No Priors 第104期｜与 Flagship Pioneering CEO兼联合创始人 Noubar Afeyan 对谈

Hi, listeners, and welcome to No Priors. Today, we're joined by Noubar Afeyan, founder and CEO of Flagship Pioneering, the firm behind Moderna and more than 100 other biotech companies. We'll talk about his approach to building biotech startups, how AI is reshaping drug development, and his theory of polyintelligence.

Welcome, Noubar. Thanks again for doing this. I look forward to the discussion. Let's start with the roots of your incredible personal journey. You arrived as a teenager after your family fled war-torn Beirut, earned MIT's first PhD in biochemical engineering, and over the past 3 decades, you've created a force with Flagship that has changed the trajectory of global health through many important biotech companies and more than $100 billion in value. Can you talk a little bit about your motivation to start Flagship originally and what you thought it might become?

I will work hard to fit the description you just gave of what I've done, or what I'm trying to do. The motivation for Flagship stems from what I was doing before. I started a company in 1987, when 24-year-old immigrants didn't start companies in this country. Instead, former McKinsey senior executives or IBM senior executives were the only ones entrusted with the massive amounts of venture capital—namely, $2–3 million per round—that went into companies. This was very early days, and I had the opportunity to start a company right out of graduate school. I ended up raising quite a bit of venture money and eventually went down a path of entrepreneurship.

Along the way, one of the things that interested me was why the entrepreneurial process was supposed to be a random, improvisational, idiosyncratic, almost emotional, game-like activity. All of those things were a bit of a put-off when it came to actually doing things in a serious, professional way. In the very early 1990s, I used to go around saying, “Why isn't entrepreneurship a profession?” If it were going to be a profession, how could it be a profession? At the time, there were largely one or two competitions giving prizes, which of course reinforced the game-like nature of it.

I started thinking about that. One way to know you're doing it as a professional is if you can do many of them in parallel. My motivation for that was investing, because venture capital is a parallel investing activity. If you could only do one of these at a time, you would have serial venture capitalists, not parallel venture capitalists. Yet people think entrepreneurs are supposed to be serial, while investors, lawyers, and everybody else can operate in parallel. The learning cycles of doing things at the same time are completely different from forcing yourself to think about the essence of what you're doing, what's reproducible, and what has to be different in each case.

I got interested in that, and parallel entrepreneurship led me to ask, “How do you do that?” In the late 1990s, alongside running my first company, I got involved in the co-founding of several other companies, where I tried this out individually. Then I realized that it was pretty hard to do as a solo player. So I created the first company-creation company. It used to be called NewcoGen when it started—it stood for “new company generation”—and that's the name we operated under for 3 years, until people told me it sounded like a disease. Then I changed it to Flagship. By the way, that's exactly what happened.

From then on, we've been on a journey to figure out not only how to do professional entrepreneurial activity, but institutional entrepreneurial activity. That means acting just as investing became institutionalized: acting as teams, with company objectives, and so on. It's a complicated way to describe the motivation, but maybe I'll simplify it by saying that I thought the most value-creating activity I know of in the current human endeavor is starting companies.

I used to teach entrepreneurship at MIT for 16 years, and then taught for 3 years at Harvard Business School on innovation. I always ask people, “What's the single most value-creating activity? What's the biggest invention that humans have made in that regard?” They start telling me about the internet, satellites, and this, that, and the other. I tell them it's the startup—the human invention—and the value created from the startup includes Google, Tesla, Facebook, Genentech, and every single one of those companies. I thought, why are we outsourcing that to some random, game-like activity? That's what motivated me to start Flagship.

What do you mean by “game-like”? Is it that it's supposed to fail most of the time, and once in a while you win and celebrate the win?

What I mean is that it's random, but not only random: there are winners and losers, and people keep score. Maybe it's the wrong word. People even call it gamification in the software space. I don't mind being playful, because if you're overly serious, sometimes you miss things, but it can't all be play.

We take hard-earned money and deploy it to do things that are damn near impossible. Once in a while, we reduce them to practice so they become not only possible but valuable. Yet people treat it like, “Well, 20 different things we tried didn't work, but one of them worked.” As an engineer by background and a scientist, I thought that what we do—especially in health care, climate, agriculture, and food security—can't be thought of as shots on goal. We've got to say, “Hey, we can get better at this.”

One thing that Flagship famously espouses is the idea of allowing breakthroughs to emerge rather than being made by people. How do you square that? How do you create an environment where breakthroughs emerge naturally but are also somewhat predictable and controllable?

First of all, if anybody listening to the podcast wants to read more about it, we finally, after 23 years of doing it, published a paper in Harvard Business Review a couple of years ago, together with Gary Pisano, that laid out the idea of emergent innovation.

If you think about human design, it's goal-based. What's the goal of creating a new company? You could say selling it. But if you said, “What was the goal of creating NVIDIA?” I don't think you can define a goal, because they thought they could create valuable products that would impact people, initially in the gaming space, and then all of a sudden the AI opportunity emerged. It wasn't predictable or in the business plan—except that it wasn't in the 5-year plan at all. The opportunity emerged.

Where else do we see activities where novelty happens in this unpredictable way? Nature. In nature, the forces of variation, selection, and iteration create unbelievable novelty and really esoteric complexity of all sorts. Life is a pretty impressive example. We still have many living things. I don't think there are that many living silicon things.

That is emergent. If you do variation, selection, and iteration in anything, you get emergence. You do it in human thought, and you get revolutions, memes, political thought, religions, and lots of other things. You do it in running shoes, and you get Air Jordans. All of this is emergent. It's just that the people who participate in it describe it as though they came up with it. They describe everything in this unbelievably, and I'm sorry to say, characteristically human way.

The fact that we have our own language, and therefore large language models that are really good at taking reality and converting it into how it is that the human agent presided over it—that's like saying whoever wins the war gets to write history.

I'm saying this having practiced it for 38 years: I don't believe anything I've been involved in innovating and creating is actually the product of my work. It's been emergent. What I try to do, and what our whole organization tries to do, is create an environment within which emergence can happen, and then have the humility not to describe it like some act of genius. You will not hear me giving interviews ever describing these things as some superintelligence that we wielded. It was an emergent thing.

Generative AI is one hell of a technology for emergence, which we're using in pretty cool ways these days, whether it's emerging new proteins or emerging new consumer products. I couldn't be more excited. It's emergence—that's what we harness and have used for 25 years, ever more systematically. But we have a lot to learn.

That's an incredibly intellectually honest point of view. Maybe one of the things you're alluding to is that the very best entrepreneurs also tend to be guilty of this: the narrative changes to tell the story in a clearer light from the beginning, as though they had much better predictive power than anybody who's been around real companies that achieved success would really understand.

I used to tell people that most successful entrepreneurs who've done unexpected things also secretly worship at the altar of chance, because they realize that, at the end of the day, their explanations of what happened do not fully account for what actually happened and why.

They'll never describe it by saying, “We did this, then we did that, then we did this, and then something happened out of the blue. I have no idea why it happened, and that's what led us to succeed.” Nobody writes books about that.

Yet if you interview lots and lots of people, they'll say, “A lot of things that happened, we got fortunate, but we were ready for it because we did everything else as best we could.” That's what I believe.

At Conviction, it's very hard as an environment to predict even what happens with the technology or what applications are relevant. But you can say there was alpha in the interest to begin with, and we can be prepared for the opportunity, be fast to adapt, and identify and cultivate that opportunity, or select down from the what-ifs, in Flagship terminology.

You used the word “adapt” very correctly. A lot of times, people ask what comes first: the variation or the selection? The answer is that the selection pressure in nature comes from the sum total of everything that exists in nature.

It's the same here. Consumer preference often reacts to what's put in front of it, which then forms consumer preference. It's really hard to separate these things. You have to create an environment in which they interact, and you get new products and new services.

Flagship, in terms of ambition and scope, doesn't just do therapeutics. You've done ventures in nutrition, agriculture, and climate. How do you think about opportunities to extend beyond medical biotech?

We're very careful in that regard. Where we have a core advantage—whether it's intellectual property we've created or maybe a daring that comes from not knowing enough about the space—we'll venture into it. For us, the first thing we do in a space informs the next 5 things we do. If all 5 things don't succeed, we'll say, “You know what? Maybe we can't get paid for the innovation in this space.”

Our methodology is all about trying to bring to life today what might otherwise exist 5 years from now. Not everything that will exist 5 years from now will be valuable. On top of that, we've got to come up with something today that's also going to be valuable, and that isn't true for every sector.

For example, we worked for many years in renewable energy. One of the most advanced ways we could make carbon-neutral liquid fuels was to engineer photosynthetic bacteria, which usually grow in the depths of oceans. We engineered these bacteria to make diesel. They could literally consume CO2, make diesel fuel, and secrete it. People thought it was impossible, and we did it. Then we created reactor systems that were dirt cheap to do this in.

The next thing you know, we go out—and this was in the 2008–2012 time frame—with a company called Joule. Guess what? It's a commodity business. The price of diesel, whether it's renewable or not, is the price of diesel. The price of carbon when we started was $50 a ton; when we ended, it was $5. When we started, the U.S. was energy-dependent. When we were done, the U.S. was gushing with energy, liquid fuels, oil, and gas.

In the end, you realize that no matter what you do in that space as an innovator, you aren't going to be able to get a premium. We learned a lesson that, at least back then, the sector didn't deserve the kind of leapfrogging innovation we could make. Since then, people have said to us, “Why don't you go back and do that?” The answer is that I still don't know if it will pay for itself, because just when you thought green energy and carbon were going to be priced in, here we go again.

You might be surprised to hear that we started companies in supercomputing back in the 2000s. We started companies in networking. It's crazy, but we did. They were one-offs where we said, “You know what? Let's see what we can do in a space.” Since we were doing it in-house, we could take our time figuring it out.

As we speak, we've recently started things in the materials space—semiconducting materials and carbon-capturing materials. We'll experiment. We do have this almost emergent mindset about where we will apply ourselves, but we know that not everything needs scientific leaps. Not everything lends itself to this kind of activity. We experiment.

Many Flagship companies are pioneering new categories that regulators don't know what to do with. We face some parallels here with some of our companies now. If I think about microbiome therapies, gene editors, or these other categories, this seems like another somewhat brutal market dynamic, along with the cyclicality you're describing. How do you think about the risk of these new markets that you have to shape through policy and public opinion? Pulling the future 5 years in isn't just a technology problem, right?

That's a very good question, and it's one of the things we think carefully about—obviously, in hindsight, inadequately, because we can't control all the variables.

I think risk is a concept best applied to adjacencies of what already exists. Imagine a circle: everything inside the circle exists today and is known today. Everything just outside that circle—the adjacent circle around it—is what's going to be known over the next interval.

I don't know which way the innovations will happen, but that's where most innovations happen: in this circle of adjacency. There is an advantage in that people sitting today can make estimates of risk and reward. That's what due diligence is all about. You ask multiple key opinion leaders what they think, aggregate their views, and make an investment decision. You judge the management, and so on.

But take that distance from the current state further and further out. At some point, people can no longer estimate the likelihood of success or the rewards of success. What do you call that? I would not call that risk anymore. I would call that uncertainty.

Uncertainty is about things to which you cannot attribute a probability of success. What we do as humans is consider those things risk anyway, because there's been an economic drive, largely by Wall Street and others, to put everything on a risk matrix. I don't believe that. Therefore, we view that as super-high risk.

Take fusion. Even though a few people claim they can do fusion in a shorter time frame than 35 years, is that risk or uncertainty? I would say it's not risk, because you can't tell me what the probability of success is.

I'm going to come to your question about market success, regulatory success, and policy, because those just add more layers of uncertainty. But we say, “Why would you think that adjacencies contain extraordinary value pools?” One thing you can be sure of is that everybody is working on adjacencies. Every startup is, every academic lab is, every large company is. Everybody is in the adjacencies.

In the scientific fields—not business-model innovations or service models; fortunately, I don't understand any of that, so I stay out of it—adjacencies have their own problem, which is the risk of commoditization. That's a risk they usually don't take into account. In our world, there's uncertainty.

So what do you do when you face uncertainty? For a subset of things that are uncertain—things that aren't yet known to be valuable or even doable—you go do the experiment. If you can come up with the right experiments to bring something to life and see whether it can be made real, you still haven't resolved some of the uncertainties and risks you mentioned, but at least you can control what you can control.

The way we look at it is, yes, if I had a choice between a really hard technical feat that, once done, had a ready market, versus one that also then had to be totally new, Moderna—which was our 18th company—is a good example of the latter. There was no mRNA drug or vaccine before, and nobody was working on it. There were academic labs that had worked on it and given up.

You need regulatory change, or at least acceptance. You need market pricing. You need manufacturing, which we had no idea how to do. But it turns out that, with or without a pandemic, there was plenty of value it could create. The pandemic distorted the path, but nevertheless, that's how we think about it.

We carefully embrace uncertainty and try to resolve it on the way to understanding what the risks are. Then we try to mitigate the risk. If you're not willing to do that, you're going to work on me-too value pools. One thing we know, and everybody who works at Flagship knows, is that we're no smarter, no harder-working, and no better connected than anybody else. What we can do is underwrite uncertainty.

One more question on Flagship: what has changed over the last 25 years? How is the Flagship approach different today than it was in 1999?

As somebody who started a fund 2 years ago, and an incubator and a bunch of things in 1999 and 2000, you have to realize that the world was being overtaken by the internet, in particular e-commerce. Ironically, at the same time, the human genome sequence was being completed, and I was involved in the company that did that—the private effort, Celera, that sequenced the genome.

It was a weird time, because all the money was being siphoned into Sunglasses.com, Diapers.com, and all these things. That was the fad of the day. It was really hard to get money to do anything with life sciences and medicine.

That's important because we realized there was nevertheless a big need for medicines. We started focusing on the intersection between biology and technology 25 years ago. From day 1, we had the notion that you can conceive of and create companies systematically. We wanted to learn how to do that and get better and better. What we didn't initially do was bet on being able to systematically make breakthrough innovations the way we've learned how to since then.

Today, Flagship is 550 people. About 200-plus are scientists, engineers, and M.D.s. We file 600–700 patents a year centrally, and every one of the things we work on has essentially no connection to what's been done before. That second innovation really came about in the late 2000s. That's what's changed.

The other thing that's changed is that we were a small organization. We were about 50 people as recently as 7 years ago. It's only in the recent past that we've brought in-house the capabilities to scale companies internally, not just conceive them. That has led us to have an internal engine through which we now have many people who know how to build companies in parallel. The learning cycles have accelerated.

I'd say what's changing much more is that we have a technology tool through generative AI. We've worked with AI-based companies for 25 years, by the way. This isn't revisionism, where everybody wants to claim they were doing it all along. We literally started a company in 2001 called Affinnova, which used machine learning and evolutionary algorithms to evolve consumer products online.

Starting in 2001, we worked for 6 years developing machine-learning tools, which were basically dynamic evolutionary algorithms. This had nothing to do with DNA or RNA, except metaphorically doing what nature does. We've gone back to doing that. Our 100th company, F-100, is now developing that with modern generative AI.

What you can do with generative AI in this kind of leapfrogging conception and hypothesis generation is remarkable. There isn't a day that goes by that doesn't change. We've also changed how we think about our space in the innovation value chain. More and more, we think we can generate breakthroughs that incumbents in a space can benefit from.

We've set up large partnerships with the likes of Pfizer, Novo Nordisk, and GSK in pharma, and with companies like Thermo Fisher, Analog Devices, and Samsung in the tech space. They're all aimed at expanding the reach of our innovations so we can have more impact. There have been lots of learnings along the way.

Let's talk about the AI piece. What do you believe are the most exciting applications of AI in health care?

If you think about AI largely as large-data-driven models that can do things that begin to look like human cognition—things we used to do with correlation, statistical analysis, and optimization, which is what machine learning was originally developed for—you start getting more ambitious in how to use it.

Among the very first things we did was use really early deep neural networks, followed by other techniques, to take data in the whole space and inform the design of the next generations of what we did, as well as how we manufactured them. Moderna was doing that on the one hand, but we also started looking at it as a way to design proteins.

About 6 or 7 years ago, we created a project called What If that asked, “What if you could computationally design a protein of any desired function?” You might say, “Of course. Now we have AlphaFold, quantum models of folding, and so on.” We didn't want to use any of that. We literally asked, “Can you show enough instances of the desired function, with the underlying DNA sequence, for a learning algorithm to generate new ones?”

People said, “No, no, no. You need to know the DNA sequence, the protein sequence, the folding structure, and this and that.” We said, “The last time we checked, every generation has DNA handed to the next generation. There's no manual that describes all these things. The DNA has no idea what a protein is and no idea what folding is, and yet the function follows.”

We said there must be an encoding of that knowledge in the DNA. There are enough data in there somewhere. There are patterns that are encoded, which we don't understand. On that whim—otherwise known as a hypothesis—we started and tried it. That's what's changed now: the incremental cost of asking that question has come down, not just with AI but also with experimental setups.

Within a couple of years, we could show that, at least for antibodies and how they bind to their targets, you could begin to make some pretty interesting leaps computationally that you couldn't make experimentally, at least not in the same amount of time.

That has led to a company called Generate Biomedicines, which was one of the very first large partnerships that Nvidia did in the biology space. That's how we know Nvidia quite well. This was 3 or 4 years ago. Now Generate has more than 15 different computationally designed antibody programs, some in the clinic and some advancing to the clinic.

A lot of people say, “Show me a drug that's been computationally designed.” The answer is that showing a drug is a lot more than the design and early testing. But we've done that in this space. We've applied it to cell models, DNA, RNA, all sorts of molecules, and lipid nanoparticle design.

Some of the very cool advanced things we're doing now are to create novel platforms that can essentially create autonomous ways of doing scientific discovery: generate hypotheses, specify experiments, run them, collect data, interpret the data, iterate hypotheses, and just do science the way Waymo drives cars.

We're nowhere near having Waymo on the streets with these things, but we can definitely show the elements and tie them together. In narrow spaces, we can show what the future of that could look like.

Think about what it looks like if, in chess and Go, you train at a level where a million people playing for 1,000 years might play enough games to reach that level. We're beginning to see hints of that.

The area that really interests us is multi-agent systems that can do emergence. We're doing that in the product space—flat-out new brands and new products. We're learning carefully about those business models. We're doing it in mental health, with early intervention, so that you can use agent-based interventions.

I don't mean training the system to do what a doctor would do. I mean letting the system interplay and learn from the dynamic between different types of agent-based models. There is very little we're not excited about. You'll notice I didn't talk about all the productivity gains—automated document writing, summarization, and so on. We're doing all of that like everybody else, but that doesn't require our extra effort. These things feel pioneering to us.

You were just describing hypothesis generation and other advancements of this type. They're leading to a quickly increasing number of candidates. But, as you said, that's not “show me a drug.” That's just top of funnel. Maybe it's higher-quality top of funnel as well as greater volume. What are the biggest bottlenecks in translating these innovations into market-ready therapies, and how do we address these challenges more effectively?

If the goal is to get more therapies to market, let's work backward. The last thing you have to do is file with the FDA a BLA or NDA to get approval. The step before that is conducting a phase 3 trial to show, at the right dose level and across a large enough population, statistically significant superiority without toxicity.

We can work backward from there. Those last steps are regulated. Unless there's a rethink that, with data-driven approaches, we don't need quite as much of the analog testing we do because we can create models of what the data is telling us, we're going to be waiting for large-scale trials that cost hundreds of millions of dollars.

The day will come when we can do things differently. Until that day comes, we're going to be waiting for those trials.

Don't you believe that should come sooner?

Of course it should, especially if it could improve the disease I'm going to die from. It's totally nuts. The way to do that, to me, is pretty straightforward: it's called Operation Warp Speed.

What we saw during COVID is clearly possible. It's a way of thinking that results when you're under an avalanche of disease threat. You get organized so that the private sector, the public sector, and the regulator don't cut corners, but essentially realize that the objective is a solution, not slowing down the process just to be on the safe side while people are dying.

When the consequence of not doing anything goes up, people act. The sad part is that the consequence of doing nothing in cancer and neurological diseases is playing out every day. It's just slower, so it doesn't feel like an urgent threat.

I don't have a magic wand, but if I did, I would at least run some experiments in how we think about data-informed ways to interpret the results of trials, including adaptive trials.

The area we're working on is understanding the state of a particular patient better. Right now, people say stage 1 cancer, stage 2 cancer, as though those stages are fixed. The medical profession created these stages, and every once in a while it changes them. It's a joke that has nothing to do with the underlying science and biology. No offense to the colleagues who work in that area; it's just a super-macro approximation.

I think we now have the tools to molecularly micro-stage what I call a patient's bio-stage disease on a trajectory that isn't 4 stages, but 75,000 stages if you want. The key is that you can then look at which mechanisms are turned on and off in any one disease during the trajectory, and identify the subset of people you should choose to homogeneously test your drug hypothesis on, so that the other people you tested it on don't defeat your trial.

You might then do a smaller trial, get a smaller indication approved, and expand that with AI tools and measurements. That could substantially increase productivity and lower the time it takes to get a drug out.

Unfortunately, it doesn't suit the interests of a large company that wants to give the drug to everybody. But for biotech companies that otherwise would be dead while waiting, this might be an interesting innovation.

We have to get regulators to be open to it. There's a lot of information in patient data that could inform us in ways that help us find the right mechanisms to go after. Accessing that patient data, given HIPAA rules and everything else, and using it to train models could tell us, “Now we know that it's this particular mechanism in a subset of Parkinson's disease that you have to go after.”

That's happening slowly, but it can also increase not just the top of the funnel in molecules, but the process of selecting down to the things worth working on. You already have human data showing that if you can do this, you know the consequence, because you have human data showing it in the genetic testing we can do. There are lots of exciting things, but it's slow.

You mentioned Operation Warp Speed, so we have to talk a little bit about Moderna. What do we do if we have another pandemic, in terms of mounting a different response?

Before the pandemic, everybody thought it took years and years to develop a vaccine. After the pandemic, there was no hurry to develop a vaccine because everybody thought it took 4 years to develop one. What's the rush?

Of course, there were technology advances that made it possible to develop a vaccine in 3 or 4 months. But people had to get their heads around the question of why not try. Unfortunately, enough people had to die, and enough calamity and economic shutdown had to happen, for people to say, “You know what? We have the money. Let's just try it. If it doesn't work, then we're no worse off.”

I hope we don't have to go through the same gestation period of debating. There were a lot of people who said you couldn't do any of this. Unfortunately, there are people to this day who think vaccines can't work or don't work. That's a problem.

In any case, I really hope that if it happens again, there will be a coordinated response in which multiple approaches are thrown at it. I feel very confident that there isn't a life form or disease for which we can't find an appropriate antidote or vaccine, if we can deploy the best technology in the most coordinated way with the right incentives.

A key thing Operation Warp Speed did was set a value outcome for people who tried. The government said, “We will buy X number of doses at this price.” So we could go to investors and say, “We don't know what the probability of success is, but there's a clear market signal.” That would go a very long way outside of pandemics as well.

Moderna made many non-obvious bets along the way. Every AI and biotech company I talk to wants to be a platform rather than a single- or two- or three-asset company. How do you think about the trade-off between investing early in a platform versus clinical assets? In a different field, as we were talking about before, investor climate and the macro changes around you mean that not every biotech survives. How do you think about that, having been through several cycles of investor climate?

Let me answer how we think about it at Flagship, and then how one should think about it more generally. We of course represent only one subset.

At Flagship, because we go to far-out places looking for undiscovered value, the notion that you do that to come back with one asset is the definition of insanity. If you're going to do that, you might as well bet on well-known, proven technologies—a slightly different version—and hope that your lottery ticket gets pulled. I'm sorry for being a bit rash, but that's how I view it.

If you're going to go after RNA for the first time, DNA for the first time, gene writing, gene editing, or computational proteins, you need to diversify because you don't know which of these will get knocked out for reasons that have nothing to do with the underlying technology.

Hence, every single one of the 100 companies we've been involved with over 25 years has been a platform. Every single one. There's no exception. We embraced platforms from day 1, but for a different reason: we wanted to go beyond adjacencies, beyond the reasonable zone, into unreasonable things.

Why doesn't everybody do it that way, even though they want to? The answer is what you said: the capital it takes to do one thing is already high. To do 2 or 3 things is even higher. To do a platform that could potentially do 15 things, even if you don't use it for that, is very expensive.

Second, investors don't properly value platforms because they don't assign value to the correlated option value of one program becoming de-risked based on 3 other programs. You don't get the credit, and you get the deduction because suddenly it's overly expensive and overly complex.

The last reason investors often don't like it is that they think it's a strain on management's capabilities. It's one thing to execute 1 program, but if you're going to execute 2, 3, or 4, that creates a lot of companies that are essentially going to be dead because that's just a number game.

In an environment like this, there's almost a mass-extinction event going on for a lot of these single-asset companies. That's been accentuated by the recent bets China has made, both through the central government and through the resulting startups. They've gone after that same space at a much lower cost and with a very different barrier to entry in terms of clinical data.

I don't know how you compete as a single-asset biotech company against a Chinese invasion of assets that we're seeing on the business-development front. Good for the pharma companies that can access those assets. Those are great assets. But remember what I said earlier about commoditization. I don't know how you avoid commoditization with the higher cost structure we have here.

That logic brings me back to platforms, because at least they give you a chance to do partnerships and try to come up with ways to stay alive. It's not that our companies aren't dying. Some of our companies are meeting an unfortunate end as well. But at least some of them are not.

That represents an opportunity for the entrepreneurs who truly have a platform, and for the investors who are able to invest through that.

Last question for you. In your annual letter for 2025, you described the concept of polyintelligence: the integration of human-, nature-, and machine-derived intelligence. Where do you think human intuition will remain most relevant?

You switched from intelligence to intuition. Intuition is basically a model. All of us generate our own models and then use them in short form. It's actually the closest thing to a computer large language model that humans have. We don't call it a large language model, but that's exactly what it is to me. Anywhere you want to use an LLM, you could use human intuition. The problem is that LLMs have a lot more data and have been trained across millions of people's worth of data, while yours is trained on yours.

I wasn't criticizing you. I was just picking up on your concept and asking where human—let's just say human intelligence—is most relevant. Is that what you mean?

Yes. I think it remains to be seen. I describe in this letter the notion that thinking the real frontier is between humans and machines misses the fact that most science happens between humans and nature. Now we have a machine that can intellectually support our inquiries into nature.

It's really a triangle. It's not a line between humans and computers. It's a triangle in which human intelligence and machine intelligence, coupled with nature's intelligence, inform one another. These 3 actors will adapt to one another.

The role of humans in that is very important. Humans compute completely differently, and they have a way to act that's also quite different from the way computers act today and from the way other forces of nature act. I choose to think of it as a 3-way thing. It's a beautiful new axis of emergence, and what's going to come out of it is the future of life.

I love it. What a wonderful note to end on. Thanks so much for talking to us, Noubar.