0:00:07 Health care is like a fifth of the American economy in growing, but if this process continues it will eventually be half the economy and then the entire economy.
0:00:15 We have some of the best doctors in the world. We have the best technology in the world. We pay the most and we have the worst outcomes. How is that possible?
0:00:20 Ultimately, is it a policy regulation issue or is it a technology issue?
0:00:23 It may be the cure and cancer as a lot easier to pack in the suitcase.
0:00:28 I think what we’re really providing patients is agency and I think that’s what was kind of missing.
0:00:42 The health care industry has captured the attention of America and it’s no wonder the confusing and complex glosses makes up about 20% of GDP and growing, clocking in at an over $4 trillion industry.
0:00:53 People seem to agree on problems, but rarely the solutions. But as we kick off 2025, it’s hard not to wonder whether the latest platform shift may offer some answers.
0:01:08 I suspect that this is the dynamic that’s going to play out in health care. The more abstract, intellectual, knowledge-driven, data-driven the task, the easier it is to get the AI to do it, the more applied, the physical, messy, unpredictable sort of all the human elements are probably the hardest thing.
0:01:19 So who will finally crack this code? And what’s really driving up prices? Could competition flip that dynamic on its head? And what will finally catalyze change? Plus, is the technology even there yet?
0:01:24 What do you each perceive as still being too hard to apply AI to?
0:01:36 Joining us to discuss and answer all those questions today are A16Z General Partners Mark Andreessen, Vijay Pande and Julie Yu. Let’s get started.
0:01:50 As a reminder, the content here is for informational purposes only. Should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund.
0:01:56 Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast.
0:02:02 For more details, including a link to our investments, please see a16z.com/disposures.
0:02:17 Let’s actually start with our quick takes on whether we believe it will be a healthcare-native company that finally cracks the nut on how technology will transform our industry, especially AI,
0:02:25 or do we think it will be a foreign body, an outsider, a non-healthcare company that will actually end up making the biggest play in our space?
0:02:34 I think it’s going to be a startup. I think it’s going to be really hard for an incumbent outside of healthcare to come into healthcare. It’s going to be really hard for healthcare to move into AI.
0:02:42 So a startup is a natural thing, and especially how that company is going to be built is that they will have to be AI-native and healthcare-native from the beginning.
0:02:48 And that’s something that’s really hard for incumbents. It’s just a straight down the middle typical disruption story.
0:02:57 Yeah, so I think the counter argument is to start by saying I’m usually a VJ’s side on these topics, and obviously we’re a venture capital firm, and so our day job is to try to fund exactly the kind of company that VJ is talking about.
0:03:07 So the easy thing to do is to just agree. But just for the purpose of conversation, here would be the argument on the other side, which is new technology entering healthcare up until now has been tools, right?
0:03:15 You go into an existing organization, an existing healthcare company of whatever kind, and you have a new kind of tool, and you basically try to get the existing organization to adopt the tool.
0:03:21 And of course, sometimes that happens, and hospitals run on hospital management systems today, and they didn’t used to and so forth, but that’s a big lift.
0:03:24 And that’s the, I think, inertia of VJ that you’re talking about, right?
0:03:27 It’s the mapping of the new technology into an existing organization. There’s the bottleneck.
0:03:35 The winning AI products might not come to market in that way. They might not be tools. They might be more almost like the equivalent of hiring workers, full stack.
0:03:42 And so I guess the question would be if the products are being delivered in a way where they almost slide into the organization as almost like the equivalent of a new employee,
0:03:47 or maybe even something bigger than a single employee, maybe as a new business unit, right, as a new organization, as a totality.
0:03:52 Does that change the dynamics of adoption such that big companies have a chance to think about this differently?
0:04:00 Yeah, or even as standalone businesses, by the way, that are powered by AI, that from the outside look like a regular provider group or a medical group that can contract with insurance companies
0:04:05 the same way that any other traditional medical group can contract with them and therefore get paid for what they’re doing.
0:04:13 But their internal economics and efficiency and their ability to scale nationwide is 10x what any traditional provider group might be.
0:04:22 So I think I would go all the way to say there might be standalone companies that are wearing sheep’s clothing in terms of the way that they’re packaged and the way they’re distributed in the market,
0:04:26 but internally look completely different in terms of the DNA of how they operate.
0:04:34 Let’s take an extreme argument where open AI creates AGI. Let’s just say it’s all the way to AGI. You’ve got the genius human being.
0:04:41 What does that know about healthcare? And I think a lot of what we have to do is we have to give the AI the equivalent of that work experience.
0:04:46 Is that something that an existing AI company will have? That’s a whole bunch of data that’s really hard to get.
0:04:55 It’s possible that an existing healthcare company will have that, but I think if you have to be really careful how you train that and whether they’ll have the AI experience.
0:05:00 So I think even still, not surprisingly, my bet’s on the startup, but I can imagine that is the way it’s going to happen.
0:05:06 But you’re making the point also that the startup has to be both healthcare native and AI native, not one or the other.
0:05:16 The other interesting thing about the way that Mark articulated this is this notion that one of the reasons I would argue that technology adoption in healthcare has been so slow in the past
0:05:28 is that the allocation of revenue within healthcare enterprises that goes towards technology, or IT as they would call it, has been in order of magnitude smaller in healthcare than other peer industries.
0:05:36 And what’s super interesting right now about these sort of like labor units that are how AI is packaged is that I’ve actually heard organizations now talk about,
0:05:47 oh, can we actually just tap into our labor budget? We have literally like a thousand open racks for nurses that we cannot fill through the existing human labor pool that is out there today.
0:05:56 Why not just allocate that one FTE’s worth of budget towards hiring the equivalent of like a hundred nurses using these AI tools that can do subclinical things,
0:06:03 but still things that are immensely important and critical to the workflows of these organizations based on what you guys are both observing across broader problem spaces,
0:06:08 whether it be creativity, education, law, software development, all the things that you mentioned earlier.
0:06:15 What do you each perceive as still being too hard to apply AI to within the healthcare domain?
0:06:19 I think we’re going to see the specialties probably be the last thing.
0:06:23 I think maybe the first thing would be the subclinical nursing and primary care.
0:06:28 Then inching into clinical primary care, I think it would not be that crazy to imagine.
0:06:34 Essentially, AI could be a great router, and especially think about like what a primary care physician has to do.
0:06:38 They have to ingest all this data and make a diagnosis and then send to a specialist.
0:06:41 That’s increasingly literally becoming a data science problem.
0:06:43 So I think that’s where we’re next.
0:06:49 The hardest thing is going to be like you take the extreme case of let’s say brain surgery is the canonical hard thing to do.
0:06:55 That’s something that sounds like it could be really hard for AI, but actually as surgeons use more robotic devices,
0:07:00 as other specialists use more digital things, you can imagine a way for it to creep in,
0:07:02 but that’s probably the furthest out.
0:07:07 Mark, a twist on that question for you is as you think about any industry where there’s like a frontier,
0:07:11 some boundary of what technology can do and not do, especially when it comes to AI,
0:07:18 do you think it’s mainly a question of data or are there other factors that you think we need to consider when it comes to how this might work in healthcare?
0:07:22 Yeah, so there’s this famous phenomenon in AI research called Moravex Paradox,
0:07:27 and it’s named after this AI researcher, Hans Moravec, and Moravec wrote in 1988,
0:07:32 “It is comparatively easy to make computers exhibit adult-level performance on intelligence tests or playing checkers,
0:07:37 and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”
0:07:39 The modern version of that is very striking, which is sitting here today.
0:07:43 We have chat GPT and its analogs, and they will paint art for you.
0:07:45 They will compose music for you.
0:07:47 They will debate abstract philosophy for you.
0:07:51 They will explain quantum physics to you, but they can’t pack your suitcase.
0:07:54 They can’t clean your toilet. They can’t cook you lunch.
0:07:57 They can’t do anything that a normal person operating in the world can do.
0:08:00 There’s actually a real evolutionary theory underneath Moravex Paradox,
0:08:05 which is the parts of us that are physically embodied, the parts of us that basically involve functioning in the world,
0:08:08 and then the sort of sensory skills and motor skills required to navigate in the world,
0:08:10 and then the ability to process unusual situations.
0:08:14 It’s like the classic, you see the demos of the robots, humanoid robots making coffee,
0:08:17 and it’s like it works right up to the point, the coffee pod, you know, the pod gets jammed in the coffee maker,
0:08:19 and the robot doesn’t know what to do.
0:08:23 All of the sort of motor sensory skills, all that stuff took many millions of years to evolve,
0:08:27 billions of years to evolve, and then at the peak of human cognition, the cortex,
0:08:30 the sort of system two reasoning was a relatively recent evolution,
0:08:33 and maybe let’s even say it’s 500,000 years versus 4 billion years,
0:08:39 and this is the exact reverse of what is intuitively you would feel like packing the suitcase is a lot easier than curing cancer.
0:08:43 It may be the curing cancer is a lot easier than packing the suitcase, like for this reason, right?
0:08:46 I suspect that this is the dynamic that’s going to play out in healthcare.
0:08:50 It’s going to be the more abstract, intellectual, knowledge driven, data driven, the task,
0:08:56 the easier it is to get the AI to do it, the more applied, the physical, messy, unpredictable,
0:08:58 the sort of all the human elements are probably the hardest thing.
0:09:01 I think a tweak on that is that AI loves the digital world.
0:09:05 It does really well, and if AI can’t be in that world, it has nothing to learn from.
0:09:12 And so to the degree there’s telemedicine and things that are virtual, whether it be a phone call or zoom,
0:09:15 that’s something that actually is a world that AI could play in right now.
0:09:19 And I think it’s not a shock that we’re seeing that as a natural first application.
0:09:24 What is interesting is as medicine itself becomes more digitized through robotics,
0:09:28 in other words, driven by people, then AI could be in that world too.
0:09:34 So it might be that there’s a parallel path of building up the digital side of medicine that AI could be in.
0:09:39 But yeah, I mean, I think it can do quantum mechanics, but it can’t unpack my dishwasher.
0:09:43 It makes sense because it’s trained all those quantum gangs has never seen a dishwasher, right?
0:09:46 The Tesla self-driving car might be the reason for optimism here, right?
0:09:49 And Tesla learned this the hard way because they tried different approaches to AI,
0:09:52 and it turned out the thing that worked was basically put a neural network in the wild,
0:09:55 embodying the form of a car, and in fact embodying the form of a million cars,
0:09:58 and then let those guys drive around and collect the data and learn about reality.
0:10:01 The more you take the rules out of the system and the more you just expose it to the real world
0:10:04 and let it gather the data and let it build up the neural network, the better it gets at self-driving.
0:10:09 That’s another form of leapfrog opportunity in the sense that the primary reason why hospitals exist
0:10:14 are because the level of complexity of the equipment and the workflows
0:10:17 and all the operations that need to exist have to be centralized.
0:10:20 It’s nearly impossible today to unbundle the entirety of that
0:10:23 and make it available to people in their home environments or in their communities.
0:10:28 But you can imagine perhaps that’s another lever here where if we are able to atomize
0:10:31 the physical components of the hospital-based experience
0:10:34 that you could actually centralize all of care delivery
0:10:38 and make it available in different sites that are much cheaper with a lower cost structure.
0:10:41 Yeah, so there’s two related concepts that are very important.
0:10:43 So one is economic growth, as you said,
0:10:46 and you generally measure economic growth through things like GDP, right?
0:10:48 Or, you know, in a company, it’s amount of revenue.
0:10:50 But there’s this other very important concept, right, which is productivity growth.
0:10:55 And productivity growth is basically the economic question of basically, can you do more with less?
0:10:58 And so if you have an economy that has high productivity growth,
0:11:03 basically what you see over time is that economy is able to produce more output with lower inputs.
0:11:06 And by the way, the sort of story of the growth of Western civilization overall
0:11:10 and economic growth globally over the last 500 years has been a story primarily of productivity growth.
0:11:13 Basically, the Industrial Revolution, figuring out ways to apply machines
0:11:17 and then new management methods around the machines to be able to do more with less.
0:11:21 And therefore, the world we live in, which in the old days, 99% of people worked in agriculture
0:11:24 and yet people were still starving because there wasn’t enough food.
0:11:26 Whereas today, 3% of people work in agriculture
0:11:29 and we’re producing so much food that everybody’s getting obese, right?
0:11:31 And that was the result of productivity growth, right?
0:11:35 And so then you go into kind of the microeconomics of the healthcare industry
0:11:38 and basically what you see is either flat or negative productivity growth.
0:11:42 By the way, the industries that are like this are the healthcare industry, education, housing,
0:11:44 and then let’s say law and government.
0:11:47 Those four basically, if you measure them, they have flat or negative productivity growth.
0:11:51 And so basically they’re getting less efficient over time, not more efficient over time.
0:11:53 And you see that directly, as you said, you see it in the graphs
0:11:56 for all four of those industries where you’re spending more and more money over time
0:11:58 and you’re getting less and less output, less and less results,
0:12:01 which is precisely the opposite of what you want.
0:12:03 By the way, the economic growth that follows from that,
0:12:05 the fact that you still see revenue growth in these industries
0:12:09 is in part a consequence of another principle of economics called Bommel’s Cost Disease.
0:12:13 And Bommel’s Cost Disease has to do with the fact that in the healthcare industry
0:12:16 you are trying to hire people who have options to work in other industry sectors
0:12:18 that are high productivity growth.
0:12:22 And so those workers have alternatives that are able to be higher paid
0:12:25 working in more productive industries and so they charge you the same labor rate
0:12:27 even in the lower productivity industry.
0:12:29 And so you see this like massive cost inflation,
0:12:31 even though you’re not getting better results.
0:12:34 And so that’s been the state of affairs for a long time in the healthcare industry.
0:12:37 Like I said, there are these other major sectors of the economy that have that characteristic.
0:12:40 At the end of the day, there are really only two things that can happen.
0:12:42 One is that process can continue indefinitely.
0:12:46 And if that process continues indefinitely, then at the limit, you eat the economy.
0:12:48 And by the way, that’s what’s happening with healthcare sitting here today.
0:12:51 Healthcare is now like a fifth of the American economy in growing.
0:12:54 If this process continues, it will eventually be half the economy
0:12:56 and then the entire economy without necessarily better results.
0:12:58 And that may be what happens.
0:13:01 Or you can break those cost curves by introducing technology
0:13:03 and then using that technology to drive productivity growth,
0:13:05 which then lets you basically do more with less,
0:13:08 which then lets you get better results and spending less.
0:13:12 In addition to technology, I think the big question is how we’re using medicine
0:13:13 and how medicine is applied.
0:13:17 If you think about it, it sounds weird that we have some of the best doctors in the world.
0:13:19 We have the best technology in the world.
0:13:22 We pay the most and we have the worst outcomes.
0:13:23 How’s that possible?
0:13:27 I think part of it is to really recognize that there’s two aspects to healthcare broadly.
0:13:30 There’s acute care, which our system is amazing.
0:13:33 If you’re in a car accident or you have stage 4 cancer,
0:13:34 I think you want to be in America.
0:13:40 But for chronic care, that’s something that I think we, for the most part, treat as if it were acute.
0:13:43 We wait for things to get bad and then we get into it.
0:13:48 So imagine a world where your house doesn’t have smoke detectors, doesn’t have extinguishers,
0:13:50 doesn’t even have circuit breakers.
0:13:52 So I said, any little thing will cause a problem.
0:13:55 And then you wait for the house to be on fire and then you call the fire department.
0:13:59 I guess what, it’s going to be more expensive and you’re going to have worse outcomes.
0:14:03 Ultimately, is it a policy regulation issue or is it a technology issue?
0:14:11 Because absent changes in our payment regulation and payment policy about how we align incentives for paying for these therapies and whatnot.
0:14:12 Is there any hope?
0:14:16 What is our source of optimism given the fact that all the industries that you mentioned are so highly regulated?
0:14:20 That’s what’s really in the way of us making big changes.
0:14:23 Yeah, so it is that they’re regulated, but it’s not just that they’re regulated.
0:14:24 And here’s what I mean by that.
0:14:26 Healthcare, education, housing and government.
0:14:31 They’re regulated in a very specific way, which is they have constrained supply and then subsidized demand.
0:14:33 Let’s take housing as an obvious example.
0:14:36 There’s a massive shortage of housing in particular where there’s lots of economic opportunity.
0:14:41 This is an example, analysts generally think that the California Bay Area is missing like two million houses.
0:14:43 This is a giant gap.
0:14:47 In other words, if two million people could move to Northern California, they could all become economically more productive.
0:14:48 The whole region would benefit and grow.
0:14:49 The country would benefit and grow.
0:14:55 But all of the constraints around housing development and zoning laws and environmental regulations and everything prevent those houses from getting built
0:14:56 and so they don’t get built.
0:14:57 And so you have constrained supply.
0:15:02 What happens in a market where you have constrained supply and increasing demand is of course prices rise.
0:15:03 But it’s not just that.
0:15:05 You also have subsidized demand, right?
0:15:10 Because what happens is you’re in a market with constrained supply, prices are rising, and then there’s political outrage.
0:15:13 Because in this, take housing, people can’t afford it by houses.
0:15:18 And so politicians respond to that, not by deregulating on the supply side, but by subsidizing the demand side.
0:15:22 In theory, that helps people buying a house in the moment because they have more money to spend on a house.
0:15:26 But because you have constrained supply of housing, what that will just do is drive up prices more.
0:15:34 And so demand subsidies have the paradoxical result of driving up prices, which then makes those demand subsidies inadequate, which then causes politicians to subsidize demand more.
0:15:36 Same thing has happened with universities.
0:15:40 Tuition and colleges universities has been rising far faster than inflation for now for several decades.
0:15:43 The price of a four-year private college degree is up to $400,000.
0:15:45 It’s on its way to a million dollars.
0:15:48 Why is that is because you have constrained supply.
0:15:53 There’s only so many universities to get access to federal student loans, which you need to do because they’re so expensive.
0:15:56 If you want to start a new university, you need to get accredited.
0:16:01 Who accredits new universities are the existing universities, which of course they run the accreditation bodies.
0:16:04 And they have no intention of causing there to be more universities, which would be more competition.
0:16:05 So that doesn’t happen.
0:16:08 And so you just have a fixed number of seats in good colleges universities.
0:16:10 And then you have the federal student loan program.
0:16:15 And as the constrained supply drives up prices, you have more and more subsidies in the form of federal student loan subsidies.
0:16:17 And then that drives up prices.
0:16:19 It turns out health care works exactly the same way.
0:16:22 There’s just a certain number of hospitals you have to get accredited as a hospital.
0:16:24 There’s only a certain number of doctors.
0:16:28 And by the way, legitimately so you have to go to medical school and get licenses to doctor.
0:16:29 There’s only so many nurses.
0:16:34 There’s always this kind of tension around the nurse practitioners because there are arguments that maybe you should liberalize that a bit.
0:16:37 And I think those arguments are actually pretty good, but there’s only a certain number of nurses.
0:16:44 And so you just have fixed supply and then you have increased population growth and you have increased percentage of people in the economy who are old and sick.
0:16:47 And then you have subsidies and you have all of the programs.
0:16:53 And look, I wouldn’t want to live in the country without some set of these programs, but you’ve got all the Medicare, Medicaid, Obamacare, all these programs.
0:16:56 So again, restrict supply subsidized demand prices go to the moon.
0:17:00 And so left on check economically inevitably what’s going to happen.
0:17:03 The political system seems like completely unable to deal with this.
0:17:07 And as a consequence, everything in these four fields will get to be infinitely expensive.
0:17:11 This where I come out as a radical on this is the only way to break those cost curves is through technology.
0:17:16 Like the only way that you can get in front of this phenomenon, I think political system is completely unable to deal with it.
0:17:20 It’s an answer that has to come from the private sector and it has to come from the introduction of disruptive technologies.
0:17:25 You think about the cost of compute, the cost of TVs, cost of anything electronic, the cost of tech.
0:17:30 That feels like it’s the reverse exponential that has to hit into this inflation.
0:17:32 And it’s one exponential against another.
0:17:36 I mean, I’m curious, Mark, how do you know which exponential is going to win?
0:17:39 You know, it’s human choice, right?
0:17:41 It’s choice. It’s the things that we choose to do.
0:17:43 We know for a fact we have technologies as a tool.
0:17:46 By the way, technology has done what I’m describing in many other industries.
0:17:49 What television sets are like a case study of the opposite direction.
0:17:52 The price of television sets has crashed as the quality has exploded.
0:17:57 The line that I use is you’re going to have a flat panel TV and 32K resolution that’s going to cover your whole wall.
0:18:01 It’s going to cost a hundred bucks, but yet it’s going to cost you a million dollars to send your kid to college, right?
0:18:06 That’s a result of tremendous technology innovation and productivity growth in consumer electronics,
0:18:10 which is, by the way, consumer electronics unregulated, no restrictions on supply.
0:18:14 Anybody can build a TV factory, at least in theory, and then no subsidized demand.
0:18:17 You don’t get money from the government or anybody else to help you buy a TV.
0:18:20 And so as a consequence, the economics drive to low prices and high quality.
0:18:26 It’s a societal level choice and then a company by company and individual by individual choice as to whether we want that to happen in healthcare.
0:18:30 Along the lines of what we’re talking about here with these bastardized supply and demand curves,
0:18:37 one of the big trends that we’re thinking a lot about and we’ve written about as well is this notion of consumers as the new class of payer,
0:18:44 where there are a number of movements in the healthcare space to empower individuals with not only the ability to shop for their own services,
0:18:47 but actually lift shift budget from these monolithic health plan products
0:18:53 that uniformly treat everyone the same to individual budgets that they can go use on their own accord to shop with agency.
0:18:59 Can you guys talk about that and is that a way for us to blow up at least a portion of kind of the way our insurance system works today?
0:19:07 It’s interesting if you think of the world in maybe a bit of a simplistic way, but as this chronic care and this acute care system.
0:19:11 The current insurance is really thinking about the acute side. It’s covered well.
0:19:19 But if you’re generally healthy, you could imagine having a high deductible plan to handle that and that might save some money for something else.
0:19:26 In a world where there’s choice and things like ICRA where consumers could put together a plan that they want,
0:19:31 you can imagine them choosing other types of providers that are thinking about the chronic care side
0:19:36 and that they could put together a plan that includes both an acute care side and a chronic care side.
0:19:42 And this way it’s still being paid by their employer, but that they’re not having to pay it out of pocket.
0:19:49 Or I think up to a certain point, even people will pay out of pocket, not for brain surgery, but for chronic care.
0:19:53 500 a year, a thousand dollars a year, I think is imaginable.
0:19:57 And the reason why is that in the end, the consumer is the ultimate payer, right?
0:20:00 Not just in dollars, but the ultimate price.
0:20:04 And so we are a long-term payer, incentive for care.
0:20:09 We will care about chronic in a way that insurance companies just can’t based on the nature of incentives.
0:20:13 Yeah, and I’d say some of these other industries that we’re talking about that have the same kind of cost curve problem.
0:20:15 There’s some, I would say, green shoots.
0:20:19 We’re looking for basically signs that there’s basically breakpoints happening or to VJ’s point.
0:20:22 There’s elements of consumer choice that are coming in that didn’t used to exist.
0:20:24 And so I’ll just give a couple of examples.
0:20:25 One is education.
0:20:29 So education in the U.S. historically has been a K through 12 monopoly, right?
0:20:30 So the government runs K through 12.
0:20:33 And there’s always been a small homeschooling contingent, but it’s not been a mainstream activity.
0:20:38 And generally in the past, it was often specialized religious communities that are separated out from the broader society.
0:20:43 And then higher ed has always been a cartel through the accreditation process of colleges, universities.
0:20:49 And really in the last basically five years, homeschooling is rising as a percentage of students and pretty significant numbers.
0:20:50 It’s far from a majority.
0:20:53 It’s still single-digit percentages, but like homeschooling is way up.
0:20:57 And homeschooling is way up in particular, not just among religious communities, but in general.
0:21:01 And in fact, among higher income levels and higher class levels, the numbers are particularly striking.
0:21:06 And then look, there’s an industry around it’s forming up and there are companies and startups and there are people running microschools
0:21:09 and there are companies doing matching and teacher training and recruitment and all kinds of things.
0:21:11 You know, activities to be able to enable this.
0:21:19 And so that’s an example where you’re starting to see consumer choice entering a market in which previously everybody just assumed you’re just doing what you’re told top down.
0:21:21 And then housing, housing is still all screwed up.
0:21:24 But you know, the interesting thing on housing is remote work, right?
0:21:29 Up until COVID, the rule of geography and work was very simple, which is you had to move to where the work was, right?
0:21:31 Because employers didn’t hire remote.
0:21:37 And in fact, historically, like in tech, for example, this was one of the main reasons for people to move to Silicon Valley and, you know, take this as an extreme case.
0:21:43 The line in Silicon Valley used to be if you live in Palo Alto, you can get 20 other job offers anytime you want without changing your parking spot.
0:21:46 It’s because there are just so many tech companies around that will hire you.
0:21:52 Well, if you’re a tech worker now and you are remote, not only can you get 20 other job offers, you can get 10,000 other job offers, right?
0:21:54 The asymmetry goes in the other direction.
0:22:06 Even if most employers are not hiring tech workers remote, there are so many who are that as a tech worker, evaluating job offers from employers that will hire remote, there are thousands or tens of thousands or hundreds of thousands of those.
0:22:12 And so, paradoxically, if you’re a remote worker, you now have more employment options than if you’re geographically co-located.
0:22:14 And again, most people in tech have not moved.
0:22:21 Most of them are still in places like Silicon Valley, but the ones who move are doing great and have job offers coming out of their ears and they’re doing fantastically well.
0:22:25 And this is relevant to housing because that means now that they can move to places in which housing is a lot cheaper.
0:22:28 They can move to places where there aren’t a lot of companies.
0:22:30 They can move to a place that actually might be a better place to live.
0:22:36 Maybe they can have a big house on a lake for less money than it would cost to get a one-bedroom apartment in Palo Alto.
0:22:41 So again, it’s this green shoot crack in the matrix thing where it’s okay, you can see how the world can work a different way.
0:22:50 And so, at least, VJ, I think basically that’s a lot of what’s going to happen in healthcare from here on out, which is basically the existing system left its own devices is going to degrade further and further.
0:22:53 It’s going to give worse and worse results at higher and higher cost over time.
0:22:56 As a consequence, some percentage of people are going to try to break out of that.
0:22:59 One of our companies, we have many examples of this, but Levels is a great example.
0:23:09 If you want to take control of your own behavioral health, your own obesity risk or diabetes risk, you can sign up to Levels and they have a whole program and software and everything that will help you do that.
0:23:12 And I don’t even know if there is an insurance reimbursement option, but it’s tech.
0:23:13 It’s not expensive.
0:23:15 You happily pay out of pocket and off in a way you go.
0:23:17 And I tend to think that there’s going to be a lot more of that.
0:23:22 Yeah, it’s also the juxtaposition of the amount of money that you’re paying for something like a Levels experience or function health.
0:23:35 If you actually look at the amount of your wages that is being taken out of your salary to pay for health insurance and then certainly in a high deductible plan, which is increasingly more common in the employer-responsive world, your out-of-pocket spend actually tends to be even higher than that.
0:23:46 And many of these companies, what they’re doing is actually doing direct contract deals where they can take 40 different services like what function is doing that individually in aggregate would have cost you tens of thousands of dollars
0:23:52 under an insurance product, but they’re negotiating a direct contract as a bundle that says for hundreds of dollars, you can basically get the same thing.
0:23:59 And it’s just totally dislocating the price curve for what otherwise beneath the traditional insurance system is just completely unaffordable.
0:24:01 So I think we’re seeing a lot of that.
0:24:04 I think what we’re really providing patients is agency.
0:24:06 And I think that’s what was missing.
0:24:09 Like why can’t you just lose weight or eat better?
0:24:18 I think there needs some help there. And I think this sense of agency is starting to really grow where people have a ability to monitor the health and then do something about it and then repeat.
0:24:21 And that cycle actually I think works for a lot of people.
0:24:25 And being in control of your health feels very different than being at the mercy of a system.
0:24:29 Let’s zoom out a little bit to some of the bigger picture ideas here.
0:24:31 The whole notion of health tech in general, right?
0:24:43 This is a fairly new space. I think much of what has been possible with health tech only came because of the pandemic where there was a relaxation of things like virtual care laws and certifications that were required to practice medicine across state lines.
0:24:51 A lot of this was entrenched in the sense that you were not even allowed to treat patients across state lines prior to some of those emergency laws being put into place.
0:24:54 And so this space is very new and it’s also speculative.
0:25:02 Like a lot of people still have a lot of skepticism about whether this is a durable trend, whether this could actually result in a completely new and big industry over time.
0:25:09 But obviously us and especially you, Mark, having seen so many waves of industry transition and transformation happen in the course of time.
0:25:12 We know that these things sometimes take decades to actually work.
0:25:30 So Mark, specifically, what advice or parables can you share with folks who are watching the digital health space play out maybe from the early days of the internet even to inform how we interpret where we are in this industry from a maturity curve perspective and just providing again some perspective on how long sometimes these things take to play out?
0:25:33 We liked Lenin Mark’s installing quotes every now and then just to make an impression.
0:25:44 So there’s an old Lenin quote, V.I. Lenin, Vladimir Lenin, not John Lenin, to be clear. And he was talking about political change, but he said there are decades in which nothing happens and then there are weeks in which decades happen.
0:25:51 And then in tech, there’s a similar kind of thing called Amara’s Law that is basically changes in technology take a lot longer to happen than you think they will.
0:25:55 But when they happen, they have much more consequence than you think. They’re much more dramatic than you think.
0:26:03 We see this all the time where it can take years and years and years and years if it seems like absolutely nothing happened. And then there’s some sort of catalytic thing that happens and all of a sudden things take off.
0:26:14 And I think there’s a bunch of things that go into this. Part is sometimes it just takes a while to get the technology dialed in. I always like to point out the first smartphone hit the market in 1987 and you didn’t get the iPhone until 2007.
0:26:20 And so it was a full 20 years. And by the way, the first iPhone actually was 2G. It didn’t make phone calls properly. It had all kinds of issues.
0:26:28 You didn’t get the App Store until I think another four years after that. So it was 25 years from the inception of the smartphone industry to actually getting the modern iPhone on a real data network or the real App Store.
0:26:38 It’s just like all the componentry that went into the iPhone, it all had to get mature. It all had to develop like the screens and the batteries and the radios and everything had to get like really good before the whole thing packaged together.
0:26:45 So part of this is on the responsibility of the tech companies themselves, our companies to be able to actually get the right product to market and sometimes that takes time.
0:26:53 The other side of it is the catalytic effect. You know, in the old days to get something new adopted, you always had to do mass marketing, mass media campaigns and hopefully people responded.
0:26:59 A lot of it is now peer to peer. It’s people telling each other and we always reference a lot in our business. It’s okay.
0:27:05 Like when we’re looking at a given thing that might be a big deal as well, is there a subreddit for it? Is there a forum on Reddit where people are talking about it?
0:27:14 And if there isn’t, it almost certainly means that nobody has any of it exists. And then if there is a forum on Reddit where people are talking about it, you can get a sense of what the early adopters are saying and how close it might be to a vertical takeoff.
0:27:21 And then of course these days, the new version of that is like TikTok or Instagram reels or tweets. Is it on the big social networks? Is it taking off?
0:27:26 My guess more and more is it’s going to be peer to peer. When these things take off, it’s going to be because people learn about them online.
0:27:29 They learn about them from their friends. They learn about them from watching videos online.
0:27:34 This, by the way, is the natural frustration. This is the Dr. Google kind of effect extrapolated up, right?
0:27:43 Which is if doctors were frustrated by patients coming in with Google printouts, which I think has been the case for the last 20 years, I think doctors are going to get increasingly frustrated because patients are going to come in and they’re going to be like,
0:27:52 well, I saw this thing on TikTok and I think I should try it. And look, there are going to be downside versions of that, things that are not valid, but there are also going to be things that actually work that actually spread.
0:28:00 Well, I mean, look, just like even fitness, fitness now is almost entirely an online phenomenon. The way people learn how to work out now is almost entirely online.
0:28:06 By the way, also the way people learn how to cook, the way people learn how to eat healthy, like that’s almost entirely happening on social networks and on YouTube.
0:28:09 So I tend to think that phenomenon will happen for more and more areas of health.
0:28:19 And I think a corollary of that is that there will be grassroots movements, even political movements towards health and that this becomes a key part of our lives that people care about.
0:28:28 And so it’s not going to be a top down, wonky, inside Washington movement, but I think you’ll see grassroots political movements and we’re already seeing the beginnings of that as well.
0:28:29 Yeah, that’s right.
0:28:35 By the way, there’s also Dr. TikTok. So physicians are sharing notes with each other also about how to subvert the systems that shackle them.
0:28:38 So I think it’s happening on both the demand side and the supply side of our industry.
0:28:41 Let’s end with a fun one. So I think all three of us have young kids.
0:28:48 I certainly have a son who’s growing up very A.I. native and I’ve observed how he uses tools like even ones that we’ve invested in like character and curio.
0:28:50 I think Mark, your son has one of these two.
0:28:59 And he’s like really developed this comfort around just talking openly about his feelings and his curiosities and his dreams that I think will actually be very healthy for him in his mental health in the future.
0:29:15 So similarly, are there things that you guys observe your kids doing or the younger generations doing in general with A.I. that you think specifically will inform what health care behaviors will emerge over time that could eventually become mainstream and crack some of the nuts that we’ve been talking about today?
0:29:19 Douglas Adams, the great science fiction author who wrote Hitchhiker’s Guide to the Galaxy, had this famous.
0:29:24 He said there’s a three part generational kind of model of technology change, health society that’s to do technology.
0:29:25 He said it’s true for every new technology.
0:29:28 So if you’re under the age of 15, it’s just natural that this is how the world works.
0:29:34 If you’re between the ages of 15 to 35 when a new technology comes out, it’s like super cool and you might be able to get a job working on it.
0:29:39 And if you’re over the age of 35, it’s unholy and against the natural order of things and we’ll destroy all of society.
0:29:41 We, of course, are the exceptions.
0:29:43 We do not fall prey to these ancient patterns.
0:29:50 My version of this was a nine-year-old that when ChatGPT first got going probably what, two years ago, less than two years ago, when it really hit critical mass.
0:29:53 Yeah, so probably in the spring of 23, so he’s like probably eight.
0:29:57 I set up his little laptop for his classwork and so I set up ChatGPT on it and got him his own account.
0:29:59 I was like so proud of myself as a father.
0:30:03 I felt like I had brought down fire from the mountain for my son, right?
0:30:04 And I was going to give him like, “Here’s ChatGPT.
0:30:05 It’ll answer every question you have.
0:30:06 It’ll teach you anything.
0:30:08 It’s going to be with you for your entire life.
0:30:11 Like this is the thing that’s going to make you like a much more advanced version of you.”
0:30:15 And it’s just like the most powerful, amazing thing I can give you as a father.
0:30:19 And I sit him down and I teach him and I’m like, “Look, you type in any question it answers the question.”
0:30:21 And he shrugs and I’m like, “What’s the shrug?”
0:30:22 And he’s like, “It’s a computer.”
0:30:25 Obviously it answers questions like, “What else would you use a computer for?”
0:30:27 I’m like basically that’s it, right?
0:30:33 And so that’s the other part of this is I think a big part of all of everything we’ve been discussing also just as like a generational change.
0:30:36 It’s new cohorts of people coming up and just people are not going to tolerate.
0:30:38 Actually, I think we saw this in the meta.
0:30:42 You tell me, I think we saw this in the health field probably 20 years ago where before the internet,
0:30:44 how would you ever look up whether a doctor was good?
0:30:48 And then once the internet emerges, like all of a sudden is there’s a different size at like rate doctors.
0:30:51 And then by the way, how would you ever as a patient with some condition,
0:30:53 how would you ever read up on it yourself?
0:30:56 Because you can’t obviously it’s not going to you’re going to go read medical journals at the university library,
0:30:58 but then you go on Google.
0:31:04 And so I think we’ve already seen probably at this point several waves of that kind of adaptation as people with the same disease, everything like that.
0:31:05 Yeah, exactly.
0:31:07 They go from being something that’s inconceivable to something that’s common.
0:31:10 And a lot of that is on a cohort age basis.
0:31:14 You get these debates like in the press about is a new technology good or bad and should it be adopted or not?
0:31:18 Those are all beside the point because the fact is like young people are just going to use what’s helpful and useful,
0:31:20 and they’re not going to have the emotional reaction.
0:31:22 And I think that that will apply to many of the things that we’ve been talking about.
0:31:26 If you’re like an executive in their fifties and you’re exhausted,
0:31:30 you don’t have time, you don’t have time to play with this thing and to learn the skills.
0:31:33 The kids will have the playful attitude, will play around with it,
0:31:37 will learn and learn things that maybe even other people didn’t think they could do.
0:31:42 I think they’re at a huge advantage to come up to speed quickly because I think you have to learn how these tools work
0:31:45 and how to train them and how to get them to do what you want.
0:31:47 And I think people are still figuring that out.
0:31:50 But our children will definitely figure out faster than we will.
0:31:51 I think they already have.
0:31:54 Well, with that, thank you both for an amazing conversation about the future of health.
0:31:55 Thank you guys.
0:31:56 Thank you.
0:32:00 All right, that is all for today.
0:32:03 If you did make it this far, first of all, thank you.
0:32:07 We put a lot of thought into each of these episodes, whether it’s guests, the calendar touchers,
0:32:11 the cycles with our amazing editor Tommy until the music is just right.
0:32:17 So if you’d like what we’ve put together, consider dropping us a line at ratethespodcast.com/a16c.
0:32:20 And let us know what your favorite episode is.
0:32:23 It’ll make my day, and I’m sure Tommy’s too.
0:32:25 We’ll catch you on the flip side.
0:32:28 [MUSIC PLAYING]
0:32:31 [MUSIC PLAYING]
0:32:33 (upbeat music)

0:00:06 In a world of abundant content, how do I allocate my scarce attention?
0:00:13 We’re so used to the internet being this place where you can publish anything and anything can happen.
0:00:16 I think you could make the argument that actually it’s a fluke.
0:00:21 The same way that you might write a piece of software and change how a million people do their jobs,
0:00:25 you could write a great essay and change who a million people are.
0:00:31 We’re now in an era where you’re not going to be able to build a billion active user ad-supported thing.
0:00:36 If what I read is what I become, who do I trust for that?
0:00:41 If you’re a listener here, you’ve likely heard of the company Substack.
0:00:45 And you almost certainly have read something on Substack.
0:00:49 But today, you’ll get to hear about the very origins of this company.
0:01:00 Why in 2017, did Chris Best and his co-founders choose, of all things, to target a technology originally developed in 1971, over 50 years ago?
0:01:03 That is, of course, email.
0:01:10 Sitting down today with Chris is A16Z General Partner Andrew Chen, as they traverse through the future of media and curation,
0:01:16 but also what changes with the latest technological wave, artificial intelligence.
0:01:21 Even before AI, we had more content than we could ever want in the whole universe.
0:01:26 Chris, by the way, was also a co-founder of Kick for nearly a decade,
0:01:32 a messaging app that he helped scale to hundreds of millions of users, so there are plenty of war stories throughout.
0:01:37 By the way, this episode was originally recorded during our last Tech Week in San Francisco.
0:01:42 So as you plan for the new year, new dates for Tech Week have already been announced.
0:01:48 We’re going back to New York in June and back to San Francisco and LA in October.
0:01:53 For more information and to stay posted on hundreds, if not thousands of events to come,
0:01:59 plus any other cities where Tech Week might be going, head to tech-week.com.
0:02:04 That’s tech-week.com, or you can tap the link in our show notes.
0:02:13 As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice,
0:02:20 or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund.
0:02:26 Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast.
0:02:32 For more details, including a link to our investments, please see a16z.com/disclosures.
0:02:43 I wanted to say one really funny thing, which was how I originally met the Substack team,
0:02:47 which was I have been writing a blog now for like almost 15 years.
0:02:52 And when these guys were going through Y Combinator, Hamish actually reached out to me.
0:02:56 So this is Chris’s co-founder, and he said, “Can I meet up with you?”
0:02:58 I was like, “Great, let’s do it.”
0:03:05 I always loved meeting founders, and then he gave me a pitch to move my blog over to Substack.
0:03:10 And I was like, “No fucking way. Never gonna do it.”
0:03:12 You don’t know how much money you could have made if you had done that.
0:03:16 Yes. Well, that’s actually very much true. I actually very much regret this.
0:03:20 And so don’t be like me. Instead, be like all the other smart people.
0:03:25 And so what ended up happening was I said no, but then years later I started to see, of course,
0:03:28 Substack all over the place on X and so on.
0:03:32 And then I approached Chris and Hamish and Jay.
0:03:36 And I was like, “Hey, I have a new job now. I can give people money.
0:03:39 It’s a really cool superpower. Would you like some money?”
0:03:42 And Chris says, “No fucking way.”
0:03:48 And actually, I think it took maybe six months for us that he finally would allow us to…
0:03:51 It was time to build. What can I say?
0:03:53 Anyway, so that is the backstory of how we met.
0:03:57 Eventually we did. And then now my blog actually does run on Substack.
0:04:00 And it’s wonderful. And I wish I had migrated earlier.
0:04:06 And so, Chris, my first question to you is, you are CEO of Substack,
0:04:12 but actually you had this whole other life before this where you were an engineer.
0:04:20 People almost don’t know that you built a multi-hundred million user products before all of this, but as CTO.
0:04:22 So tell us about the whole journey.
0:04:25 Well, I went to University of Waterloo for college.
0:04:30 We did the co-op internship program where you have six real jobs when you go to school,
0:04:34 which is nice because then you can afford to go to school and you do real things.
0:04:37 I got kind of like a taste for building.
0:04:42 I was sort of bored at the fake learning thing and always really liked the really making something thing.
0:04:44 And I got into the kick thing by accident.
0:04:47 I was just hanging out with Ted Livingston, who was the CEO.
0:04:50 And we were just tinkering. We were just building some stuff.
0:04:52 We weren’t really sure exactly what we were making at the start.
0:04:56 We started out building a music player for the Blackberry and…
0:04:57 Very Canadian.
0:04:59 Very Canadian, very Canadian story.
0:05:03 And we ended up building a messenger that had an absolutely wild arc.
0:05:07 I learned a lot. It’s funny, I think, being sometimes the technical co-founder.
0:05:09 You experience things in a weird opposite land.
0:05:16 We went through a period where we were doubling our user base every four or five months for several years.
0:05:21 And the way that I experienced that at the time was like, “Oh, this hurts.”
0:05:23 And I’m not getting to sleep at all.
0:05:27 And we’re all dying and we’re 23 and we have no idea how to scale a service like this.
0:05:32 And we’re flying servers around because we didn’t know to use AWS because we’re idiots.
0:05:34 And it was a bazillion years ago.
0:05:38 And of course now I’m like, “Oh man, that’s the luckiest thing we ever could have had happen to us.”
0:05:42 But at the time I was like, “This sucks. I don’t get to sleep. We’re working hard.”
0:05:48 And when you started, you were actually coding, how did the job evolve and change?
0:05:52 And how did you stay involved like close to the metal or not or become a manager?
0:05:54 Tell us about your technical journey there.
0:06:01 At Kick, I started out just writing the code, ended up building the engineering team as we grew,
0:06:05 and then gravitated towards the product end of the world.
0:06:11 Because for me, I realized the impact of the things that we build on people is the part of the challenge
0:06:15 that motivates me. That’s where the fascinating problems wind up being
0:06:20 and then being able to bridge that into the physical realities of what do we have to build?
0:06:22 How can we make this thing real?
0:06:26 And even at Substack, I’m the CEO this time, which is a very different story.
0:06:28 But I built all the early code.
0:06:32 Jay and I wrote the vast majority of the product for the first over years.
0:06:37 When we finally relented and did that series A, it was just the three of us working out of my apartment.
0:06:39 And it was Jay and I that built everything.
0:06:44 I think some people have that choice where they want to go down the super IC really deep
0:06:49 versus a thing where you fork and you’re going more engineering manager and leader.
0:06:52 And then of course, a CEO, how did you know which one was right for you?
0:06:54 And obviously Jay is like a genius, right?
0:06:59 And so you obviously ended up with a very strong technical co-founder that you met at Kick originally.
0:07:05 I never really even thought about it too much as like the management track versus the IC track.
0:07:09 I think when you’re a founder, you tend to really blur those things
0:07:12 and you wind up having to do both in order to do a good job.
0:07:16 The way that I think about it is it’s like a cliche in our industry,
0:07:21 but it’s deeply true is you have to start with the user, start with the product,
0:07:25 start with the thing that matters in the world and be willing to work backwards to everything else.
0:07:29 Work backwards to what should our stack be?
0:07:31 How should our servers be architected?
0:07:36 And also, what should our org structure be and how should we hire and what should our business model be?
0:07:42 And ultimately, when you’re building a company, it is sort of a challenging, fun engineering problem.
0:07:50 If you look at it the right way, I think weirdly, maybe the founder path is a way to do both of those things in a pretty unique way.
0:07:54 And maybe before we transition, you can tell us about Peek-Kick.
0:07:59 Yeah, we started, we were sort of like WhatsApp before WhatsApp blew up for a minute there.
0:08:01 We were growing very fast. They were unknown.
0:08:04 And then Blackberry sued us and tried to kill us and almost succeeded.
0:08:05 And then we built it back up from nothing.
0:08:09 And in that second phase, I think probably the most interesting part,
0:08:14 there was at one point we were like 40% of U.S. teenagers were using kick.
0:08:17 We raised money at a billion dollar valuation from 10 cent.
0:08:23 The world was our oyster, except we didn’t know how to make money.
0:08:26 Yes. And then you guys have hundreds of millions of active users.
0:08:28 Hundreds of millions of users.
0:08:29 Yes, exactly.
0:08:30 It was big.
0:08:31 Really impressive.
0:08:33 And yeah, you were in your 20s when you built it, right?
0:08:35 This is truly like out of college.
0:08:37 This is one of these really rare, interesting experiences.
0:08:41 Yeah, I started building it in college instead of going to class.
0:08:44 Okay, so after kick, you have the messaging app wars, et cetera.
0:08:49 And then eventually you decided to move on and talk about early days of,
0:08:51 okay, you’re going to start a new company.
0:08:53 You want to do a bunch of things differently.
0:08:55 What were some of the lessons you took out of kick?
0:09:00 And how did that inform how you picked the sub stack idea in the first place?
0:09:02 There’s a lot of positive stuff I took from kick.
0:09:07 Even I think having been somewhere where we built something that had that kind of
0:09:13 traction made me believe that it was possible and understand like what pieces
0:09:17 have to come together for something like that to happen and understand the impact
0:09:21 that you could have when you make one of these online systems that everybody’s
0:09:22 spending their life in.
0:09:26 Like I gained a big respect at kick both for human nature.
0:09:30 Like the people using your software at scale are going to do all kinds of
0:09:34 strange things and can’t change who they are too much and you shouldn’t try.
0:09:39 But nonetheless, when you’re designing the systems that people are spending their
0:09:43 lives in online, you can make a heaven or a hell out of the exact same people.
0:09:46 And so the way you set up the rules, the way you set up the incentives,
0:09:49 the way you set up the norms, the way the system works.
0:09:54 I guess I became a believer in the kind of social technology that we’re building
0:09:59 at sub stack at all, which I think was a big part of what probably spurred me to
0:10:00 do it ultimately.
0:10:02 We did do a bunch of things differently.
0:10:05 We also had a business model in mind, a way to make money.
0:10:08 This is part of why we were always saying no to the money is we had a way to
0:10:09 make money, which was nice.
0:10:11 It’s very inconvenient for us.
0:10:12 I recommend it.
0:10:15 It gives you a lot of leverage whenever you need it.
0:10:16 I don’t know.
0:10:17 We built the thing fast.
0:10:20 Many of you guys know in HBO Silicon Valley, there’s the Russ Hanuman thing
0:10:24 about how it’s so great to be the no revenue company because then if you have
0:10:27 the no revenue company, then there’s no multiples.
0:10:28 There’s that.
0:10:29 It’s just the dream, right?
0:10:30 And there’s sort of an advantage of that.
0:10:31 You guys probably saw that.
0:10:33 We’re like, I hope everybody believes this shit.
0:10:35 Yeah, that’s right.
0:10:36 That’s right.
0:10:37 Well, it’s funny.
0:10:40 I mean, I was recently just making the argument that we’re now in an era where
0:10:45 you’re not going to be able to build a billion active user ad supported thing.
0:10:46 That’s an idea from 15 years ago, right?
0:10:49 And instead, because we’re in a much more competitive environment, there’s a lot
0:10:53 more going on, much easier to think about how do you build and monetize something
0:10:54 that’s very valuable?
0:10:56 That’s like a hundred million user environment.
0:11:03 I often ask founders now, hey, if someone wants to spend $5,000, how do they do that
0:11:04 in your product?
0:11:07 Because it’s almost like the deeper the monetization, the better.
0:11:12 And I know with Substack, there are certain people that spend a lot of money.
0:11:14 I forgot what the word of magnitude is, but a lot of money.
0:11:17 And then tell us about day one of Substack.
0:11:21 I think you had like Bill Bishop, you’re quite a famous writer, and he was involved
0:11:22 quite early.
0:11:25 I think he was like one of the seed investors and he was making good money on
0:11:26 Substack from the early days.
0:11:29 I’ll just imagine what like kind of year one was like.
0:11:30 Yeah.
0:11:33 Well, I’ll give you the potted founding story, which was I wasn’t actually planning
0:11:35 to build a company yet.
0:11:36 I was taking a year off.
0:11:39 I’d planned to give myself basically a sabbatical.
0:11:41 I’d been grinding pretty hard at the last startup.
0:11:44 I wanted to live my life and pursue my hobbies.
0:11:47 And I’d always been an avid reader.
0:11:49 I grew up in a house full of books.
0:11:54 And I’ve always thought that what you read, the media you consume, shapes how you
0:11:58 think, shapes how you see the world, sort of shapes who you are.
0:12:01 And so great writing and great culture is valuable.
0:12:05 The same way that you might write a piece of software and change how a million
0:12:09 people do their jobs, you could write a great essay and change who a million
0:12:10 people are.
0:12:13 And that thing is this inherently valuable, wonderful thing.
0:12:17 And my answer to that was, I should be a writer.
0:12:18 I should start blogging.
0:12:19 Cool.
0:12:22 This is going to be one of my things that I do in this sabbatical that I take.
0:12:26 And I wrote what was going to be an essay or a blog post or a screen or
0:12:27 something.
0:12:30 It was like outlining my frustrations with the media economy on the internet.
0:12:32 I was kind of like, wow, wow.
0:12:34 The internet’s killed the business model for traditional media.
0:12:39 And maybe social media is not, you know, an unalloyed good, et cetera, et cetera.
0:12:43 And I sent it to my friend Hamish, who’s a really a writer who became my co-founder
0:12:45 and he let me down very gently.
0:12:47 And Hamish, by the way, he’s like a published author.
0:12:49 He like wrote a book on Tesla.
0:12:52 He worked to Tesla.
0:12:55 He has worked for a lot of the top publications in the world.
0:12:58 And he was like, dude, it’s 2017.
0:13:02 And your point here is maybe newspapers are in trouble.
0:13:05 And maybe social media is a mixed blessing.
0:13:08 Sometimes we know this is not the brilliant new idea.
0:13:09 That you think it is.
0:13:14 However, a more interesting version of this thing might be so what, let’s say that all
0:13:18 of these things you’re whining about in this dumb essay, you’re writing are true.
0:13:20 What would you even do differently?
0:13:22 And we started arguing about that.
0:13:26 And that’s the thing that became the kernel of the idea for Substack.
0:13:32 And as we talked it over, we realized that it was simultaneously like a very ridiculously
0:13:34 ambitious grandiose idea.
0:13:39 Like maybe we could reinvent the business model for culture on the internet and create
0:13:45 an entire new economy of writing and media and video and things based on the idea that
0:13:49 people would pay for things they value that could cause all these new things to exist.
0:13:51 Blah, blah, blah, blah, blah.
0:13:58 And also there was this very simple MVP core of it that was like the very first instantiation
0:14:02 of that idea that was basically paid email newsletters.
0:14:08 And that thing was the very simplest version of this very large idea.
0:14:13 But that thing, I was like, I bet you I could build that in the weekend.
0:14:17 And he was like, I know this guy, Bill Bishop, I know a couple of people, like a couple of
0:14:22 my friends would probably use that tomorrow if only it existed.
0:14:28 So we had this like small group of people initially for whom the simple thing that we
0:14:31 could build was like the cure for cancer.
0:14:33 And of course, I couldn’t build in a weekend.
0:14:34 You’re always a little bit optimistic.
0:14:37 It took me a few weeks, but we built an early version of it.
0:14:41 We actually started out the very first act we did was publishing a manifesto.
0:14:45 I don’t recommend this, but we on this little CMS that I was hacking together, we like wrote
0:14:49 our founding document of what we were doing and why.
0:14:53 And I think Hamer sent that to Bill and a couple of other people and it hooked us up with
0:15:00 our first customer who was someone who had been writing email newsletter for years for
0:15:05 a big international business and government audience and had been wanting to charge money
0:15:06 for it, but couldn’t fit.
0:15:08 He was like the perfect early customer.
0:15:12 And we launched with him in October, 2017.
0:15:14 And it instantly worked.
0:15:17 He made a hundred grand his first day.
0:15:19 And I was just sitting there being like, is this good?
0:15:21 By the way, startups always work like this.
0:15:23 It turns out you built something in a weekend.
0:15:25 You make a hundred K in the first day.
0:15:26 You know what it is though?
0:15:30 If you ever talk about professional gamblers, every professional gambler has a story where
0:15:34 at the start of their career, they go on this huge winning tear.
0:15:38 And the reason this is true, obviously, is because everybody who starts their careers
0:15:42 a professional gambler on a losing tear does not go on to become a professional gambler.
0:15:45 And so we did this thing where we’re like, I wonder if this crazy thing we’re working
0:15:47 on is going to go anywhere.
0:15:52 And our first customer, we got super lucky in addition to being very smart and hardworking.
0:15:54 And it really worked.
0:15:56 And we were like, oh, shit, we’re doing this.
0:15:58 And we like got into YC.
0:16:00 And I remember very clearly in the YC interview.
0:16:04 So first of all, when we applied, we put on the form, we have $0 in revenue.
0:16:06 And then we interviewed four days later.
0:16:08 And they were like, it says you have $0 in revenue.
0:16:12 And I got to say, oh, actually, there’s been an update in the past four days.
0:16:13 There’s been $100,000.
0:16:15 And that seemed very good.
0:16:18 And I remember the thing that they said to us, they were like, oh, this is going to be easy.
0:16:23 If you just get four or five more customers like this, you’re going to like raise your demo day money.
0:16:25 The whole thing’s going to go like you’re off to the races.
0:16:28 We’re like, great, this is going to be like smooth sailing from here on out.
0:16:32 It took us like three years before we ever got another customer that was that good.
0:16:34 They were the best customer.
0:16:38 They were the biggest customer and the fastest launching customer for so long.
0:16:43 And there was such a long time where that early win overshadowed everything else we did.
0:16:47 And we were gradually like, oh my God, did we make a terrible mistake building this thing?
0:16:48 Is it going to work?
0:16:53 But that initial success, I think, gave us enough energy momentum to keep going.
0:16:58 And of course, eventually it turns out there’s lots and lots of people who go on to make millions of dollars on Substack.
0:16:59 And that thing has really worked.
0:17:07 But if we didn’t both, I think, find the right hole in the universe and then get a little bit lucky,
0:17:09 I don’t know if we would have done the thing.
0:17:19 One thing that was so fascinating in the early days was your choice of technology platform because this was way after the mobile first thing.
0:17:25 And I remember when we first had it, I was like, OK, it’s a website and email list when you didn’t do mobile.
0:17:29 And then, by the way, this guy was so against all the algorithmic anything.
0:17:31 This was almost like a Luddite startup.
0:17:33 You know what’s a sexy technology to work on?
0:17:34 Email.
0:17:36 Yes, exactly, exactly.
0:17:38 And so maybe two things on this.
0:17:43 Why pick from a tech standpoint something that was where you started as one.
0:17:47 And then in the early days, as you’re hiring and building out your engineering team,
0:17:55 how do you get people excited about the mission and everything when maybe folks come in and they’re like, OK, well, what’s the technology problem in this?
0:17:58 And tell us maybe how the engineering vision has evolved over time as well.
0:18:01 Technology is a means to an end.
0:18:16 And the thing that was so cool about email to us was this is one of the last places on the Internet, especially on your phone where you can have a direct connection with your audience without having to install an app.
0:18:21 And it’s not mediated by one of the big social algorithms.
0:18:23 Yes, there’s like a Gmail priority inbox algorithm.
0:18:34 But email is sort of like this really fun hack where early on you could be a writer and have a direct connection with your audience in a way that did actually work on the phone.
0:18:40 And it didn’t require that you install a new app because in the early days of Substack, nobody was ever going to install an app.
0:18:44 But if you have one newsletter you’re subscribing to, and it’s get this app to get it, that’s never going to work.
0:18:50 So we knew that we needed, especially early on, to have a growth loop that didn’t require putting an app in that thing.
0:18:52 We wanted it to be just as easy as possible.
0:18:54 We wanted it to be like, hey, you want this thing?
0:18:55 What’s your email? Cool.
0:18:57 If you want to pay for it, give some money. Great, you’re done.
0:18:58 So we built that.
0:19:00 And this is like working backwards from the customer.
0:19:02 We’re like, what’s the thing that’s going to be great?
0:19:04 What do we need to build to make that happen?
0:19:09 And what’s a great fun engineering challenge is scaling a service that has lots of users because you made something good.
0:19:18 That’s a better challenge than figuring out some really cool deep technological problem and then realizing that you didn’t make something that anybody cares about.
0:19:21 And then the tech problems are going to come with scale naturally.
0:19:22 And now we get to do all that stuff.
0:19:29 Yeah, we’re building like crazy machine learning stuff for our recommendation engine and we’re doing AI video editing for our video platform.
0:19:32 And we’ve got an app and it’s wonderful and we get to do all that stuff.
0:19:38 But we get there by not focusing on what’s a sexy technology, you focus on what’s a sexy product to build.
0:19:42 And then how are we going to let ourselves move very fast to do that?
0:19:57 I wanted to build on one point that you mentioned about the open Internet because I think a very interesting argument that we’re so used to the Internet being this place where you can publish anything and anything can happen.
0:20:08 I think you could make the argument that actually it’s a fluke because prior to the Internet, the primary platform was Microsoft Windows and a little bit Macintosh and Mika and like all these other things.
0:20:18 But Microsoft was a closed system and Microsoft did a lot of things with their power that caused them eventually to be sued by the DOJ and they became a convicted monopolist on all those things.
0:20:24 And then the Internet was open and then mobile is really a duopoly of two companies.
0:20:29 And then this new thing that we’re creating right now on top of it, which is all the AI things.
0:20:35 You’re like, well, is this going to be a couple really wealthy companies or is this going to be something that can be open again?
0:20:46 And so one thing with the Substack team that I always admired and something that got me excited about the company beyond just the business being great is that the mission is actually so amazing.
0:20:52 And you guys have such a strong philosophy beyond this as an amazing business model for people.
0:20:55 You’ve gotten to a point where there’s actually been real scale.
0:20:58 There is a mobile app that’s going big time.
0:21:02 What are you excited about from a technology standpoint now that you have hit that scale?
0:21:06 And presumably you are not coding and making all the tech decisions these days.
0:21:13 Yeah, funnily enough, I’m still like one of the top three contributors to the code base, even though I haven’t done anything useful in three years or something.
0:21:16 Yeah, so I mean, we’re over 3 million paid subscribers.
0:21:20 We have tens of millions of active users.
0:21:28 Some of the people who write on Substack are among the most read columnists in the U.S. in politics and a bunch of different topics.
0:21:35 I think there’s an increasing cultural footprint and intellectual footprint that runs even ahead of the economic footprint.
0:21:43 It’s easy to see ideas that started on Substack making their way into the traditional media and the rest of the media.
0:21:45 You can see that thing happen in real time.
0:21:51 And there’s a lot of frontiers of Substack that are very exciting to me because at its core, Substack is not about just email.
0:21:53 It’s not about just writing.
0:21:55 The center of it is this business model.
0:22:03 It’s this Promethean idea that as a writer, as somebody who’s making culture, you should own your connection with your audience.
0:22:05 You should have creative freedom.
0:22:07 You should get the editorial independence.
0:22:08 You should get to do the thing you want.
0:22:12 And you should be able to make money and potentially a lot of money.
0:22:24 If you can make a living or a fortune making something that people deeply value, the magic of that is not just, oh, I can do the same thing that I could do anywhere else and make some more money from it, although that’s nice.
0:22:28 The magic of that is the kind of stuff you can make gets better.
0:22:37 And then the opportunity that creates for us is now there’s this whole universe of really differentiated stuff on Substack.
0:22:47 And you can get the Substack app and read all the things that you’re subscribed to and have a wonderful experience for that and also discover everything else in that universe.
0:22:52 And then that thing adds, of course, to the network value for all the creators who are there, all the writers who are there.
0:22:57 And it’s become possible to build machine learning recommendations.
0:23:02 You can do all of this stuff that you used to have to have a team of 100 really smart people to make.
0:23:05 You now need a team of four really smart people to make.
0:23:16 We have those people at Substack and they’re getting to make this entire new universe of media on the internet with these sci-fi amazing tools from scratch that tons of people are reading and are making their favorite writers.
0:23:19 Filthy rich, that’s cool as hell.
0:23:24 Another thing we’re doing is expanding into audio, podcast, into video.
0:23:26 We have people launching video native stuff.
0:23:31 We’re at the forefront of all like we’re using all these AI tools where you can like automatically edit.
0:23:32 You can find a clip.
0:23:34 You can transcribe in another transcription.
0:23:45 There’s all of these things that are giving people tremendous new creative leverage that make things that you never could have made before suddenly become possible into a world where now we have a way of doing it.
0:23:47 Now we have a way to distribute that.
0:23:49 Now we have a way to make money from that.
0:23:51 It’s a very exciting time to be working on it.
0:23:54 We touched a little bit on AI a second ago.
0:24:04 And obviously one of the most interesting and compelling use cases for everything that’s happening in UI is creativity and being able to edit and a lot of other stuff.
0:24:08 Say a little bit about what you think, how it might play a role in the product.
0:24:19 And then to the extent that you’re skeptical about any roles that I might play, meaning ways you don’t think people want to consume content or ways that maybe tools won’t be useful.
0:24:22 I’m a big AI bull in general.
0:24:28 I think we are in the early phases of the most important technological revolution of my lifetime.
0:24:33 I think there’s going to be, as with any new technology, there’s like a hype cycle and a reality cycle.
0:24:37 And those two things don’t always match up and they don’t always have a predictable pattern.
0:24:38 That’s all true.
0:24:41 But I’ve seen enough this thing is going to change everything.
0:24:47 And the way that we think about it, I said it before a little bit, is like creative leverage.
0:24:58 I think what’s going to happen is that the cost of creating content of where things are fussy, when I need to produce things, when I need to like do work.
0:25:01 I used to have a big team that had to realize this creative vision.
0:25:11 The cost of doing that thing is going to plummet to zero and you’re going to have like a massive supply of content of every kind that you could ever want.
0:25:13 By the way, that’s not a new trend.
0:25:15 That’s just a continuation of a trend that we’ve already been in.
0:25:21 Even before AI, we had more content than we could ever want in the whole universe.
0:25:28 And the reason that people subscribe to a sub stack is not, oh, gee, I don’t have enough emails to read and I wish I could pay money to get some more.
0:25:36 The problem you have in a world of abundant content is like, how do I allocate my scarce attention, right?
0:25:38 You know what AI has not given us yet?
0:25:40 Is more seconds to be alive.
0:25:43 And so how am I going to spend my time in my life?
0:25:45 What kinds of things do I want to put in my mind?
0:25:49 If what I read is what I become, who do I trust for that?
0:25:52 What is the point of this culture that I’m consuming?
0:26:00 That stuff, it gets to like really interesting philosophical territory, but my answer to all of this is you’re still going to care what other people think.
0:26:06 You’re still going to want to have a connection to the human culture and human society that you’re a part of.
0:26:10 You’re going to want to not just consume that, but you’re going to want to act back upon it.
0:26:14 But then all of the tools that make that stuff are going to get so much better.
0:26:19 You’re going to be able to have two people write and make a feature film by themselves.
0:26:24 That’s better than anything that could have existed before or whatever the new format that replaces the feature film will be.
0:26:34 You’re going to have people that can sit there on their phone and just talk to each other and have it turn into the best TV show or the best series of clips or the best essay if you wanted to.
0:26:38 You’re going to have weird things where you write a science paper and it turns into a podcast.
0:26:48 Like, I don’t think we can predict exactly how it’s all going to bounce, but the like cost of creating really high production value slick, well done.
0:26:51 Whatever you want from it, content is going to drop.
0:26:57 Actually, that’s not only not going to lessen, it’s going to increase the value of your trust in relationships, the things that you really care about.
0:27:02 People always talk about like, how is AI going to impact, substack the product and what are we building?
0:27:04 There’s a bunch of interesting stuff there we’ve talked about.
0:27:05 It’s all creative tools.
0:27:06 It’s all that stuff.
0:27:14 The thing that I think is maybe less thought about, but just as important to us is how is AI changing like making a company?
0:27:23 Because my view on this is, I think there’s going to be like a great extinction event where there’s sort of like a free AI world and a post AI world.
0:27:29 And companies that are getting started now are going to be native adopters of all of the things that change as that happens.
0:27:31 And that happens pretty quickly.
0:27:40 And companies that came before by default are going to die, I think, because by force of habit, you’ll keep doing the old thing.
0:27:49 And unless you’re willing to remake yourself pretty substantially, you’re going to get laps by new entrants that are using these tools natively.
0:27:58 And so we’re pretty obsessed at substack of like avoiding that fate by being constant, early testers, early adopters.
0:28:00 I think of it as like testing the fences.
0:28:05 Because at any given moment, there’s like hype about 20 different things in AI that’s, oh, everyone’s going to be doing this now.
0:28:10 I’m like 19 times out of 20, it’s not there yet. It’s bullshit. People are getting hyped about something.
0:28:12 You can do a cool demo, but it’s not real.
0:28:15 But then one in 20 times, it’s fucking real.
0:28:18 And we cannot afford to miss the one in 20 times.
0:28:25 We got to be like, we have a Slack channel called model behavior where we’re just constantly talking about, is this possible now?
0:28:26 Does this work now?
0:28:27 Can it write tests for us?
0:28:28 Can it do this?
0:28:29 Can it handle our customer support?
0:28:30 Turns out, yes, it can.
0:28:32 And I think that’s true in our careers too.
0:28:35 I think this is true if you’re like a more senior engineer.
0:28:41 It’s very tempting to feel like, oh, yeah, I played around with a co-pilot. It’s not that good, whatever.
0:28:42 I’m not even faster with it.
0:28:44 This is all fake. It’s not going to be fake.
0:28:45 You got to test the fences.
0:28:55 And just to add to your point about companies, I think it’s absolutely true that we can now really conceive of using LLMs to write essays and ultimately books.
0:28:58 I think that’s like a linear extrapolation on what’s going to happen.
0:29:02 And we often talk about the two people that get together and they make a Marvel movie.
0:29:04 Today would cost hundreds of millions of dollars.
0:29:08 Tomorrow it’s something that you can do on the laptop and same with video games and so on.
0:29:11 But I very much agree that companies themselves are almost like a format.
0:29:16 How do you run something that feels like a thousand person enterprise, but by yourself?
0:29:18 That’s a very interesting vision in this.
0:29:21 And it’s absolutely true that we don’t know how companies are going to be organized.
0:29:27 And in fact, if you go back to pre-industrial revolution, I mean, every business was basically a family business.
0:29:32 You have the blacksmith and you teach that to your kids and it was all based on like the labor of your family.
0:29:34 Then factories were very different.
0:29:39 And then even the concept of a limited liability corporation had to be invented and so on and so forth.
0:29:44 And so yeah, one of the things we have to talk about is in the future, do we even know what the job titles are going to be?
0:29:45 Do we even know what?
0:29:47 You know, I think it’s very hard to extrapolate.
0:29:49 So it’s cool that you’re leading into all this.
0:29:56 And I go back to the thing I said before, if the thing you’re doing is starting with the value in the world that you’re making
0:29:59 and working backwards, so how do we do that thing?
0:30:06 And you’re good at that and you’re focusing on that and you’re willing to like shift up how you do that as you learn new things.
0:30:09 I think this whole thing is going to be a huge win for you.
0:30:13 It’s only going to make the people doing that more powerful, more effective.
0:30:17 We’re going to have such wealth and abundance that gets created.
0:30:22 And you just have to be willing to use whatever tool becomes available to make great things.
0:30:29 Do you think people want to subscribe to and follow like fully AI content creators?
0:30:35 Or do you think that it matters that there’s always a human in the loop for that authenticity aspect of it?
0:30:37 I’m not going to say anything’s not going to happen.
0:30:46 I have enough epistemic humility and I’ve been surprised enough times by things that have happened that I’m not going to say this is not going to happen or that’s not going to happen.
0:30:52 When I try to think about this kind of from first principle, the thing I come back to is like, what is the point of culture?
0:30:56 What are we doing here? What do I want from Substack?
0:31:01 What do I want from whatever thing I’m using to experience the Internet and put media?
0:31:04 What’s the point of media? Like, why do I consume media?
0:31:09 And I do think one answer is basically solely for its effect on me, right?
0:31:14 Like I use media like a drug. It’s like I watch this thing and I feel good and that’s what I get from it.
0:31:18 And I think AI is going to be able to do that really, really well.
0:31:23 I think it’s going to be like science fiction wire heading, like we already have things that are like this.
0:31:31 There’s already people that treat TikTok this way where you just switch your brain off or reality TV or any kind of thing, giving me a feeling, making me forget the world.
0:31:34 That will happen. There’ll be big businesses built on that.
0:31:37 That’s one of the real motivation, giving me what I want.
0:31:42 But the other point of media to me is learning what I should want.
0:31:48 It’s being part of a wider society of trying to figure out who I want to become.
0:31:54 Am I becoming more like the person that I think I should be? Am I understanding the purpose of my life?
0:32:02 Am I getting to contribute back positively? Am I connecting with other people who I can live in this society with?
0:32:08 I basically think both of these things are going to be enormous businesses and we’re building the second one.
0:32:12 Tell us about the next chapter of Substack.
0:32:18 What are you the most excited about now that you have scale, now that you have that like momentum going?
0:32:21 What is the next chapter? What are you the most excited about?
0:32:27 I think there’s probably three related things that we’re working on right now that I think are going to change what’s possible.
0:32:30 The first is the Substack app and destination.
0:32:32 Like you sort of asked like, why did we never build an app?
0:32:34 It’s like, well, we couldn’t at the start.
0:32:46 Turns out we can now and there is a place, you know, we want to be like the first class place on the internet where you can go and have a real direct relationship with the people that are making the things you deeply value.
0:32:51 And then that enables some other things, right? It enables the multimedia stuff.
0:32:53 We’re pretty deep into podcasting now.
0:32:55 Video is starting to really take off.
0:33:05 The fact that this model doesn’t only work for writers, but it also works for all kinds of creators who are starting to have the same kind of complaints.
0:33:12 A lot of people feel about Instagram today, like writers felt about the New York Times in 2019.
0:33:14 I don’t have enough control over this thing.
0:33:16 I’m not going to make the thing that I actually want.
0:33:19 I’m not making as much money as I wish I could from this.
0:33:29 It’s sort of like class of creative people beyond exclusively writers who are starting to discover substack in weird and wonderful ways that I never would have predicted.
0:33:38 There’s a thriving like fashion scene on substack, which could not have been further from Hamish and my original network that we were making.
0:33:47 And then related to those two things, the last thing is just like making it radically easier for the next generation of creators.
0:33:52 Within the last month, we’ve made it possible where you can just pull out the app and publish something.
0:33:55 And we’ve already had people who like pulled out the app and publish something.
0:34:00 And then three weeks later, they’re making like 50K and they’re like, oh, this is very interesting.
0:34:02 And I think you tie them to the video stuff.
0:34:04 You tie that into all this AI stuff.
0:34:08 I think we’re going to live in a world where like I can talk to my phone.
0:34:13 And if the things I say are important enough, smart enough, good enough, I can get famous.
0:34:15 I can make a million dollars.
0:34:22 Every other thing that I would have had to worry about besides making something great, I don’t have to worry about anymore.
0:34:27 And that’s going to dramatically increase the amount of great culture we have.
0:34:28 Last question for you, Chris.
0:34:41 One of my favorite things about substack when I talk to people about it is substack is one of those products where people will literally come up and tell me this has changed my life, especially for creators, right?
0:34:47 And so I’d loved to maybe close on maybe an anecdote or something, a very interesting creators experience.
0:34:49 I knew you have a lot of creators.
0:34:51 Well, there’s one person in particular that kind of jumps to mind.
0:34:59 I won’t say who it is, but there’s this early period where like a butterfly could have flapped its wings and they wouldn’t have started a substack, right?
0:35:01 It’s like at the very beginning, I’m not even a writer.
0:35:03 I don’t know if I want to do this.
0:35:05 I don’t know if I have anything worthwhile to do.
0:35:06 I don’t know if I can do it.
0:35:09 And sometimes you get to just give a little like push.
0:35:23 And so there’s this one woman who told us after the fact that she couldn’t figure out how to work our stupid website to start a substack to the point where she actually paid someone else to set up her substack for her and then started.
0:35:34 And then she went on to make over a million dollars a year on substack to the point where her husband was able to quit his job and spend time with the family.
0:35:39 And it like completely transformed their life.
0:35:43 And she’s like doing the kind of work she really believes in.
0:35:53 There’s so many people that have something very valuable to give the world if they had a way to make culture and have a real business model for it.
0:36:10 And the more that we can take this kind of Promethean act of giving them the power to do that, making it radically easy, especially at the beginning, the amount of stuff that can get created is the reason it exists.
0:36:12 All right, that is all for today.
0:36:15 If you did make it this far, first of all, thank you.
0:36:23 We put a lot of thought into each of these episodes, whether it’s guests, the calendar touchers, the cycles with our amazing editor, Tommy, until the music is just right.
0:36:29 So if you’d like what we put together, consider dropping us a line at ratethispodcast.com/a16c.
0:36:31 And let us know what your favorite episode is.
0:36:34 It’ll make my day, and I’m sure Tommy’s too.
0:36:36 We’ll catch you on the flip side.
0:36:39 [MUSIC PLAYING]
0:36:42 [MUSIC PLAYING]
0:36:44 (upbeat music)

0:00:04 – Upwards of 300,000 clinicians left the workforce
0:00:05 in the year 2021.
0:00:09 – The large banks right now up to 15% of their staff
0:00:10 is just in compliance.
0:00:12 – You go from having 5% net margins
0:00:14 to 30% net margins.
0:00:18 – We’ve just never seen the rate of technology adoption
0:00:20 that we’re seeing now ever in the history of health care.
0:00:21 – If you don’t do this right
0:00:24 and you don’t figure out like a very repeatable motion
0:00:25 for these smaller acquisitions,
0:00:27 it’s gonna be really challenging to build the company.
0:00:29 – How long are we willing to wait
0:00:30 to solve this problem?
0:00:33 The everyday person should really care about this.
0:00:36 – Hello everyone and welcome back to our big idea series.
0:00:38 We’re here hours away from 2025,
0:00:41 which will mark nearly 22 years
0:00:43 since the Human Genome Project was completed,
0:00:46 21 years of Gmail, 19 years of Twitter,
0:00:51 15 years of Instagram, and even 17 years of the App Store,
0:00:54 which released 500 apps when it first came out
0:00:55 and now has millions of apps
0:00:59 that have generated developers over a trillion dollars.
0:01:03 It’s been a wild ride and that ride is far from over.
0:01:05 And that’s why our partners who meet daily
0:01:07 with the folks who are building our future
0:01:11 are sharing their big ideas for 2025.
0:01:12 Last year, we predicted
0:01:15 – A new age of maritime exploration.
0:01:17 – Programming medicines, final frontier.
0:01:20 – AI pro schemes that never end.
0:01:22 – Democratizing miracle drugs.
0:01:24 – And so far in this year’s series,
0:01:26 in part one, we covered the growing hardware
0:01:28 software intersection, and in part two,
0:01:30 how large companies and startups
0:01:32 are fighting for AI dominance,
0:01:34 but also how countries are thinking
0:01:36 about its impact on their sovereignty.
0:01:39 Today, we talk applied AI
0:01:41 and how this platform shift changes the game
0:01:44 in healthcare, fintech, and even opens up a new opportunity
0:01:47 to rethink growth in traditional service industries.
0:01:49 Now, if that wasn’t enough,
0:01:51 you can check out the full list of 50 big ideas
0:01:54 over at a16z.com/bigideas.
0:01:59 As a reminder, the content here
0:02:01 is for informational purposes only.
0:02:02 Should not be taken as legal, business,
0:02:04 tax, or investment advice,
0:02:06 or be used to evaluate any investment or security,
0:02:08 and is not directed at any investors
0:02:11 or potential investors in any a16z fund.
0:02:13 Please note that a16z and its affiliates
0:02:14 may also maintain investments
0:02:17 in the companies discussed in this podcast.
0:02:19 For more details, including a link to our investments,
0:02:22 please see a16z.com/disclosures.
0:02:30 First up, AI applied to healthcare,
0:02:33 an industry fraught with dissatisfaction,
0:02:35 too slow, too expensive, too clunky,
0:02:37 but also too few staff.
0:02:39 – Healthcare is dealing with the mother
0:02:41 of all clinical staffing crises.
0:02:43 We’re short hundreds of thousands of doctors and nurses
0:02:45 relative to the level of rapidly growing demand
0:02:47 for clinical services that is projected
0:02:49 in the next five years.
0:02:51 That was Julie Yu, I’m a general partner
0:02:54 on the Bio and Health team here at Andreessen Horowitz.
0:02:57 – So can AI in the form of specialist models
0:02:58 realistically turn this around?
0:03:00 Here is Julie’s big idea.
0:03:04 – My big idea is what we call super staffing for healthcare.
0:03:06 And what that means is we’ve talked obviously
0:03:08 about just the rampant challenges
0:03:09 that exist in our healthcare industry.
0:03:11 So much is broken about it,
0:03:14 but really one of the sort of underlying fundamental issues
0:03:17 that we face is this supply-demand mismatch
0:03:20 between the amount of clinical supply that we have
0:03:22 in the forms of doctors and nurses
0:03:24 and other clinical service providers
0:03:27 versus the level of demand that we as patients
0:03:28 represent in the industry.
0:03:32 And it’s not just that the demand itself is high,
0:03:33 it’s that the nature of the demand
0:03:34 keeps getting more and more complex.
0:03:37 We’ve talked a lot about how the chronic disease burden
0:03:39 in America is increasing over time.
0:03:40 The number of diseases that any individual has
0:03:42 is growing and growing.
0:03:43 The number of drugs that we’re all taking
0:03:44 is growing and growing.
0:03:47 And so again, not only is it that there’s just
0:03:48 a shortage of labor relative
0:03:50 to the number of humans who need it,
0:03:52 but also the nature of the work
0:03:54 is also getting more complex over time.
0:03:55 – So let’s dive into that.
0:03:58 Just how short are we in this clinical staff?
0:04:00 – There’s a lot of different reports and stats
0:04:01 that have a wide range of numbers here.
0:04:03 But I think generally speaking,
0:04:05 when you look at, for instance, the number of doctors
0:04:07 that we have estimated a shortage in,
0:04:12 it’s anywhere between, let’s call it 60 to 100K doctors today
0:04:14 in terms of the relative demand that we have in our system.
0:04:17 And roughly about 75K to 150K nurses
0:04:19 in a commensurate sense.
0:04:21 But I think probably one of the more interesting issues
0:04:24 to unpack here is what do we actually mean by shortage?
0:04:26 So there’s largely two schools of thought
0:04:27 of how people define this.
0:04:30 Simply the number of human clinicians that we have
0:04:32 is one way to look at it.
0:04:33 The other way to look at it though
0:04:36 is given the number of nurses and doctors we do have,
0:04:39 are they actually fully performing
0:04:40 to the full extent of productivity
0:04:44 that number of laborers actually represents?
0:04:45 And I think it’s probably, at the end of the day,
0:04:47 a combination of both,
0:04:48 but those are two different ways
0:04:49 that people have cut that pie.
0:04:50 – And how did we get here?
0:04:52 What is leading to this shortage?
0:04:54 – If you zoom out the lens entirely
0:04:56 on this highly regulated industry,
0:04:59 there’s actually a structural and regulatory constraint
0:05:02 on the amount of supply that we allow into the industry
0:05:03 on any given time basis.
0:05:04 – Right, through education you’re saying?
0:05:06 – Education is one, certainly the number
0:05:08 of accredited medical schools is constrained,
0:05:10 both in terms of actual accreditation,
0:05:11 but also in terms of funding.
0:05:13 Standing up of medical schools is actually funded
0:05:16 by federal dollars and there’s just a physical limit on that.
0:05:18 The other version of that is licensure.
0:05:20 So once you get through your training,
0:05:22 you have to be licensed at the state level
0:05:24 to be able to practice medicine.
0:05:26 And again, there’s only a certain number of licensures
0:05:27 that are approved on a given basis.
0:05:29 You have to renew that every few years.
0:05:31 So there’s a structural constraint in the industry level
0:05:33 that constrained the supply side.
0:05:35 In addition to that, once you are a doctor,
0:05:36 there’s all sorts of challenges,
0:05:38 some of which you alluded to,
0:05:41 of the job itself that have led to constraints.
0:05:43 One is certainly the aging population.
0:05:45 There’s a ridiculous stat that in the last couple of years,
0:05:48 the percentage of doctors that are over the age of 60
0:05:51 has reached like 45%, which is crazy.
0:05:54 And so there’s this huge silver tsunami dynamic
0:05:56 of what happens when all those doctors retire.
0:05:57 There’s gonna be a huge gap just purely
0:05:59 on the basis of age.
0:06:00 There’s also the burnout issue.
0:06:01 Just given everything that happened
0:06:03 during the pandemic in particular,
0:06:05 we put so much burden on our frontline staff
0:06:07 and what we were asking of them to do
0:06:08 was obviously very challenging.
0:06:11 So doctors and nurses have left the workforce actually
0:06:14 in droves because of that burnout issue.
0:06:15 There’s lots of surveys that show
0:06:18 that upwards of 80% of doctors and nurses
0:06:20 will say the number one issue in their lives
0:06:22 is burnout due to their jobs.
0:06:24 And again, another ridiculous stat,
0:06:28 upwards of 300,000 clinicians left the workforce
0:06:29 in the year 2021.
0:06:30 – This is in America specifically?
0:06:31 – In America alone.
0:06:33 And even in the last couple of years,
0:06:37 about 7% of the country’s total active physician workforce
0:06:40 left their day jobs as being clinicians.
0:06:43 Even if we’re able to train like 100% of the capacity
0:06:45 of our medical system, of our licensure system,
0:06:46 the folks are just not sticking around
0:06:49 because the job itself has so much overhead.
0:06:51 And the last thing I’ll say is that doctors
0:06:52 are increasingly taking on more and more
0:06:55 administrative functions and academic functions.
0:06:58 So another amazing stat is I used to live in Boston.
0:07:01 Boston is actually the highest density of physicians
0:07:02 in the entire country,
0:07:04 but they have the worst access issues
0:07:07 because the vast majority of doctors there
0:07:09 are working in academic environments.
0:07:11 So more than half of their time might actually be used
0:07:15 on research, not on actual clinical activity.
0:07:17 So I can go on and on, but these are some of the reasons
0:07:19 why even if you were to look at the number
0:07:21 of doctors and nurses, there’s a lot of other factors
0:07:23 that play into the actual capacity
0:07:25 that’s represented in that pool of labor.
0:07:26 – Maybe we can also just quickly speak
0:07:28 to what the patient feels.
0:07:31 Is it longer wait times at the hospital?
0:07:33 Is it more expensive procedures?
0:07:36 Is it not being able to even get healthcare in some cases?
0:07:38 Can you speak to the second, third order effects
0:07:40 that ladder down to the everyday person?
0:07:43 – I think the most obvious issue is wait times.
0:07:45 And there’s studies that show that the average wait time
0:07:46 for let’s say a specialist appointment
0:07:49 across the country is like 50 days.
0:07:51 And it can range from 27 days on the bottom end
0:07:53 and up to 90 days on the high end
0:07:55 and even worse for certain subspecialties.
0:07:57 And so that alone is obviously an issue.
0:08:00 The other related issue is that even when you’re able
0:08:02 to book an appointment, let’s say a month out,
0:08:04 there’s data that shows that after the second week,
0:08:07 like after 14 days, basically the no show rate
0:08:08 goes way up, right?
0:08:10 Because your likelihood of actually committing to that date
0:08:12 and not canceling or not just showing up
0:08:13 goes through the roof.
0:08:15 And so one of the exacerbating issues
0:08:17 is that these slots are booked
0:08:20 and therefore the doctor is saving that time.
0:08:21 But a lot of that just goes unused
0:08:23 simply because the patients aren’t showing up
0:08:26 because they’ve waited so long for those appointments.
0:08:28 To your second point, if I have an issue
0:08:29 and I need to see a doctor immediately,
0:08:32 the longer I wait, the longer the chances is
0:08:33 that it gets worse, right?
0:08:34 I might end up in the emergency room,
0:08:37 which obviously can be 10, if not 100 times more expensive
0:08:41 to our healthcare system than that initial physician visit.
0:08:43 Another sort of issue, especially just given
0:08:44 how much is happening around the consumer
0:08:46 healthcare landscape these days.
0:08:49 The other challenge is that the less easy it is
0:08:50 for you to see an actual doctor,
0:08:52 the more likely that you’re going to go online
0:08:55 and try to self diagnose and get advice from folks
0:08:56 who might not be medically positive.
0:08:57 – Doctor Google. – Exactly.
0:08:59 And so those are some of the major issues,
0:09:00 but I think the wait time piece
0:09:02 is the underlying fundamental challenge
0:09:04 that leads to all these negative outcomes.
0:09:06 – Yeah, and I mean, you just spoke to this idea
0:09:08 of Doctor Google and that’s where technology maybe
0:09:09 is not playing the right role,
0:09:13 but it sounds like technology maybe can play a better role
0:09:14 in fixing this whole thing.
0:09:16 And I know there’s many different facets to this,
0:09:17 but I just love to hear your take on that.
0:09:20 Is it realistic that technology can actually come in
0:09:21 and curb some of these issues?
0:09:23 – I mean, I think the way to think about it
0:09:24 is absent technology.
0:09:26 How would we solve this problem?
0:09:27 And I think it boils down to like,
0:09:30 how long are we willing to wait to solve this problem
0:09:32 through just simply growing the supply side of our market
0:09:35 and also at what cost to take those seven to 10 years
0:09:37 to like fully train a doctor is not a cheap endeavor.
0:09:39 And so that’s really where I see the opportunity
0:09:40 for technologies to say, listen,
0:09:44 what if we just assume that there are a fixed number
0:09:45 of clinicians in our system?
0:09:47 And if we just have to go with the number of doctors
0:09:48 and nurses that we have today,
0:09:51 how can we increase supply and capacity?
0:09:52 Well, there’s two ways.
0:09:55 And this is where this super staffing concept comes in,
0:09:58 is how do we super staff by enabling
0:10:01 every individual provider to be able to be more productive
0:10:02 in their day to day work?
0:10:05 So that could come in the form of things like co-pilots.
0:10:07 How can I ensure that the amount of time
0:10:10 that a given doctor is taking to make any given decision
0:10:13 at the point of care in real time can be shrunken down
0:10:15 such that they are just far more efficient
0:10:16 with their decision making
0:10:20 and that the caliber of the decision that’s being made
0:10:23 can be at parity with, again, the best of their peers
0:10:26 on the basis of great data and great decision support.
0:10:28 So I think co-pilots is one way that you could imagine
0:10:30 unlocking additional capacity
0:10:34 within the current existing pool of clinicians.
0:10:35 And then a second vector,
0:10:37 which is a different form of super staffing is,
0:10:40 how do you augment the current labor pool
0:10:42 with autonomous agents?
0:10:45 And largely the form that’s taken today is,
0:10:48 if you were to unbundle the job of a doctor or a nurse,
0:10:50 there is a very significant portion of what they do
0:10:52 on a day to day basis that has nothing to do
0:10:53 with patient care,
0:10:54 that has nothing to do with the use
0:10:56 of their unique clinical judgment.
0:10:59 So can you take those tasks largely administrative in nature
0:11:02 or just communications-based and unbundle those
0:11:05 and give those to an autonomous agent
0:11:09 to then do a much better scale with much better accuracy
0:11:11 with much better reliability, such that, again,
0:11:13 the clinicians can actually focus on what they do best.
0:11:15 – And do you have a sense of what portion that is?
0:11:18 I know it will differ based on specialty and doctor,
0:11:21 but are we talking that this administrative portion
0:11:24 is like 3% or is it 30%?
0:11:24 Do you have a sense?
0:11:26 – Yeah, there’s actually studies that look at high motion
0:11:29 of what percentage of a clinician’s time is being spent
0:11:31 on non-clinical patient-facing care,
0:11:33 and it can be upward to a 50%.
0:11:37 Like half of their job is mired in these clerical tasks
0:11:38 that, again, take their time away
0:11:40 from actual care delivery.
0:11:42 Imagine them was doubling the capacity
0:11:44 of a given clinical labor pool
0:11:46 simply by applying some of these unbundling tactics.
0:11:48 – Yeah, and are we seeing tools like this already
0:11:49 on the market?
0:11:53 Where are we in the trajectory of LLMs or AI in particular
0:11:56 entering this industry and actually reshaping it?
0:11:57 – Yes, and this is what’s so exciting right now
0:12:00 is that we are truly in like a golden era
0:12:02 of healthcare technology adoption.
0:12:05 And one of the dynamics that we believe is driving
0:12:08 this rapid adoption cycle of these new AI tools
0:12:10 is what we call this kind of leapfrog dynamic
0:12:11 that we’ve talked about in the past,
0:12:14 where healthcare historically has been a laggard
0:12:16 with respect to software-based technology.
0:12:18 And we’ve always thought that was to the detriment
0:12:20 of our industry and everyone else has these great workflow
0:12:22 tools, these SaaS products.
0:12:23 Because we didn’t do that as an industry,
0:12:26 we don’t actually have the sunk cost bias
0:12:28 that other industries are facing right now
0:12:30 where they say we just spend billions of dollars
0:12:32 laying out these workflow tools.
0:12:33 We trained an entire generation of our workforce
0:12:35 to use these workflow tools.
0:12:36 And now we’re being faced with the decision,
0:12:39 okay, gosh, do I need to spend yet another billion dollars
0:12:42 on these new AI products completely rip and replace
0:12:44 this last generation of technology capabilities
0:12:47 and retrain my workforce on a whole new paradigm
0:12:49 of how they do their work versus in healthcare
0:12:51 where we’re just simply leaping from,
0:12:53 we can’t throw enough bodies at the problem
0:12:55 to how do we actually augment our bodies
0:12:58 and extend through these AI capabilities.
0:13:00 And so we believe that’s driving
0:13:01 a lot of the rapid adoption cycle
0:13:02 that we’re seeing right now.
0:13:04 There’s a few examples of this.
0:13:07 One is, so I mentioned co-pilot versus autonomous.
0:13:10 On the co-pilot side, there’s this class of products
0:13:12 called ambient scribe products,
0:13:15 where it’s a way for a doctor and a patient
0:13:16 to be having a visit.
0:13:19 And there’s an ambient scribe literally listening
0:13:21 so the doctor can actually make eye contact with you.
0:13:22 They’re not sitting there on their laptop
0:13:25 and that AI scribe is specifically tuned
0:13:28 to be able to translate that conversation
0:13:30 into medical documentation that is compliant,
0:13:33 that facilitates billing in the appropriate way,
0:13:35 that completely captures the clinical context
0:13:37 of that particular encounter
0:13:38 and then can be saved directly
0:13:39 into the electronic health record.
0:13:41 And so that is a class of products
0:13:43 that has really taken off.
0:13:45 Doctor friends that I talked to are all using,
0:13:48 one of these products is completely transformed their lives.
0:13:50 They’re getting back hours of the day
0:13:53 that are no longer being spent doing manual documentation,
0:13:55 which is a requirement of their job.
0:13:59 Another example is a class of autonomous communication
0:14:01 platforms where one of the biggest ways
0:14:03 by which the healthcare system communicates with us
0:14:05 is through call centers, right?
0:14:06 I think certainly in senior populations,
0:14:09 you tend to see that folks want
0:14:10 some degree of personal connection
0:14:12 and obviously they’re much more used
0:14:13 to using the phone modality.
0:14:15 And so there’s companies like Hippocratic AI
0:14:19 in our portfolio that has built an army of super staff,
0:14:21 of folks who can actually play the role
0:14:23 of either an outbound call center agent
0:14:26 that is trained on very specific protocols
0:14:28 that are infinitely patient
0:14:30 and can actually execute a number of tasks
0:14:32 that otherwise would rely on humans to do.
0:14:34 And what we’re seeing is it’s remarkable.
0:14:39 Number one, most hospitals have about 10 to 20%
0:14:42 of their headcount open at any given time.
0:14:44 And they simply just can’t find enough people
0:14:46 to fill those roles, namely in the call center setting.
0:14:47 And so what they’re saying is, gosh,
0:14:49 if we can’t hire people fast enough,
0:14:51 like how amazing is it that we could spin up
0:14:54 not just one or two, but 10 or 20
0:14:55 of these autonomous AI agents
0:14:57 that are actually again doing the same job
0:14:58 as these humans would do.
0:15:00 And what’s also incredibly interesting
0:15:02 and part of the why now for why we think this trend
0:15:05 is gonna be at a tipping point is they’re starting to say,
0:15:07 okay, why can’t we just pay for this
0:15:11 out of our labor budgets instead of our IT budgets?
0:15:12 And one of the historical constraints
0:15:14 with healthcare technology has been
0:15:18 that the portion of the budget of let’s say a hospital
0:15:20 that’s spent on IT has been a tiny fraction
0:15:22 of what it is in other industries.
0:15:24 The average is like two to 5% of budget
0:15:28 goes towards IT and healthcare versus like 15 to 30%
0:15:30 in like banking or financial services.
0:15:33 And so now we have this huge unlock opportunity to say,
0:15:36 rather than be constrained by that tiny IT budget,
0:15:39 what if we tapped into the 60, 70% labor budget
0:15:42 and actually be able to create additional leverage
0:15:44 by deploying AI in that fashion.
0:15:46 So these are some of the fascinating trends
0:15:49 that we’re seeing already happening in market
0:15:51 that demonstrate the sort of rapid adoption cycle
0:15:53 of these new AI products.
0:15:55 – Is it just that there is such clear demand
0:15:58 to fill these roles that we’re seeing something different
0:16:00 today or why does it seem like there has been this adoption
0:16:03 or like jumping to participate in this AI wave
0:16:05 when we didn’t really see that in the past wave?
0:16:08 – There is just a very high burden of integration.
0:16:09 At the end of the day, if you’re gonna be introducing
0:16:12 a new software tool into someone’s workflow,
0:16:15 it has to deeply integrate with an EHR type system
0:16:16 or whatever other practice management system
0:16:17 they might have.
0:16:19 And those systems are 40, 50 years old,
0:16:21 they don’t always have great APIs.
0:16:23 And so just the amount of work and time
0:16:26 and customization that’s necessary to get those in
0:16:29 has been a huge barrier to adoption in the past.
0:16:30 The second thing I would highlight
0:16:32 is that these are just magical products.
0:16:33 We have a company called Ambience Healthcare
0:16:34 in our portfolio that creates
0:16:35 one of these ambient scribe tools.
0:16:37 And when you look at the testimonial videos
0:16:38 that they have from their doctor users,
0:16:40 it’s unbelievable.
0:16:42 I mean, these folks are like literally in tears
0:16:44 saying that they’ve been a doctor for 30 years
0:16:47 is like the first time that they’ve actually enjoyed
0:16:48 the job of being a doctor
0:16:51 because it’s just allowed them to focus on the patient.
0:16:53 The last waves of technology products
0:16:55 that have been given to them
0:16:58 have just increased the amount of complexity of their jobs,
0:17:00 whereas these are doing exactly the opposite.
0:17:02 – Are there challenges that you’re seeing on the ground,
0:17:03 whether it’s the tools today
0:17:05 or the things that you see on the horizon?
0:17:06 We were talking about healthcare.
0:17:08 So it could be regulation or some other thing
0:17:11 that you think maybe actually producing friction
0:17:12 for us to implement this.
0:17:14 – I think the obvious thing is just
0:17:15 the stakes are much higher,
0:17:16 obviously in the healthcare setting
0:17:18 and the hallucination issues that you see
0:17:20 with kind of the generalist foundation models,
0:17:22 like that is just a non-starter
0:17:24 if you have issues like that in this particular setting.
0:17:27 And so what we have seen and perhaps why there was
0:17:28 maybe a little bit of lag relative to some of the
0:17:30 generalist foundation models
0:17:31 in terms of adoption and healthcare,
0:17:33 but these companies that I’m describing
0:17:36 have built specialist models in healthcare.
0:17:38 One of the issues with the generalist foundation models
0:17:40 is that, I mean, not an issue.
0:17:41 It’s obviously a huge feature
0:17:42 that they are trained on internet data,
0:17:45 like the entirety of all publicly available human knowledge.
0:17:48 But the reality of healthcare is that the majority
0:17:51 of the esoteric protocol specific,
0:17:53 compliance specific information
0:17:55 tends to not just be online.
0:17:57 It tends to be in these proprietary systems
0:17:59 behind the firewall,
0:18:01 embedded into paper-based binders and things like that.
0:18:03 And so because of that,
0:18:04 it is necessary for these companies
0:18:06 to build specialist models
0:18:08 that are trained on a different set of data
0:18:10 than they might extend some of those generalist models,
0:18:13 but they do need to ingest a different set of information
0:18:15 and specifically be tuned for safety.
0:18:17 And so people do need a specialist approach
0:18:19 to tuning the foundation models
0:18:21 to perform appropriately in this setting.
0:18:23 Another one would be exactly what you said,
0:18:26 the regulatory compliance issues associated with
0:18:28 the use of autonomous AI in particular
0:18:32 and how that bumps up against FDA approval criteria,
0:18:35 under what circumstances do you need to get a AI tool
0:18:37 approved by FDA or not?
0:18:38 – Do you?
0:18:40 – Yes, in fact, one of the benefits that healthcare has had
0:18:43 is that we are the only industry that actually has
0:18:47 an existing regulatory framework for approving products,
0:18:48 especially in the clinical setting.
0:18:52 So there are already hundreds of AI radiology models
0:18:55 or other models that sort of help automate the assessment
0:18:58 and diagnosis of a given set of patients
0:19:00 based on a certain area of medicine.
0:19:02 But how does that apply to a gen AI model?
0:19:04 There’s a lot of work being done right now by the FDA
0:19:06 to really extend those models to be appropriate
0:19:08 for these far more dynamic systems
0:19:10 that just have a lot more uncertainty
0:19:13 than your standard ML predictive model modality,
0:19:16 which is really what that system was designed for.
0:19:18 The last thing I’ll say is reimbursement.
0:19:21 There are some early signs that insurance companies
0:19:25 are creating actual billing codes for AI,
0:19:29 where they now say that when AI enabled visit
0:19:31 or an AI enabled diagnostic
0:19:33 can now be reimbursed differentially.
0:19:35 And you can get credit for using these AI systems
0:19:37 in a different way than if you were just not.
0:19:38 And so that’s really exciting
0:19:41 to see these new paradigms persist.
0:19:42 But I think there’s still very much a long pull
0:19:45 until we have the ability to say, okay,
0:19:48 if I have a fully autonomous like AI therapist,
0:19:50 Slingshot AI, which is one of our portfolio companies,
0:19:54 how can we ensure that those will be systematically reimbursed
0:19:57 at least that parity with what a human therapist
0:19:58 might get reimbursed for,
0:20:00 that has not yet been defined in our industry.
0:20:03 So I think there’s a lot of good early signs
0:20:06 that there’s a willingness by the system
0:20:08 to embrace these new paradigms,
0:20:09 but still a lot of wood to chop
0:20:11 in terms of defining a comprehensive set of ways
0:20:12 to get that in.
0:20:13 – Do you think that this has the potential
0:20:16 to even reshape the ecosystem that already exists
0:20:17 and not just augment it?
0:20:19 – Yeah, I think there’s a lot of really amazing things
0:20:21 to consider like what this unlocks
0:20:23 in terms of both the existing clinical paths
0:20:25 and then things that we don’t even do today
0:20:27 that are now possible to do.
0:20:29 So one really interesting one is just part of the constraint
0:20:31 of our current clinical supply
0:20:33 is that we do have state-by-state licensure, right?
0:20:35 So if I become a doctor,
0:20:37 I’m basically constrained by my geography.
0:20:39 And if there’s a poor patient
0:20:42 who happens to need my services in Montana
0:20:44 and I’m not licensed there, but I do have capacity,
0:20:45 I’m not able to serve that person.
0:20:48 I think what this whole kind of AI world unlocks
0:20:52 is imagine that you had national scale capacity
0:20:54 and the sort of the routing challenge
0:20:56 becomes a really unique one
0:20:58 that actually could leverage AI, right?
0:21:01 So you have this sort of national scale routing opportunity
0:21:03 to just simply take advantage of the latent capacity
0:21:05 in the system at national scale.
0:21:06 There’s this new concept
0:21:08 that actually one of our portfolio companies introduced
0:21:11 called asynchronous medicine, which means, okay,
0:21:13 the typical way we think about seeing a doctor
0:21:16 is you show up at their office, it’s a 30 minute visit,
0:21:17 it’s face-to-face and live,
0:21:19 and it’s highly transactional though, right?
0:21:22 So you’re in and you’re out and then months might go by
0:21:24 or years might go by between visits,
0:21:26 whereas asynchronous care is something
0:21:28 that completely flips that paradigm on its head
0:21:31 and says what if your doctor was in the cloud
0:21:34 and anytime of the day, 24/7, you could text them,
0:21:37 you could send them a message and they would respond,
0:21:40 but the level of ongoing engagement
0:21:42 that you can have like the friction to anyone encounter
0:21:44 with your doctor goes way down
0:21:46 because you can text them at any time of day.
0:21:49 And that’s a new paradigm that again is purely enabled
0:21:52 through these AI-powered clinic models
0:21:55 where you can have these doctors who are available 24/7,
0:21:58 who are using AI as co-pilots to enable them
0:22:02 to be hyper responsive to any question that comes in.
0:22:04 And so I do think that the nature of maybe how doctors
0:22:06 do get trained and like what specialties actually exist
0:22:09 that don’t exist today that are purely available
0:22:11 in the post-LLM world.
0:22:13 I think a lot of those contours
0:22:15 are very dynamically changing as we speak.
0:22:17 – Maybe just looking to 2025 in particular,
0:22:19 since this was your big idea for next year,
0:22:21 is there anything that you’re specifically looking out for
0:22:23 or excited to see?
0:22:26 – We’ve just never seen the rate of technology adoption
0:22:29 that we’re seeing now ever in the history of healthcare IT.
0:22:30 We’ve been waiting for these moments
0:22:32 to take what we know to be the problems
0:22:34 and we know what the solutions can look like,
0:22:37 but there’s just been again like this sort of a barrier
0:22:39 between the problem space and the solutions.
0:22:41 And I think what we’re seeing now are signs
0:22:42 that perhaps that wall has been broken
0:22:44 and we’re at a breaking point
0:22:45 where the industry needs these solutions
0:22:48 and therefore we’re gonna see a much higher adoption
0:22:49 than we’ve ever seen before.
0:22:51 I think it will come in the form of these super staff.
0:22:53 I think we will increasingly see adoption
0:22:56 of these AI tools in the form of labor
0:22:59 and coming from completely different budgets,
0:23:00 from completely different buyers
0:23:02 than we’ve historically seen.
0:23:04 I think this is all to lay down the groundwork
0:23:06 to truly unlock the latent capacity
0:23:08 that we have on the supply side of our market
0:23:10 that’s been sitting there all along,
0:23:13 but just hasn’t had the right technology modality
0:23:14 to unlock it at scale.
0:23:19 Julie just talked, AI applied to healthcare,
0:23:22 spawning a fundamental rewrite of buyers and budgets,
0:23:26 but AI may also have the potential to rewrite compliance.
0:23:29 Companies in banking, insurance, and healthcare industries
0:23:31 spend countless hours in millions of dollars
0:23:32 staying in compliance.
0:23:34 Today, banking and insurance regulations
0:23:36 spend tens of thousands of pages.
0:23:40 SBA landing documentation alone exceeds 1,000 pages.
0:23:44 That was Angela Strange, general partner at A16Z.
0:23:46 Here’s Angela’s big idea.
0:23:50 – In 2025, regulation is gonna become code.
0:23:52 Regulation is one of the biggest costs
0:23:54 to existing companies, banks and insurance companies
0:23:56 and healthcare companies specifically,
0:23:59 and a huge burden to new entrants coming in
0:24:00 and to innovation.
0:24:03 And I think in this country, we have a system
0:24:05 where we efficiently add regulation,
0:24:06 but we never take away regulation.
0:24:08 And so this problem keeps getting worse.
0:24:11 And finally, with the advent of AI,
0:24:13 technology can help solve this problem.
0:24:15 – You mentioned the fact that we add regulation every year.
0:24:18 Just how lofty are these regulations?
0:24:21 – If you look at just federal banking regulation,
0:24:25 right now there are 50,000 different rules and regs
0:24:26 across different agencies,
0:24:29 and that’s up from 10,000 in 1970.
0:24:31 – Has compliance always been this burdensome
0:24:34 or like what’s actually driving that increase?
0:24:35 – Here’s how it’s done today.
0:24:39 You’ve got a large thousand page PDF documentation.
0:24:44 A software company will make a fairly brittle workflow tool.
0:24:46 And then you will have to hire many people
0:24:48 to work inside brittle workflow tool
0:24:50 and figure out is this compliant or not compliant.
0:24:53 So for instance, anti-money laundering.
0:24:54 We’re gonna compliance monitoring tool.
0:24:57 It sends up alerts if they think Stephanie is laundering money.
0:25:00 And then somebody who works in that compliance department
0:25:01 is gonna have to do a bunch of research
0:25:02 across a bunch of web sources
0:25:04 and other different databases
0:25:06 and write up a small report, yes, no.
0:25:09 When that goes wrong, it goes really wrong.
0:25:11 And most recently, big headline,
0:25:14 TD Bank was fined $3 billion
0:25:17 for having a long backlog of these compliance alerts.
0:25:19 – If you actually break down
0:25:21 what they were failing to monitor,
0:25:25 we’re talking about tens of thousands of detection alerts,
0:25:26 thousands of investigation cases.
0:25:28 These are really big numbers.
0:25:30 And to your point, the kind of things
0:25:32 that are really hard for startups to manage.
0:25:34 So maybe we can talk about that.
0:25:36 How does that impact the economics?
0:25:38 Whether it’s the compliance officers they need to hire
0:25:39 or just like dealing with these fines,
0:25:41 how does this actually work out?
0:25:44 – The large banks right now up to 15% of their staff
0:25:46 is just in compliance.
0:25:48 You went into the fourth fastest growing job
0:25:50 in the United States in the last 20 years.
0:25:51 – Compliance officers?
0:25:52 – Compliance, surprise, I glad the witness said, right?
0:25:53 – Yeah, yeah, yeah.
0:25:55 – So you can’t hire enough people.
0:25:56 They’re completely underwater, right?
0:25:59 I’m sure the compliance team at TD was working very hard.
0:26:01 So we need a technology solution.
0:26:03 So how is this going to happen?
0:26:05 I believe that every white-collar worker
0:26:07 eventually is gonna have a co-pilot.
0:26:10 And some of those roles will be replaced by agents,
0:26:11 but we’re talking regulation.
0:26:13 So we’re not just gonna let LLMs out into the wild.
0:26:15 We’re gonna use them to turbocharge
0:26:17 very hardworking compliance people.
0:26:20 And that’s exactly where some companies are entering.
0:26:22 So for instance, you would get an alerts.
0:26:24 We think so-and-so is money laundering.
0:26:25 They can go do the web searches.
0:26:26 They can check the databases.
0:26:29 They can bring all of this information together.
0:26:31 And now instead of a compliance officer
0:26:33 having to do hours of research,
0:26:35 they can look at this report generated
0:26:37 by their friendly compliance co-pilot.
0:26:39 Yes, no, that’s a five-minute task
0:26:41 versus a several-hour task.
0:26:43 Why do you think LLMs in particular
0:26:45 are uniquely suited for this?
0:26:50 ‘Cause they’re able to apply some form of judgment.
0:26:52 And so let’s take a very specific example
0:26:53 near and dear to our hearts.
0:26:55 Anytime we want to tweet something,
0:26:57 we are a registered investment advisor.
0:26:59 We have a great compliance team.
0:27:00 We have to send our tweets over
0:27:02 and they will tell us if we are compliant
0:27:04 with the SEC marketing role.
0:27:06 I imagine nobody listening has actually read that rule
0:27:07 is 400 pages.
0:27:09 So our choices are you could have a software
0:27:14 that creates an if, then, this, that type decision tree
0:27:19 or you can have a LLM ingest that rule
0:27:21 and they’ll be able to make those judgment calls
0:27:23 based on what’s in that document
0:27:24 if my tweet is compliant or not
0:27:27 and then it could come back with specific modifications.
0:27:29 And there’s companies, for instance, Norm AI
0:27:31 is one that’s doing exactly that.
0:27:33 It is codifying specific regulation
0:27:34 such that you can send it things
0:27:36 and it’ll come back and tell you the answer.
0:27:39 – Do you feel like from the types of products
0:27:41 that you’ve seen that they are at the point
0:27:44 where they can sufficiently interpret those gray areas
0:27:47 that sometimes do exist in compliance?
0:27:50 – So LLM’s hallucinating for my consumer colleagues,
0:27:51 that’s a feature.
0:27:53 It comes out with great creative fun
0:27:55 and enhancing things in the world of compliance
0:27:56 that is definitively a bug.
0:27:58 Smarter entrepreneurs are getting around this
0:27:59 in two different ways.
0:28:03 One, it’s much more of a co-pilot versus a full agent.
0:28:07 And then they’re able to test enough edge cases around
0:28:11 that the likelihood of hallucination is very, very, very low
0:28:12 based on how they fine tune the model
0:28:15 to that specific regulation.
0:28:17 The other piece is a big part of compliance
0:28:20 is being able to show that you’ve gone through
0:28:22 all of the steps to be compliant.
0:28:25 And so versus a black box coming back with,
0:28:28 yes, we’ve onboarded this business in a specific way,
0:28:30 it’s gonna show exactly all of the checks that you’ve done,
0:28:32 all of the documents you’ve gotten,
0:28:33 what you’ve pulled out of it.
0:28:37 And so it’s a very clear delineation of what’s been done.
0:28:39 But again, it saves the officer having to do all that work,
0:28:40 they can just check the box.
0:28:41 – That makes sense.
0:28:43 And are you seeing these tools on the market?
0:28:46 Give us some examples of what already exists.
0:28:48 – Notoriously banks, insurance companies,
0:28:51 even fintechs can be slow to sell into.
0:28:52 I think there’s a couple things going on.
0:28:56 One, this has become such a burdensome problem
0:28:59 that buyers are very open-minded to new tools.
0:29:01 And two, companies of all sizes have woken up
0:29:03 and said that AI can make a significant difference
0:29:04 to our business.
0:29:06 And so the sales cycles are moving much more quickly.
0:29:10 And so they range from thinking of co-pilots
0:29:12 that go along with your compliance officers.
0:29:14 I mentioned a couple of examples there.
0:29:19 The other thing that’s happening is these older system
0:29:22 of record tools, so take transaction monitoring
0:29:26 that are built pre-AI, they used to be seen
0:29:28 as too risky to rip and replace.
0:29:32 And now what we’re seeing is the new Gen AI first tools
0:29:34 are 10x better versus 2x better.
0:29:36 So for instance, Sardine is a company
0:29:39 that’s got a much more modern transaction monitoring system
0:29:42 and they are seeing customers rip and replace,
0:29:44 optimize, it’s been around for 20 years.
0:29:46 And that’s just not something that we would have seen
0:29:50 five years ago, but it’s because it’s just that much better.
0:29:52 – I assume, as you said, many of these industries
0:29:55 are a little slow to adopt technology.
0:29:57 What stands out in terms of what may actually
0:29:58 stop this adoption?
0:30:02 – It is moving a lot more quickly than I would have thought.
0:30:03 I’ve been investing in and around the space
0:30:06 for the last 10 years and the momentum
0:30:09 and even just the ACVs that some early companies
0:30:11 are getting is quite fast.
0:30:13 I think the two things that come up with the blockers
0:30:16 where the smart founders have gotten around are one,
0:30:18 how do I prevent this from hallucinating?
0:30:20 And I think the two ways are one, showing the work,
0:30:22 and then two, proving that you have fine tuned your models
0:30:24 close enough to this regulation
0:30:25 that that actually doesn’t happen.
0:30:29 And a backup plan as a human does actually resolve it.
0:30:33 But I think the most important can be wedging in
0:30:36 at a pain point that is very acute with the right buyer.
0:30:39 And personas that weren’t sold into that often before
0:30:41 have become very popular.
0:30:41 – Oh, like what?
0:30:44 – Chief compliance officers, like BSA officers.
0:30:47 Like I imagine all of these potential buyers
0:30:48 are now soon getting a lot of interest
0:30:50 from very technical solutions.
0:30:51 – That’s great.
0:30:53 If we think maybe a little more long term
0:30:54 about opportunity here, right?
0:30:57 Given the overhead that exists today,
0:31:01 what are we unable to do because of all of that compliance?
0:31:05 – So think of just a consumer’s life in financial services
0:31:08 and how much better some of these new entrants
0:31:09 have made your life.
0:31:12 So trading now is free.
0:31:13 I would argue that’s in large part
0:31:16 because Robinhood came along, made trading free
0:31:19 and put pressure on all the large brokerages.
0:31:22 For a large portion of the US population,
0:31:25 it was very difficult to get a free checking account.
0:31:27 Time came along, enabled that to be possible
0:31:29 and put pressure on the incumbents.
0:31:32 And so it is good for everyone, for new entrants
0:31:34 to be able to come into this space.
0:31:37 And one of the biggest barriers is just regulation.
0:31:39 And so the easier that we can make it
0:31:43 for smart people of all resources
0:31:44 to be able to start new companies,
0:31:45 it’s gonna be better for consumers,
0:31:47 it’s gonna be better for businesses.
0:31:49 – And as we looked at 2025,
0:31:51 what are you looking for in particular?
0:31:54 – Yeah, I love two categories.
0:31:58 One are old systems of record
0:32:00 that you can do just infinitely better
0:32:03 and especially ones that have an extreme economic impact.
0:32:05 So I mentioned before, small business lending.
0:32:09 Many businesses get loans from the SBA.
0:32:11 It takes on average 90 days.
0:32:12 Why?
0:32:15 Part of that is there’s a 1,000 plus page documentation
0:32:17 about how do you make this loan compliant, right?
0:32:20 So what is that LOS in that process?
0:32:22 It’s very much gonna streamline that.
0:32:24 And it is great for the economy for our small businesses
0:32:26 to be able to get the credit that they need.
0:32:28 So systems of record, 0.1.
0:32:31 And then 0.2 would be any area
0:32:35 where a company needs to hire dozen plus people
0:32:37 to do a task where those people would rather
0:32:39 do something a little bit more strategic.
0:32:41 That can be a very interesting wedge in
0:32:43 for a co-pilot type opportunity.
0:32:44 And then you can start expanding
0:32:46 and doing my task from there.
0:32:47 – So as we look to these changes,
0:32:49 it might be obvious for a big bank like TD
0:32:51 how this can improve their operations.
0:32:54 But why should the everyday person care about this?
0:32:56 – The everyday person should really care about this.
0:32:58 And there’s three main reasons.
0:33:01 One, it’s a big barrier to new innovative companies
0:33:03 coming into the space.
0:33:05 And I’ve mentioned examples like Robinhood, like Chime,
0:33:08 like others that really upend some of the business models
0:33:10 of the larger banks and provide free trading, free checking,
0:33:13 just better consumer services.
0:33:17 And then it’s a hidden cost to a businesses
0:33:19 or consumers everyday life.
0:33:21 It is longer for a business to get a loan
0:33:22 because there’s a lot of regulation
0:33:24 and it’s very difficult to comply with it.
0:33:27 It is more difficult to get a loan modification
0:33:29 that you desperately need to be able to pay your mortgage
0:33:31 because somebody needs to be trained
0:33:34 in the thousand plus any main mortgage servicing guide
0:33:35 so that they’re able to decide,
0:33:37 can you get that loan mod and how will they do it?
0:33:40 So improving regulation, making it code,
0:33:42 making it easy to be compliance with
0:33:45 will improve everyone’s lives and economy.
0:33:46 But if you just look at the chart,
0:33:48 it just keeps growing and growing and growing.
0:33:50 And so this problem is going to get worse, not better.
0:33:52 That’s just federal.
0:33:55 Every state is following exactly the same trend line.
0:33:59 And the opportunity for AI to disrupt
0:34:03 how businesses operate does not stop with compliance.
0:34:06 – AI is automating jobs across traditional service industries
0:34:09 like insurance, law, real estate, and IT.
0:34:11 Though many of these businesses have historically been
0:34:14 low margin and difficult to scale,
0:34:16 some are now leveraging LLMs,
0:34:19 particularly to automate roles involving voice,
0:34:22 email, or messaging to transform
0:34:25 into high margin, scalable models.
0:34:25 – That was.
0:34:27 – Joe Schmidt, I’m a partner on the apps team
0:34:29 here at Andries and Orwitz.
0:34:30 – And here is his big idea.
0:34:33 – My big idea is romanticizing organic growth.
0:34:35 And really what that means is I see the next evolution
0:34:37 of private equity being possible
0:34:40 where small businesses can leverage technology
0:34:43 to basically automate parts of their services organization
0:34:45 and then go and use a better business model
0:34:47 to acquire other small businesses.
0:34:49 – Obviously LLMs, the talk of the town,
0:34:51 penetrating every industry.
0:34:52 But tell me what you’re seeing on the ground
0:34:55 in terms of how these LLMs are reshaping industries.
0:34:58 – It’s just amazing to see how fast adoption has been
0:35:00 across whether it’s horizontal applications
0:35:02 or vertical applications.
0:35:03 And just for this conversation,
0:35:05 maybe we’ll focus on the vertical.
0:35:07 Obviously recently announced our investment in 11x,
0:35:09 which is automating the sales function.
0:35:11 And I don’t think I’ve ever really seen demand
0:35:14 like what we’re seeing for basically automating
0:35:15 the role of an SDR.
0:35:16 And so you’re already seeing it
0:35:17 within the sales organization.
0:35:20 If you think even on a more vertical specific basis,
0:35:21 you know, I see it in like freight
0:35:23 with companies like happy robot,
0:35:25 basically automating the call center operations
0:35:27 around what a freight broker does,
0:35:29 whether it’s like calling and booking a load
0:35:31 or dealing with a new carrier
0:35:33 that a freight broker has to deal with.
0:35:34 We see it in healthcare with another one
0:35:36 of our portfolio companies, Tenner,
0:35:38 which is automating the back office processes
0:35:39 associated with a small health core practice.
0:35:41 And so I think we’re seeing a bunch
0:35:42 of these applications pop up
0:35:43 and frankly have growth rates
0:35:45 on anything I’ve ever seen again.
0:35:46 And it’s, I think it’s just a really exciting time
0:35:49 to be building software and to be building startups.
0:35:50 – Yeah, and I love those examples
0:35:52 ’cause you’re touching so many different industries.
0:35:54 And maybe if we zoom out just a little bit
0:35:56 and think about maybe like a framework for disruption,
0:35:59 how do you think about where maybe this could be applied?
0:36:00 – If you’re interested, go read our blog
0:36:02 because we’ve got a bunch of ideas on it.
0:36:03 – Yeah, and we’ll link to that.
0:36:03 – Yeah, we’ll link in it.
0:36:05 But I think that like the way to just zoom out
0:36:06 and think about this opportunity is like,
0:36:08 where are there a bunch of humans
0:36:10 that are basically dealing with whether it’s voice
0:36:12 or paper processes
0:36:14 where there’s a bunch of unstructured data
0:36:16 that they need to basically synthesize
0:36:17 and then make sense of
0:36:19 and then generate some sort of output.
0:36:21 Obviously that’s what LLMs are super good at.
0:36:23 And so like the very common way of thinking about that
0:36:25 might be like, hey, there’s a call center somewhere, right?
0:36:27 That’s like the healthcare or the insurance example
0:36:28 or the sales example
0:36:30 or maybe you have some sort of,
0:36:31 we call it the messy inbox problem here
0:36:33 where there’s a bunch of different pieces of paper
0:36:35 that are coming in and you need to synthesize them
0:36:37 and put them into some database.
0:36:39 But oftentimes that’s like an internal team.
0:36:41 But then also you could look for wherever there might be
0:36:44 DPO spend like business process outsourcing spend
0:36:46 where maybe a consulting firm is doing this for you.
0:36:47 So those are the hot buttons
0:36:49 is like looking for work like that
0:36:51 where it’s basically something that is,
0:36:53 a bit’s not Adam’s industry,
0:36:54 like you’re dealing with paper,
0:36:55 not physical world things
0:36:59 and something that is conducive to be automated by LLMs.
0:37:00 – Yeah, I have to admit,
0:37:01 when I was reading your big idea,
0:37:03 I thought, isn’t this just like private equity?
0:37:04 So convince me otherwise,
0:37:07 how is this different from traditional private equity?
0:37:09 – First, like what is private equity?
0:37:12 To oversimplify, like private equity is a buyer,
0:37:15 purchases a company using some combination of equity and debt,
0:37:17 traditionally like a good deal of debt
0:37:19 and then basically tries to improve earnings
0:37:21 in the relatively short term.
0:37:23 Normally private equity has like a three to five year
0:37:25 whole period and then sell it to the next buyer.
0:37:28 Private equity firms typically are optimizing for IRR
0:37:29 in general, we’re gonna return
0:37:31 and they’re less focused on ultimately
0:37:34 like a multiple on money return,
0:37:36 basically saying they’re not focused on the biggest
0:37:38 possible version of the company that they can build,
0:37:41 rather they’re focused on making incremental improvements
0:37:42 in this time period,
0:37:43 ’cause that’s like how they think about liquidity
0:37:45 and looks how the model works.
0:37:46 The problem with that is it creates
0:37:48 like a little bit of a perverse incentive structure
0:37:50 where typically they’re focused on one of two ways
0:37:51 of making money.
0:37:53 First is like the most obvious,
0:37:54 like when people hear like private equity,
0:37:56 it’s like, oh no, people are gonna get fired.
0:37:57 But obviously what they’re trying to say is,
0:38:00 hey, this hasn’t been run in the most lean and efficient way,
0:38:02 maybe it possibly could have.
0:38:03 And so, hey, the accounting department
0:38:04 doesn’t need to be this big
0:38:07 or maybe we can consolidate these two different teams
0:38:10 or whatever it might be, that’s the first bucket.
0:38:11 The second bucket normally is saying,
0:38:14 hey, we’re gonna go acquire other small businesses
0:38:15 or large businesses, doesn’t matter,
0:38:16 we’re gonna acquire other companies.
0:38:18 We’re gonna use more equity and debt
0:38:21 and we’re gonna basically buy the next incremental company
0:38:22 for a lower multiple,
0:38:25 roll it into our company and have multiple expansion.
0:38:27 So these are the two very common ways
0:38:28 that private equity firms make money.
0:38:31 These aren’t designed to have these businesses
0:38:32 become the best versions of themselves,
0:38:34 it’s designed to make money.
0:38:34 – In the short term.
0:38:35 – And it does.
0:38:37 So I think what’s interesting about this model
0:38:39 is that it’s actually addressing
0:38:41 like a third leg of the stool in my opinion
0:38:43 where traditionally private equity
0:38:45 isn’t focused on like fundamentally shifting
0:38:47 the way that the business works.
0:38:48 What I’m talking about is actually building
0:38:51 like a technology business and the technology core
0:38:54 to a traditional services company, right?
0:38:55 So in the past, the private equity firm would say,
0:38:58 hey, like we don’t need that many people in accounting,
0:39:00 I’m talking about automating like a large swath
0:39:01 of the company.
0:39:03 And to do so, you need to have a very different view
0:39:06 on returns and you need to have a very different view
0:39:09 on short-term earnings because what will end up happening
0:39:12 is it turns out engineers pretty expensive, right?
0:39:14 And it requires like significant investment
0:39:17 to get this business to a more automated state.
0:39:19 And so you have to not think about IRR
0:39:20 over the next two or three years.
0:39:21 And instead you need to be thinking about
0:39:25 what’s the largest long-term valuable business
0:39:27 that I can build and how do I invest in that future?
0:39:30 So that’s like the very different way of building businesses.
0:39:32 I think you have to structure these businesses
0:39:34 from the start, not as private equity companies,
0:39:37 but as like long-term permanent,
0:39:39 like we’re gonna build the best version of this company ever
0:39:41 and like a company that can go to public.
0:39:43 And frankly, I think the best versions of this company,
0:39:45 like there’s a couple of examples on the public markets
0:39:48 that are a little bit more holed-co
0:39:49 traditional software versions of this,
0:39:53 but one is a very well-known company called Danerher.
0:39:54 There’s another one called Tyler Technologies
0:39:56 based out of Texas.
0:39:57 These are like probably the best examples
0:39:59 of these businesses historically.
0:40:01 But again, they’re not AI native
0:40:04 and they’re basically focused on the first two buckets
0:40:06 of this, which is like financial engineering
0:40:08 and/or like operational efficiency,
0:40:11 not building an AI to their businesses.
0:40:12 – That’s super interesting.
0:40:13 And I think just for the listeners,
0:40:15 let’s try to be really concrete here and paint a picture
0:40:17 of what this could look like,
0:40:19 this idea or opportunity of quote,
0:40:22 AI-powered vertical specific service startups.
0:40:23 – Yeah, that’s a melt ball.
0:40:24 – Yeah, I was gonna say, it’s a lot of Ss,
0:40:26 but what would that look like?
0:40:27 Let’s try to really paint a picture here.
0:40:30 – Okay, so let’s imagine that staff has an insurance agency
0:40:32 in Columbia, Missouri.
0:40:34 Okay, so let’s call it a small insurance agency.
0:40:37 You do three or $4 million in gross written premium.
0:40:39 That generally will mean you do a few hundred thousand,
0:40:42 call it half a million dollars in revenue, that revenue.
0:40:44 Let’s say you have single-digit margins on that revenue.
0:40:46 So you’re living a nice life in Columbia, Missouri,
0:40:48 but this is a small business.
0:40:49 There are thousands, tens of thousands
0:40:51 of these businesses across America.
0:40:54 You’re selling, imagine home, auto, life insurance,
0:40:55 and a little bit of commercial insurance.
0:40:57 At your staff insurance agency,
0:41:00 maybe you have one or two or three people in the back office.
0:41:01 Maybe there’s a CSM that’s dealing
0:41:03 with like net new business that you’re writing.
0:41:05 There’s someone in the back office
0:41:06 that’s dealing with all the paperwork, right?
0:41:08 ‘Cause when you write an insurance policy,
0:41:09 it’s like, hey, I gotta get a bunch
0:41:10 of information from the customer.
0:41:13 I gotta go get the quote from my legacy system,
0:41:14 and I gotta write all this stuff down,
0:41:17 and I gotta deal with all these back office processes.
0:41:20 It’s painful, and it’s a lot of people moving paper around.
0:41:22 That’s the reason your margin is so low,
0:41:23 and you’ve tried to fix this, right?
0:41:25 You’ve tried to buy software,
0:41:26 but the only software in this industry
0:41:28 is really legacy, right?
0:41:30 There’s these two big businesses in insurance.
0:41:32 You don’t really have a great system here,
0:41:33 and so you’ve just been hiring these people,
0:41:35 and you’ve been living this, frankly,
0:41:37 hell for your entire career.
0:41:40 And so what I’m talking about now is I come to you
0:41:42 and I say, hey, Steph, wouldn’t it be great
0:41:44 if I could give you like basically automation
0:41:46 that could take all of that paperwork
0:41:50 that you’ve been paying five, 10, 15 plus percent
0:41:53 of your net margin for, and we can automate it?
0:41:54 And what does that require, right?
0:41:57 It’ll require you going and getting deeply integrated
0:41:58 from the technology side,
0:42:00 deeply integrated into the core systems of record,
0:42:02 leveraging AI agents to basically have
0:42:05 some of those conversations with the customer,
0:42:07 whether it’s on new business, on renewal,
0:42:09 or on client servicing,
0:42:12 and then actually resolving these claims in real time,
0:42:14 or like resolving these service issues in real time,
0:42:16 which is a super interesting opportunity
0:42:18 because then you start to think about, okay,
0:42:20 now Steph’s insurance agent doesn’t need four people,
0:42:23 it might need two people, it might need one person,
0:42:25 and you go from having 5% net margins
0:42:26 to 30% net margins,
0:42:29 and you go and you put a pool in, or whatever it might be.
0:42:31 Or frankly, what I think is really exciting
0:42:33 is you use that extra cash,
0:42:35 and you buy John’s insurance agency that is street,
0:42:37 and now you see what the opportunity is
0:42:38 to run this playbook.
0:42:39 – Yeah, because basically what you’re saying is that
0:42:41 what you put into the technology
0:42:44 can be replicated for insert other business here.
0:42:46 Tell me a little bit more about why you need
0:42:48 to buy the company though.
0:42:49 Could I not create a tool that just sells
0:42:52 into John’s insurance company, Steph’s insurance company?
0:42:53 – Yeah, and I think that’s,
0:42:54 you actually made a really good point
0:42:55 at the end of what you just said,
0:42:58 which is, hey, this can be applied to the next business.
0:43:00 And if you build this software, I think in the right way,
0:43:02 in the right industry vertical,
0:43:04 once you’ve built those core integrations,
0:43:06 in the example that we were just doing,
0:43:07 it’s into the applied epic system,
0:43:09 or whatever core system of records
0:43:12 exists across the small business you’re thinking about,
0:43:14 those systems are, in many cases,
0:43:16 in these legacy industry are pretty ubiquitous.
0:43:18 And so you can repurpose this technology
0:43:20 across each one of these small businesses.
0:43:22 I’m not saying it’s gonna be simple or easy,
0:43:23 it’s all required implementation,
0:43:24 but that’s the opportunity here
0:43:26 is to stay vertical specific
0:43:29 and build this deep kind of understanding of the systems.
0:43:31 But it’s a great question on why not
0:43:34 just sell software to these companies.
0:43:36 Frankly, if you’ve ever talked to a small business
0:43:38 across middle America, like these mom and pop companies,
0:43:41 they are not the most technology savvy customers.
0:43:43 And it’s not that they’re not great businesses
0:43:45 or super educated, it’s just they just aren’t accustomed
0:43:47 to running their companies this way.
0:43:49 If I’m a software vendor, and frankly, I’ve done this,
0:43:51 I’ve gone and I’ve tried to sell software
0:43:54 to mom and pop insurance agency in Columbia, Missouri,
0:43:56 trying to drive adoption is super tricky.
0:43:58 And so in buying the business,
0:44:01 instead of trying to do business development exercises
0:44:02 and convince someone that like,
0:44:04 “Hey, the pot of gold at the end of the rainbow
0:44:06 “is worth it for you.”
0:44:07 Instead, we can now say,
0:44:10 “Actually, this is what we’re gonna run the business.”
0:44:12 Operationally, this is our new cadence and our new workflow,
0:44:14 and here’s how we’re gonna get to the next stage.
0:44:16 – So you’re saying basically, in purchasing the business,
0:44:19 it gives you the leeway to actually shift things fundamentally,
0:44:22 which is really what this is articulating.
0:44:23 Maybe another question is just,
0:44:26 you seem to think this is a really big opportunity.
0:44:28 PE itself is a huge industry.
0:44:29 So why do you think this is actually
0:44:32 maybe a bigger opportunity than traditional private equity?
0:44:35 – So I think that this all comes down to capital efficiency.
0:44:36 And so for example,
0:44:38 buying one of these small companies that I just described
0:44:41 isn’t actually like that expensive of an endeavor.
0:44:42 And if you can do what I’m talking about,
0:44:44 you impact earnings and all of a sudden,
0:44:45 like the first one, two or three,
0:44:48 maybe you don’t immediately pay back using cash.
0:44:51 But after the first few of these transactions,
0:44:52 you start to throw off cash
0:44:55 and that cash can then be used to acquire more
0:44:56 and more of these small businesses.
0:45:00 And so that becomes this beautiful like flywheel effect
0:45:01 to the acquisitions that you can make.
0:45:03 And then that new business is you acquire.
0:45:06 And so unlike traditional private equity
0:45:07 where you’re constantly raising debt,
0:45:09 maybe you need more equity.
0:45:11 And again, it’s much more financial engineering.
0:45:13 Instead, you can actually just compound your investment
0:45:16 over a long period of time and build value that way.
0:45:17 So I think that’s the opportunity here.
0:45:18 – Yeah. And I guess what you’re getting at is
0:45:22 the margin change is so significant where you can do that.
0:45:24 But instead of you’re doing some of the financial engineering
0:45:27 that traditional P relies on,
0:45:30 you’re often not changing the margins from 5% to 30%
0:45:32 or to 50% even in some cases.
0:45:34 – Yeah. And I think there’s a nuanced point here where,
0:45:36 again, if you do this right in the right vertical
0:45:38 and the businesses that you end up acquiring
0:45:41 are worth more to you than to any other buyer,
0:45:43 that’s when you can start to do really interesting things.
0:45:44 Right? So I don’t know if this makes sense,
0:45:46 but let’s say I’m doing something
0:45:48 and Mark is doing something in the same space
0:45:49 for those insurance businesses.
0:45:50 We’re both running the same playbook.
0:45:52 He’s a private equity firm.
0:45:53 I’m running a technology business.
0:45:54 And I go to you and I say,
0:45:57 hey, I have this wonderful technology platform
0:45:58 when you like to come and grow with me
0:46:01 and I can take your earnings from 5% to 20%
0:46:04 versus Mark says, hey, I’ll just pay you a bunch of money
0:46:05 and you’re going to stay at 5%
0:46:07 and maybe I’m going to fire Susan
0:46:08 who runs your accounting, right?
0:46:10 Whose value prop sounds better?
0:46:12 And so in this way, you can one,
0:46:15 like your business is actually worth more to me, right?
0:46:17 Because I can impact earnings in the short term
0:46:18 based on my technology.
0:46:20 And two, you want to work with me
0:46:22 because you see, hey, I’m not going to have to go in
0:46:24 and fundamentally do a bunch of these things
0:46:25 that are really cutthroat.
0:46:26 Obviously, maybe there’s certain people
0:46:27 that are a little redundant.
0:46:29 So I’m not saying this won’t happen,
0:46:30 but I do think that it’s a much better story.
0:46:31 And it’s a growth story
0:46:33 versus just a traditional efficiency story.
0:46:35 – Yeah. And speaking of that growth story,
0:46:36 it sounds like a lot of the growth
0:46:38 is inorganic in this case.
0:46:40 At least that’s what this story is about.
0:46:40 So tell me more about that.
0:46:43 Why the inorganic growth strategy
0:46:45 versus maybe a more organic approach.
0:46:47 – Yeah. So I think that the way to think
0:46:49 about these more traditional services businesses,
0:46:51 oftentimes they’re local enterprises, right?
0:46:53 So again, like in Columbia, Missouri,
0:46:56 you have a reputation as being like a trusted source
0:46:57 of insurance services.
0:46:59 And so people want to come and work with staff
0:47:00 for that reason.
0:47:02 And it’s actually traditionally pretty hard.
0:47:05 It’s not just insurance, but think about real estate, law,
0:47:08 like all of these industries, they’re traditionally local.
0:47:10 And so it’s pretty hard to go add new customers
0:47:11 and grow organically.
0:47:13 If you’re doing that, like it’s pretty special.
0:47:14 In my opinion, the best version
0:47:16 of this inorganic growth strategy,
0:47:18 it’s like you’re acquiring a very sticky,
0:47:21 a very local, geo-specific customer base.
0:47:23 And then you’re improving the business that’s serving them
0:47:25 and freeing up that person who already has
0:47:27 like all of these wonderful connections
0:47:31 in whatever space they are to go do more of that, right?
0:47:32 So now you’re not focused on,
0:47:34 hey, shoot, I’ve got to deal with all of this stuff
0:47:35 in the back office.
0:47:37 I can go to the little league field
0:47:40 or wherever the heck you go to sell insurance.
0:47:41 And basically get to know people
0:47:42 and do business development.
0:47:45 And so now not only are we impacting like earnings,
0:47:48 we’re actually impacting the net new organic growth rate,
0:47:49 which is pretty clever.
0:47:50 The way to think about this in my mind
0:47:51 or the way that I think about it is,
0:47:54 hey, if you were running an enterprise software company,
0:47:56 right, and you were in the public markets
0:47:59 and you’re saying, hey, I could have 18, 20, 24 month paybacks
0:48:01 on a great enterprise customer logo,
0:48:03 that’s the way to think about these small businesses.
0:48:07 Can you find a way to have like strong paybacks
0:48:09 on these new business acquisitions
0:48:11 and treat them much more as just like
0:48:13 an enterprise software sales motion
0:48:15 rather than a traditional private equity firm
0:48:16 acquiring some small business
0:48:18 and doing a bunch of financial engineering on it.
0:48:20 – But obviously this is not easy.
0:48:21 – No.
0:48:21 – This is not operationally easy.
0:48:22 So tell me more about that.
0:48:23 Like as you-
0:48:25 – As I sit here in my venture capital ivory tower,
0:48:27 all these small businesses are simple to automate.
0:48:30 Yeah, I can’t wait for my tweets at me.
0:48:31 – Well, tell me more about that.
0:48:32 So as you think through the challenges,
0:48:35 what is most stark, what’s most clear
0:48:38 as the thing that someone pursuing this would need to solve?
0:48:40 – Yeah, I think the biggest one is just understanding
0:48:42 that like there’s gonna be a bunch of human problems
0:48:42 when it comes to this.
0:48:44 And so you have to be like eyes wide open.
0:48:45 You have to be a great operator
0:48:48 and you have to deal with scaling a human driven business.
0:48:49 So I think that’s the first one
0:48:51 and all the implementation problems around that.
0:48:53 There’s gonna be a bunch of hairy legacy systems.
0:48:55 You’re gonna need to navigate like all the process mining
0:48:57 and figuring out what do I need to build here
0:48:59 to actually drive incremental value.
0:49:01 So that’s the second piece, human level issues.
0:49:02 Once you get through those,
0:49:04 then you can start to do the process mapping,
0:49:05 which is also exciting.
0:49:06 And then the third piece is,
0:49:08 I’ve talked about this a little bit before,
0:49:10 but a lot of these legacy services businesses,
0:49:12 while they may be bits oriented,
0:49:15 like at some point there’s like an atom involved,
0:49:17 someone has to go do the business development.
0:49:19 Like it’s not like you can just fully automate these companies,
0:49:20 but it could be even more complicated.
0:49:22 Imagine you bought saying in the insurance example,
0:49:24 maybe you bought a third party administrator,
0:49:27 which is someone who deals with like insurance claims.
0:49:29 So you could deal with all the back office stuff around
0:49:32 handling insurance claim, accepting all the documents,
0:49:33 figuring out what’s what,
0:49:35 but AI is not gonna drive the truck to the crash.
0:49:38 AI is not gonna figure all that stuff out.
0:49:40 And so I think that’s the other tricky part of like
0:49:42 of these spaces is like identifying
0:49:45 where is the right kind of combination of bits and atoms,
0:49:46 primarily hopefully bits,
0:49:48 and then trying to figure out how to navigate
0:49:50 some of these human problems and implementation problems.
0:49:52 But it’s definitely gonna be hairy.
0:49:53 I think if you don’t do this right,
0:49:56 and you don’t figure out like a very repeatable motion
0:49:58 for these, again, smaller acquisitions,
0:49:59 I think it’s gonna be really challenging
0:50:00 to build a big company.
0:50:01 So that’s the opportunity here.
0:50:03 – What are you looking out for in particular?
0:50:04 Is it a certain type of founder?
0:50:07 Is it a certain motion that you see someone doing
0:50:08 that you think is really intriguing?
0:50:10 And you think, okay, they’re on the right track here.
0:50:12 What are you really looking for?
0:50:12 – First things first,
0:50:15 I’m looking for an entrepreneur or entrepreneurs
0:50:17 that have bottoms up understanding
0:50:18 rather than tops down understanding.
0:50:19 And what does that mean?
0:50:21 I wanna find entrepreneurs that have earned secrets
0:50:22 in these end markets, right?
0:50:24 So somebody who has maybe sold
0:50:26 to the thefts insurance agency
0:50:28 or someone who has run thefts insurance agency
0:50:30 or whatever the corollary is
0:50:31 for each one of these markets.
0:50:34 It’s very hard to come in and know the shorthand.
0:50:36 You have to trust me that I’m gonna come in
0:50:37 and help you bring your business
0:50:39 and take your business to the next level
0:50:40 if I’m gonna be that acquirer.
0:50:42 And so I think you need to be a very special person
0:50:43 on that front.
0:50:45 The second piece is an understanding of,
0:50:48 hey, can this industry be automated?
0:50:51 Isn’t it sufficiently bits oriented
0:50:53 to the point where AI can actually have an impact on that?
0:50:53 – Like what percentage?
0:50:54 – Right, exactly.
0:50:55 Is it the right kind of work
0:50:58 that is conducive to being automated by AI?
0:50:59 That’s another piece.
0:51:00 And then I think the third piece
0:51:04 is hopefully there’s a very large fragmented market
0:51:07 of acquisition targets such that like you’re not
0:51:09 basically going in and having to compete
0:51:11 with the larger mid to large size
0:51:13 traditional private equity deals.
0:51:15 And so if you can find one of these end markets
0:51:18 where there’s a significant number of mom and pop businesses
0:51:20 where you can go and run that cack-minded playbook
0:51:22 of acquiring and implementing
0:51:24 and building a repeatable motion around that,
0:51:27 I think there’s just insane room to run
0:51:29 and massive businesses to be built.
0:51:30 And so those are the combination
0:51:31 of factors that I’m looking for.
0:51:34 And hopefully someone out there is considering this
0:51:35 the way that I would run it
0:51:37 is I would go find a design customer,
0:51:39 start building software and then show yourself
0:51:40 and frankly everyone else
0:51:41 that you can actually start
0:51:43 to move the needle on earnings.
0:51:44 And that’s when I think you know
0:51:45 you got the tiger by the tail
0:51:47 and it’s time to go and run fast.
0:51:49 – All right, I hope these big ideas
0:51:52 got you geared up and ready for 2025.
0:51:55 Stay tuned for our final segment part four
0:51:56 where we dive into the big ideas
0:51:59 straight from our A16Z crypto team
0:52:03 who submitted 14, yes, 14 big ideas for the new year.
0:52:04 We’ll see you then.
0:52:07 (gentle music)
0:52:09 (gentle music)

0:00:06 In 2024 alone, I think there were more than 700 pieces of state-level legislation that were AI-specific.
0:00:10 The average query on perplexity is 10 to 11 words.
0:00:14 The average search on Google is 2 to 3 keywords.
0:00:19 You’ve had an enormous amount of NVIDIA’s purchasing orders come from the balance sheet of governments.
0:00:22 You can actually reimagine the AR experience.
0:00:28 If you’re a founder like that who has the guts, then your impact on humanity ends up being quite generational.
0:00:33 Here we are again, inching even closer to the end of 2024.
0:00:38 And as we near 2025, here are a few dates to give you some perspective.
0:00:46 We’ve had 24 incredible years of Wikipedia, 18 years of the iPhone, and 16 years since the Bitcoin White Paper release.
0:00:54 So as we look to 2025 and the speed of innovation is only increasing, we continue our coverage of A16Z’s big ideas.
0:00:59 Together with the dozens of partners who are meeting daily with the people building our future.
0:01:01 Last year, we predicted.
0:01:03 A new age of maritime exploration.
0:01:06 Programming medicine’s final frontier.
0:01:08 AI through schemes that never end.
0:01:10 Democratizing miracle drugs.
0:01:14 On deck this year, closing the hardware/software chasm.
0:01:16 Game tech powers tomorrow’s businesses.
0:01:18 Super staffing for healthcare.
0:01:22 And throughout our four-part series, you’ll hear from all over A16Z,
0:01:26 including American dynamism, healthcare, fintech, games, and more.
0:01:30 However, if you’d like to see the full list of 50 big ideas,
0:01:34 head on over to a16z.com/bigideas.
0:01:40 And of course, if you missed it, check out part one, all about the intersection of hardware and software.
0:01:45 As a reminder, the content here is for informational purposes only.
0:01:48 Should not be taken as legal, business, tax, or investment advice,
0:01:51 or be used to evaluate any investment or security,
0:01:55 and is not directed at any investors or potential investors in any A16Z fund.
0:02:01 Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast.
0:02:07 For more details, including a link to our investments, please see a16z.com/disclosures.
0:02:15 Today in part two, we’ll be talking about the topic of the day.
0:02:19 The month, and quite frankly, the year. Artificial intelligence.
0:02:22 There’s certainly AI, AI, AI.
0:02:28 And the race is on, whether it’s across companies like Google and the Disruptors Chasing Way 10 Blue Links,
0:02:32 or sovereign countries trying to capitalize on the next frontier,
0:02:37 or even device companies figuring out their role as AI meets the edge.
0:02:39 That is all on deck today.
0:02:43 There’s a bit of an innovator’s dilemma here, and I’m excited to watch it play out.
0:02:48 That was Alex Zimmerman, and I’m a partner here on The Growth Fund.
0:02:50 Here’s his big idea.
0:02:53 The search monopoly ends in 2025.
0:02:59 Google controls 90% of U.S. search, but its grip is slipping.
0:03:05 Its recent U.S. antitrust ruling encourages Apple and other phone manufacturers
0:03:08 to empower alternative search providers.
0:03:13 More than just legal pressure, GenAI is coming for search.
0:03:18 ChatGBT has 250 million weekly active users.
0:03:25 Answer Engine Perplexity is gaining share, growing 25% month on month,
0:03:28 and changing the search engagement form.
0:03:33 Their queries average 10 words, three times longer than traditional search,
0:03:36 and nearly half lead to follow-up questions.
0:03:44 Claude, Grock, MetaAI, Poe, and other chatbots are also carving off portions of search.
0:03:52 60% of U.S. consumers used a chatbot to research or decide on a purchase in the last 30 days.
0:03:57 For deep work, professionals are leveraging domain-specific providers
0:04:01 like Causally, Consensus, Harvey, and Hebbia.
0:04:05 Ads and links historically aligned with Google’s mission.
0:04:11 Organize the world’s information and make it universally accessible and useful.
0:04:17 But Google has become so cluttered and gamed that users need to dig through the results.
0:04:20 Users want answers and depth.
0:04:26 Google itself can offer its own AI results, but at the cost of short-term profits.
0:04:32 Google as a verb is under siege. The race is on for its replacement.
0:04:37 Maybe start by setting the stage. So how big is the search market today?
0:04:42 The search market is enormous. Anyone listening to this uses search.
0:04:45 Virtually anyone with the internet uses search.
0:04:49 But to put some numbers around it, Google that I just mentioned, they’re the biggest game in town.
0:04:53 They’re approaching $200 billion of revenue annually.
0:04:56 They’re still growing double digits, highly profitable.
0:05:00 Microsoft Bing, which has been the number two player for a long time,
0:05:06 mid-single-digit market share, so pretty small, they have $12 billion of revenue.
0:05:10 This is a massive, massive market, one of the largest out there.
0:05:14 And it does feel like there are forces reshaping this industry, so tell me about those.
0:05:17 So there’s certainly AI, AI, AI.
0:05:23 But before we jump into that, we can level set with the legal pressure that’s mounting on Google.
0:05:26 So earlier this year, Google was declared a monopoly.
0:05:32 The court ruled that they’re spending billions of dollars to phone manufacturers.
0:05:37 In the case of Apple, tens of billions of dollars is monopolistic.
0:05:42 It’s anti-competitive and it’s preventing their competitors from gaining share in the marketplace.
0:05:46 They are the default search engine on all these phone manufacturers.
0:05:50 And it basically makes it impossible for any of the others to gain share.
0:05:55 So the exciting technological change, of course, is around AI.
0:05:59 The Gen AI search providers are fantastic.
0:06:05 And the market has been dominated by Google for close to 25 years at this point.
0:06:08 And as a monopoly, they’re not really innovating.
0:06:11 They have no incentive to until now.
0:06:17 If you think about the Google experience as it is today, it’s just a long list of links.
0:06:21 And the first few are sponsored, they’re ads.
0:06:26 And I as a user, when I make a query on Google, I then need to make the decision.
0:06:31 I have to sift through the information on that site and find the answer that I’m looking for.
0:06:33 That’s actually a pretty long process.
0:06:37 Instead, with Gen AI, I can just get the answer.
0:06:44 So on chat GBT, perplexity, Claude, or when I’m chatting with my AI friends on character AI or on Poe,
0:06:46 I get an answer immediately.
0:06:50 And that’s just a much, much better experience.
0:06:54 And I hear tons of people saying they’re trying these new tools, but maybe you can give us a sense.
0:06:58 Is this shift really stark? Are people really moving over?
0:07:00 People are definitely moving over.
0:07:02 And I’d say there’s four main reasons.
0:07:07 One is the one we just talked about, the shift from links to answers.
0:07:16 A second one would be how they are personalized, how the answers feel interactive, how they are conversational.
0:07:21 The average query on one of these new services has a follow-up question.
0:07:27 So it’s not just about that initial engagement, it’s about the ongoing conversation.
0:07:33 The third difference is about how these new services engage with complex queries.
0:07:37 So the average query on perplexity is 10 to 11 words.
0:07:41 The average search on Google is two to three keywords.
0:07:47 As you can imagine, their ability to leverage that information and get you what you want is much higher.
0:07:50 On Google, you’re going to get a list of links.
0:07:52 One of those links might have one side of the debate.
0:07:54 Another link may have the other side.
0:07:56 You may never make it to that second side.
0:08:00 On perplexity, on chatGBT, they’re going to synthesize both sides.
0:08:02 They’re going to give me both perspectives.
0:08:06 And then the fourth, they’re just not cluttered with ads.
0:08:09 That may change over time, but putting it all together,
0:08:14 it should be no surprise that these AI-native services are gaining share.
0:08:20 For 10% of consumers, they’re now referring to chatGBT as their search engine of choice.
0:08:25 Perplexity queries are growing 25% to 30% month over month.
0:08:33 And I saw a survey last week that 60% of consumers for purchase decisions in the last 30 days used a chatbot.
0:08:37 All of these services are growing massively.
0:08:41 So if we look at some of the tools that exist today that are coming for that market share,
0:08:47 you take a perplexity or you take a chatGBT, those are also broad-based search engines or chatbots.
0:08:53 But there are these other players you mentioned, Consensus or Habia, for example, that are verticalized.
0:08:57 Does it surprise you that the last wave of search engines weren’t verticalized?
0:09:05 It does not surprise me that the last wave or the next wave will result in a winner-take-most dynamic.
0:09:08 It’s a pretty monopolistic market.
0:09:11 Search is very much a distribution game.
0:09:13 Google has an incredible brand.
0:09:15 They have incredible direct traffic.
0:09:18 They are the default search engine on most browsers.
0:09:22 They are the Kleenex, the Band-Aid of online search.
0:09:28 It’s important to note that search engines definitely benefit from network effects.
0:09:32 More users coming to Google provides more data around preferences.
0:09:37 For Google, that means the next search has more relevance and brings more users.
0:09:44 It also means more users means more advertisers, more profit dollars that you can invest back in the service
0:09:49 or into ensuring that Apple puts them as the default provider.
0:09:50 Absolutely.
0:09:56 And so as we think about this next AI wave, it sounds like you think maybe a similar dynamic might be at play.
0:10:01 Or how do you think about maybe these smaller players actually finding a wedge or differentiating?
0:10:08 One framework for thinking about how search could fragment or verticalize
0:10:13 is just looking at why does vertical software beat horizontal software in some cases.
0:10:21 And it’s typically the vertical requires a specific user interface, proprietary data,
0:10:24 features, workflows, compliance, etc.
0:10:31 And that’s why Viva in life sciences and pharma can beat Salesforce in horizontal CRM.
0:10:36 For the average query on chat, GBT, perplexity or Google, they’re pretty good.
0:10:40 You don’t need a vertical specific application.
0:10:47 But for deep domain research, I can imagine a world in which standalone apps thrive.
0:10:53 In the case of our portfolio company, Hebbia, they have a unique interface.
0:10:58 It looks like a spreadsheet which is native to financial services and their customers.
0:11:04 It brings in public filings, earnings transcripts, but also private data.
0:11:07 It can bring in research, it can bring in survey results.
0:11:14 You can query and then the output can also be specific to that industry.
0:11:20 Not just look like a spreadsheet, but populate a meeting agenda so you can go in immediately prepared.
0:11:25 And so when I think about vertical search or vertical apps, it’s not just about the search,
0:11:27 it’s about everything around the search.
0:11:28 That’s a good distinction.
0:11:33 And as we think about how this maybe continues, the progression of this industry,
0:11:38 if we go back to the last wave of search, we did see a bunch of search engines get traction to start,
0:11:40 maybe Alta Vista or Ask Jeeves.
0:11:41 We all pick on Ask Jeeves.
0:11:45 You know, all the Gen Zers will not know what the hell any of that is,
0:11:49 but they had some traction and then Google exploded and took over.
0:11:53 Do you expect that same kind of consolidation or is this time different?
0:11:59 I think consolidation should be expected in general purpose search because of distribution,
0:12:01 because of network effects.
0:12:07 Google won the last time around in part because of their page rank algorithm.
0:12:09 It produced superior results.
0:12:11 It had a minimalist UI.
0:12:12 It was really simple.
0:12:16 Ask Jeeves, Alta Vista, they had tons of ads.
0:12:19 It was cluttered links on the homepage.
0:12:21 No one wanted to use it.
0:12:27 The irony is that today, what we’re all complaining about with Google is that their pages are cluttered with ads on the results,
0:12:31 and that creates the opportunity for these new search engines.
0:12:34 I would expect consolidation again.
0:12:40 And I think something else that’s interesting is what you’re pointing at with some of these verticalized solutions
0:12:46 is that it’s not just for the everyday consumer, but also for the lawyer or for the academic researcher.
0:12:49 Do you think that is also part of the fragmentation?
0:12:51 Do you expect that to continue?
0:12:56 Search historically has been thought of as a consumer product and for good reason.
0:13:00 I make a lot of searches that are related to my personal lives,
0:13:04 but I use, as do all the professionals you named, search at work.
0:13:09 I probably make more Google searches at work than I do around personal matters.
0:13:14 But there is a category of enterprise search in a market that has existed,
0:13:20 and you can think of that as querying box, Dropbox, Salesforce, all in one.
0:13:23 But I think those two worlds are going to blend together.
0:13:27 Consumer search should not just be limited to what’s on the web,
0:13:31 and enterprise search should not be limited to just proprietary data.
0:13:35 It should have both, and a lot of these AI-native services are working on that.
0:13:40 Yeah, that’s a great point. And as we think about those two models maybe blending together,
0:13:42 we think of consumer search for sure.
0:13:45 It’s always been an ad-based model, or at least today.
0:13:49 Google is this massive economic engine, but no one’s paying a subscription for that.
0:13:53 The new entrants do seem to have a more subscription-based model.
0:13:57 Is this a temporary thing, or how do you see those two dynamics playing?
0:13:59 So the subscriptions could stick around,
0:14:03 but I think this is going to continue to be a digital advertising-focused market.
0:14:07 And I think digital advertising is going to grow because of these AI services.
0:14:10 So these AI services today, they don’t have ads.
0:14:12 I imagine that’s going to change.
0:14:16 It’s been in the news that Perplexity is already talking to advertisers.
0:14:20 They have a high-income, highly educated user base.
0:14:23 That should be attractive to these advertisers.
0:14:28 But as we discussed earlier, the queries on these services, they’re longer.
0:14:32 They’re more complex. They’re more detailed. They’re more personalized.
0:14:37 And because of that, there’s greater intent, which should be more helpful to advertisers
0:14:41 in producing the best results on their end and creating an even larger market.
0:14:46 But in the meantime, these subscription business models, they make a lot of sense.
0:14:50 They bootstrap the business, they cover costs, and from a personal perspective,
0:14:55 I’m very happy to pay $20 a month for chat, GBT, Perplexity, and Poe.
0:14:57 I mean, they provide so much value.
0:14:59 A lot more than $20 a month.
0:15:03 Yes, yes. So obviously, this was your 2025 big idea,
0:15:07 and I know this is going to be a many-year, maybe even decade-long progression.
0:15:09 But what are you paying attention to in the next year?
0:15:11 What opportunities are still on deck?
0:15:15 As I looked at 2025, Google it may be on decline,
0:15:18 but I don’t expect them to go down without a fight.
0:15:21 Google and Meta, I think they can be big players.
0:15:26 Google AI Overviews already has a billion monthly active users.
0:15:30 Meta is not far behind. They have 500 million monthly active users.
0:15:34 Again, this shows the power of distribution for search.
0:15:39 And maybe to that end, Apple, if you could call them a dark horse, is a dark horse.
0:15:46 They are not investing the CapEx like Meta and Google and Microsoft and Amazon,
0:15:49 but they control a central node to the consumer.
0:15:52 And if they wanted to build a search application,
0:15:55 the next day, they could have a billion users.
0:16:00 We just heard how quintessential winning AI is to multi-trillion-dollar companies
0:16:02 like Google and Meta.
0:16:04 But what about nation-states?
0:16:09 In the race for AI dominance, compute has become critical national infrastructure,
0:16:13 but not every country is equipped to compete in that race.
0:16:14 That was…
0:16:18 Manjaneh Mida, I’m a general partner here at A16Z, where I focus on AI infrastructure.
0:16:20 And here’s his big idea.
0:16:23 My big idea is infrastructure independence.
0:16:29 The idea that a lot of countries and regions are starting to realize that modern AI,
0:16:32 deep learning-based AI, generative models,
0:16:36 are a form of what have been called general-purpose technologies.
0:16:43 In the history of humanity, we’ve only had maybe 20 or 22 or so general-purpose technologies.
0:16:48 And these are usually types of technologies like electricity, the printing press,
0:16:52 that have very broad-based applications in society.
0:16:57 Not being largely horizontal economic multipliers and progress multipliers
0:17:01 across a whole set of pillars and domains in society.
0:17:04 There are usually two moments in the adoption of a general-purpose technology
0:17:07 where first countries, nation-states start asking,
0:17:12 “Are we going to welcome this technology or are we going to be hostile to its development?”
0:17:13 Okay.
0:17:17 And that’s the first step that becomes pretty important in a country’s progression
0:17:19 or in a nation-state’s progression or a region’s progression.
0:17:20 Do we even want to adopt it?
0:17:21 Do we allow this in, regardless of whether we own it?
0:17:22 Right.
0:17:23 Do we embrace it or not?
0:17:24 Yeah.
0:17:26 And then the second is, do we build or buy?
0:17:27 Right.
0:17:31 Which is, can we trust somebody else to provide it for us?
0:17:34 We’re well past the stage of, do we embrace it or not?
0:17:37 We’re already well into billions of people around the world now,
0:17:39 having already embraced it.
0:17:41 So the governments don’t really have a choice, so to speak.
0:17:44 So in a sense, AI is already percolated throughout society
0:17:48 at one of the fastest diffusion rates of any general-purpose technology,
0:17:51 because it’s piggybacked off of years and years of digital infrastructure.
0:17:54 And so now the question everybody’s asking is, do we build or buy?
0:17:55 Yeah.
0:17:58 It’s the single largest, probably purchasing decision
0:18:04 that’s going to happen in the next 24 months is, do nation-states start buying it?
0:18:05 Do they build or buy?
0:18:06 Do they build or buy?
0:18:07 Yeah.
0:18:10 I love the parallel of companies because there are many companies that do choose to build,
0:18:12 but also companies that choose to rent or buy.
0:18:13 Right.
0:18:17 So as you think about that, there’s large nations around the world
0:18:20 like the United States, which are clearly building.
0:18:21 Right.
0:18:26 But talk about the argument for smaller nations or the 190 plus
0:18:28 that should be thinking about buying.
0:18:32 The good news here is that we’ve got hundreds of years of human history
0:18:35 to look at for clues about what happens next.
0:18:39 If you’re a small country and you were in the early 1900s,
0:18:44 you were watching the modern electrification of the developed world,
0:18:46 the United States or Europe.
0:18:48 If you chart what happened with many of those countries,
0:18:52 many of them decided to actually enter into what were called joint venture agreements.
0:18:55 It starts with a joint venture with a country that’s at the frontier.
0:18:56 Right.
0:18:58 These countries at the frontier of AI is what I call hyper centers.
0:19:02 These are countries that have the ability to develop, train,
0:19:04 build and host their own frontier models.
0:19:08 I call them hyper centers mostly as an homage to the word hyperscaler.
0:19:09 Right.
0:19:11 Which is that there have been a handful of companies that have had the compute
0:19:13 and the talent to actually build frontier AI.
0:19:17 And now I think what we’re seeing is a shift from just those companies
0:19:20 driving a bunch of frontier AI to countries and regions driving it.
0:19:22 And so if you’re a small country and you’re going,
0:19:29 well, we certainly believe that it’s important to have our own AI infrastructure.
0:19:33 We want to be independent, but we don’t have all the compute
0:19:36 required to train these models or we don’t have all the talent locally.
0:19:39 So what you enter into as a joint venture with a country
0:19:43 or an overseas partner that matches your values.
0:19:46 And this is the really important thing about AI and AI models
0:19:49 and how they’re different from infrastructure like electricity
0:19:52 is there’s a fundamental encoding of human values in AI models
0:19:54 because they’re trained on data.
0:19:55 Yeah.
0:20:00 And the data has these local norms and cultural values embedded.
0:20:05 And so if you happen to train a model on a bunch of internet data collected in the US,
0:20:08 the models are just generally American.
0:20:10 They’re encoded with that.
0:20:14 And if you’re trained the models on data in France,
0:20:19 they actually subtly have a bunch of different values encoded in the models
0:20:21 that reflect those cultural norms.
0:20:23 And so I think step number one, if you’re a small country,
0:20:27 is actually being a little bit crystal clear about which value systems
0:20:30 you align with most out of the hyper centers.
0:20:33 Now it’s not lost in people that the way the internet worked out
0:20:35 was they essentially ended up being two internets, right?
0:20:37 The Chinese internet and the rest of the world.
0:20:38 Yeah.
0:20:40 AI may not end up looking that different.
0:20:41 And if you’re a small country,
0:20:45 what you really have to figure out is whose values align more with yours.
0:20:50 A good historical precedent to look at here is the technology of money.
0:20:52 Money is a pretty general purpose technology.
0:20:57 And what happened in the early 1900s with the modernization of finance
0:21:00 is a number of countries started to ask the same question is,
0:21:02 do we build or buy our own currency?
0:21:04 Do we rely on the dollar?
0:21:05 Right.
0:21:07 Or do we have our own currency?
0:21:09 And that led to the modern day currency regime
0:21:12 where the dollar is a single global reserve currency.
0:21:14 And that happened through a bunch of allied cooperation
0:21:17 where a number of countries realized they did not have
0:21:22 the local resources required to hold the peg of gold right to the dollar.
0:21:26 And so I think what we’re going to end up seeing is the emergence
0:21:29 of very similar to what happened with currency flows,
0:21:32 where you have a couple of large countries
0:21:35 that control their own sovereign currencies, right?
0:21:38 You have the US, you have China, you have India.
0:21:41 And then you have a number of smaller countries that decided
0:21:43 they wanted to be flow points.
0:21:47 So you have Singapore and Ireland and you have Luxembourg and Zurich
0:21:51 that become massive global leaders in modern finance
0:21:56 because they decide they want to ally with one of those power centers.
0:21:58 So if you think about in the AI world,
0:22:00 let’s call regions at the frontier hyper-centers
0:22:02 and then we have compute deserts.
0:22:05 And these are places that have literally no install base
0:22:07 of compute capacity to even be relevant.
0:22:09 All the smaller folks have to figure out
0:22:11 which of the hyper-centers they want to align with
0:22:13 and how do you become a modern day Singapore,
0:22:17 Ireland, Luxembourg, etc. for the world of AI infrastructure.
0:22:19 And so it starts with deciding
0:22:21 whether you want to be a compute desert or not.
0:22:23 And if you’re not and you’re going to actually embrace
0:22:25 AI infrastructure as a government,
0:22:28 I think you’ve got to figure out which hyper-center you want to align with most.
0:22:31 And then it becomes actually quite easy to reason about
0:22:34 how to be a valuable ally.
0:22:35 That’s such a good parallel
0:22:38 because a lot of people think about resources in terms of the farmland that you have,
0:22:40 the people who are working in that economy.
0:22:43 But what you’re pointing out is that countries for a long time
0:22:46 have offered value or offered a resource in other ways.
0:22:48 And as we think about AI,
0:22:51 there’s a few things that you’ve pointed out that countries can invest in,
0:22:53 whether it’s the compute capacity that they have,
0:22:57 the energy resources to power AI and forward thinking policies.
0:22:59 So maybe we can break down each of those.
0:23:01 How do you think about each of those blocks
0:23:04 and how countries should be maybe maneuvering or investing in those things?
0:23:07 The good news is that there’s only three or four ingredients here that really matter.
0:23:09 The first is compute, which we’ve talked about.
0:23:12 The second is abundant and low-cost energy,
0:23:14 which powers the data centers.
0:23:17 The third is data, just the availability of really high quality
0:23:19 tokens for these models to learn on.
0:23:21 And the fourth is regulation.
0:23:22 So that’s the good news.
0:23:25 Now, the bad news is that world is pretty unevenly split up.
0:23:28 Some countries just have dramatically more compute than others.
0:23:32 Others have dramatically more energy than others
0:23:34 because of their natural reserves.
0:23:36 So if you are in the Middle East,
0:23:39 you may not have massive data centers yet,
0:23:42 but what you do have is vast reserves of oil.
0:23:47 And how you translate that into becoming a hypercenter is quite simple.
0:23:49 It’s the law of comparative advantage.
0:23:50 You’ve got energy.
0:23:55 You should use that to attract the world’s best teams and companies
0:24:01 and foundation model labs and so on by trading what you have with what they have.
0:24:07 And so I’m quite bullish on allied ties between countries
0:24:10 that recognize what their strengths are
0:24:12 and then partner with other countries to fill that gap.
0:24:16 And by countries, I mean private companies too from other countries.
0:24:19 One of the things we may end up seeing in the coming years
0:24:21 is jointly trained models between countries.
0:24:24 Basically, I think for most countries, it’s impossible
0:24:27 to have total infrastructure independence at all parts of the stack.
0:24:31 What is much more feasible is to be great at one part of the stack
0:24:35 and then collaborate with another sovereign or another country or region
0:24:41 to achieve joint independence from a value system that you don’t subscribe to, like the CCP.
0:24:46 And so I actually think what’s more important is for countries, regions
0:24:50 and frankly, in some of the world’s largest companies that operate at nation scale
0:24:54 to assess which parts of the stack are critical to them
0:24:56 that they must have independence from.
0:24:59 And the answer there is the function of what asset they already have, right?
0:25:00 It’s their strengths.
0:25:03 And then if they’ve got a critical gap to fill, it’s to go and buy that.
0:25:06 Now, in the long term, you might be able to build things out.
0:25:09 But with infrastructure, especially of this kind,
0:25:12 you can often take years if not a decade long scale.
0:25:16 So as an example, lower down in the stack from the model layer, you have the chip layer.
0:25:19 And even below that, you have the lithography layer, right?
0:25:22 There’s a company in Tallinn called ASML
0:25:25 that builds literally the world’s most important machines.
0:25:26 How many machines do they make per year?
0:25:28 It’s some very small number.
0:25:30 Each machine costs about $200 million.
0:25:34 And I think they did 23 billion in revenue this year,
0:25:37 80% of which came, by the way, from China,
0:25:40 because China was stockpiling ASML machines
0:25:42 before a bunch of export restrictions kicked in.
0:25:44 And they’re the only company that can actually make these new lithography.
0:25:47 They’re the only company that can do UV lithography of this precision.
0:25:51 Now, is it feasible for the US to say we’re going to build our own ASML like tomorrow?
0:25:53 No, I mean, it’s just going to take 10 plus years, right?
0:25:56 UV lithography just takes a really long time.
0:25:59 On the other hand, is it feasible for a smaller country to say
0:26:02 we’re going to train our own local models at the frontier?
0:26:06 That’s a little bit easier to do over the quarters of timescale
0:26:09 if you’ve got a leading research team.
0:26:10 If.
0:26:11 If, and that’s a big if, right?
0:26:14 There’s only a handful really of research teams globally that are capable of this.
0:26:17 And so to answer your question, yes, I don’t think sovereign AI
0:26:21 or infrastructure independence means you have 100% ownership
0:26:24 over every part of the stack that’s infeasible over the short term.
0:26:29 It means that you don’t rely on somebody for a critical part that you don’t trust.
0:26:30 Right.
0:26:32 Can we talk about private companies for a second?
0:26:33 Sure.
0:26:34 Because you’ve brought them up a few times.
0:26:38 How do you think about that dynamic where, as a nation state,
0:26:40 you’re saying, okay, you need this sovereignty.
0:26:41 Right.
0:26:44 But at the same time, can you rely on that sovereignty through the companies
0:26:47 that exist within your nation, just using America as an example?
0:26:48 Right.
0:26:50 Does the government really need to be involved
0:26:54 or can they just let anthropic or open AI kind of command that part of the stack?
0:26:59 Or how do you think about the difference between government versus private enterprise?
0:27:04 The line is pretty stark in a few countries and is more blurry in others.
0:27:06 So in China, the line is very clear.
0:27:10 There’s a law called the PRC 2017 National Intelligence Law
0:27:15 that says Chinese individuals and entities are required
0:27:20 to support PRC national intelligence work by law,
0:27:26 which means if there’s any technology that a PRC company has access to,
0:27:29 they are automatically obliged to make that available to the government.
0:27:31 And that’s not the case in the United States.
0:27:32 Right.
0:27:36 There are some covered types of technology like dual use technology,
0:27:39 like classified defense technology, where if you are developing it,
0:27:42 particularly if you’re funded under a defense program,
0:27:44 then you’re required to make that available to government
0:27:46 because the government’s paying for the development of that technology.
0:27:47 Right.
0:27:49 But by and large, the private sector in the United States
0:27:52 and most other allied countries is by default protected
0:27:54 from having to make its technology available to the government.
0:27:56 It’s not the case in the CCP.
0:27:59 So I think that the question becomes for most countries
0:28:01 is where on that spectrum do you want to exist?
0:28:06 There’s a general framework through which most infrastructure is categorized.
0:28:09 Every country approaches it slightly differently,
0:28:13 but the G5, the Five Eyes, the US, Canada, UK, Australia, New Zealand,
0:28:18 we generally have a joint approach or framework to categorizing this infrastructure.
0:28:25 And by and large, AI models have not been categorized as being dual use
0:28:27 or protected under national security.
0:28:30 The short answer is the history of technology has largely shown
0:28:37 that if you’d like to win, then unlocking the best talents of a country
0:28:41 with as few bureaucratic slowdowns usually ends up winning.
0:28:45 Well, if we think about wanting to keep America at the frontier
0:28:49 and we think about the different layers or ingredients that we talked about earlier,
0:28:54 are there any high risk areas that we think or that you think we’re falling behind?
0:28:59 I think we go back to the four ingredients we talked about earlier of the frontier of AI,
0:29:03 which is compute data, energy, and laws.
0:29:07 Now on the compute front, I think the private market in the United States is doing a pretty good job.
0:29:09 It’s pretty responsive to market demand.
0:29:13 And I think there’s no coincidence that the largest infrastructure businesses
0:29:16 in the United States are chip companies and computing companies,
0:29:21 because I think we’ve generally done a pretty good job of letting the market feed that demand.
0:29:25 I think on the data side, things are extraordinarily tough,
0:29:31 because one, the Biden executive order last year was a starting gun that said,
0:29:34 “Oh, AI is important. Please do something about it,”
0:29:36 and left it to the states to figure it out.
0:29:42 And the states have all taken a complete patchwork of approaches to data regulation.
0:29:48 In 2024 alone, I think there were more than 700 pieces of state-level legislation that were AI-specific,
0:29:53 and a bunch of those laws, if you look at it, are really well-intentioned,
0:29:57 but atrociously implemented ideas for data regulation.
0:29:59 Right, and impossible to adhere to.
0:30:00 Basically impossible to adhere to.
0:30:05 And so I think one area where we’re just handicapping ourselves is that there’s no unified framework
0:30:11 in the United States at the federal level yet for data, especially around training.
0:30:13 And I think we needed that yesterday.
0:30:19 Overseas in a number of countries where rule of law, especially on copyright and IP and so on,
0:30:24 is just less stringent, those labs are happy to just race ahead,
0:30:28 whereas our companies here are trying to figure out what they should even comply with,
0:30:31 and that greatly hurts you more than actually a laissez-faire approach.
0:30:33 I think our companies would be totally fine.
0:30:35 The best founders at the frontier of AI would be,
0:30:37 find the United States being compliant.
0:30:39 They just want to be told what to comply with.
0:30:43 Not across 50 different states with different regulations that are changing and unclear,
0:30:44 and in some cases, impossible.
0:30:47 Right, and there’s also the fundamental scientific problem that
0:30:50 they’re just very real data walls that these models run into.
0:30:54 And I do think one of the things that hurts frontier research in the United States
0:31:00 and allied countries is a lack of government support in collaborating across borders
0:31:03 to make more data available to allied regions.
0:31:04 So that’s number two.
0:31:08 On energy, I think we’ve obviously hamstrung ourselves in the United States with nuclear.
0:31:11 France, for example, is an embrace of nuclear 20 years ago,
0:31:14 has positioned them to have extraordinarily efficient data centers today.
0:31:15 Whereas in the United States,
0:31:17 I think we’ve basically shot ourselves in the footer on that.
0:31:20 And then lastly, I think around inference regulation,
0:31:26 what we’re not doing enough of is making it clear who the liability rests on.
0:31:30 I’ve seen a number of proposals ahead of legislative sessions next year
0:31:36 that want to hold model developers liable for the outputs of the inference,
0:31:38 even if the misuse is being done by somebody else.
0:31:39 And what does that do?
0:31:42 That drives those very important developers elsewhere.
0:31:49 Essentially forces most startups to lose much needed ground to big tech companies
0:31:52 and that entrenches incumbents more.
0:31:55 So as we think about 2025, whether it’s in the US or elsewhere,
0:31:59 because I mean, this idea really is truly global.
0:32:00 What are you looking out for?
0:32:04 Or what should maybe let’s say a legislator or let’s say the head of a nation,
0:32:06 what should they be thinking about?
0:32:08 And what are you looking out for in some of those decisions?
0:32:13 Are you looking for countries that are buying GPUs or building out new energy centers?
0:32:14 What are you paying attention to?
0:32:16 The leading indicator is definitely compute.
0:32:21 If you think about the AI supply chain, the first mile starts at the data center.
0:32:24 That’s the new atomic unit of sovereignty, I would say, which is a new thing.
0:32:28 We’ve never actually had nation states think about atomic units of an AI data center
0:32:30 as a thing that countries should be purchasing.
0:32:36 And I think about 24 months ago, we started seeing nations reason about that first mile as being important.
0:32:42 So you’ve had an enormous amount of Nvidia’s purchasing orders come from the balance sheet of governments.
0:32:48 Just unprecedented demand they’ve been seeing from nation states realizing that they want to be hyper centers.
0:32:55 And that starts with them placing orders 12 to 36 months in advance to take delivery of GPUs
0:32:58 because if you don’t get in front of that line, it’s over.
0:33:00 You’re getting it after everybody else.
0:33:02 So that was step one.
0:33:05 The second thing I look for is founders who are deeply both technical,
0:33:11 who often come from deep research backgrounds and scientists who’ve led frontier model development already,
0:33:13 often inside of large hyperscaler labs.
0:33:16 So an example is Arthur Mench who started Mistral.
0:33:21 They worked at DeepMind or Guillaume Lomp who led the initial Lama family at Metta,
0:33:27 who are deeply mission led and believe that they can help solve a bunch of these infrastructure problems for the world’s largest governments.
0:33:33 So there’s a new class of founder who’s both primarily technical and has their training in academia,
0:33:38 but is motivated to solve all the really hard problems that come with having to deliver,
0:33:44 solve a bunch of these infrastructure problems for really large nation states and regions.
0:33:51 But I think if you’re a founder like that who has the guts, then your impact on humanity ends up being quite generational.
0:33:56 But now let us convince you that the future of AI may not be so straightforward.
0:34:02 Instead of models running in the cloud, perhaps you’re bound for a future where many more applications will run on device.
0:34:07 I expect smaller on-device AI models to dominate in terms of volume and usage.
0:34:13 This trend will be driven by use cases as well as economic, practical and privacy considerations.
0:34:17 That was Jennifer Lee, I’m a general partner on infrastructure team.
0:34:19 Here’s her big idea.
0:34:25 My big idea is on-device and smaller generative AI models will become more popular in the next year.
0:34:30 If you’re a frequent user of Uber, Instacart, Lyft, Airbnb applications,
0:34:33 I’m sure there are many, many machine learning models already running on your device.
0:34:41 Very easily when you load up an Uber screen, it’s 100 models that’s coordinating routes and giving you a real-time price.
0:34:45 What I’m more referring to is the generative models that are creating.
0:34:52 Image, voice, video will become more prevalent in the same way to run on device and within your applications,
0:34:55 similar to these other traditional machine learning models.
0:34:58 The models that we’ve seen in the last few years do take a lot of compute.
0:35:04 Can you square that with how much compute we can get from something like a smartphone and also these models,
0:35:07 whether they’re getting smaller or how this kind of comes together?
0:35:10 Yeah, first, never underestimate the compute power.
0:35:16 Our smartphone is probably as powerful as a computer 10 or 20 years ago, thanks to moral law.
0:35:23 At the same time, the models are for especially smaller sizes of two billion, eight billion parameter models.
0:35:29 That’s enough compute for them to run on device and it can generate and create very robust experience already,
0:35:31 be it text or image or audio.
0:35:37 And some of these models, if they’re diffusion models, they’re intrinsically smaller than large text models to be very capable.
0:35:44 And there’s another new set of tooling and also technology developed around distillation is if you have a very powerful large model,
0:35:51 can be distilled to a smaller parameter size model and still maintain a lot of the capabilities the large model contains.
0:35:56 So both on the infrastructure side and also on the device compute power side,
0:36:00 it’s a perfect setup for the smaller models to be more popular.
0:36:02 Totally. So I’m hearing a few things.
0:36:05 I’m hearing that the smartphones are becoming more powerful.
0:36:10 Some of these models are becoming more efficient, but that kind of brings us to the question of why?
0:36:12 So why would we want to run these models on device?
0:36:15 What are the advantages of that and also the disadvantages?
0:36:22 As consumers and day-to-day users, we’re already spoiled by real-time and very performant applications.
0:36:25 If you’re talking to a chatbot, if you’re talking to a conversational AI,
0:36:30 if you’re adding filters to your video and images on Instagram or TikTok,
0:36:33 you don’t want to wait for multiple seconds to load a new filter.
0:36:37 You don’t want to wait for multiple seconds for the chatbot to respond to you.
0:36:42 Those are many real use cases that can really delight and improve user experience.
0:36:44 Also optimization for compute.
0:36:50 There are a lot of harder, more complex questions or video processing that requires going into the cloud.
0:36:58 But largely, if it’s again changing user experiences and improving the visual and sound effect of things,
0:37:02 it doesn’t have to route through multiple servers going through a network.
0:37:10 So both from a user experience and efficiency perspective, it’s a much better design to run some of the models on device.
0:37:13 And then the last part is just privacy.
0:37:16 Users do care about if my meeting notice is taken locally,
0:37:24 I probably would use this meeting to take it much more often than knowing some of the data is being sent to a server at their processing.
0:37:30 A lot of my private conversations, so it depends on the use case again for the application.
0:37:32 I think that also improves that option.
0:37:37 Absolutely, and that has my wheelspinning for sure in terms of maybe this unlocks new applications.
0:37:41 So on that note, you mentioned a few already, but where might we see applications pop up
0:37:45 or where perhaps are we already seeing applications with these on-device models?
0:37:48 First come to mind is real-time voice agents.
0:37:52 It’s a very popular topic and it’s something I’m very excited about.
0:37:55 We invested in and work very closely with this company called Eleven Labs,
0:38:02 and that’s one of the areas they’re spending also a lot of efforts on is not just having the human like synthetic voice
0:38:08 but being able to handle conversations fluently with end users and to get the latency down
0:38:14 and also to think about what type of real-time exchanges you want to have with your AI companion,
0:38:18 your support agents, or any sort of life coach.
0:38:26 I think we do need to think about the modality and the latency in a much more, I guess, improved fashion.
0:38:34 So I won’t be surprised if some of those inference workloads are running locally coming into the next 12, 18 months.
0:38:39 Absolutely, and as we think about how maybe these different models also interact with other parts of a smartphone,
0:38:44 let’s say the camera, do you expect this to also maybe change user behavior and what we can do?
0:38:52 100% you can actually re-imagine the AR experience of if I point a camera to this room
0:39:00 and I want to see a new surface and wallpaper and furniture, the technology is already there.
0:39:07 We can actually leverage both generative AI and the camera and also prompting interaction
0:39:12 to create new experiences already of how we interact with real physical life.
0:39:17 And that’s where I also think a lot of on-device models will play a big role
0:39:21 of how to interact with the 3D world, how to interact with physical world,
0:39:25 and not just using the camera for capture but also using it as a projector.
0:39:32 Definitely, and let me ask you about economics then because a lot of the models that exist today do rely on inference
0:39:35 and sending that inference up to the cloud and that costs money.
0:39:42 Do the economics change if you all of a sudden have these models running on device on the smartphone compute that already exists?
0:39:47 Do the economics actually shift or can we come up with new ways of monetizing in this new world?
0:39:51 Yeah, it’s a great question and I honestly don’t really have the answer
0:39:56 because even for larger models, the inference price has been dropping really significantly.
0:40:00 Further optimizations to be done if it’s a very workload-intensive compute,
0:40:04 let’s say using your computer or phone, I think we’ll still have economic benefits
0:40:12 but I don’t think it’s a very direct answer of it’s going to substantially reduce infrastructure costs for some of these applications.
0:40:16 But architecting and structuring sort of the whole tool chain,
0:40:22 it does change sort of economics on the developer efficiency and sort of iteration speed.
0:40:28 There is pros and cons when shipping in the cloud where it can launch more continuously on device
0:40:34 as its own challenges because you’ll have to go with the updates with the application and with hardware.
0:40:42 So there is that side of economics that I think will have impact from how teams are being structured in launching models in a hybrid mode.
0:40:47 So I would encourage teams who are thinking of leveraging this technology consider it more holistically.
0:40:52 Super interesting. And as we think about that world, are there any players that you think really succeed here?
0:40:59 Like in one sense I could see maybe the phone manufacturers, I could also see maybe the manufacturers of wearables
0:41:05 being able to introduce all kinds of new applications, think of wearing an Apple watch, Fitbit, Woop, things like that.
0:41:11 Is it Nvidia that benefits in some way? Who do you think actually benefits from this idea of the models becoming more efficient
0:41:15 and these on-device models becoming a thing?
0:41:24 Right now I’ve seen more interest and enthusiasm from the hardware development side, whether it’s chips, it’s the phone makers.
0:41:34 I do think there’s also a lot of interest from the model developers as well of just like proliferating the model adoption across different setups and devices.
0:41:38 But I think over the long run it’s probably going to impact the whole supply chain.
0:41:44 We talked about some of these macro trends throughout. How do you specifically see those trends shaping up in 2025?
0:41:47 And is there anything in particular that you’re putting your eye toward?
0:41:54 This will sound more like a consumer investor, I’ve been like a hardcore investor, but I am very excited about the mixed reality
0:42:01 where generative models, 3D models, video models that really again makes the reality of what we’re seeing today
0:42:04 and through the camera lens, through the microphones.
0:42:10 Much more creative world even when sitting at home or when going on the ride.
0:42:13 That’s the type of experience I’m very much looking forward to.
0:42:18 I think the foundation model technology is pretty mature, the infrastructure is getting ready.
0:42:22 So I’m personally very excited about sort of the new consumer experience.
0:42:28 All right, I hope these big ideas got you geared up and ready for 2025.
0:42:30 Stay tuned for parts 3 and 4.
0:42:38 And again, if you’d like to see the full list of 50 big ideas, head on over to ASXMUSY.com/bigideas.
0:42:39 It’s time to build.
0:42:41 [MUSIC PLAYING]
0:42:43 (soft music)
0:42:46 (gentle music)