E39: Navigating the AI Supercycle with Will Summerlin of Autopilot

0:02Welcome back to Turpentine VC, a podcast where we discuss the art and science of building successful venture firms, VC to VC. Today's episode is with Will Summerlin, Managing Partner at Autopilot. In this conversation, we cover AI's impact on various industries, the importance of strategic investments in potentially monopolistic markets, and the evolution of AI models. Will is also a host at Turpentine with the Autopilot podcast. We'll include a link in the description so you can check out his interviews with AI founders backed by Benchmark, Greylock, YC, and more. will welcome to the podcast thanks so much for joining yeah thanks for having me eric excited to talk about autopilot let's get into the new fund first just break it down for us what what is the thesis of the new fund why this and why now yeah it's a great question now i think if you look at returns in venture capital they often follow the start of a super cycle.

1:02We saw this with cloud and companies like Twilio and CrowdStrike. We saw this with mobile and companies like Uber. And I think we're definitely at the start of a new super cycle with AI. And our thesis, if you go back and zoom out to a 30 ,000 foot view, we're getting to a point where AI can do cognitive tasks that humans do not only faster and cheaper, but oftentimes a lot better. I think the origin of this and what caught most people's attention, if you go back several years was AlphaGo and DeepMind's program beating the best Go player in the world. And it was a really hard game for computers to learn and compete in.

1:41But the way that it beat the human was really interesting. It didn't just play a human strategy. It basically realized humans have been playing in this small box of possible strategies. There's this giant box outside of that of potential solutions, of potential ways you can play the game. and it wouldn't explore all these different paths that humans had never thought of. And so it was able to execute that strategy and be the best go player in the world. And now we're seeing that across a lot of other domains. If you look at Tesla's full self-driving capabilities, it's 10 times safer than a human driver, and it's only getting better.

2:14I can go through, continue on down the list of all these ways that AI is better than humans at many complex tasks. But if you think about the opportunity set, we currently pay humans $32 trillion a year to write software code, to review contracts, to do these tasks that AI can increasingly automate. And so I think we're at the start of a new super cycle where the technology, if we paused where it is today and it got no better, it would still be an incredible super cycle. If you look at accounting, for example, we looked at an analysis, about 50 % of common accounting workflows can be automated with technology as it is today.

2:48But it's not going to stay where it is. It's going to continue proving at an exponential rate. So if you think about the opportunity set within the next decade, I think it is going to be perhaps the most value creating super cycle and the history of modern society. And so based on that, it's a pretty good time to launch a new firm and be investing in startups. Sam Lesson came on this podcast and he talked about why he is, well, he's bullish on AI. He's bearish on it for startup investment opportunities, especially at Seed, because he says incumbents will take on a lot of the advantages. It's going to make the players with data have much more capabilities and be able to use that data.

3:27And then also the foundation models themselves. And it's kind of too late to invest in them or too expensive to invest in them. And then just more broadly, everything is priced very highly. What say you to that? I actually agree in a lot of ways. So I was at a firm called Arc Invest prior to launching Autopilot, Kathy Wood's firm. And ARK was among the largest shareholders and a lot of big companies like UiPath and Square and Twilio and got to know these businesses really well. And especially the founder-led ones, they know the opportunity. They have the capital to invest and compute, right? And we've seen scaling laws.

4:07I mean, the larger a model is, the more compute you use to train it, the better it performs. If you're a startup, investing$100 million in a GPU cluster is pretty cost prohibitive. But if you're a big company that's doing three, five billion year revenue, you can make that investment. They also have proprietary data, to your point, right? Not only the data to train these models, but also the feedback loop. They have instant scale. So if they roll this out to 100 ,000 customers or 100 ,000 users, they're automatically instantly capturing reinforcement learning with human feedback from 100 ,000 people.

4:36So the incumbents really do have the opportunity to win. And in most markets, they should win. I think it really comes down to a market-by-market discussion of how aggressive, innovative, and competent are the incumbents. Would I compete directly against CrowdStrike? Probably not. Would I compete directly against Tesla? Probably not. But there are a lot of markets that AI is going to touch and transform where the incumbents are pretty stale. They're not innovative companies. They don't know what cloud is. So if you look at, for example, manufacturing, and you look at a subset of manufacturing called tool and die making, where companies basically build the molds and the dyes that go into a factory to produce bumpers and lots of other parts, that's a really stale industry.

5:25Their technology is probably 30, 40 years old. And they have some software, but it's designed in the late 90s. And so are those incumbents that don't understand technology going to all of a sudden innovate and hire a bunch of people out of DeepMind? Probably not, right? And so I think there are a lot of these stale markets where the advantage really does go to startups who can think from first principles, who can recruit great talent, and who can build AI native solutions. And so for us, we think that innovative incumbents are in the advantage position. And a lot of these markets, I think that people see as obvious, like AI co-pilots for sales or an AI accounting startup are probably going to be pretty crowded on the startup side and also be pretty competitive from the incumbent side.

6:08But if you go look at markets like tool and die making, the question around startup versus incumbent, I think, becomes much more clear where the startups definitely have an advantage. Yeah, that's well said. Let's dive into that a bit further. You shared some markets that you're somewhat bearish on for new AI startups because incumbents are strong, like security with CrowdStrike or self-driving cars with Tesla. But then you also said, you know, some spaces will be crowded like AI sales and accounting, but use some opportunities in subsectors of manufacturing. Why don't you flesh out, let's go into more markets where you're bearish and more where you're bullish.

6:55And I also want to plug your great podcast that you do with us, Autopilot, that gets into some of these spaces as well. Yeah, it's a great question. So I think there are two ways to look at this. The first is what does the competitive landscape look like today and what will it look like five years from now? And I think I sort of believe like the Peter Thiel-ism of competitions for losers. And if you look at a market like AI co-pilots for sales, really useful technology. I can very clearly see the value and I think it's going to grow very quickly that market. But there are quite literally a thousand startups and a dozen innovative incumbents competing for the same dollar.

7:39And so when you have that sort of dynamic, your customer acquisition cost goes up and your pricing power really starts to erode. But if you look at a market like injection molding and tool and dye making, the incumbents, going back to what we were talking about a minute ago, they're not very innovative. And there are one or two interesting startups in that market because it is really complex and it's hard to underwrite. And it's a much harder problem to solve and a much harder technology stack to build than just building an AI co-pilot for sales that's going to grow 500 % year over year out of the gate.

8:12So I think that's criteria number one is what does competition look like today and what is it going to look like in five years? Criteria number two is pricing power. And the most valuable companies on earth have structural competitive advantages that compound with scale. And you can go back in time and look at standard oil, right? Standard oil benefited from the classic economy of scale, where they had all these advantages so that as they grew and got bigger, their cost got lower. And so they got to a point where J.D. Rockefeller could sell oil profitably at a price point that would put all of his competitors out of business.

8:47His competitors, say the arbitrary number of a dollar a gallon, that was profitable for him because his costs were so much lower because of his scale. But the competitors would be losing money at a dollar per gallon. And so he had economies of scale, right? It created a monopolistic outcome. If you look at Meta, right, the Facebook ecosystem, they have network effects, right? So the more people that they add to the network, the more value the network creates for everyone on the network. And so you have this like compounding mode. And so I think we look for companies that can essentially monopolize an industry and develop this pricing power over time.

9:19And so whether that's network effects or economies of scale, that's great. I think with AI, you have a new form of a compounding mode and reinforcement learning with human feedback, where the more feedback you get from usage of your models, the better you can tune those models and train the next generation. The better your models, the more people you attract or people you attract, the better the models become. And so I think you do have this new kind of compounding mode that we see in some cases with AI, but those are the two criteria we look for. So in a market, is it not very competitive and can one company essentially monopolize it by developing some structural competitive advantage?

9:52So that's the framing. And I guess to give a couple examples of where we see opportunities and counterexamples of where we don't, I think AI co-pilot for sales is an example of an area that we don't see a lot of opportunity. I think horizontal writing assistance is another, where we've seen a lot of startups that essentially are building wrappers on top of GPT-4 and helping you write paragraphs or write marketing copy. Well, guess what? There are hundreds of other startups doing that. They all have access to the same underlying technology. And now Microsoft, Google, Notion, every word processor also is embedding those features.

10:27So those are markets that don't really meet the criteria. I think where we get more excited, we mentioned injection molding as an example. I think that's an area where there's very clearly a need for automation. And so the labor force in tool and die making has declined 50 % in the last 25 years while the market's grown at an 8 % hacker. So you really need automation. It's clear that there's not a lot of competition. The incumbents aren't going to do it. And there are a couple of startups that are interesting. And you can see how economies of scale take hold here. And the company that wins develops pricing power over time and has sort of a self-fulfilling flywheel.

11:02So I think for us, that's really interesting. We're spending a lot of time in industries that are maybe today a little bit more services driven. So like revenue cycle management within healthcare, that's a huge industry. And right now it's pretty much done by, all the work is done by humans. And we've done some work here and probably 70 % of what those humans do today can be automated with off-the-shelf AI technology. So you can see how there's very clearly a story there and the incumbents there run services businesses. They don't know how to build technology. That's an example of an area we're really excited about.

11:36And I keep going down the list, but I think we're looking for more of these stale industries where the incumbents can't compete. There are a couple of startups and you could develop pricing power over time. Yeah. I love that framing. Maybe give us one or two more, or maybe I'd position it this way. Do you have a request for startups or in the sense of if you could guide really talented entrepreneurs with the right set of skills and right set of relationships to work in any specific space or spaces? What's another space where you see opportunity that you'd want a founder to go pursue and then call you?

12:13Yeah, that's a great question. I think we're getting really excited about things that touch the physical world. If you look at the economy, the global enterprise software market today is a trillion dollars. So companies spend a trillion dollars on software. And a lot of the AI startups that we see are competing for some IT budget. They're selling software to incumbents. And I think that's fine. There's going to be some opportunities there. Some of those markets will be too crowded. Some of those markets will find pockets of opportunity. But I think one of the more interesting strategies we're seeing in AI is not to sell software to the incumbents, it's to vertically integrate AI and compete with the incumbents.

12:50So we talked about revenue cycle management. Instead of trying to sell AI to those legacy companies that don't understand technology, just build a new revenue cycle management business from the ground up and make it AI native. right? And you'll be able to run it at probably 60 % gross margins, not 85 % gross margins, but you can build a really big business. And I think this is sort of AI kind of shifts the opportunity set here. And a lot of investors, I think still are dogmatic about only investing in software and like afraid of services, which I think is crazy. I think a lot of these markets, you can really attack and to some degree monopolize if you're willing to build a business that has some services component, but really relies on AI to automate most of the work that would otherwise be done by humans.

13:34So I guess my request for startups is if you're a great entrepreneur and you're looking at the opportunity set on the horizon, look at these industries that are relatively stale, that are services heavy, where you can use AI to automate a lot of what humans do today. And don't be afraid to vertically integrate and compete with the incumbents. Yeah, well said. More broadly, what's your take on the broader sort of the wrapper debate in terms of where will value be accrued to the foundation models or where independent startups will be allowed to sort of prosper and not get disrupted, not only by incumbents, but by the foundation models themselves?

14:12What's the right way of thinking about that? Hey, we'll continue our interview in a moment after a word from our sponsor. I think LLMs specifically are going through some level of commoditization, right? And we've seen this like Lama 3. It's almost as good as GPT-4, but your inference costs are going to be 10 times lower if you're using an open source model like Lama 3. And so is it good enough for most use cases? Probably. I think what we're also seeing is a lot of companies are opting to train smaller proprietary models for very narrow use cases. So if you're trying to build an open-ended chat bot that can answer any question about any subject, you probably need to use a big model like GPT-4 or LAMA-3.

14:55But if you're trying to solve a very narrow subset of tasks, you don't need GPT-4 or LAMA-3. It's sort of like putting 1 ,000 horsepower in a golf cart. It's kind of unnecessary. And so if you end up training a smaller proprietary model that's more domain-specific, your quality of answers will be the same, maybe a little bit better, but your inference costs are going to be a lot lower, like 100 times lower in some cases. And so instead of spending a dollar for an answer, you'll spend a penny for an answer. And I think that really matters as you get to scale. The inference speed can also be a lot better.

15:32And so instead of waiting a full second for a response, you get a response in milliseconds. So we're starting to see companies train these smaller proprietary models that are domain-specific. So I think those two things together, the open source phenomenon combined with this trend towards domain-specific models, is going to create some headwinds for the really big closed source models. I think there's an argument that that changes a little bit as we move to multimodal models and eventually AI agents. And we talked about scaling laws for a second a while ago. So as we move towards like AI agents and multimodal models, there are really three things that matter.

16:12There are first is talent, right? You need smart researchers who can actually train these models. The second is compute. So scaling laws hold that the more compute and more data you use to train a model, the better it performs across a wider range of tasks. And the third is data, right? You need some level of proprietary data, especially as we move into like multimodal models where you need a lot of vision data, a lot of audio data. So I think as we see the shift from LLMs to multimodal models and eventually to AI agents, there's an argument that the closed source model providers will have some advantages.

16:45And if you're meta, maybe you can spend$100 billion training a model and open source it, but meta might be the only company in the world that can do that. And so I think if their priorities shift in any way, I think it creates an opportunity for the foundation model providers as we move from LLMs to multimodal models and agents. Yeah. That makes sense. Have you looked into any of the companies trying to be sort of dominant, like Lindy AI or any of the companies trying to be agents directly? Yeah. I don't think we're there yet. The opportunity is incredible. I think one way to look at this is like, look at UiPath.

17:25UiPath has built an incredible business with robotic process automation, which is sort of like the early version of what an AI agent looks like, right? They have technology that can automate pretty simple processes, like taking data from an application that runs on a mainframe and moving it into a modern SaaS application, doing some screen scraping and other techniques. And that version of RPA is somewhat cumbersome to implement. It could be expensive to implement. It can break pretty often. but they went from one to 600 million in revenue in five years with this technology, which is incredible.

18:02And so I think that shows the demand for AI agent-like capabilities. And as we move from where we are today to functional AI agents in the next couple of years, I think that they're going to be a lot easier to implement than RPA. They're going to be able to do a much wider range of tasks than RPA, and they're going to break a lot less frequently than RPA. And so I think if we look at that one to 600 million growth trajectory in five years, that's really just scratching the surface on the opportunity for AI agents. But I think given what we just talked about, it's really going to come down to proprietary data and compute, as well as some level of talent.

18:38And I think if you're a startup, it's going to be challenging to compete with the big guys, OpenAI, Meta, XAI, who are going to have access to probably a lot more data, but certainly a lot more compute. And so unless you're raising billions of dollars out of the gate to go build giant GPU clusters and find ways to acquire a lot of proprietary data, it's going to be difficult. You're going to be facing an uphill battle. If you could have done OpenAI or Anthropic a round or two ago, are you long-term bullish on these companies that you could have developed the conviction there or unclear? Yeah. I mean, look, at ARK, we invested in Anthropic pre-revenue.

19:24And that investment's done very well, right? And I think the company's crushed it. I think they've grown from zero to 350 million in ARR in 18 months, which is insane. And so they're crushing it, really smart team. And they're focusing more on the enterprise. And so what they've realized is a lot of enterprise customers, they want to fine-tune their models. So they want a base model like Claude2, and they want to fine-tune it on their own enterprise data and sort of host that within their own enterprise environment. So I think Anthropic has found an interesting niche within the enterprise. Open AI, their structure scares me.

20:06Like, you know, we saw this governance issue that happened earlier this year, and there's still a lot of unresolved questions around that. So I think they've crushed it. Like Sam Altman's obviously, you know, an end of one founder or CEO. And I think there are a lot of reasons to be excited about what they're doing. And I think they're also aggressively pushing into multimodal models and AI agents. And so whether or not LLMs become commoditized could be less relevant five years from now for OpenAI's story. But I think given some of the other structural challenges there, I think we consciously sort of looked at that and said, the risk reward here maybe doesn't make as much sense.

20:44Yeah, that makes sense. Is it fair to summarize your AI thesis in the sense that it will both make the most skilled people much higher leverage, but at the same time, it will replace labor for a large percentage of people and thus make businesses just much more effective and productive, and especially in industries that don't have that much of it, there's a massive opportunity there? Yeah. I think going back to history, we talked about this just before we started recording. It's really hard to predict the future, but often the future rhymes with the past. And I think we can go back in history and try to draw parallels to what's happening today.

21:29Going back to the Industrial Revolution, the tractor increased the productivity of farmhands 17-fold. And you would think like, okay, well, farms will just hire 17 times fewer people. But that's not what happened. The individual became 17 times more productive. And humans are creative and as a capitalist society, find ways to make that productivity boost, generate more and more revenue and eventually more profit. So what happened is the wage rate for those people went up because they were able to produce 17 times more goods or 17 times more harvest in an hour. I think we're seeing the same thing happen a little bit here.

22:07If we look at software development as kind of an early indicator, we're already seeing AI make software developers five times more productive. And so I think if you go to most companies, think about a startup, if you have developers that are five times more productive, are you going to hire five times fewer developers? Or are you going to write five times more product or develop products five times more quickly? And I think there is this trade-off or this question of like, is the job category a cost center or a revenue center, right? And I think you can argue sales, for example, is certainly a revenue center.

22:44Like if your salespeople become five times more productive, you're not going to fire 80 % of them. You're just going to generate five times more revenue, which is awesome. I think the same could be true with like software developers to some degree. They write product faster that allows you to unlock more pockets of revenue more quickly and compete with others more effectively. But then I think there are cost centers. If you look at compliance within a company or you look at call centers, they're purely cost centers. And if you can reduce your costs by 80 % by automating most of what humans do today, that's what you're going to do.

23:19So I think we will see some job categories like customer support probably shrink. But I think we're going to see these revenue centers just become a lot more productive. And I don't think people will get laid off. I think people will actually end up earning higher wages because of their productivity increase. That's well articulated. I want to zoom out and have you described the more fund mechanics and fund strategy in terms of how have you thought about where to play? Because it's much earlier than you're playing at ARK. ARK was much wider, of course. And so how you thought about sort of that narrowing to some degree?

23:57How do you thought about fund size, et cetera? When do you get into the sort of mechanics of autopilot? Yeah, for sure. So I can't talk about our fundraising plans. I've been slapped by our lawyers for doing that. So I can't talk about active fundraising plans, but I can't talk more broadly about where we see the opportunity set. Yeah, at Ork, we had a lot of flexibility, which is great. We were in the public markets investing in the most innovative public companies. On the private side, we were investing from pre-seed to pre-IPO, so stage agnostic, and really got to flex into where we saw the opportunity set.

24:32I think that was great and had a lot of advantages. For autopilot and looking at our skillset, where we see the opportunities, we're really narrowing in on series A and series B. We think that the pre-seed and seed market has gotten really crowded. There were 5 ,000 sub-hundred million dollar funds raised in the last five years, which is crazy. So there's this flood of capital at pre-seed and seed. And the Series A, Series B funds haven't kept pace. We've seen these mega-multi-stage funds like Andreessen just raised, and that's great, right? And that brings more capital to the ecosystem. But I think boots on the ground, we do see a lot of opportunity at Series A and Series B.

25:12From a valuation standpoint for AI specifically, we see a lot of seed companies raising it kind of five on 25 with no product, no revenue, with just an idea. And then they'll turn around and raise a series A at maybe 70 or 100, but they're already generating two or 3 million in ARR. And the loss ratios at series A and series B are much lower than pre-seed and seed. So there's a significant de-risking that happens in that step up. And I think that to us is more attractive right now. I also think given our skillset, we're predominantly public market investors and did a lot of growth pre-IPO stuff.