From OpenAI Employee to Venture Capitalist | with Jenny Xiao, Leonis Capital

0:00Venture is strangely confident that AI will transform every other knowledge profession but venture capital. We actually did backtesting with our own algorithm on 20 ,000 companies and we found that our algorithm is actually better than us at picking winners. The reason why Antwrapa can have almost a 2x premium vis-a-vis OpenAI is almost 100 % because of model performance. The really outlier founders are the ones who are thinking in terms of like where the next model capability jump is and like how do we build ahead of the curve.

0:33Jenny Xiao: Jenny, welcome to Going Solo. Thank you very much for joining me. Yeah, really great to, you know, really grateful that you're having me, Dan. It should be interesting. I've been following you and your work and your writing for a long time. So I've been looking forward to this. To start with, just give a quick background to you and your firm and we can go from there. Cool. Yeah. So Leonis Capital is a research-driven fund. We're based here in San Francisco, investing primarily in seed and pre-seed stage companies. And we take a pretty research-driven approach to sourcing, diligencing, as well as supporting all of our portfolio companies.

1:04And my background before Leonis is actually in AI research. Previously, I was an early employee at OpenAI and did my PhD at Columbia. I did a lot of research in the early days of LLMs, especially like benchmarking and evals, and decided to work on this fund with my partner, Jay, towards the end of 2022, right after Chad GPT took off. And it's been a good four years so far, a scale from Fund 1, Fund 2, and now to Fund 3. It has some pretty exciting portfolio companies in our Fund 1 and Fund 2 portfolios already. So pretty excited to be here.

1:33Jenny Xiao: What kind of companies do you look for generally? Is there like a founder profile that you look for typically? Yeah, I think when most VCs are underwriting AI companies, they tend to focus on the product, whereas like for us, we very much focus on the underlying capability curves. We think a lot of AI products are just manifestations of the underlying model capabilities. And if you look at like the best products, they're usually a little bit ahead of their time. So like Cursor started working on AI coding back in 2021. if you look at like lovable they started you know with gpd engineer in vibe coding and only after sonnet 3.5 did lovable take off so like the best products including the ones in our portfolio if you look at like victor that's what that's in what our fund 2 portfolio they started working on ai employees even like back in 2024 and they had this vision of building end-to-end ai employees you know two years ago like two years ahead of the curve and the real the real takeoff point for them is really like opus 4.5 or 4.6 where models are capable of running long horizon tasks so we're really looking for founders who have the technical foresight to see where the next model innovation comes from whether this is like you know better tool calling capabilities or longer horizon task capabilities or better reasoning capabilities like what does the next generation of models unlock and how do we build the infrastructure for that so that when the next model capability happens, like we're able to capture all of it.

3:01So those are the types of founders that we're looking for. And that's why we'd like to back founders who are working on products that seem a little bit ahead of their time.

3:09Jenny Xiao: That's a fascinating thesis. It's kind of like skate to where the puck is going if the puck is like the capability of the models. How like this may be getting into the weeds a little too quickly, but how much certainty do you think founders can have about like what the capabilities of the models are going to be like three four or five years out i think uh three four five years out very hard to predict but if you're looking at like where model capabilities are heading um in the next like 12 to 18 months that's actually a very um reasonably easy thing for people to do if they have a research background i mean that's why a lot of our founders come from like very deep like technical and research backgrounds.

3:48And the reason is that we think those types of founders are more likely to succeed in the age of AI. And if you look at the top AI founders, we actually have a list of the top 100 AI founders. We did it back in like 2005, so last year. And among the top 100 companies, about 40 % of them have at least one co-founder who has a research background. And that number is up significantly if you compare it with the Unicorn Club founders, where almost no one comes from a research background and now you know almost like half of these companies have at least one co-founder who was previously a researcher and that's the kind of like um the research taste and the research like um ability to predict where the next model capability curve um you know is headed towards and for example like our team the victor team they actually came from metas llama 2 and llama 3 pre-training and post-training teams and they were you know at the cutting edge of like the foundation model layer before starting an application layer company.

4:47Jenny Xiao: I have to ask this question. It's probably one you get a lot, but you're there at OpenAI for a year. You see the launch of ChatGPT and that incredible moment. And then you decide, I'm going to go be a VC. Yeah. When, you know, in a way, there's like a massive green flag for me because you could stay an AI researcher and like make tons of money doing something in that field. But you chose to go down the VC route instead at a time when I think the industry really benefits from people who understand the technology investing as well. But like, what was the rationale for that decision? Yeah, I think basically two reasons.

5:23Number one was when I was at OpenAI, one of the biggest takeaways for me was that VCs really don't understand the technology. And this was like back in 2021, right? All the VCs, whether back in your like NFT companies and like this chain and that chain. And I was working on LLMs and I was like, wow, like no one's paying attention to LLMs. No one's paying attention to AI. People's conception of AI is still like traditional machine learning. No one knows about foundation models or like the transformer architecture. And like all my VC friends, what they're talking about is like, oh yeah, let's invest in like this new cryptocurrency or this new exchange.

5:59So that was very much the zeitgeist of like 2021 and 2022 when I started thinking about VC as a career path. And the motivation for me is that I think in the AI era, there needs to be a new generation of VC funds that really understand the technical foundations of AI. And I think for VCs, at the time at least, there was at least a three to five year gap between like what VCs are seeing versus what researchers are seeing. Even today, if you look at the hottest concepts in AI investing today, like recursive self-improvement, our own environments, like those are very old terms in the research world.

6:35We've been doing those for like almost like 10 years. And suddenly like, you know, like the venture world is like, wow, this is a hot new thing. Never heard of it. Let's like, you know, put a ton of money into it. And I think there needs to be a new type of fund that, you know, like really resolves that. And then the other motivation for me to go into venture is actually like looking at like Sam Altman's background. So when I was at OpenAI, I was like paying a lot of attention to like who actually has influence over the organization. And one of my biggest takeaways is that like it was not the researchers, right?

7:04It was really Sam. And what did Sam do before OpenAI? Well, he was a VC. And I was like, okay, maybe there's a thing or two that I could learn from a VC career. And that's why I decided to like, you know, do this fun with my partner, Jay. And it was a massive learning experience, given that I didn't have any investing background before going into this.

7:23Jenny Xiao: I can imagine. I first got to know you when you wrote an article, I think it was in 2024, about valuing models, which is really interesting because at the time I was also in valuation, private market valuation. And you had a really interesting framework about how to think about the way the valuation of companies shifts as they gain or lose the kind of edge with the model releases that they have. Has your thinking on that updated at all in the last year or two? How do you think about it now? And generally, any thoughts you have about like valuing AI companies would be really interesting. Yeah, no, for sure.

7:56And for those of you who haven't like, you know, read that article before, the main argument that we're trying to make in that article is that there is a zero value cutoff line for foundation model companies. Like if you're not able to outperform the best open source models, the enterprise value of your company basically goes to zero. And at the time, we were really looking at how do we value a foundation model company, right? What kind of like discount factor do we need to apply to foundation model company valuation frameworks to really account for the threat of being cleared to zero by an open source model?

8:33And I think like back in early 2025, that's when the article came out. The motivation for it really was like DeepSeek. DeepSeek V3 came out like earlier that year. And then we were thinking about like, what does this mean for the rest of the AI? industry. And the obvious question is that what does this mean for foundation model companies like OpenAI and Anthropic, whose entire valuation depends on their ability to outrun open source models? And what does this mean to like a wave of Neolabs that were getting funded back in 2024 and early 2025, and even to today, that open source models are getting so good?

9:09How do you survive as a Neolab, right? Like those were the questions that we had at the time. And I think One of the technical bets that we were making at the time is that open source models would catch up a lot faster in the post-training reasoning and also like the agentic era compared with the pre-training era. And the reason is pretty structural because in the pre-training era, like model performances come from how much data you can put into the model. And like closed source labs have a huge advantage there compared to like open source labs. However, when you're going into the agent era, that gap is being narrow because so much of the data comes from actual model usage.

9:49So an open source model, especially in fields like coding, can get a lot of user data very quickly and train on top of that user data. And also because agentic performances are much more about engineering and how do you make all the little components work well instead of how do you discover a new model architecture or pile data into it. And that's why we thought that in the current day and age, right, like in the emergent, you know, like reasoning and also agentic era, open source models would be able to catch up a lot faster compared with the pre-training era. And that's exactly what we're seeing today with like Kimi, Kimi 3 and then the latest like, you know, Ox Alpha model.

10:27Like open source models are like catching up really, really quickly. And especially in 2025, in late 2025, when you saw the rise of model harnesses, like the biggest impact of model harnesses is that they reduced the gap between open and closed source models and their performance. Right. Like, you know, you can get a better performance with a harness with like Opus 4.5 or 4.6, but you can get a hell of a good experience out of like a cheap model, like a deep seek. Right. With model harness. So, you know, all in all, we think like open source models are becoming a lot more competitive. And this is actually a really good thing for application companies, because if you're relying on open AI or anthropics models as an AI application, you're basically paying a model tax every single time your users use your model.

11:15Like only with open source models can you get your performance margins, your margins up to like 60, 70 percent, as opposed to like 30, 40 percent. So that's a huge brain dump on like our view of open source and closed source models and what this means for the whole industry. but happy to like, you know, dive into any of those.

11:32Jenny Xiao: I guess the main question as someone who isn't very technical on this is, is there a point at which open source models overtake? Because obviously there's like the cost saving argument and they're definitely narrowing the gap. But is it, you know, a gap that reduces by percentages, but we'll never actually, they'll never exceed it? Or is there a future where they actually come out ahead? I think coming out ahead is probably a little bit hard, especially in the medium, short to medium term. And the reason is because open source models just like don't they don't have the money to like buy as good of, you know, like our environments or post training data compared to like the foundation model labs, the closed source labs.

12:08But I think the gap is like getting smaller and smaller. And then I actually don't think it really matters if open source overtakes closed source models in terms of like performance. I think it's whether open source models can cross a bar where like the output is good enough that people can't really tell the difference. So, for example, in my day-to-day workflow, unless it's code A or anything technical, I just use GLM 5.2 or Kimi 3 or one of the Sonnet models. I don't really need an Opus model or a Fable for the highest end work. You're seeing this in Anthropics usage data. Fable actually didn't contribute that much to Anthropics overall revenue because the use cases for it are pretty narrow and it's a very expensive model.

12:51So I think increasingly we're going to see people route less important tasks or daily enterprise routines to open source models instead of closed source models.

13:01Jenny Xiao: Is there, you know, I think that there's maybe an incentive at the moment where the big labs want people to consume tokens so they don't route tasks as efficiently as they could do. But is there a future where that changes and that starts to become a priority for them as well? So you're saying like the foundation model labs because they want to know people to use more tokens and because your token usage becomes training data for them and therefore they're subsidizing token usage. They're 100 % doing that, right? Like if you look at, you know, what a$200 subscription can get you from like Anthropik and OpenAI tokens, I think the ratio is like you can get 8 ,000, 10 ,000 worth of tokens from like a$200 subscription or something.

13:42And I think there's definitely like that factor of like, oh, you know, we can get more training data out of our users. But also like these foundation model labs don't really care about individual users. Individual users are a bad trade for them, especially like users that, you know, do token maxing. They really care about enterprises. And for enterprises, especially if you're in like very token heavy workflows, you have to use their API instead of from your subscription. So that's where the money is made. people routing the models to the most efficient one? I don't think so. And I think this is what a lot of enterprises are talking about today.

14:18Like how do we route the vast majority of our tasks to open source models and only reserve the highest end tasks for closed source and the top tier closed source models? So I think that's absolutely a trend that we're seeing today.

14:31Jenny Xiao: Bit of a tangent. There's not been as much talk as I thought there would be about how AI changes the venture industry itself, what the impact of AI might be on investing, selecting, screening, all of these decisions. Do you have a view on that? Yeah, I think venture is strangely confident that AI will transform every other knowledge profession, but venture capital. Like if you talk to them at VC, and they're like, oh, yeah, AI is going to transform consulting and banking and whatever. And you're like, you know, what models are you guys using? And they're like, well, we're just using ChatGPT or Clot, right?

15:06They don't really have like an AI native tool for venture. And then if you look at like what their tech stack looks like, they're still using their very old CRMs and like, you know, traditional email tools and all that. We think the job of venture capitalists can definitely be automated by AI. And we've actually built out our internal system to do that. Most of our tasks are handled autonomously, which is why despite scaling the fund, like, you know, several times and going from like fund one to fund three, our team is really really lean because we just get so much out of ai like but before you even talk to a company right like our ai helps us choose which companies to like talk to or not um and also we have our internal scoring system for founders um and helps us predict like the success rate of a company even before we talk to the founder and we actually did back testing with our own algorithm on um 20 000 companies and we found that like our algorithm is actually better than us at picking windows.

16:00So I think VC can be automated. If you look at our roles, our jobs, like sourcing, market mapping, reference gathering, like competitive research, all of that can be automated with AI. But I think the biggest issue in VCs, you know, that's inhibiting VCs from adopting AI really is their tech stack. Like they don't have their information all in one place. It's scattered across like Notion and Google Drive and like some Word documents here and there. what we've done is like we put all of our funds contacts into GitHub. We have a GitHub repository called the Leonis Alpha. Everything about this fund is like on that repository, including like all of our CRM data, all of our, you know, like investment memos, like everything is just like in one single repository.

16:45So like when the agent needs something, it can just like go and like query it instead of like ask us like, hey, where did you put that investment memo from like 2015? teen. So I think, you know, knowledge accumulation is really important. Like, unless you have that, like, infrastructure for capturing all the knowledge and judgment within a fund, like, otherwise, you know, a venture fund is more like a consulting company. And then our insight really is we want to build a fund that works like an AI company, where, like, we have a data flywheel. Like, we are training our own models. And, you know, we're making our AI more intelligent.

17:18And we're, We're like absorbing, you know, we're distilling our judgment into an AI agent. So that's our philosophy. I think a lot of the industry suffers from a long time people saying things like venture

17:33Jenny Xiao: is an art, not a science. And therefore, because of that, they haven't been careful with the data or the tools and data hygiene where they keep things, all this kind of stuff becomes very important. But for the firms who've maybe been around for, you know, 30, 40 years who haven't been looking after this they have a bit of a problem maybe and then also i don't really buy this argument that like venture is more art than science i think it's a balance of both but also like whenever you say it's more of an art than science i always say that that's because you don't have enough data once you have enough data everything is a science right like i like to make an analogy with like machine learning the bitter lesson in ml and ai is really that like with more data and compute almost always beats better human handcrafted architectures.

18:18And I think the same applies to venture as well. Like if you have enough data and high quality data, your AI is going to perform better than human intuition over time, which is why like prediction algorithms work way better. Like if your friend recommends a movie to you, high likelihood it's not as good as a movie recommended to you by Netflix because there's just so much data with these algorithms.

18:40Jenny Xiao: This is true. This is true. To try and push back, to give the other side of the case a little bit, maybe, there's a good paper by a researcher called Maxime Bonelli called Data-Driven Investing or Data-Driven Investors. and he makes the case that essentially because the training data is backwards looking it it does actually produce uh better short-term success in some some ways like better markups better survival rates but it actually doesn't like data-driven investing doesn't produce the same rate of like really big outcomes because it misses the you know the crazy little outliers because they they can't be found easily through a data-driven approach what do you think about that I agree.

19:24And that's why I think, you know, it should never be an AI making decisions on our behalf. It should be like the AI helping us make decisions. And I think accounting for outliers is something that data is very bad at capturing. But to be honest, like most humans are very bad at, you know, accounting for outliers as well. Right. I think if you compare like data-driven investing with like pure human intuition-driven investing, I think, you know, if you do like a one-to-one like comparison there, I think you might realize that like data-driven is actually overall a better deal compared to like human, you know, like human intuition.

19:57But I think in terms of outliers, I don't think either approach is very good at capturing the outliers.

20:02Jenny Xiao: Yeah, it's an interesting study. I remember four or five years ago, I think it's published in MIT's online magazine. They took a big group of angel investors in this case, and they had some angel investors who just made their own decisions, some who were augmented with an algorithm, and then some entirely algorithmic decisions. And they basically found the algorithm outperformed almost everybody except for the small group of angel investors who had a lot of experience and really understood the biases in investing and could control for that kind of thing. That group outperformed everybody, but even then not by a huge margin.

20:42Jenny Xiao: So there's, you know, there's room for AI to make a big difference, I think. And a lot of it is processes that take away the biases that affect a lot of our judgment, maybe. No, I think I 100 % agree, right? Like the way that we use our algorithm internally, a lot of times it's more about like search and retrieval. So for example, like when we come across a company, right, I always like feed it into Alpha and I ask Alpha like, hey, you know, of our data set, out of our data set, like, you know, are there other companies that work in a similar domain? Did they pivot out of that domain? If so, what are the reasons why they pivoted out of that domain?

21:18and also like are there venture scale outcomes within that within that field that this company is operating in and I think that's really important data for us when we're making a decision on whether to back a company or not or at least like it's a good question for us to ask in a due diligence process right whenever we talk to a founder we're like hey our algorithm basically like surfaced five companies that worked on something very similar all five of them failed one of them got acquired you know like uh what's your take on it why did all the other companies fail but like what why why would you guys be successful right and i think um oftentimes that's a very good like you know conversation with the founder because it really tests like whether the founder has done their homework and understanding the market um and i think it really helps us with

22:02Jenny Xiao: like founder judgment as well yeah that's where i was gonna go next actually so that's a good transition i was going to mention um you know a lot of vcs will say things like the the number one trait to look for in a founder is like tenacity or grit how do you mix the really like human qualitative side of of like evaluating a founder and like the diligence aspect of it with the really technical side it seems like a like a bit of a balancing act to do both i actually think technical due diligence fundamentally is about founder due diligence, right? Because when we're looking at founders, the number one quality that we look for in founders is whether they have the technical foresight to see, you know, where the models are going and like, how do we build ahead of the models?

22:46Because if you talk to founders, like 99 % of them will say, hey, models can do this. Therefore, let's build this capability on top of models. And this is how you get a company that's going to be commoditized by foundation models in the next like six months or 12 months. And this is how you get like, you know, a very crowded, like thin wrapper company, or like there's 10 companies working in the same domain. The really outlier founders are the ones who are thinking in terms of like where the next model capability jump is and like, how do we build ahead of the curve? And if you look at the interviews with like Boris from like Cloud Code, that's actually his advice to founders as well.

23:21Like think about where the next model capability unlock is and build ahead of the curve before AdWrap, before OpenAI. And this is how you build a durable company. So I think like, you know, technical due diligence for us in a very large and very large degree is evaluating whether the founder has that kind of technical foresight. Well, I pressure test them on a lot of different like, you know, questions around a defensibility of their product and like how they're thinking about whether their product gets like thinner or thicker as like foundation models improve. And I think the best founders usually have a great answer to that.

23:53And just going back to the question of like being research first and, you know, technical first versus being founder first, I think a lot of VCs will say like, oh, yeah, it's just about the founder, like back good people. I think, you know, for me, I feel like founder first has mutated into like founder only for a lot of investors because they're almost using it to justify stupidity in other domains. They're like, oh, yeah, I can, you know, I can basically not understand anything else as long as I have a good read of the person, right? but that's just so um that's like so hard to quantify and it's like so hard to like you know describe like you know how how do you even tell people like i'm really good at reading people right like are you like reading their palms or something so i'm always a little bit worried when people are saying like oh yeah i only focus on the founder i'm like you know i have a really

24:42Jenny Xiao: good founder to use i completely agree unless you're a fortune teller right unless you're like i'm a professional fortune teller um otherwise i don't really trust that yeah no i i completely agree. There's another good paper I like called Entrepreneurs as Designers of Problems Worth Solving. And it talks about understanding the quality of a founder through the problem they've chosen, but also how they think about that problem, how they approach designing a solution to that problem, and really getting to the quality by looking at the work they produce, which I think it makes so much sense to look at things that way.

25:20Jenny Xiao: What do you think about investing in, for example, just to pick a slightly spicy topic, Neolabs. So Julian Beckett-Sequoyer said recently, investing in a Neolab now is a bit like investing in Quora or stumble upon after Facebook was around. It's maybe a bit too late, I think is the implication. Do you agree with that? I don't think our biggest takeaway is that I don't think the problem is that investing in Neolabs today is too late. I think the problem is that VCs don't really understand what they're investing in. And I think there's basically not one type of Neolab, right? There's like four types of Neolabs.

25:57And we actually have a paper on this that's going to come out in the next week or so, but I'll give you a little preview here. Our thesis is that like, if you look at like Neolabs on two different axes, on one axis, you have like, how much capital do you need to build this thing, right? On the other axis, you have like, how fast can you make money from this thing you can basically divide neolabs into like you know a two by two where um on the top left quadrant you have like venture shaped neolabs these are like neolabs that fit the venture capital model very well you don't need a lot of capital and then you can monetize really quickly and then you have like grand shaped neolabs that don't require a lot of capital but also like don't you know commercialize very quickly and then you have like growth equity shaped neolabs like um that, you know, take a lot of capital, but also like can monetize relatively quickly.

26:46And finally, you have sovereign shaped Neolabs. And these are Neolabs that need a lot of capital, and they don't make money very quickly. And if you look at like where venture money is going, right, 60 % of the Neolabs that are being funded today, and we have like a market map of 119 Neolabs, 60 % of them, the vast majority of them are sovereign shaped. Like only 8 % of them are venture-shaped. And this means I think venture capital is chasing the wrong types of neolabs because they're like, well, I made money from OpenAI and Anthropic. And these are, you know, like initially when these companies were getting started, they required a lot of capital.

27:22And then also they were very slow to commercialize, but they have commercialized a lot faster than people anticipated. And therefore, let's invest in like other neolabs that look like OpenAI and Anthropic. And I think this is where you're going to see a strategic misalignment between the types of neolabs and the kinds of capital that's like chasing after them our whole thesis is that you know remember the the saying that we wanted flying cars but we got 140 characters all right and now we have our flying cars like venture is back to backing flying cars and those flying cars today are called neolabs but we don't really know how to finance them we don't really know how to price them and we don't really know if like venture capital is even the right category for them i

28:02Jenny Xiao: like the the positioning that you have which is your background as a researcher allows you to look for non-obvious ideas within ai so it's kind of the best of both worlds to an extent you're on the frontier but you're not backing the really like frothy stuff so my question would be when you look at you know without without naming any names obviously but when you look at the the deals being done and the stuff that gets the headlines is there like a rough percentage of stuff that you think is really just like is not going to end well like is it half of the deals you see i think are very misaligned or more or less it's hard to give a number but i think um there's a few factors here right i think um the reason why a lot of these new labs can command really high valuations is because researchers are worth a lot of money.

28:55And if you can, you know, raise a ton of money and, you know, hire 100 really good AI researchers, like someone's going to acquire you and you're probably going to get your cost of capital back. And it's a totally different game than traditional venture investing because traditional venture investing didn't really have that safety net. But Neolabs do, right? Someone's going to want to acquire like, you know, So Ilya's company or someone's going to want to acquire like Mira Morati's company, right? Because of the researchers that are in those labs. So I think there's definitely like a safety net effect there.

29:27But at the same time, right, like if you look at how a Neolab can actually drive interest skill returns, I actually think it's a very, very, the math is very hard. If you invest in a Neolab, it's valued at$30 billion. That company has to be at least like$150 billion for you to like actually make a buck on it like even if it gets acqui-hired for like you know i don't know like 10 billion dollars or something like you're not really making a whole lot of money you might be losing money even even with a really large outcome and that's why i think um like sovereign funds are really the best like funders for these neo labs because a whole like a big thesis is that we have to have our national ai champion or like you know we want to build our own data centers or you know we have to fund this type of, you know, national strategic research, I think that's a really good source of capital for Neolabs.

30:16And I think like venture capital is like a little bit misaligned for like the capital requirements as well as like the returns of Neolabs.

30:26Jenny Xiao: You've been in venture since 2022. So you and you kind of you've been in tech for longer than that, obviously. So you saw the, let's say the the rocky years of nfts and metaverse and all that fun over that time what do you think is the most valuable lesson that you've learned as an investor i think the most valuable lesson that i learned as an investor from my you know like investing in ai is that the biggest enemy is not other startups like the biggest competitor to AI startups really is the foundation model, which is why we pay so much attention to like model capability increases and like how, you know, what kind of strategies foundation models are, you know, are pursuing.

31:13And that's why we did a deep dive on OpenAI. We're doing another deep dive on Anthropik. We pay a lot of attention to like what's happening in AI research and people are like, how is this relevant for your startups? Because you guys are investing at the application layer and I'm like, well, it means the world to us because foundation models are the number one biggest competitor to like all this, all these startups out there. And they determine your technical roadmap. They determine like whether you have an edge, they determine how big your mode is. And that's why I think like people in my position to really pay attention more to like AI research, especially like the work that the closed source labs and also the open source labs are pursuing.

31:54Jenny Xiao: Do you have any thoughts about the the potential outcomes of these the two big labs going public might be and maybe in in the case that it all goes very well and obviously then it's a huge financing moment for them and it's access to even greater growth capital and so on but also perhaps like maybe if it doesn't go so well how does the industry look have you have you given that any thought yeah we actually built our own like model of like anthropics revenue from like 2026 projecting out to like 2030. And one of our biggest takeaways is that for Anthropik to really grow, like keep its momentum and also like be valuable as like a public company.

32:35And I'll just use Anthropik as an example because it's a leader in the industry right now. For Anthropik and OpenAI to grow and maintain its position and for the AI industry to be healthy, we need to constantly find use cases outside of coding. Because if you look at AI revenue growth from like 2025 all the way to today, it's almost like 70, 80 % driven by coding. Non-coding revenue is such a small part of AI revenue. And the biggest challenge is like, how do we find industries that look like coding and can use a lot of tokens and can run like autonomously for like hours to really like keep up the growth, right?

33:19Because if we're really only relying on coding, there's several issues. Number one is like coding is a big market, but it's going to tap out somewhere. And I think the second part is that like coding is also an area where open source models are catching up the quickest. So you're going to face the erosion from open source models fastest from coding. So how do we like pursue domains outside of coding? And our belief is that it's going to happen in like three phases. Like coding is the first phase for AI revenue takeoff. The second phase are like fields that are coding adjacent. So like data science or finance, like anything that's quantitative.

33:52And, you know, you can very easily verify whether the outputs are right or not. And then after that, you have like the huge like judgment heavy knowledge work. And this includes venture. This includes like banking. This includes like sales. Right. Like, but there has to be like the next, you know, a next curve after coding for these like giant closed source models to float these giant closed source models. And I think the next generation of AI products have to look very different for that to come true. And I think if you're just building a tool to help VC source deals, that's just not going to burn that many tokens, right?

34:32Like you're not going to burn that many tokens to sustain revenue growth for the foundation model companies. you really have to push towards like 24-7, always-on agents that are always running in the background, that are like, you know, like doing long horizon tasks for hours or even like days. And I haven't seen that kind of use case for, you know, domains outside of like coding or technical domains because I use like Cloud Code and Codex like daily. I have like seven, eight agents running in the background, even as we speak, right? I have agents that are running for five, six hours, often overnight, working on coding tasks.

35:10But in my day-to-day role as a VC, I actually don't really burn that much tokens. But I think in order for AI revenue to grow, there needs to be use cases for non-technical folks that are long-horizon tasks.

35:28Jenny Xiao: That's a very interesting take. I have a bit of a theory of my own, which overlaps with this a bit. I'd be curious to hear what you think about it. It's essentially like looking at the upcoming IPOs and the valuations of these companies. And perhaps the surge that Anthropic has had in the last year is actually going to end quite badly for it, because it's going to end up pricing in a lot of future growth that may not happen for the reasons you've described. Whereas on the other hand, if open AI ends up having a slightly slower and slightly more cautious approach and maybe goes to market at not such a crazy valuation, long term, that could end up being a better outcome for them because over time, that performance will probably converge between the two of them, in which case they're the better investment.

36:17Jenny Xiao: I'm very worried that people are assuming that Anthropics growth is going to keep going and it seems unlikely. I think in AI, right you know people always overestimate how much the number one player in ai has a lead versus the second number two player and i think a year ago people were talking about gemini as like the top model and like google's back you know it's the king of ai it's overtaken at throp again open ai and people thought that was forever it lasted for like three months or so right like two years ago people are talking about like gpd4 or like um you know gpd like um o1 or whatever as like the king of ai and that lasted for like maybe half a year and now people are talking about anthropic as the king of ai and i don't think it's going to last for like as long as people um think it will and i think like ai revenue is very highly correlated with ai model capabilities um the reason why anthropic can have like almost a 2x premium vis-a-vis open ai is almost a hundred percent because of model performance.

37:15And the reason why engineers are using Anthropite versus other models is almost 100 % because of performance. So I think the lead in AI revenue is very fragile because it fundamentally depends on your model capability lead.

37:29Jenny Xiao: Yeah, which ties back nicely to the article that you wrote, which we talked about earlier. Perfect. We have a couple of standard questions that we ask at the end. Be curious to get your thoughts on these. The first is, if your general view of the world and your thesis is correct, what do you think the world looks like 10 years from now? Do you have any kind of picture of what that might be? Yeah, I think software spent the last 40 years recording human decisions. And I think the next 10 years are going to be about software making decisions on behalf of humans. I think the future of AI enterprises, enterprise AI is going to look more like AI running entire divisions within an enterprise instead of like, you know, AI being supportive to humans.

38:20And I think if you look at the history of industrial revolutions, and that's a topic that I'm very fascinated about. We read our whole team read a whole lot about like the history of the first and second industrial revolution. And one of our biggest takeaways is that even if you have a technological revolution, like social change and organizational change comes much, much slower. So I think the next 10 years really is about building AI native organizations so that we can rework our entire workflows to make AI capable of being in the driver's seat. instead of like AI being permanently handicapped by the fact that like it just isn't a first-class citizen in an organization.

38:58That's something that we're doing internally as well, right? Like by just building our entire tech stack so that AI can like read and write files like really easily, really quickly. It could, you know, get context to like the entire organization. I think more and more organizations are gonna have to go through that change for AI to really have a huge effect on like our productivity. I think there's one data that's really, really interesting is that in the last three years, AI usage has gone from zero to 53 % for LLMs. However, if you look at organizations that actually have AI in their workflows, it's very, very, very small, especially if you go outside of tech.

39:39Very few people use AI in their day-to-day workflows. And I think it's because there's organizational inertia. I like that their workflow is designed around humans instead of like around AI. So that's the biggest change I'm worth seeing in the next like 10 years.

39:56Jenny Xiao: Have you come across Carlotta Perez, the economist? She writes about technological revolutions and how the technology gets dispersed across the economy eventually. And, you know, it's a similar kind of phenomenon to what you're describing. Her book lives on my desk constantly. I think it's one of the most important things to read, particularly at the moment. Other question is, what's the best piece of advice you've been given as an investor? Yeah, I think this is more general. This is the best career advice that I was given when I think when I was like 21 or something. I was going through the Teal Fellowship application.

40:31I ended up not doing it because I ended up working for OpenAI, so I didn't really need a fellowship. I was looking at previous Teal Fellows and looking into what helped them get the Teal Fellowship. And I came across this interview with Lucy Guo, one of the co-founders of ScaleAI. And in that interview, she was talking about, you know, career advice she would give to like an ambitious 20 something. And she said that the best career advice for anyone who's working in tech is join a well-known company for like six to 12 months. Have that on your, you know, like CV resume and then start thinking about like what are some new like company ideas?

41:09What was a company that you want to start? And then rinse and repeat, right? If you don't have a good idea, go earlier, join an earlier stage startup, work for like one or two years and then think about like new ideas. And that was exactly what I did in my career. I joined OpenAI, left like after a year and I was like, OK, well, I have like that stamp of approval. I have like the big name, you know, big company name on my resume. Now I'm going to go out and do something on my own. So I think it's a great piece of advice for anyone who's entrepreneurial. And yeah.

41:40Jenny Xiao: Well, Jenny, it's been a real pleasure. You mentioned the report you're going to be publishing in a week or so. If you send me that, then I'll make sure we have a link in the description down below so people can check that out as well. But yeah, I'm a big fan of you and what you're doing. I think your approach is great. And I'm always a fan of VCs who really understand the technology they're investing in. I think it's quite important. So thank you for taking the time. It's been a great conversation. All right. Thank you so much, Dan. And same here. I'm really a big fan of your research. I've been following you on Twitter for a while.

42:13So really excited to have this conversation.

42:17Jenny Xiao: Likewise. Let's do it again sometime.