[0:00:00] To watch the video version of this episode, visit incumbentsandinsurgents.com. This podcast is presented by the Towerbrook Healthcare Institute. Incumbents and Insurgents in Healthcare with Eric Larson. This presentation is for informational purposes only. Please review the full disclaimer at the end of the episode. Ian, nice to see you. Jack Jack, how are you?
[0:00:22] I feel, A, overdressed, which is always the chronic thing, but I don't have my obligatory private equity venture capital vest. I do have one here in the room, actually. There are no monograms here. This is just a normal vest because it's 23 degrees here where I am. So I'm actually wearing it legit, and the heat has not kicked on yet this morning. So thank you for that, though. Yeah, I'm in Utah today. It's 29 degrees, so it's all good. All right, let's get to this. I'm super excited about this. After a little bit of a hiatus, you know, we compressed.
[0:00:51] We compressed four recordings into a week, you know, with Dario and Vinod and Alex Karp and Tom Holovick. And we'll sort of meter them out because it's kind of this crazy embarrassment of riches because these are all brilliant humans and deeply authentic people that care about healthcare. But let's start with... So that was the coming attractions. That was the coming attractions, I guess. Coming attractions. Yeah, just kind of hook everybody. All right, perfect. Let's talk about our guest for today, Dario Amadei, who is the founder, co-founder...
[0:01:21] And CEO of Anthropic, which doesn't really need a big introduction, but it's one of the two fastest growing companies in the history of capitalism. And there's a real gravitational pull that I know that you and I have to Dario. And I've got some thoughts just to kind of frame for our group who he is and why I think he's such a force for good in the world. But Ian, I'd love your thoughts first. So with Dario, you've, listen, you've spent more time.
[0:01:50] With Dario individually and obviously Anthropic, but the time that I've spent, which has been incredibly valuable, and the day that we had in July with him and his team and with 12, 14, 16 of the 150 of the various CEOs of systems and payers in U.S. healthcare was incredibly profound. And I think the reason it was profound, not only
[0:02:16] From a demo perspective, but the appreciation of the orientation that he's turning Anthropic towards, which is this concept. And you guys talk about this, you know, in your session together, which I think is really important, which, you know, I had to wait towards the end of the discussion for it to get there as you guys hypothesized about humanities and the future of the world and the impact of generative automation and super intel.
[0:02:44] But really this concept of the diffusion of the technology and the fact that he is turning Anthropic towards enterprise transformation, enterprise application. And he said one thing that I loved, which is that the speed of transformation is actually the quotient of or equals the advance of the technology times the ability to diffuse it.
[0:03:07] It's almost a challenge that happens at the end of the discussion that you guys have, which is it's one thing where you recognize is, hey, I'm telling the world, I'm telling the 150 about the technology, the capabilities and showing it to them.
[0:03:21] But it's incumbent upon us at Anthropic to work with you, the 150, to build the bridge to help you understand how to diffuse it within your organizations. And I still know that leaves us with a, oh, what's next for many CEOs and many people, hopefully, that will be listening to this discussion. It's almost as if he threw down the gauntlet and almost welcomed those into his lair to actually engage and have that conversation. And I know we've begun to do that with some of our portfolio companies with them.
[0:03:49] Yeah, I think that's super well said. And, you know, Dario is arguably one of the two or three most consequential people and architects of this paradigm change, right? And I think he can lay claim to being one of the real sort of designers of this almost speciation event, right? Of this new synthetic intelligence that is becoming polymathic, that is spilling over into every domain.
[0:04:18] And what I really appreciate about Dario is his authenticity. He's a computational neurobiologist. He cares with deep sincerity about healthcare, democratizing access, bringing deflation to it. And how do you, you know, he wrote quite eloquently in his Machines of Loving Grace essay last year, this notion of returns to intelligence.
[0:04:43] And biology is the domain that arguably has the highest returns to intelligence. And, you know, this sort of vision of this compressed 21st century of 100 years of biomedical and scientific advances telescoped into a five to 10 year period. It's really a captivating vision. It was interesting when he explained to you how he came up with 10 years. You know, why 10 years versus something longer or versus something shorter? It was very interesting. And I love that. Yeah, I love that, too.
[0:05:11] Let me ask you a quick question, though. Given how much Anthropic has grown, and if you think about Frontier Labs, I think you have to think about Anthropic as an incumbent. Is Anthropic an incumbent or an insurgent? And how come there's not a counterparty in this discussion? Or not a counterparty, but an and party? A partner, yeah. This thing is kaleidoscopically changing by the minute. And the fact is, two years ago, Anthropic...
[0:05:39] was an emergent company. Now it's, again, the second fastest growth company in history. And I do think can lay some claim to being an incumbent, but that just underscores sort of the exponentiality of the tech and how quickly the game board changes. I really just wanted to do this, and you and I debated about, do we pair Dario with somebody or not? I just think that there's so much subtlety, nuance, and complexity in the dialogue about what he's building.
[0:06:08] And that we really just wanted to honor that and just, you know, have a sort of an unfiltered back and forth. You know, the one thing I'll say is that, you know, Dario's gotten a lot of attention and really kind of got villainized a few months ago when he came out and said, look, I'm looking at the trajectory. There is going to be major job dislocation here. And he actually put numbers on it. He said 50% of all entry-level jobs are going to get disintermediated. And that may translate to 10 to 20%.
[0:06:38] And he just got pilloried for that by people I otherwise really respect. But I just think that there's this real reticence in this moment to talk about the fact that there may be job dislocation. And it may be at an industrial or an economy-wide level. And one of the things I really value about Dario is the fact that...
[0:07:05] He's willing to take on popular positions and he's willing to voice sort of the risks on mechanistic interpretability and export controls with China and possible job disintermediation. And so I think with that, Ian, we should probably turn to the conversation. I think he's such a force for good and a great partner for the 150. And obviously he's been a great partner to you and me working with our R1 company and some of our other.
[0:07:33] Close allies and portfolio companies. And so we've seen it firsthand. So I'm excited for this. And I personally think it's just a really revelatory kind of set of insights he shares. And listen, you guys have part of the conversation when you talk about the restoration of the humanities. It's quite an interesting part where when you think about the work that's going to be automated that we believe with so many white collar jobs. But do those people lose those jobs or do they level up?
[0:08:00] Do they level up? And are there just the interpersonal skills, the negotiating skills, the various things that, you know, do people actually become more productive in that regard? It's going to be very interesting. I'll be curious to see, you know, what kind of feedback you get. So looking forward to the discussion. Awesome. Good to see you, my friend. Thanks for having me. Well, thank you for having me. I mean, we're here in your awesome San Francisco downtown offices.
[0:08:25] Which is a very cool affect. It's got a good vibe. Yeah, glad you like it. We filmed a lot of stuff here and I've become very fond of the space. We've only been here a little over a year, but I really like it. So Dario, first off, I'm super grateful to you. I know you're like frenetically busy. I mean, I think Anthropic is only the second fastest growth company in the history of capitalism. So it's not like you've got some other stuff to do. But the reason I was so excited for this is...
[0:08:50] I know you have almost the same, if not even a greater evangelism than I do about how, what Anthropic and a very small group of protagonists here in Silicon Valley are creating. What the implications are for biology.
[0:09:08] Broadly, but the transformational impact for U.S. healthcare. And you and I have had the privilege of doing a couple of these conversations in front of healthcare executives, and you've been super gracious, and we've hosted some of the leaders in establishing healthcare here. You've heard me sort of kind of repeat this, but I think there's probably about 150 CEOs in U.S. healthcare that all transformation needs to go through. They kind of refract.
[0:09:37] And so the purpose of this dialogue is to just kind of dimensionalize what do they need to know about this moment, the exponentiality of the tech, the implications for agentic capabilities on workforce and organizational change, and then just this transformation in synthetic and computational biology, which you wrote about so eloquently in Machines of Love and Grace. So the first question I want to ask you.
[0:10:06] When healthcare leaders sit down with you and they ask you to kind of demystify the technology, like how do you explain to them the emergent and the existing and the future capabilities, what the models can and will be able to do?
[0:10:23] Yeah, you know, I think one of the things that I point to, one of the things I emphasize first is the rate of progress of the technology. I think that's one of the key things because there's a natural tendency, I think, especially when you're looking at AI for the first time, you can play with it. You want to get a sense. It's the right thing to do, right? You want to sit down and get a concrete sense of the technology and you say, okay, the models can do this and they can't do that.
[0:10:49] And, you know, that's an accurate way to get a sense of what's happening right now, right at this moment. The known wisdom, the received wisdom about what the models can do and what they can't kind of gets surpassed and gets redefined, right? To give an example,
[0:11:02] Far from healthcare that we don't even work on. You know, there was this idea a couple years ago that AI image generation models couldn't make hands, right? That they would make hands with six fingers or seven fingers or something like that. And so that this was a reliable way of telling apart AI from non-AI. And it was for a moment in time. And then it changed. And so just the velocity, the fact that the field is always in motion and that it's in motion so fast.
[0:11:30] That two or three years ago, you know, the models would be almost always inaccurate if you ask them for medical. It was just downright dangerous, right? Because they just didn't know what they were doing. And now we're more in an area where, honestly, they're right more than they're wrong. And they, you know, they might be better than humans, right? I have some stories about Danielle and my sister, you know, finding the right medical treatment with AI when a lot of expensive human doctors couldn't help her. Maybe we'll get to that story.
[0:11:59] Now it's at the point where it's like, you know, they might be better than humans at average, but they still make mistakes fairly frequently and we have to understand those mistakes. And so I think when we think about, you know, sitting down with especially the healthcare side of it, where, you know, there's just, there's so much more process, there's so much more regulation, there's so much more to do, you know, even then the kind of scientific discovery side.
[0:12:29] It's going to take us a long time to really build this into our health systems. So if it's going to take us a long time, let's try to only do it once. And let's try to, as we pay that big one-time cost, skate where the puck is going. I love that. And one of the kind of recurrent themes in our conversations has been this notion of the most powerful technology in history of...
[0:12:53] You know, civilization, right? It's a multiplication of intelligence. And in many quantifiable ways, it's already, you know, the synthetic intelligence supersedes our biological intelligence in some demonstrable quantifiable ways, right? But the big theme for me is that US healthcare has the greatest susceptibility to disruption from this technology for very specific reasons. And I'll sort of suggest three reasons, and I want to kind of decompose them with you.
[0:13:20] You know, the first is this notion of labor intensity, right? So U.S. healthcare is the single most labor dependent sector in the U.S. economy, employs 23.8 million people. And you've been super articulate about thinking through the potential labor dislocations from this technology. So I want to come back to that in a minute. The second is this notion of, you know,
[0:13:44] Biology and data superabundance, right? Like one third of the world's data, like 60 zettabytes is in healthcare and it's increasing 36% per year. As you articulated in Machines of Love and Grace, this sort of returns to intelligence, a country of geniuses in a data center, like that quantum of data has great returns to intelligence. Let's come back to that. And then the third one is...
[0:14:08] You know, here we are in Silicon Valley, right? And there's almost this sort of mutual incomprehension between Silicon Valley and establishment healthcare. And as a result, like the 150 CEOs have been impenetrable to every tech phase shift to the past generation, internet, mobile, social, cloud, big data and analytics. And so I want to look at it through that lens and start with labor.
[0:14:29] Because you very courageously came out several months ago and said, hey, I'm following the exponentials, right? And we're going to see massive labor dislocation. In fact, 50% of entry-level jobs may get vaporized, right? And that may lead to 10 to 20% unemployment. And what was really interesting was to just see the reaction to that comment. There's almost this Orwellian kind of conspiracy that thou shalt not talk about labor dislocation in this moment.
[0:14:58] So let's start with the labor part for health care, given the labor intensity in health care. How do you think about the capabilities you just described sort of ricocheting through the most labor addicted sector in the U.S. economy? Well, I mean, first of all, you know, I would say there isn't a sector that illustrates the stakes here and the moral dilemmas better than health care. Right. Because, you know, you know.
[0:15:24] As you say, it's an incredibly labor-intensive field, and yet it's also a field where the benefits to the end users, the patients.
[0:15:34] are so great compared to almost anywhere else, right? We basically have to do this. If we can deliver the better care, we have to prioritize the patients. We should be thinking about the patients first and the employees who deliver care to them second because our primary duty is to the patients is very clear. But at the same time, we don't want to create a society where a bunch of people are thrown out of work with no solution to offer them with no other thing that they can do.
[0:16:03] No way to reallocate them to make the productivity even greater. So that's really the dilemma in a nutshell. And that's the thing we have to address. Going back to, you know.
[0:16:13] Why do I think there will be this large disruption? So if I look across healthcare and other areas, there are basically two factors that drive forward the technology. One is the underlying improvements in the models, right? Every few months, we release a new model, and the new model is just smarter than the previous model was. It can do new tasks. In a healthcare setting, again, you know, previously it was like, you know, this thing really wouldn't say things that you trust at all, to the point where it's like, oh, you know, it's a little like the medical information you look up.
[0:16:42] to where, wow, this thing is almost like it's offering a second opinion to where, oh, it's actually better at processing information than humans' doctors are, at least on average, even though it makes mistakes. So one factor is the improvement of technology, and the other is what we call diffusion or adoption of the technology, even if we froze the quality of the technology, right? Even if we took today's AI, today's LLMs.
[0:17:08] There are lots of industries in which it's not being adopted as widely as it could because of various barriers, right? And, you know, healthcare is really one of those. There's all this data we'd like to have access to, right? We'd like the model to read.
[0:17:22] Patient records and help to express the best treatment. We'd like it to have access to all these types of data. But there are laws, often for good reason, why we can't just do that. And so the integration is difficult. And consumers react, patients react by just uploading their information to the LLMs, right, kind of outside the system. But the fact that people are doing that shows that we have all these barriers to adoption. The prediction I made was basically a prediction.
[0:17:51] And when you compound those two factors, the adoption of the technology compounded with the underlying improvement, those two together create enough of a powerful force that I said, look, I don't know exactly the pace at which it's going to proceed. The thing I know better is the underlying improvement of the technology. That factor is going very fast. And the models are going to be capable of that in a year or two.
[0:18:19] Now looking even beyond healthcare, if I look at banking, consulting, law, you know, finance, sales and communication, like the entry level, the lower level tasks, if I just think about it, it's like they're tasks that require processing information and require cognitive labor. They don't require a huge amount of like super original thought, right? There's something routine to it. Everyone starts out that way.
[0:18:45] And if I look at the LLMs, they're getting to where they're capable of this. And so I look at it and I say, okay, it's going to take time to diffuse. The models still need to get better and more reliable. But you put that together and boy, the models are really going to do it. And of course, the humans are going to try and take a step up. The law of comparative advantage is powerful in that even if the humans only do 10% of the task, there's a world where, of course, it augments the humans. I think at the end of the day,
[0:19:14] There will still...
[0:19:15] Be places, and I think it may be a lot of places where there may be less need for human labor rather than there is more. And that it's a problem across the economy. And so I don't want to give the Panglossian story of, oh, humans will just adapt. Some of that will happen. We should make as much of that as we can happen. But that's not a full solution. And so that's why I felt I needed to warn and not just say everything's going to be okay. And I'm very grateful you did that because it strikes me as super unconstructive when we did.
[0:19:45] Not just the possibility, but I align with you the probability that you're going to see some technological unemployment. And the reason I think healthcare has a greater vulnerability to this than even some of the other industrial verticals you talked about, consulting, finance, legal, etc., is because there's such an accumulated tech debt.
[0:20:07] Technology is intrinsically deflationary, right? It augments productivity and it's augmented productivity in just about every other industrial sector. U.S. healthcare is the only industrial vertical to see negative productivity growth. And so you having the sort of forthrightness to call that out, I think is really important. Now you've got 150 CEOs, right? That if they look at their $4.9 trillion sector, $2.7 trillion is in labor.
[0:20:34] And it's going to have massive implications. And the good news is, like, I don't believe in the lump of labor fallacy, right? Like, there is going to be new work created. And there's also 2 million unfilled jobs in healthcare. I think if we don't have the courage to call this out, how can we possibly prepare for it? And so I want to double down on one thing you said. Because...
[0:20:58] I think they are heroic.
[0:21:26] They are mostly altruistic. They want to serve. And they took the Hippocratic Oath, right? And yet, if you think about doctors in this country, there are a million of them. They're the single highest compensated occupation in the country. Nine of the top 10 and 20 of the top 25 highest compensated occupations are medical. And yet what we're talking about is going to be demonetizing, right? Like 1 billion people in the world have access to a primary care doctor. 5 billion people in the world have access to this. Absolutely.
[0:21:55] How should we prepare our clinicians that are also the most respected profession in the country? What advice do you have to clinicians to think about this tech shift?
[0:22:10] Maybe I'll start with the story I alluded to. Daniela is my sister and co-founder. She took a few months off when she was pregnant. She just had her second child a few months ago. And while she was pregnant, she had a series of recurring infections because, of course, pregnancy can weaken the immune system.
[0:22:30] And, you know, she went to a bunch of, you know, very, very fancy doctors in line with what you're saying, all of whom wanted the best for her, all of whom were trying to do everything they could. But, you know, they're busy. They have a bunch of patients, even the top of the line guys, right, have a bunch of patients to deal with. They only have so much time. And the data is complex, right? You know, you need to take this culture and, you know, you need to do this imaging. And there's a long history, right, where there's different specialists and you talk to them.
[0:22:59] This specialist three weeks ago, and then there was this infection, but no, no, it was a different, there was a different time that you went in and you talked to someone and it's just, it's hard to keep straight, right? It's not, it's actually not something humans are inherently good at. And so all of these doctors assume that the infections that she had were viral.
[0:23:17] So she just uploaded everything to Claude and Claude methodically went through and, you know, thought about it. And finally it said, you know, I think you have a bacterial infection. By the way, based on your symptoms the last 24 hours, I'm worried that your infection might be going systemic, which is obviously a very dangerous thing. And so...
[0:23:35] You know, she saw this and she was like, okay, I really got to get an antibiotics. And, you know, it probably would have been caught anyway, but Claude caught it earlier than it would have been caught. And so I am very grateful to Claude. She is very grateful to Claude. Her husband is very grateful to Claude. It's just really amazing that Claude was able to do it. And I think of it as like a second opinion who's more different than different physicians are from each other, right? On one hand, that points towards some of the weaknesses of today's
[0:24:04] Clinicians that it's hard that they can miss things. It's hard to think outside the box. It's hard for them to process all the information and to kind of, you know, to kind of give their full attention across so many things, whereas AI is designed for it.
[0:24:17] I think there's a story.
[0:24:47] of complementarity here. You know, again, I don't want to have this Panglossian view where it definitely is going to be. It definitely is going to be demonetizing, but we have some choice here, right? We can go to some world where it's, you know, potentially good for clinicians and good for patients, gives a greater degree of access.
[0:25:08] Increases the quality. Humans do what they're best at. AIs do what they're best at. I don't think it'll necessarily turn out that way by default, but I think if we work hard, we can make it turn out that way. I share your optimism. I think about how augmentative it's going to be for the clinicians. And candidly, there's so much that physicians do that is...
[0:25:32] Bureaucratic and administrative. In fact, the average primary care doctor spends 30% of her time laying hands on patients and the other 70% fighting with insurers. How much of it is billing, insurance, documentation? I mean, that stuff should just be all AI. Like that is so repetitive. And I think, you know, the administrative simplification side is like a no-brainer, right? And I think it's going to restore joy to the practice of medicine because it's epidemic, the burnout and the disengagement among clinicians.
[0:26:02] And I think at some point, if we start the conversation now, we should think about how do we adjust compensation, right? I mean, the typical primary care panel is like 1,000 to 2,000 patients. I mean, you could theoretically have 1,000 to 10,000 patients and do a lot asynchronously or do a lot agentically. Like Vinod, who we did this podcast with, talks about 80% of 80%.
[0:26:25] Yeah. Is there a way to kind of...
[0:26:54] Expand the ratios, the number of patients per doctor, while also allowing them to be more empathetic, right? That's kind of the positive some world. Again, we have choices to make here, right? We have to like find the positive some world. But yeah, there's some world where doctors can see more patients. They're less rushed.
[0:27:13] More people get medical care. It's higher quality. It's more human. The doctors, you know, somehow make more money. We're expanding the pie. Like, I think that's possible if everything goes just right. I totally agree. And, you know, not to get super technical, but there are payment mechanisms that will encourage and reward that, such as capitation or a prepayment, right? Where you really get off the fee-for-service treadmill. And we should talk about that. I think one of the most beautiful parts of your Machines of Love and Grace is you said, look,
[0:27:43] I approach these topics with like infinite curiosity, but with humility too. And we ought to convene others to dialogue on these questions. And it starts with being courageous and acknowledging what are the dislocations that will happen. I mean, if you look across the history of technology, they tend to accrue to the benefits of society, but the dislocations tend to be very specific, right? And part of what's motivating me in this conversation and in this whole podcast series of incumbents and insurgents.
[0:28:10] Try to confront some of these inevitable or highly probable dislocations and not be Panglossian about it. Not just assume it's all going to end up in the best of all worlds, right? We got to really think about that. Another beautiful part of your essay was this notion of the compressed 21st century, right? Where you take a century of biomedical and scientific advances and concentrate it into a really telescope small period of time, five to 10 years.
[0:28:36] And you actually did a little retrospective on some of the seminal medical breakthroughs. And you said they come kind of metronomically once a year. And the mapping of the genome or cell therapy, especially CAR-T or immunotherapy or advances in microscopy or advances in CRISPR. And one of the arguments you made is that if you look at a lot of these, like mRNA,
[0:29:02] That existed since the 1960s. CRISPR we knew about since the 1980s. And the returns of intelligence to biology are very high. So I'm hoping for this group, give a little bit of an articulation of the virtual biologist.
[0:29:17] Yes, exactly as you said, right? The CRISPR example is always the one I take where, you know, it was like we knew since the 80s that, you know, there was this weird property of the bacterial immune system, right? And we knew that we were trying to do gene editing, but the idea that these were connected to each other, that one was a mechanism for doing the other, somehow that languished for a long period of time, you know, until a small set of brilliant people figured it out, right? I talked to some of the inventors and, you know, it was literally like,
[0:29:46] I was reading about X and then I went to a seminar, I heard about Y and my kind of brain made the connection, right? So it's great that brilliant people could do this, but think about how much AI specialize in being able to draw these together. I think the breadth of AI systems is already broader than it is for humans. They're not as smart. They're not as able to draw the right connections and reason, but they're quickly making up.
[0:30:13] For that, I could ask it about the history of Japanese samurai. I could ask it about...
[0:30:18] The history of baseball, I could ask it about how to program GPU kernels. And I know maybe a half out of three of those, right? The AI knows about those and a thousand other things. It isn't quite as good as reasoning through things as I am, but it's quickly getting there. And so if I look at what I did during my former career as a biologist, I just think about everything, right? Reading the papers, proposing the experiments, actually controlling
[0:30:47] I'm controlling the lab equipment to do things, looking at the previous literature and its gaps, thinking about my own common sense, looking things up in the protein databases, having other people read my work. Every piece of that is a piece that LLMs, as they get smarter, can do if we have the right.
[0:31:08] What if you have a thousand Feng Zangs or something like that, right? You know, one of the inventors of CRISPR. Could we have 10 CRISPRs every year, right? I mean, you know, these inventions are all different. You can't...
[0:31:33] Turn them out like it was a factory. But there are a lot of them to figure out.
[0:31:38] And some of the bigger ones really drive things forward, right? Think about not only how many medical treatments come from CRISPR, but, you know, just doing basic biology, right? You know, just doing basic controlled trials where you say, you know, okay, you know, the mice with these genes and the mice with these genes. That is much easier with CRISPR than it was with previous techniques. And we have a level of targeting and a level of accuracy that we didn't have before.
[0:32:08] And so it's this broad-based benefit across the field. And I look at how many of these are there. And again, they could have happened earlier than they did, right? They wouldn't have happened unless someone really smart figured them out. And if there were more smart people around, they probably would have been figured out years or decades earlier. So that kind of makes me think that the whole enterprise really can be sped up.
[0:32:33] And then, yeah, it would be this kind of virtual biologist. You know, we've taken little steps towards this. We released this Claude for Life Sciences a few weeks ago that, you know, it's being used from academic researchers to folks at, you know, pharmaceuticals and even some in the healthcare side of things. So that's just a tiny step. Our work with Benchline is a tiny step. You know, our work with...
[0:32:56] Pharma companies to simplify clinical trials and paperwork is kind of a tiny step. But I think there's a lot of acceleration here. The counterargument that I always get, which is one that should be taken seriously, is, well, okay, you can speed up the science and the discovery, but then you have this horribly complex clinical trial system that has regulation and you just...
[0:33:20] Are you doing experimentation on people? So some of the regulation actually makes sense. And, you know, rules that you have to follow and should follow when you do this experimentation so that people don't get hurt. And then beyond that, there's the actual scaling of the delivery of care. So that is definitely going to slow us down, right? And so when I said, how long will it take us to get 100 years of biomedical progress?
[0:33:42] I chose 10 years very specifically. I said we won't be able to compress 100 years into one year. But my hope is that if we can get all this basic biomedical progress, and, you know, some of it will make it easier to predict clinical trials. Some of it will make them go faster. But we still have to do one or two go-arounds, right? It's like maybe we can invent a bunch of stuff and then push it all through the pipeline roughly in parallel, right? That would be the best of all possible worlds. And so if I think about...
[0:34:11] How long does it take to push things through the pipeline? Once is probably too much to do everything in parallel. But, you know, give it three or four go-arounds. You know, like if we were so smart that we could push cell therapies and gene manipulation and, you know, everything that we can possibly imagine. Aliens came down and, you know, they gave us alien tech for manipulating biology and we had to push it all through the clinical trial system and make it happen.
[0:34:37] Probably would take us three or four iterations to get it right. And how long does that take? Maybe 10 years. That was how I got that. But, you know, it has its limits, but it's exciting. 100 years in 10 years is a lot, right? That's, I think, the right mix of hope and skepticism. And it turns out to something very radical. But, you know, it has its limits, but it's something pretty radical if we can get it right. And it's a captivating vision, and it really does capture the imagination.
[0:35:04] And I want to kind of decompose it with two kind of dimensions. First is a little bit of a philosophical or maybe even an epistemic question. And then is the practical question of how do our linear institutions adapt themselves for this exponential tech? But the kind of philosophical or epistemic question is, it's really interesting because what you described.
[0:35:22] It's almost this sort of polymathic approach, like looking interdisciplinary and seeing connections among seemingly disparate things. And there's something intrinsic about the models in that they think analogically and that they are polymathic and they have such a breadth. One of their strengths, yes. And to me, that's not just a replication of biological intelligence. It's something that is superior. And it's almost the philosophical question is, if that's true.
[0:35:48] If predominantly great ideas come from analogic thinking and connecting things that are seemingly disparate, right? What is the role of human creativity in this process? And in a very real way, all of our advances, like the most advanced cognitive technology since the Enlightenment is almost, I'd think about it as the scientific method, right? And scientific method emerge because of two things, like cognitive limitations and data scarcity.
[0:36:18] And suddenly we've kind of solved.
[0:36:21] Both. So if I'm a researcher or a scientist, like how should I think about my role in this sort of generative sort of epistemological world now? I think there's the short run and the long run, right? So, you know, in the short run, I would go again to comparative advantage where, you know, I would think of it as, you know, everyone who's currently a PI, you know, can get a thousand grad students and everyone who's a grad student can also get a thousand grad students. And so it's about.
[0:36:49] Thinking about what are the most, using your knowledge of the field to what are the most promising directions, where can you set people to work? We're already seeing that in the software industry now where people at Anthropic, for instance, and at some of our customers don't write a lot of code ourselves. We kind of set the models to work.
[0:37:06] And we become kind of higher level supervisors. So I think a lot of the key discoveries are going to happen in that way. Now, do we get to a point where we totally close the loop and kind of the models are doing it themselves? Probably eventually we will get to that point. And then I think it becomes more kind of art than science. We have these agents that are very autonomous, but they're acting on our behalf. And I think our role is then at a very high level to think about what problems should they solve.
[0:37:35] What should society be trying to solve? I have this interesting actual thesis I've been thinking about for the last...
[0:37:43] This is actually new. I don't think I've said it in public before, which is over the last 30 years or so, we've really seen a shift from kind of, you know, humanities to tech as where kind of the action is happening. You know, science, science and technology, in many ways going back, you know, hundreds of years or more since the industrial revolution. But I think a lot of these humanist questions are going to come back to the fore.
[0:38:08] Which is like, what do we prioritize? How do we live our lives? How do we educate people? What's the good life? What's the makeup of a good society in which the AIs can provide most things, but how do we make the pie go around everyone? How do humans find meaning?
[0:38:25] So I think there's almost going to be like a revenge of the humanists or something. Oh, I like that. And by the way, I think that's totally true. I've been telling my kids who are college age, like, you know, study the humanities, get the broadest possible culture. Because when you're going to interact with the alien intelligence that is internalized, the digitized sum of human knowledge, the broadest culture will extract like the greatest insights. But it's a really interesting sort of musing that you're offering.
[0:38:53] You know, there's this article in 1930 written by John Maynard Keynes, and he said, economic possibilities for our grandchildren, right? Okay, so, and he predicted with actually like technological unemployment, yes. And he predicted it a century out. So he wrote it in 1930, and he predicted it by 2030. But the interesting thing for me was like, he sort of said like, humans are built for scarcity.
[0:39:14] We're not great at leisure. And he looked at the indolent British aristocracy, and he's like, that's our future. Like, we're kind of screwed. We've got to do better, yeah. So what he prescribed, and he actually articulated it like a prescription. He's like, 15 hours of work per week just for the spiritual nourishment of it. If we don't deal with it now before it actually materializes, it's going to be a crisis, right? But if we start thinking about it now, I think we can approach it with some degree of a prepared mind. Well, let me flip from the philosophical to the practical.
[0:39:43] If your vision comes true, and it's a vision I subscribe to, you know, where we're going to have this sort of like renaissance of like biomedical and scientific advances, and you've even sort of postulated we could double human lifespan and maybe in the next couple decades add 10 to 20 years to human lifespan. Here you've got the most exponential technology in history, almost science fiction possibilities that we're considering.
[0:40:05] Slamming headlong into the most institutionalized, regulatory-intensive, incrementalist sector in the economy, which is U.S. healthcare. And I say this lovingly, but our institutions in U.S. healthcare are built for permanence.
[0:40:22] They're not flexible. And you talked about the clinical trials. If you have this sort of Cambridge explosion of invention that slams headlong into this hyper-bureaucratized process. And so my question to you is this. You're interacting with 150, right? We've been bringing them and you've been having these dialogues. What do you tell them?
[0:40:43] To prepare for what you're describing, what can they do today? My partner, Ian Sachs, would chastise me if I didn't try to land something tactical. What is your counsel to them? I mean, you described the dilemma well. It basically is kind of like an unstoppable object hitting a movable post. And we can't predict quite what's going to happen. So I would divide it up into a few effects.
[0:41:08] One effect actually, and it will happen on its own, but it's good for the 150, as you described, to kind of be aware of it because it can maximize it, is the optimistic part of this is that I think...
[0:41:20] Even though the institutions are kind of slow moving and I think built to be reliable and not make mistakes rather than to maximize innovation, which can make some sense when people's lives are at stake, right? But some of this will get better on its own just from the technology, right? So if we look at the difficulty of clinical trials, the ambiguity, all the documentation that we have to do, all the statistical analysis, all the...
[0:41:50] We ran this for a year and then we stopped it for futility and there were 10 arms to the trial. I mean, I'm caricaturing, but when you get something that actually works really well, you don't need a lot of statistics to show that it works really well. You don't need a large trial, right? When something really works, I mean, we have some institutions around this like accelerated approval, but can we push that?
[0:42:14] Can we push that even further where, you know, when the effects are really clear, you know, even the bureaucratic institutions, right? People love to make fun of the FDA or whatever. But, you know, at least from my vantage point, looking at it, when something works really, really well and you don't need a lot of statistical analysis, they do kind of get it through relatively fast. Maybe not as fast as they could, but they get it through relatively fast. So can we go to a world where we predict things better, larger fraction of trials work?
[0:42:42] Effect sizes are big, so trials can be relatively small. All of this will come from better technology and better ideas. And so just knowing that that's coming and kind of preparing for a world where that is the case so that we can accelerate it even more and we don't get slowed down by our kind of old ways of thinking. So that's the optimistic part of it, where I think some of it will get better on its own just by the force of the technology.
[0:43:10] But I think there are other areas where we will be tragically held up. And so, you know, I think the 150, you know, here's where I'd like to encourage some friendly competition and rivalry to say you can move faster than the other guys. And then if all of you think that way, all of you will move fast. Here I would say the silos of data, the protocols that don't make sense.
[0:43:32] When something works very fast from a Hippocratic attitude, not wanting to ever cause harm, the balance of that is about to change. And so, do we really want to hold back something that could...
[0:43:48] Give you 90% cure rates because of some little obscure thing where there's a tiny chance that this kind of privacy could be compromised or we're interpreting this law in this very arcane way that was designed for an era of technology 40 years ago. Is that really the right trade-off we want to make? Is that really the right risk benefit? Or not even there's a real, can someone just...
[0:44:17] Someone's worried that they could be sued for it. But even from a self-interested perspective, does the benefit exceed the cost? Can you push to take a little more risk because all of this innovation is coming? So I think there's a lot of stuff around data, around operations. We're not going to be perfect, but boy, is the benefit going to exceed the cost. Yeah, I love that. And I think it's a real call to action. I mean, you know, because the...
[0:44:46] Observation is I think we hold technology to an inhuman, hypocritical standard. We hold it to the standard of perfection, not human equivalency, let alone human superiority. And there's an exogenous-like factor here, too, that you've written about, which is this sort of Sino-US dynamic, right? And it's a philosophical point, but also a geopolitical point. I worry that in the installation phase of a technology, and this happens to be the most consequential technology we've ever seen,
[0:45:15] Autocracy and monarchy do much better at the diffusion than we Western liberal messy democracies do, right? And the geopolitical implications of this, which you've written about and you've advocated for export controls, which I'm also an advocate of, I worry that I look at the biotech companies that are doing synthetic and computational biology, where are they doing their clinical trials? They're not doing them in the United States, they're doing them in New Zealand, Australia, and China.
[0:45:41] Because they run things really, really sketchy way in China. But, you know, sometimes they get things done faster that way. We need to find a happy medium where we can deregulate, where we can embrace the asymmetric upside in the tech. And one of the reasons I'm so grateful to you is that you're helping articulate what could go right and really dimensionalizing, like, what are the questions we need to ask? And what are the 150 need to do? I want to ask you a little bit about the evolution of Anthropic, right? And you guys are, you know,
[0:46:10] You guys are architecting one of the most proficient frontier models. And in certain domains like coding and autonomy, it's the most proficient model, right? And it's really permeated on the enterprise side. But what we're finding is that, you know, with all of the sort of superhuman capabilities of the model, the installation is really hard, right? And, you know...
[0:46:36] We almost need a translator and this sort of notion of a forward deployed engineer. And, you know, and Alex Karp and I did this podcast last week and he sort of originated the phrase. And, you know, I tested on him a theory I've been working on, which is, you know, I think we've actually internalized all the wrong lessons from the industrial revolutions, like mechanization, electrification, computerization. This is going to be the most consequential one, agentification.
[0:47:00] But we sort of mythologized the inventors of the technology, you know, Watt and Edison and Gates. But we sort of neglected the most instrumental facilitator, which is the engineer. So the unsung hero of mechanization was the mechanical engineer. The unsung hero of electrification was the electrical engineer. The unsung hero of computerization was the software engineer. And the unsung hero of agentification is going to be the forward deployed engineer.
[0:47:24] And I understand that Anthropic has been building up its sort of musculature in that domain. And so my question for you is, as you continue to serve healthcare, how are you envisioning sort of the productization around that? Is it a group of domain-specific forward deployed engineers? Because if I'm a CEO, I'm saying, Dario, I get it. I'm captivated by the vision.
[0:47:46] And I get your recommendations on what we do, but I need help. Am I characterizing this correctly? Absolutely. So I just couldn't agree more with this conception. So if we go back to the speed of this technology is the multiplication of the advance of the underlying technology and the diffusion of the technology. So if we put this together, in some ways, it's the first that's holding back the second, which is that.
[0:48:11] Every few months, we train a new model that's better. And because these models are things that we grow more than we build, they're already different from other technology. And we're already in the position where when we release a new model, we ourselves don't know all of its quirks. If I raise a child and someone went to me and said, can you predict what they're going to do in every situation? You raised them, right?
[0:48:35] It doesn't work like that, right? And it's just the same with the models. And then you layer on top of that, you have new models every few months. We are in a position where we as the creators of the technology don't fully understand it and all the best ways to use it. We're trying to figure out as fast as we can. And we have information across lots of different customers and lots of industries, but we're still struggling to keep up. And so we have a lot of empathy for our customers who, if we can't fully figure it out, if we can't fully keep up, how can we reasonably expect them to?
[0:49:05] We need to use all the effort we can to tell them about
[0:49:09] And so, you know, whether we call them forward deployed engineers or something else, I think the way we think about it is, you know, we have our kind of account executives who talk to folks and, you know, help them understand the technology, you know, and then we have the kind of applied AI and solutions engineers. And we're going, I think, especially heavy on the second, which is just the practical implementation installation, as you call it.
[0:49:35] Oh, I think we're going to have a lot of these. I think we're going to have thousands of these. It's probably going to be the most labor-intensive part of the company. And I completely agree with you that these folks are unsung heroes, along with the product engineers who can kind of...
[0:49:51] I don't know what the word is, like Appify things, right? Because things are easier to use when there's kind of an easy interface for them. And even with that, you need forward deployed engineers and folks to tell you how to use them. So I think those two working together will help to diffuse the technology. And it'll be a little bit different in every industry, right? Where if I think of like, what does an intelligence agency need?
[0:50:18] Versus, you know, what does a hospital need? Those things couldn't be more different than each other, right? If I try and make things designed for one, for the other, you know, I'm just going to make... Even though the underlying AI is the same and maybe even the underlying apps are the same, the needs are so different that if I'm kind of calibrated for one, I'm going to be comically miscalibrated for the other. So we are increasingly going industry by industry, right? We've had...
[0:50:44] Claude for financial services. We've had Claude for life sciences. We have CyberClaude. All of these are slightly different models that hook into slightly different data, people with different expertise to help use them. And, you know, because we're still a relatively small company and we're growing fast, we're just at the beginning of building these teams. But they're growing very fast and we're getting expertise.
[0:51:14] Fast pace, right? So, you know, I've always known the science side of biology, but the health side of it and even the clinical trial side of it, I never worked on that when I was a biologist. So I'm learning about that for the first time. And I'm certainly learning about manufacturing for the first time. I'm certainly learning about retail for the first time. I'm certainly, you know, I'm certainly learning about intelligence collection for the first time. All of these are kind of new areas. And I myself, along with the company, am on a learning curve.
[0:51:43] Absolutely determined to serve each of these industries in its own way. I love that. And I'll say in closing, kind of where I started, first off, gratitude to you. You know, I think we have an opportunity here that is unprecedented specifically in healthcare. Because I almost think like U.S. healthcare, the establishment, has, for good and bad reasons, almost abdicated its co-creative responsibility, right?
[0:52:08] You know, we've had this almost Silicon Valley techno solutionism that never works in healthcare. It sort of bounces off the battleship. And so, you know, with your instrumentality and advocacy and just teaching, how do we bring the 150 into a dialogue where we can co-evolve and maybe have some co-creative co-design?
[0:52:32] That'll accrue to the benefit of patients. It'll democratize access and will be a force for good. And you're a force for good. Very grateful to you, my friend. Thank you for this. Thank you so much for having me.
[0:52:57] or as an offer to sell or solicitation of an offer to buy any securities or related financial instruments. The statements and opinions expressed are those of the hosts and guests and not necessarily those of Tower Brook Capital Partners, LP. Tower Brook makes no representations or warranties regarding the accuracy, completeness, or applicability of the content. Tower Brook expressly disclaims any and all liability or responsibility for any direct, indirect, incidental, special, consequential,
[0:53:25] Arising out of any individual's use of, reference to, reliance on, or inability to use this presentation or the information presented therein.