The Untrainable AI Moat Is Not What You Think | Conviction Vs. Discipline | SpaceX's $23T Lesson

This is the lesson I think VCs always relearn. We relearned it with Google. We relearned it with Facebook. I think we've relearned it with NVIDIA. The big companies end up being much bigger than you think because usually there's some combination of tailwind and timing that pushes them to a place that was hard to imagine before. And I think so long as that's the case, there will be companies that are worth paying for. But I don't think that means that on average price and ownership don't matter. You know, we only have one SpaceX. We have one Anthropic and one OpenAI. I just don't know that you should just throw all of the playbook out of the window.

It's the week of June 22nd. Welcome to The Learning Corner, a Precursor Ventures podcast. Welcome to The Learning Corner by Precursor. I'm Mia Farnam, and I'm joined by our GP and managing partner, Charles Hudson, where we will be talking through the articles and Goodreads that caught our team's attention over the past week. Our first piece this week is from Hallie Teko's Substack, Massively Better Healthcare, and it explores one of the more contested questions in venture right now. Does ownership discipline still matter when AI is enabling companies to grow faster and raise at higher valuations than ever before. The piece maps out a really clear split between two camps of healthcare investors. The first are the convicted. They've essentially retired ownership thresholds as a binding constraint, underwriting each deal on its own terms and stretching if the conviction is strong enough. The second bucket is the disciplined, and they argue that when markets get frothy, the ownership bar matters more, not less, and they're sourcing earlier and writing bigger checks to hold the line. Where both camps agree is that no one is bargain hunting. Everyone believes this AI moment is structurally different from prior bubbles, and the clearest signal of a durable business is the multi-year enterprise contract in healthcare. That's never changed. The open question here neither side can fully answer yet is whether healthcare's structural realities, the longer sales cycle, the complex regulations, finite market sizes, will ultimately bound AI's upside in ways that make ownership discipline matter more here than anywhere else. Obviously, this piece focuses on the healthcare of it all, and it's just a different industry than I think what we see in the media when we see companies race to $100 million in ARR so quickly. They're typically selling to more traditional B2B customers with much larger budgets and room for experimentation than customers in healthcare, who you and I both know from lived experience investing in those places, where you're typically floating maybe a single-digit to negative margin if you're on the health system side. You're a payer, there's just longer cycles. So maybe zooming out a bit more broadly, we can talk about the healthcare of it all, of, of what this dynamic looks like for that specific industry. But then how does this dynamic of the convicted versus the disciplined, does that take a different form in different industries as it pertains to AI and the pricing that we've been seeing?

I think software has always had people in the like price doesn't matter crowd and that you should just be in the hot company. I feel like healthcare as someone who doesn't invest in as much healthcare as others, I've always felt like it was more of a monolithic community. It was like, hey, this stuff just moves slowly because it's regulated and because the big players can move slowly, that there were more guardrails on what you could pay, how much you need to own, and the scale of outcomes. And I think AI has changed the math for those folks too. Maybe not as much on the realized exit side like we've seen on SaaS, but I think on people's perception. I think this is the question in venture and I read through it and I was like, the place where I've landed on all this is if the outcomes are really going to be bigger, and it appears that they will be at the extremes, if you find a truly great company, it's probably worth more than you think. This is the lesson I think VCs always relearn. We relearned it with Google. We relearned it with Facebook. We, I think we've relearned it with NVIDIA. The big companies end up being much bigger than you think, because usually there's some combination of tailwind and timing that pushes them to a place that was like hard to imagine before.. And so I think so long as that's the case, there will be companies that are worth paying for. But I don't think that means that on average price and ownership don't matter. You know, we only have one SpaceX, we have one Anthropic and one OpenAI. I just don't know that you should just throw all of the playbook out of the window. And I think what the Scrub Capital folks said in the article makes sense to me, which is like for every deal, and what Hunter said too, for every deal you have to look at the combination of entry price, market size, team, ambition, all of these things and figure out, do I think this one's gonna pencil for me? Yes or no? Yeah. And I think the idea of never stretching seems crazy, but the idea of saying price and ownership don't matter at all because everything will be bigger also doesn't make any sense to me.

People believe that they're either disciplined or convicted, but how can you actually prove out if that's the camp that you're actually in versus that you're telling yourself day to day?

I think it's the like, this time it's different versus it's always the same crowd. I think that this time it's different crowd is being disciplined is silly, not because discipline is a poor trait. But discipline is leading you to make the wrong decisions about what to invest in because you have, you know, I've been saying this a lot internally, you haven't updated your firmware and that entry prices are different, exit prices are different. So the prices you can pay and the structure you can pursue needs to be updated with the scale of outcome. And if you're still focused, we'll just make this, if you're still focused on 10 posts as some magic number, that ship has sailed and you're running an old model. That doesn't apply. The flip side is everybody says the same thing at the top or in a bull market. This time it's different. This is why multiples can be this high. This is why you can pay this revenue. This is what everything said in 2021. SaaS, it's the golden era of SaaS. Like, sure, you can pay 100x forward revenue. What's the problem with that? These companies are growing so fast, the market's so large. And it turned out at some point, like, gravity set back in. I think the unspoken thing here is I think the outcomes I've learned in venture are very different for these two strategies. I'm not convinced that the rewards for discipline from, in terms of what LPs reward, are as great as they are in the public markets. The public markets, if you're disciplined and you hold and things come your way, or you can go short, you can make money. I think in venture, if you're too disciplined, you follow your model and nobody cares. And I think the other thing I've learned is if you are convicted, it is this weird asymmetry. If you're convicted and right, you make a lot of money and then it kind of doesn't matter at like an individual level whether you're ever right again. If you're convicted and you're wrong, but everybody else is wrong, you kind of get to hide in the herd. So there is the part of me that says like increasingly it feels to me venture is really a momentum business in a way that maybe it wasn't before. It's a momentum business. And being in the hot companies and being on trend has all of these other ripple effects around fundraising, brand, notoriety, visibility, and that maybe outweigh the benefits of being— it's kind of like if you're trying to be a value investor in a bull market, it's pretty tough. You're kind of on your own island for a long time until the tide comes in. And I think in publics, being on that island can be more attractive than it is in privates.

But how does this feel different from cycles before, Charles, of being on your own island and being able to have conviction that I'll see it through?

I think everyone always says this time it's different, and sometimes it is. The internet was different. It just wasn't different right away. Mobile was different. It wasn't different right away. And look, there are also in these bubbles, there are companies that get acquired or who are perceived as being valuable. If you were in those companies and can get in and out of them, it's great. If you were a website, if you were in broadcast.com in the first bubble, that's not a company of consequence today, but you did quite well in that one. I can name a bunch of other companies from the mobile era that if you were in them, they were consequential. And I'm sure we'll look back and say, wow, there were some companies in the AI era that we thought were going to be very consequential early that turned out to just not matter. And that's, that's part of, I think, the convicted person's personality too tends to be slightly more gambling oriented. There's a willingness to say like, I'm going to bet on this trend before we figured out the outcomes. And bet that if I'm in the companies that matter, that some of these things are going to work out.

And then just to bring it back to the healthcare of it all too, for this specific article, I think that's where I feel personally for my friends who sit on more of the growth side of healthcare investing, the risks of not being part of the biggest plays within healthcare right now, the biggest companies are so massive because, and Holly says it throughout the article and the piece that you wrote, the sales cycles are so slow. There are so few decision makers at each of the big customers that if you miss out on the one that's big, I don't know when there will be another one. And especially in this AI era, it just feels so consequential.

I think the other thing, just in hearing you say that, I think this is different math for being a generalist than a specialist. I think if you're a specialist fund, you have to be in the companies in your category that matter, and your options are like pay a little bit early or pay a lot later. And so I've always felt like if you're a specialist fund, the math is different because I don't think you can afford to be completely disciplined as a specialist fund because part of what your LPs are paying you for is category access. It'd be like if you were an AI fund and you're like, I'm not in any of the model providers, I'm not in any of the big inference companies, and I'm not in any of the big high-flying application companies. None of those things might work, but not being in them would look really weird if I was like, hey, I put money into an AI fund that's in none of those companies. Yes.

Well, our second piece this week is a recent newsletter that Sam Lessin from Slow Ventures sends out, which made the provocative argument that the SpaceX IPO is proof that the VC model is no longer the only globally scalable story of value. His thesis is that the internet has made it possible for alternative belief systems around volume. So think crypto, you can think meme stocks to scale to truly trillion-dollar size in ways that simply weren't possible before. SpaceX is obviously the clearest version yet because it's backed by a genuinely futuristic narrative, complete with the $23 trillion TAM in their S-1 that Sam describes as, quote, so ridiculous. It was brilliant marketing, and we obviously saw that come true almost 2 weeks ago. The deeper point isn't really about SpaceX at all though in his piece. It's that we may be entering a capitalist multiverse where investors are no longer all playing the same game, holding assets not just for cash flow, but for identity, for story, optionality, and the chance to participate in what Sam calls the swerve to infinity. He's careful to hedge that history suggests this probably ends badly. But argues that post-internet world is structurally different enough that modeline capitalism with one universally accepted story of value may actually be over. Though Sam's argument is genuinely that we can no longer assume all investors are optimizing for the same thing. So from where you sit, Charles, is that actually a new argument or have this different investor bases all been playing different games and now we're just creating a name for it?

I think it's a description. I think Sam left out one really important thing, which is entertainment. I think some of this is entertainment. And I was very perplexed by the stonks, GameStop meme stock stuff until I started meeting people who were part of that community. And it's a lot of the things Sam mentioned. It's identity. It's striking back at the man.

Yeah.

It's entertainment. It's fun. They were like, this is fun. Like, this is fun. There's some money to be made. There's a little bit of nihilism. They're like, I don't know how else I'm gonna make money. Probably not. Like, I'm never gonna be able to buy a house with my 9 to 5. And I do think I read a lot of pieces which are like, SpaceX isn't worth this because I ran this model and look at this price to earnings and EBITDA. And I'm like, but no one's valuing the company based on that. They're valuing it because it's Elon and it's space and it's cool. And there's a lot of money propping it up. I think there's all these other things too. Financial markets have changed so much with high-frequency trading, with index funds, with all of these things. The idea that stocks move because individual stock pickers are doing DCF analysis, I think that world has been over for a while. And I think I was expecting a unified paradigm to replace it. And I think instead what you end up with are these fractured, this multiverse like you described. Some people are still looking for stocks where the math pencils out. Some people are just, ooh, there's momentum. There's momentum around NVIDIA. Well, what's the next choke point in the AI story? Oh, it's memory. Quick, let's go buy Micron. Other people are like, wouldn't it be funny if we took a shoe company and made it an AI company? Cool. Let's rebrand Allbirds as an AI company. None of this stuff is real and serious. Except that it's money. And I think if you're going to participate in markets, you have to understand the motivations and behaviors of other market participants. It's too easy to just say these people are being irrational, like, but they're participating. And I think the same is true in venture. I've been guilty of this sometimes. It's like, this price for this company doesn't make any sense. Well, someone else paid it. I've read something that said, like, you know, conditional on getting to $10 billion, the probability of getting to $100 billion is quite high. I'm like, that feels true. If you get to $10 billion, however you get there, there's a lot of late-stage money that says, well, let me find the stuff that feels de-risked, but the stuff that feels de-risked might not actually be as de-risked as we think it is. Those companies might still have a lot of money, a lot of work to do, but we live in a world where so much money is trying to find the thing that's going to go from— if you said, what game would you rather be in? We'd rather be in the $0 to $100 million game, the $100 million to a billion dollar game, or the billion to $100 billion. I think a lot of people are like, you know, I'd rather find the assets that already have gotten some momentum and hope that instead of going from 1 to 10, I have something that's going to go from 1 to 100. You get the notoriety, the name brand value, the visibility, and you can still make the same kind of multiple.

It was hard not to think about— they're obviously quite different in the spaces that they are, but when WeWork filed the S-1 describing this whole universe that will exist because of the company, and that's what you're investing in, you're not investing in this company, it's a tech company.

Yeah.

It's gonna be a school, it's gonna be a village, it's gonna be everything. The public rejected that notion, obviously, entirely before it even hit the scenes. This one, because of the space, I think Elon has far more ground than Adam did. This one was very embraced. And people, even the folks who participated in the public markets when it went out, I think they all agreed that I'm not participating because of what the company is today, but if Starlink takes off, this is what I'll be part of. If we actually get commercial travel into space, this is what I get to be part of. That distinction is, is interesting that people have widely accepted and wanna be a part of that versus WeWork when they read that being like, ooh, I actually do not wanna be part of that.

I think some people are buying SpaceX because they believe in the business. I think some people are treating it like a vote for a version of the future that they want to see. I think there's some people who are just like, this is a good trade. Other people are going to be in it. And, you know, I look at it, I'm like, Starlink is an interesting business. SpaceX is an interesting business. And then you have the whole XAI. But I think Elon is very good at attracting attention to his companies. I read all these like Tesla bear cases. I'm like, based on the fundamentals, you all are completely correct. But the support for that stock has nothing to do with the fundamentals and has everything to do with communities on the internet who are not reading your research. That are making different decisions. So you can write all the research you want about how the fundamentals don't work. Fundamentals only matter when there's a corrective force. And if there's no corrective force, things can remain expensive, ludicrous, crazy for a very long time.

And that's what I think I always thought of the public markets as that would be the corrective force. But now we're just seeing a completely new thought enter in. And you're right, it might have existed always previously, but this feels— it just felt palpably, palpably different.

I think ventures in many ways, that was my takeaway from 2021 is you would have thought a lot of people would have gotten fired for the excesses of 2021. And it didn't really happen. And I'm like, well, maybe there isn't much of a corrective force in our business in the short term on the part of LPs. Because, you know, they're thinking about this manager had one bad year, had one bad fund. Do I want to fire this person and jeopardize what could be great returns in the future and great returns that we've gotten in the past over a lapse in judgment that was done by many of their peers? I think my view is the collective LP answer is I do not want to.

Well, our last piece is from Sarah Guo on an essay in her Substack titled The Untrainable. And it's a more grounded take on AI and venture defensibility, the central argument of it is that anything you can measure, you can train against, which means anything measurable is already on its way to becoming a commodity. The real moat, what Sarah calls is the untrainable corner, is work where correctness exists only in private. So it's inside a specific firm's data, behind a specific firm's liability, inside the habits and the trust that a specific user base has built over years. She walks through why this matters practically. A majority of American doctors now open Open Evidence, which is one of her investments, every day, and no amount of compute compute buys that kind of adoption. You can't gradient descent your way into a physician's daily workflow. Her framework boils down to a 2x2 of how saturated a task is versus whether its correctness is public or private. And the prize is the unsaturated work with private ground truth, the stuff that can't be scored from the outside. The piece closes with a useful challenge for builders and investors alike. The absorption frontier keeps rising. You can't find a defensible spot and rest. You absolutely have to keep stepping towards something that hasn't been scored yet. Her closing argument on the fact that you can't find a dispensable spot to rest and you have to continuously move forward makes sense, but her whole moat framework is built on trust and habit and things that by definition take years to earn and, and can't be rushed. So how can founders actually hold both of those truths at the same time?

I like this piece a lot and it made me think about, I've been using agent loops to build some things in Cloud Code., and I was like, oh, it's very similar. I know what right looks like. It's actually pretty cool to be like, this is the answer. This is the canonical answer for these tasks. And I've built some things that are much better than the last version of things I've built, but it's only because I know what right looks like. And I read this and I was like, I think this is part of why the AI ROI conversation is so hard is because of many of the things that people are building internally that work. It's hard to observe from the outside that those things are working. And it's also hard to know like, well, what's, what's that thing worth? And I think what you said is true. Spending an expensive token and model run on a question with a trivial answer is a dumb thing to do. You should use the cheapest, simplest model. And I think this is the evolution that we're going through right now after the lunacy that was token maxing. People are like, hey, if we're going to get value out of this stuff, it's got to work on hard problems and it's got to be cost-effective. And I think also about the Kirkland and Ellis piece we talked about before, which is like, they were like, we don't get paid for average. We get paid for high-end work. And I think there are a lot of firms who are just like, well, the answer for us is we've got to find a way to do the quality of work that we have signed up to do for our clients, whether you're a physician or an attorney or whatever, assisted by AI. And that's hard. It's hard to do. You know, most models can't tell you, is this contract good enough to ship to a client? Tell you maybe, oh, it's error-free. It addresses all the issues. It doesn't mean it's good enough to ship. In the same way Claude Design can build you a presentation, builds a lot for me, doesn't mean I would ship the output directly. And so I read the piece and I was just like, we are getting into a more nuanced understanding of where the opportunities are in AI. And it changes. And this is the thing that I find interesting about this current moment is in public markets, if you decide that Micron is overdone and memory is no longer the big story, you could just sell it and go rebuy some Nvidia, or you could go buy Google because you think Gemma. In private markets, you can't do that. Yeah. Like you're stuck holding that position. And that company better figure out how to navigate and migrate or else they're going to get crushed. And the stakes are high because of the dollars that are going into these companies. I think it was a good framing to think about it. Also, the untrainable stuff is, I think, the stuff that companies are most afraid of handing over to the models too. So they're like, hey, the only way for me to work on the untrainable is probably private inference. Models that I train and control where I can observe and manage the outputs. And I think this notion that, you know, ChatGPT is still the consumer app of choice. Claude does a pretty good job. So does Gemini, but in the same way that Claude is that, how do I know that my work would be any different if I just said, I'm not using Claude anymore, I'm gonna use Gemini for everything. I don't know that. I'm not gonna find out. For all I know, if I swapped Gemini in for Claude, maybe my life would be exactly the same, but that's not, that's not what I'm doing.

And I think her piece calls into this focus of it kind of does take two to make this all work. It does take the company itself building and shipping product, but then it also takes the customer where they are on the range. So to your point, I, as the large enterprise, need to be willing to let go or figure out a compliance and security protocol that works in order to take hold of this innovation. And if I don't do that, then I'll never basically unlock the growth. And she had a good quote from, I think, Noam Brown, who said the only way to surely evaluate an agent over a 1-year horizon may actually be to just run it for a whole year. And I think that's where enterprise customers have yet to kind of see that through because they are trying so many different pilots and different attempts at this simultaneously and maybe cutting pilots 3 months in if they're not seeing the returns on it.

I read this and I was like, this explains why I think my initial skepticism of the AI timing, not the order of magnitude, was really that technology can move faster than people. I remember in the early days of the internet, my friend was like, we have this website. I'm like, what happens? He's like, oh, people fill in this form. It's so cool. What do you do with the form? Oh, we print the form out. I'm agreeing with you. He's like, well, people fill out the form. It's cool. We have an internet form. But then like we print the form out, then everything after that form is the same old process. And I'm like, well, you're not going to get very much gain if all you're doing is digitizing the very first step. And I think that's sort of what we're working through with AI. And this is why I suspect the low ROI numbers are probably accurate. And it's not because the technology's not good. It's because the total workforce process transformation that has to happen to get value from it, most people are unwilling or unable to do all of that today. Yeah. And as they, they look in customer success and customer service, my friends are like, oh, we've moved almost all of our first line stuff to chat, the savings are real. They're like, we're not sure if the satisfaction is the same, but the cost savings are real. And I think as latency and voice agents gets better, there are all these things. I'm just like, if you are willing to make the process redesign, that will increase the odds that you will get the value. But that's transformation work and you can't do that. That doesn't look like this. It doesn't look like the revenue curves of these companies. That takes time. It takes people.

We're still early on. I will say, like, I met the first person in a conversation who bragged about how much tokens their company gives them for usage. And I was like, ah, we're still, we've still got some time to figure this all out.

I know, I know. Everyone has to learn.

We're all out of time today. Thank you so much, Charles, for stopping in. Thank you to our listeners for hopping in as well. Please like and subscribe, and we will talk with you all in a couple weeks. We're taking off for 4th of July, but we will see you soon. Thank you. Bye.