AI Safer Than Human Doctors This Year | Superpower CEO Max Marchione

AI is likely going to far outperform doctors in the next year. In the next five years, it won't be even comparable.

Health is not dying. Health is enhancing ourselves.

I don't think doctors are good at reasoning. If 'clinically relevant' means I have not thought for myself and I listen to a textbook created by pharma...

Peptides are going to be one of the most impactful things for human health of the past 50 years. I think probably 50-plus% of the population ten years from now will be taking some sort of daily enhancement therapeutic. The modality of that therapeutic...

Michael, excited to be here and looking forward to chatting.

I think there are very few themes that matter for humanity, and very few themes that have always mattered and will continue to. I think health is one of them. Health is one of these things that pre-AI matters, post-AI matters, is very Darwinistically primitive. Health is not dying. Health is enhancing ourselves. And we have this very timeless theme intersecting with several technological breakthroughs, intersecting with some cultural moments. On the technology side, the cost of diagnostics are approaching zero. OMIC sequencing, genomic sequencing has dropped faster than Moore's law since 2000, when we first sequenced the human genome. Data is becoming ubiquitous. We have wearables and all sorts of data sources. And then therapeutics are being used for things other than sickness, which massively expands the TAM of therapeutics and allows people to achieve things that previously were really hard without that therapeutic. GLP-1s and peptides being kind of the harbinger or hero products in that theme, but really them just being an indicator of where we're going. And then we combine this with what's happening in culture, where health is in vogue, where the whole looks-maxing theme is in vogue, where health is social, where we have like the decline of alcohol and the increase of gym attendance and yoga, and we get the fervor we have today, which is health. Health is having a moment that is both cyclical and, I think more importantly, structural and here to stay.

I think it's reflexive. So I think there's like a structural theme which is up and to the right, and that is because of technological breakthroughs and this fundamental need for humans to continue living and improve the quality of our lives. And then I think there are like cyclical ups and downs, and the cyclical ups and downs are where culture intersects with technology. Peptides was on the down two years ago, then six months ago it was on the up. Right now it's probably on the down or neutral, and then it will be on the up again in three months. So it's kind of like structural, up and to the right, and then cyclical has this up and down along that curve, and that's culture.

Well, if I look at healthcare, a lot of it is quality and cost. If I look at the constraint on the cost side, there are two primary drivers of cost in the US, to simplify. One is doctors, right? The human cost, physician cost, is a lot, and it's very hard to solve that. It's supply constrained, so it's hard to solve that by just hiring and training more doctors. One could argue that AI does solve that, and AI has a marginal cost closer to zero rather than $300 an hour. And then on the other side of the cost challenge, we have all the intermediaries that exist in the US, all the perverse incentives, the insurers who make more money the higher their cost structure, the hospitals incentivized to perform surgeries, the PBMs and pharma companies looking to mark things up as much as possible. But if we actually look at the physics cost—like the physics cost of a drug isn't much, the physics cost of a diagnostic isn't much. So I think on the cost side, there's also the opportunity to bring the actual cost structure of healthcare closer to the physics cost by removing intermediaries. So there are still cost problems in US health, and then there are quality problems, and one might argue that a quality problem is again constrained by the use of the human doctor to perform compute, right? Like health is maybe an informatics problem. We have the database of the person and of medicine, and we need to compute between the two, and doctors are not fantastic at that. AI is likely going to far outperform doctors in the next year at that, arguably where we are now. In the next five years, it won't be even comparable. And then on the quality side, we need more biotech, right? Like a computer knowing what is going on is one thing. Inference is one thing. But actually having the tools at our disposal to modify biology or modify human life is another thing. And we're starting to get there. But I think bio is still early, and it's going to be a really important trend of the next decade.

I think there are some external forces which make it easier to do now, which is LLMs are a thing. So they make it easier to replace the human. Health is cultural, which essentially lowers customer acquisition and earns you the right to get to the next milestone. There are all sorts of APIs that now make it easier to start a health company. The cost of diagnostics has come down to zero. And then part of it is just, can you execute, right? I don't think anyone has actually executed well per se. If we executed the same as we are executing today five years ago, we would still likely be in a good position—not as good of a position, because of the external forces—but I think the primary variable is still execution and team. And then the secondary variable is kind of market and timing.

I think ironically, people think we're good at marketing. I don't think we're that good at marketing. I actually think it's probably what we're weakest at. The product today is incredible. There are other companies who look like us in the space because their marketing is similar. There is no company that even gets remotely close to the product we have built. We have—I'm not exaggerating—hundreds of members, thousands of members who have said we changed their life, hundreds who have said we saved their life, and I get these stories every single week. And a big part of that is probably—there are probably two big parts of what has made that possible. One is the intelligence layer, the AI inference, the ability of the AI to connect the dots between all sorts of different data sources, the ability for it to structure that data and check in and hold someone's hand is just incredible. Like I have spent hundreds of thousands on concierge doctors, and I thought I knew most of what there was to know about my own biology, and the AI is discovering new things. I thought we'd get there eventually when we had more scale and more capital and more compute and bigger models and more data, but we're already kind of there, and it's hard to communicate that. It's more one of those things that you get it when you see it. And then I think the other reason we've executed well at the level of product is we brought a lot into one place in a very seamless way, which is non-trivial to do both of those things—aggregating a lot and then doing that seamlessly. Look, I think our execution—I think the average execution is like a three out of ten. I think typically best in world is like six or seven out of ten. And I love this framing because it means that if you execute the best in the world, you're six or seven out of ten. And the implication is there's like an opportunity to execute so much better, right? So do I think we're executing best in the category? Yes. Do I think we're executing exceptionally well? Certainly not. Far from it.

I think that most of execution is like culture and is like strategy, culture, and talent. Strategy is like, are we focused on the right things? And are we relentlessly cutting things, saying no to more and more and more to focus on the most important? I think we do an okay job of that today. I don't think we do a great job of it. The second thing is culture. I think we have a strong culture, a stronger culture than 95% of companies I've come across. I don't think we've got XAI-level culture yet, and I think we need to get there. So culture is the second thing. And then the third thing is talent, right? Like most of what we do is a function of who's on the team. What does the team create? And we're just constantly upgrading our talent. I spend most of my time recruiting, chatting with potential candidates, convincing people to join, headhunting, seeking the best people in the world. And we've done a really good job of that, but again, like we can just always do better.

No, I think there are probably five kind of muscles that we need to simultaneously be good at. Number one is infrastructure: owning pharmacies, labs, manufacturing, which we're doing more of over time—only a small amount today. Number two is kind of AI research. Does our AI outperform any other AI on very stringent health evals? Right? Is it detecting things? Is it connecting dots in really accurate and truth-seeking ways? Number three is, do we do a good job of the consumer interface? And consumer interface is where we talk about the UI, the experience. What does that look like? What does it feel like? I think there are three consumer interfaces of the future of health. There's digital. We do a good job of that today. There's at-home. We do a mediocre job of that today. And then there's physical—there are clinics or centers for diagnostics and therapeutics you cannot get at home. And we don't do that yet, but we will in the future. So there's infrastructure, there's AI, there's consumer interface. Then there's diagnostic and therapeutic IP. We've made our first kind of meaningful acquisition of therapeutic IP, and we'll go live with that next year. We haven't acquired diagnostic IP yet, but we're quite deep exploring a couple of opportunities. And there's the final, the fifth thing, which is the consumer application. The consumer application is like what people know us for, right? The kind of hero consumer application we have today is blood test into ecosystem, and then the second consumer application we're launching late this year or early next is peptides. And they're kind of the five things: infra, AI, consumer interface, diagnostic and therapeutic IP, and then consumer application.

A diagnostic is a commodity, right? And we want to commoditize commodities. To commoditize a commodity means providing at a low cost in a seamless way. And then the slightly less commoditized things, the intelligence on top. We never just did blood testing. We probably should have, but we launched out of the gates reasonably feature complete. Like we had dozens of different diagnostics. We had prescribing capabilities. We had a medical care team. I think we're known for blood testing because it's how we compete at the level of customer acquisition. But like post-acquisition, it's always been just the tip of the iceberg.

I don't think anyone's been able to do what we've done, even remotely close. Being able to wrap a blood test and sell it as a DTC product experience—great. Go do that. Be a supplement company called a blood test, right? I haven't seen anyone build an AI that performs the way ours does. I haven't seen anyone aggregate all these things into one place. I haven't seen anyone have the kind of tooling to engage someone and put them on a health journey. I haven't seen anyone—Hims has made bets around vertical integration. I haven't seen anyone make bets around acquiring bio IP. So, yeah, I just haven't seen people doing what we're doing yet. I've seen a blood test sold as a CPG product, but that's most of what I've seen.

It's related to peptides is most of what I'll say.

Earlier today, I was catching up with someone who was like to me, 'Your AI saved my friend's life. They were noticing a bunch of symptoms and they spoke to it and it said, you should go to the ER. And they went to the ER and they were having a stroke.' And that was striking. The only reason I wear my Oura ring now is because I want our AI to have the data. And it messaged me the other night—we give it access to email and SMS—and it messaged me the other night and it said to me, 'We see when your respiratory rate is higher, you sleep worse at night. And when it's lower, you sleep better. This could be allergies. Do you want an allergy panel?' And I said yes, and I've got a nurse coming to my home next week to draw my blood to assess for allergies. So that was kind of cool, how it was able to structure that data over a long period of time and then observe the correlation autonomously. There are dozens of things like this every day. We have many, many, many member stories of people saying, 'I've been grappling with this thing for a long time. I've been feeling fatigued, and the AI connected this kind of SNP or thing in my genome to this in my microbiome to this in my blood, and I took XYZ action and now I feel alive again.' So we get those stories more often than I would expect. I thought we'd do that eventually, but I thought that would take us quite a lot longer than what it has taken us, which is partially a testament to our team, but also partially a testament to the foundational models, right? These large models, open source and closed source ones, are pretty impressive.

Yeah, we'll aggregate all of the different data we can on someone. So you can connect your wearables. You can pull in your past medical records. You can upload any past lab and it will automatically visualize. Then you can chat with the AI. You can ask it questions. It's pretty good at processing the data. You can ask it, 'What do you see in my sleep data? What should I look out for? How does my genome relate to my gut microbiome?' You can ask it questions, but then it will also proactively text you and email you. Text is in beta. So today for most people, it's email them. It'll send them push notifications. Then it'll put that in the app as well. And that's where we don't tell it what to do. We don't tell it when to message someone. It just makes up its own mind. And people love that because many of them learn new things about their health. And that often results in them taking actions that change their lives and their health.

Yeah. Yeah. 100%.

Yeah, we don't do like pre-training, right?

I didn't think any—like, look, there's three parts of an AI stack. There's compute, a low IRR, massively capital-intensive, a good place for sovereigns to park capital, a good place for SoftBank to park capital. Then there's kind of all the pre-training work, the large model work, and OpenAI does a pretty good job of that. Anthropic does a pretty good job of that. Qwen is doing well at the moment. Kimi is all the rage right now. They're all the kind of foundational model pre-training companies. And then there's the application layer. It's like, what makes a good application? And there's post-training, there's multi-agent orchestration, there's context and memory. There's the harness you build. There's the way you structure data and allow the AI to process data. There's like dozens and dozens and dozens of micro decisions that go into creating a breakthrough application. I wish it was one thing I could just be like, 'it's just fine tuning.' It's like, no, it's actually literally a hundred small decisions that add up to allow an application to be breakthrough in terms of how someone feels when using it, and then also its performance on evals.

I think in the short term most moats are execution moats, right? And like the execution moat is where I've been spending the past two years with like a genius team optimizing this thing. And like I think someone else could copy it. I think it's probably going to take them two years with a really smart team because it just takes reps to learn what to do. You need to iterate many, many times. In the long term, the moat is scale and data. Scale because of the ability to vertically integrate, and then data because of the ability to allow kind of our models to know things that others cannot.

I don't think Apple's core thing is application layer, right? And like Apple's not going to write your prescriptions and order your tests and put a doctor in the loop to validate. Like they're not going to do that. So I think the probability that Apple is the one who does this is very low. OpenAI might try to do that. So I think if I were to make a bet on a company to be the health OS, it's probably OpenAI. OpenAI, us, maybe Oura. No, no one else is an obvious contender to me right now. Oura are definitely trying. I've been using Oura for like four plus years, and like the data is super helpful in the right context, but also it's kind of overwhelming, and I think they've been definitely struggling with how do they recommend actions on top of the data. And they have even integrated an AI assistant recently of sorts, but it's not that useful. It's kind of like what Google is doing in their products, like strapping Gemini on top and trying to push it to use it, but there is not a very clear link, I think, between that and outcomes, at least for now.

Yes, I think the access to testing is a commodity. The access to a wearable is a commodity. The aggregation of all of that data is not a commodity—like no one has been able to aggregate that. And then the ability to have intelligence that understands that is not a commodity, at least today.

Yeah, so we do a lot of eval work, like where we're constantly evaluating: is the AI accurate? Is it giving someone responses we deem as good responses? Obviously, nothing is classed as medical advice. Like, we sit in a similar zone to ChatGPT or Claude, where it can provide information related to health based on questions you ask, but it can't tell you what a doctor would tell you, right? So we need to escalate to—if you want a prescription, we're not going to say you should take a GLP-1. We're going to say this could be something to explore. So I think most of it is just like eval work, right? Can we provably increase accuracy over time by performing evals?