It Exist: Science Meets Markets

Jackson: (1:24:26) Yeah, yeah, yeah. I guess it comes back to longitudinal datasets again and again and again. A few more things: obviously, like in so many ways, you are playing across nonprofit and more traditional academia, science world, open-source stuff, and then like proprietary, commercial business stuff. And it seems that you've been really thoughtful about, like, the ways to use all of those shapes in tactical ways to get outcomes. You said somewhere researchers don't think in products, they think in papers. I guess I'm curious what your encouragement might be to maybe the academic side of the house around thinking about the commercial side of things in a way that isn't like, "Let's just commercialize everything," but in a way that is like thoughtful and nuanced and very ultimately focused on sort of like outcomes.

Patrick: (1:25:24) I think in academia you're incentivized to think, "Can something exist?" right? Markets tell you if something will exist, right? And a lot of the kind of non-academic side of my life has been an attempt to understand those variables, right? You know, so when I kind of look at, build, or invest in companies, I think a lot about business genetics, right? And it's actually very funny, because, you know, different types of companies have different genetic substrates and profiles and signatures, right? And they interact in weird and nonlinear ways to create emergent phenotypes, much like in cell biology, right? You sort of are doing comparative genomics when I look at a deal in defense, and in robots, and in healthcare, and in, I don't know, like RL environments, and in cybersecurity. They all have different markets, they have different competitors, they have different players, they have all these kind of different variables that you have to learn. And it's, you know, kind of similar or analogous to all the variables of kind of the different kind of genes and gene programs and how they interact. And so I think I kind of have this comparative genomics view of how I look across business genetics, right? And so I think the kind of analytical quality that you have to solve these biomedical problems are in a way very transferable into the for-profit sector. It's just with a very different set of kind of like variables and players, but the kind of core way that you kind of frame and look at the problems are similar, yeah.

Jackson: (1:27:08) What about when it comes to solving a specific problem? Like, I don't know, on one end of the spectrum you have FROs or nonprofit structures or whatever. On the other end of the spectrum you have, whatever, Thrive spinning out Isomorphic from Google with a crazy amount of capital. I imagine there's a broad middle ground there, too. In some sense, Arc is embodying this, which is, at least as I understand it, a huge part of the goal here is eventually to start spinning things out that are commercialized. But what is the right model to think about the nonprofit side of the spectrum into the—in some sense, commercial and obviously practical side of the spectrum?

Patrick: (1:27:49) I think it's all downstream of capability breakthroughs.

Jackson: (1:27:51) Okay.

Patrick: (1:27:53) Like OpenAI was and is a nonprofit, right? The OpenAI foundation is the largest foundation in the world that owns 25, 26% of the company of the PBC, right? And you know, I think with Arc, our genuine hope is that we'll be able to bring to bear different flavors and sources of capital and expertise on our, you know, basic science breakthroughs over time in order to actually turn them into real products. We have a number of programs that are very translational that have come out of Arc Labs lately, right? So, Isha Jain, for example, at Arc and at UCSF has invented small molecules that can tune how oxygen binds to red blood cells, right? So, you can do like altitude therapy in a pill.

Jackson: (1:28:41) Oh yeah, I read about this. Cool.

Patrick: (1:28:42) That's like, you know, insane. And she's using it to treat neurodegenerative diseases, to treat cancers.

Jackson: (1:28:48) And that's already been commercialized?

Patrick: (1:28:51) They've tested it in animals, and they're testing in larger animal models, you know, creating drugs out of this that have like real drug-like properties, right? There's a bunch of machine learning models that our investigators have been developing, like Hani Goodarzi at Arc has been training all kinds of models for designing RNA medicines, for example, and there's a number of other companies that are being incubated at Arc that will, you know, kind of become public to the world soon, that kind of take the principles that we used to try to recruit everybody. And there's like a lot of net-new generative kind of projects that we couldn't have predicted when we founded Arc, but have happened through the unique mix of people and ideas that we brought together. So this is like real time. This is happening, right? It's not like theoretical stuff. So, yeah, it's quite fun.

Jackson: (1:29:47) What do you think would cause the biggest inflection to Arc in the positive direction before 2030, if you could figure it out, if you could get it right, if you could change something?

Patrick: (1:29:58) More GPUs.

1:29:59 Biofuturism: Interoception, Peptides, Muscle Modulation, Sleep, and New Forms of Life

Jackson: (1:29:59) Okay. A few questions about maybe just some futuristic stuff, I guess. One, you talked about interoception a couple of times in my research, which, as I loosely understand, it's kind of this more Eastern philosophy around how the body and the brain are connected, and when you feel something in your stomach and then it turns out to be true, or whatever. Any updates on this broadly? I don't know, it seemed both really interesting and like not obviously in this wheelhouse, and so that's why I was curious about it.

Patrick: (1:30:32) Yeah, yeah. This is actually one of our craziest discoveries so far that hasn't come out yet, but it's a profound interoception modulator, right? Which sort of the biology of the gut-body-brain axis and how different organs communicate with your brain and nervous system. And so there are like deep biological mechanisms for how, you know, like cognition creates stress, right? And there's actually a mechanistic circuit for how they create, you know, sort of the nervous system innervates the gut and can create the release of pro-inflammatory cytokines that leads to astrocyte inflammation that creates ulcers, for example. And so it's, you know, like we've always sort of known if you're hyperstressed, you can get ulcers, but our PIs have discovered exactly how this works, right? And, you know, vice-versa, right? Kind of inflammatory signals from the gut can lead to brain inflammation. And, you know, we've been figuring out what are modulators of these pathways. And, you know, I think these will be—

Jackson: (1:31:35) Why has there been so little progress on this historically? I mean, maybe they're—maybe I'm ignorant, but it seems like that this is not something that has been a huge area of discovery?

Patrick: (1:31:42) Because it's multisystem, right? People are used to studying individual organs. And these are really complicated multisystem models that are built on top of decades of infrastructure that we've created around different types of knockout animals. Being able to—you need to use very specific models to cut nodes of connectivity between the gut and the enteric nervous system, the enteric nervous system and the central nervous system, the brain and then back, right? And sort of we kind of needed the right transgenic drivers to map each of these nodes, which we kind of now have combined with the high-throughput genomic and proteomic and spatial transcriptomic technologies to view what's happening, and then the ability to synthesize and test those perturbations or modulators, right? So we're, like, doing a lot of work to discover new peptides at Arc, for example. And, you know, I think they will be much more science-backed certainly than the things that people are injecting in their bodies every day.

Jackson: (1:32:47) Maybe speaking of peptides, what do you think is not obvious that sort of will be consensus in, let's call it five years around peptides?

Patrick: (1:33:01) I think there are a tremendous number of bioactive peptides, like really tremendous. Most of them don't have drug-like properties. So they're released spatially in very specific places locally and they signal, sort of, you know, to things that are nearby, and they degrade really rapidly in circulation, right? And so a lot of the sort of work to create peptides you can take as drugs involves advancements in peptide chemistry, which is like this multi-decade long-running story where we don't have deterministic, simple ways to just make peptides last in the bloodstream really potently or to control their biodistribution or what receptors they bind to, right? And I think as we start to kind of tame these nodes, we'll be able to create them as an entirely new class of drugs, or like a large class of drugs beyond the kind of individual kind of success stories that we have today, right? Outside of GLP-1s, there aren't that many super successful peptide drugs, right? But I think that, you know, market share will really increase, yeah, if we solve those chemistry problems.

Jackson: (1:34:22) And you think those are near-term solvable, at least on some level?

Patrick: (1:34:26) We have figured out key nodes that seem to be repeatable that work well for some peptides and not others.

Jackson: (1:34:37) Can you give me an example of something that would be maybe not obvious, that could be like if you could solve some of these problems would be like a really miraculous peptide drug?

Patrick: (1:34:45) Okay, I'll give you some—I'll kind of go down one level. Attaching a fatty molecule to a peptide can really increase its half-life. A lot of peptides also have proteases that bind and cleave them and cut them up. And so figuring out what those protease sites are and then blocking them, right? So these are sort of two of our major nodes for increasing peptide, like pharmacodynamics, right? Also predicting what receptors peptides bind to, which is very imperfect today, but with really good protein structure prediction, we should be able to solve this problem with AI. It will help us take advantage of all the single-cell data that we have to map what the likely biological mechanism of that peptide is based on what receptor it's targeting. And if we had working virtual cell models, we could figure out what that receptor's sort of signaling network is intracellularly, so you have like a functional fingerprint for every new peptide that you discover computationally. This has been traditionally like wet-lab experiments done over the course of many, many years. And the more of this reasoning we can push into computation is how this geniuses-in-a-data-center idea can run across many different peptide sequences in parallel, right?

Jackson: (1:35:57) Yep, yep, yep, yep.

Patrick: (1:36:03) That's sort of like how we make, I think, unprecedented progress in this area across many different sequences.

Jackson: (1:36:11) To go back up a level, can you give me an example of like a type of peptide that people, if we pull this off, could be—I don't know.

Patrick: (1:36:20) Yeah, sleep peptides is a great example.

Jackson: (1:36:21) Which is just straight up like you inject and you just can sleep less?

Patrick: (1:36:26) Yeah, I think for modulating wakefulness, for example, for, you know, creating better and higher quality deep sleep or REM sleep would be another example.

Jackson: (1:36:37) So this is sort of an unbounded, like, things we want to modify in the human body, we might be able to use peptides for like a really, really broad—

Patrick: (1:36:46) I think there are ways that we can make sleep modulators that could be peptide drugs or they could be small molecules, they could be antibodies. I'm much less predicated on the drug class and much more on the product, right? Yeah, and by understanding the mechanism by which we can modulate sleep biology, right? Like there are these orexin agonists that are small molecules. There are also orexin agonists that are peptides. But like whichever one dukes it out better in the market kind of doesn't matter. The end patient wants something that can safely and in a long-lasting fashion modulate wakefulness where you feel like a vitality of attention and wakefulness as opposed to this sort of drug-induced, like, speedy wakefulness, right?

Jackson: (1:37:33) Huh. We don't seem to have many drugs for musculoskeletal injuries, and it seems like at least when we talk about it in longevity, it's actually like one of the things that most significantly impacts people's feeling of feeling young or feeling healthy.

Patrick: (1:37:47) There's a lot of work on muscle modulation.

Jackson: (1:37:50) Really?

Patrick: (1:37:50) Yeah, yeah, with targeting activin and myostatin pathways. So there are these sort of—if you just Google "myostatin cow", right, there are these like insanely jacked Arnold Schwarzenegger cows, right? Because they sort of are mutated in this sort of regulator of muscle synthesis. So when you lose it, you know, they get insanely jacked. And so there's work now to try to pair these with GLP-1s, right, in order to preserve lean muscle mass as you're losing weight. The really interesting, at least early clinical data so far is that it's not so simple, where muscle is both a metabolic tissue, but it's also a functional tissue, whereas fat is mostly metabolic tissue in the sense that even if you have larger muscles, if you're not training them, right, then it may not be functional, right? And so it's not quite as simple. Like you can preserve more lean muscle mass, but to actually be stronger, you also have to work it out in the right way.

Jackson: (1:38:52) Or to solve your back pain, you have to actually not just sit all day.

Patrick: (1:38:57) I actually think pain is another extremely exciting area. Chronic pain, non-opioid kind of pain modulators—lots of really interesting work here. Yeah, we haven't even gotten into like brain-computer interfaces and how you can use these to modulate pain pathways non-invasively with ultrasound or whatever.

Jackson: (1:39:17) Any—maybe it's brain-computer interfaces, but I'm curious if there are any sort of like out-there moonshots that come to mind as being maybe you are disproportionately excited or optimistic about that we could see in the next, let's call it decade around human biology.

Patrick: (1:39:35) In the next decade.

Jackson: (1:39:37) Something like that seems like the right relative. I don't know the difference between five and 10 years, frankly, on some of this stuff, especially given the rate of progress.

Patrick: (1:39:48) Yeah, yeah. I'm personally super excited about BCI. I'm also really interested in sleep, and how actually the union of BCI and sleep, right? For example, we spend, you know, eight hours a day totally unconscious. What are, like, more functional ways that we can use that time? Like can you much more, like, deterministically write memory consolidation, right, during a sleep period?

Jackson: (1:40:17) Ah. You think sleep is inefficient?

Patrick: (1:40:20) I don't think we know nearly enough about sleep to have an opinion one way or another.

Jackson: (1:40:25) Yeah, yeah, yeah. So it might actually be very load-bearing, but it would be nice to know.

Patrick: (1:40:28) Yeah, yeah. I mean, I think there's been a lot of work from Maiken Nedergaard and others showing that when you sleep, you significantly increase the rate of cerebrospinal fluid flow in your brain, which seems to be sort of flushing out extracellular proteins that are secreted by neurons as waste. Sort of like Recology, like flushing the toilet for your brain when you're sleeping. It's a way of flushing out the toxins, right? That's been like one molecular mechanism that has been worked out over the last decade, right? There are probably many other complementary mechanisms that we haven't really worked out yet, but, you know, obviously in addition to toxicity modulation, it's also clearly involved in learning and memory. And, you know, understanding that better allows us to write those memories much more effectively, yeah. I love biofuturism, by the way. Like we're only just getting started. This could be a whole other hour.

Jackson: (1:41:26) Yeah, yeah, yeah. We'll do this on part two. Any—maybe given the CRISPR past, any use of gene insertions that you think are sort of like, maybe similarly placed, like not real, but maybe on the edge of viable, that could be really, I don't know, amazing?

Patrick: (1:41:45) I think a lot about—there's an entire community of synthetic biology that's working on, like, the minimal genome. Right? Sort of how much fat is there in our genomes? Can we make them as small as possible to figure out the most essential minimal components? I guess I find this actually, I think, a very important field, but fundamentally less interesting than what are like the net new synthetic functions that we can add. And for example, there are these studies that have shown like a single enhancer can be toggled to create a tail or not have a tail, right? So there are just like all these insane biological programs, some of which are vestigial, that are just sort of encoded in our body. And like maybe having a tail is sort of like a silly, sort of lurid example, but can we do something much, much more functional, right? I think the combination of models that can think, these genetic manipulation tools that can synthesize long information and recombine and rearrange them in the right ways, is going to allow us to move beyond, you know, 10,000 base pair genomes to 100,000 base pair genomes to million base pair genomes to 10 million over the, I don't know, N years to come, right? And as these models get higher fidelity, I think we will create, like, evolutionarily distinct forms of life, at least by sequence, right, that will be as distinct from existing known species as, you know, species that we have sequenced today, right?

Jackson: (1:43:30) Do you think that will happen soon?

Patrick: (1:43:31) Yeah. It's already happening in a way, right? And the question is how different can those functions be? And, you know, how complex can those organisms become?

1:43:47 Fun, Obsolescence, Going Slow, and the Future

Jackson: (1:43:47) You have said a few different places—maybe one of the first kind of instances of it I found is you talked about being in that wet lab at a university at like age 14 and knowing you loved it after like two weeks. And you've talked about in a bunch of places this sort of through line of making sure to have fun. Somewhere else you referenced the Scottish proverb, "Be happy when you are alive, for you are a long time dead." I don't know that you operate in fields that are always full of people who you would describe as having a ton of fun or being radically self-expressive or whatever. Where does that come from, and how do you sort of maintain that and keep that sort of at the center of things?

Patrick: (1:44:33) I think science is really hard, right? It's such a long journey and it's mostly failure. So if you don't have a way of bringing the energy and the vibe and the vitality, it's going to be pretty miserable. I also think the profession incentivizes and rewards a rigor of criticality that is not great in other domains of your life, like relationships, for example. Learning when and where to use these traits or skills—like everyone has superpowers that make them great at what they do and who they are, but there's a dark side to all of those superpowers that are also your foibles. So I think being really kind of self-aware of when and where to use those things is really important because absolutely my scientific training has made me much more picky and much more critical. But you have to know when and where to use that. Similarly, I think you need to figure out a way to make the discovery process fun because you're mostly eating shit.

Jackson: (1:45:50) That's a good answer. Anthropic just announced that they are giving Claude access to a wet lab, I believe, as I understand things. Any thoughts?

Patrick: (1:46:01) Yeah, yeah. They've been working on this for a while. I'm a fan of a bunch of the people working on that team; several of them trained with me. I'm excited about the vision of Claude-maxing science. I think the default way that we—we're going to use these agents for everything. It's going to be a default interface by which we interface with scientific knowledge. One of the kind of big ironies of writing papers is I don't really know if I'm writing them for fellow humans anymore. If you think about the long arc of history—

Jackson: (1:46:41) Tyler Cowen talks about this with the books too and all the blog posts, yeah.

Patrick: (1:46:44) Yeah, like they're just ingesting all of this. So the way that I would write for my colleagues and the way I was trained to do that, I don't think is going to last the test of time because—I don't know—if you think about the FLOPs of intelligence that will be used to process any thought I've ever had will be primarily from these agents. That's really trippy. It pressurizes the scientific literature and scientific publications and journal enterprise like never before. What does it mean to record and disseminate human knowledge? Who's going to be the one who's thinking about these things? I struggle with that a lot.

Jackson: (1:47:28) Yeah.

Patrick: (1:47:29) And that, if anything, emphasizes the importance of having fun. That only matters more. Yeah.

Jackson: (1:47:39) Do you fear becoming obsolete?

Patrick: (1:47:43) The profession is one of self-invention and reinvention. Yeah, like if you simply do nothing, the default is to become obsolete, right? And you see this across—lives are long and careers are long. So many generations of scientists became prominent for some discovery or some innovation in an area that had its moment. And then another trend came, and then you are obsolete. So it's not even a fear; it's actually just reality, right? And so being able to be bold, to be fearless, to be risk-taking are actually—they're kind of survival traits, right?

Jackson: (1:48:32) I realized I never actually asked for clarification. You said at the top you have three jobs. Is that Arc—kind of running Arc—Arc, your lab, and Thrive? Is Thrive not counted in that? What is the—

Patrick: (1:48:44) I'm also a professor at Stanford, right? So Arc, Stanford, Thrive are my three jobs, yeah.

Jackson: (1:48:49) Right, right, right. Okay, okay. How does the Stanford—I would assume mostly teaching—work and the investing, which is obviously very broad and you do a wide range of things, how does that inform or inspire what I would presume is the main job?

Patrick: (1:49:08) The trainees are so good, right? I mean, I just took a PhD student from Stanford—one of the Stanford graduate programs—this cycle. We'll have more over the years. And these are just such brilliant, ambitious young scientists who want to do ambitious things. The really interesting thing is I think in the age of AI, the folks who decide to do a PhD are weirdly more ambitious, right? You'd sort of think, "Oh, like maybe the best talent in the world is getting starved and they're all just going to AI labs or something." But in my experience, that's not really true, because the big labs are able to do a very specific type of research, right? And we, in moments of great change, tend to overrate the—the way that, I mean, like people are concentrating into these areas very sensibly, but there is going to be a huge amount of capability that's unlocked downstream of the models that folks who are focused on the long tail of research are going to be the most well-positioned to capitalize on.

Jackson: (1:50:15) Yeah, it's more of a long-termist view in some sense.

Patrick: (1:50:17) Yeah. Yeah.

Jackson: (1:50:18) If you could spend one year sabbatical doing anything else, but you can't do something you're doing now, what would you do?

Patrick: (1:50:24) Ooh, I would do something in the real world that's really like esoteric and physical in nature.

Jackson: (1:50:31) Okay.

Patrick: (1:50:33) Maybe I would—I don't know what that is. I think I would learn a completely different domain of science.

Jackson: (1:50:39) Cool. Yeah, cool. Okay. Maybe someday. Maybe when the robots make this part obsolete someday.

Patrick: (1:50:45) Exactly.

Jackson: (1:50:47) You have, I think it would be fair to say, gone very fast in your life. Where do you think it is a virtue to go slow?

Patrick: (1:50:57) Actually, knowing thyself well enough to identify your gas and brakes is really important to live a fulfilling life and to live a fast life in the first place, right? Because if you want to be high-performing or whatever—if that's, you know, like not everyone should be or needs to be or wants to be—but you have to be able to intuit and understand your energies, because I can't go ham the way I did in my teens, right? I have to understand those energies to avoid burning out. If anything, I kind of specialize in how to not burn out.

Jackson: (1:51:40) Throttle.

Patrick: (1:51:41) Right. Yeah, yeah. Because you can—yeah, if you're just ripping at full throttle all the time, it's very easy to break something, right? So knowing— Yeah, you have to know when to go slow a lot of the time. Like, you know, there's sort of that adage, like, "Slow is smooth, and smooth is fast." It's very, very true. Yeah. Yeah. Yeah. Yeah.

Jackson: (1:51:59) I think it was in the conversation with the founder of Benchling. He asked about your superpower, and you said you have a deep internal sense, a compass that is like a right direction to point. Where are you pointed?

Patrick: (1:52:14) I think there's a lot of discussion, especially, you know, these days with cybersecurity and Millennium Prize, etc., about the inflecting capability of these models and how scaling laws don't seem to be anywhere in sight, or the end of scaling laws doesn't seem to be in sight. And I feel like the interface of this model superintelligence with the physical world, right, is the kind of thing that the market hasn't really priced in yet but is really inevitable. I'll give you an example. I was recently watching a video of a robotics model that was controlling a robot hand to try to do a complex dexterity task. And the model was hooked up to a 3D printer that could print new appendages and limbs

Jackson: (1:53:05) Oh, wow.

Patrick: (1:53:05) in order to change the way it could physically interact in a real dexterity sort of engagement, and it can sort of RL with the ability to change the physical substrate, right? So I found this really trippy, obviously. And that sort of core concept you can lens across many different physical tasks. If you could do directed evolution in the physical world across many domains, things are going to be really interesting, right, and really crazy and trippy and hard to predict. And so I spend a lot of my time trying to think about that 2030 and beyond future. A lot of my work today involves thinking through, like, what are the bottlenecks, what's essentially the supply and demand side of creating superintelligence, right? My arb is really how do we apply this in the domain of biomedicine, right, and be at the frontier of doing that, right? And then, you know, how does that kind of change the rest of the physical world? And so I spend a lot of my time pondering that.

Jackson: (1:54:21) Directed evolution is a good answer from a biologist. Patrick, that's all I got. Thank you.

Patrick: (1:54:25) Thanks so much. That was a joy.

← Back to all episodes