OpenAI Researcher turned VC on what we’re missing on the AI bubble | Term Sheet

0:00We're the Hartford, with decades of experience insuring millions of unique small businesses. When it comes to your small business insurance, Thank you. one size absolutely does not fit all. Get a quote or find an agent today at thehartford.com slash smallbusiness. Get in the game with the college-branded Venmo debit card. Wreck your team with every tap and earn up to 5 % cash back with Venmo Stash, a new rewards program from Venmo. No monthly fee, no minimum balance. Just school pride and spending power. Get in the game and sign up for the Venmo debit card at venmo.com slash college card. The Venmo MasterCard is issued by the Bancorp Bank NA.

0:38Select schools available. Venmo stash terms and exclusions apply at venmo.me slash stash terms. Max$100 cash back per month. A lot of business observers just assume that AI progress is going to happen like in a very linear way. But in reality, AI progress happens in lumps. Hello, hello. Welcome to Term Sheet. I'm Allie Garfinkel. And this is the podcast where we talk about the weird world of private capital, tech and startups.

1:08Now we're undoubtedly in for another year of chaos in the private markets. And of course, a key source of that chaos is going to be you-guested AI. But our guest this week is someone who has a very unique perspective on it. Jenny Xiao is the founder of Leonis Capital, and Jenny is a former open AI researcher turned VC. She has a very academic lens on AI, and she is a person who really understands the international landscape. When I want to understand AI in China, I go to Jenny, and I'm really excited to bring her to all of you. Before we get to Jenny, we need to talk about Sunday night. Fed Chair Jerome Powell came forward and said that he is under investigation by the Department of Justice.

1:48Now, the contours of exactly what this all entails are still evolving as we speak. But I think it's worth saying this out loud, number one, and going back to a couple of basics. The first is this is genuinely shocking. Powell is a famously thoughtful, measured figure in a lot of ways, whether you agree with him or not. So to see him in that video coming out swinging like that, I was floored. This is about whether the Fed will be able to continue to set interest rates based on evidence and economic conditions or whether instead monetary policy will be directed by political pressure or intimidation.

2:24Trump and Powell have had a lot of tension for years. But what surprised me in that video was the timing and the tone. Very clearly, this is a situation that isn't actually just about Powell himself, I would argue, but about the Fed as an institution and probably about the independence of Powell's eventual successor. It's worth also talking about this in the context of the private markets. VCs in particular who I talk to will often say, ah, yes, what the Fed does doesn't really affect us. because it's long term. We are investing on such long time horizons, to which I say the long term stability of the United States economy actually does matter.

3:03And if we get to the place where the Fed is responding to short term political pressures, that fundamentally changes the economy in ways that affect every single investor and every single startup, ultimately. Now, in a more fun piece of news, this week I published an inside look at Strava. For those you who aren't familiar with it, Strava is a fitness tracking app. They say it's about discipline. I think it's about routines. That has a long history, but actually is en route to going public. I had a really interesting conversation with the CEO, Mike Martin, who's been the CEO for about two years, about the company's turnaround and how in certain ways Strava was brought back to life, brought back into this new era by female Gen Z runners.

3:47And it's the sort of thing that when Strava does go public, they officially won't comment, of course. But when that day does come, it will be a tech IPO, but it will also be an IPO at the intersection of a lot of trends around wellness, which I think are going to be a bigger story in 2026 than we even know right now. And that's the news for this week. Now, in an interview recorded at Brainstorm AI, here's Jenny Hsiao, founder of Leona's Capital. Jenny, thank you so much for being here. Thanks for having me. I'm excited. Jenny, I am so excited you're here. You've seen AI from so many different perspectives as an academic, as a researcher at OpenAI, and now as an investor.

4:26A lot of people are really worried about the existential risks. What is your take on balancing innovation with safety? I actually think this is a topic that's not being discussed enough in Silicon Valley. I mean, there are small pockets of researchers here and there, especially at OpenAI and Anthropik, that are thinking about existential risk. But I think it is largely contained within those small communities and not discussed enough in big tech, in the startups, and also in the VC ecosystem. No one really talks about it. So it's kind of niche right now. I think it is more niche than it should be.

4:59But I think it really should be like, you know, one of the top three topics that the broader AI community talks about. And I think the way that, you know, we handle AI safety very responsibly first comes from the foundation model labs, because these labs face a lot of competitive pressure from each other to release products really quickly. And oftentimes a lot of products are released without proper safety testing. We've seen, you know, really terrible results previously from Google, Gemini, very early days of AI. And then later with Grok when, you know, Grok is doing all sorts of crazy things and saying racist stuff.

5:33and I think a lot of AI labs are pressured to release their models faster and faster because of market expectations without properly testing those models for safety and then from an application perspective I really think VCs and startup founders should pay a lot of attention to safety and safety here really means that we're handling customer data responsibly and also you know like adhering to AI regulations. Well it's the sort of thing too that part of what you're saying it sounds like is that the pace of change is not just about the technology itself it's about the competition that's happening between top labs in part no absolutely i think there is well anthropic always says says that if they're the best player in ai safety there might be you know a race to the top eventually because if they said the safety standards and everyone's going to want to imitate them but we're seeing there's actually a race to the bottom happening in terms of safety standards because a lot of the newer labs, especially like Grok and XAI, they're releasing models really, really fast without even testing them sometimes.

6:35With sometimes radical consequences. Yes, radical consequences. And I think the AI labs in China are also releasing their models super, super quickly. And they're not necessarily that concerned about AI safety or like doing a proper safety test before releasing. What do really good safeguards look like in AI? that's a very hard to define thing because you also always have to be careful about safeguards because a lot of people have complained about ai models becoming quote-unquote dumber if you have too many safeguards and if also if you safeguard thing um a model in a wrong way it might have other consequences that are not predictable like what um i remember listening to you know dario um talk about this on another podcast like he's he's a ceo of anthropic and he says that he has a lot of Anthropi users asking him, hey, can we just make Grok respond in a certain way?

7:25Can we inject a certain prompt into Grok, sorry, into Claude. Yeah. And, you know, make it do a certain thing. But he's like, and large language models are very unpredictable. So you can't just say like, hey, don't talk about this. And it's just going to completely filter it out. It's going to have like a lot of negative consequences that are unpredictable. When you think about the role academics have to play in this, like I was thinking about nuclear earlier today, for example, a situation where academics played a key role. What role are academics currently playing in AI in the conversations around where AI is going?

8:00I think the biggest role that academics are playing is educating the next generation of AI researchers. And this is where all of the AI talent comes from, right? Very few people go into AI and say that, hey, I'm just gonna like, you know, become an AI researcher without going through an academic program, either an undergraduate or, you know, postdoc, PhD program. So I think academics play a huge role in doing that. And also we're seeing a lot of professors come out with startups like Fei Fei Li, for example, started World Labs and a lot of Stanford professors are coming up with startups in AI, in robotics.

8:34And we think increasingly you're gonna see professors and academics come up as startup founders. And, but overall, I think academics are playing a smaller role in AI today than they were like 10 years ago The main reason is that academia doesn't have the resources for model training. And this is something that the AI community has been talking about for a really long time. I remember reading this thing from one of Anthropics co-founders, Jack Clark's blog, many years ago when he was talking about how academia just doesn't have GPUs for doing large-scale AI training and also for doing experiments.

9:10and his proposal there was why don't we create like a national GPU cluster and give it out to academia and non-profits so for them to like be able to do frontier AI research because right now the amount of research that you can do in academia is relatively limited and most of it's limited by GPU resources. Well I was going to say it sounds like part of what you're saying is that academics play a key role educating the next generation but also a lot of people drop out of their PhD programs to go join one of the frontier labs where they actually have compute. which creates a situation where a lot of the frontier research or all of it pretty much is happening at private companies that's largely true i think there's still a lot of theoretical work that can be done in academia but outside of theoretical work or like small scale research you can't really do large scale experiments or train large models in academia and i think that's the biggest um that's the biggest challenge that academia faces nowadays but you go to any academic institution, especially the elite ones, almost all the professors and the PhDs there are affiliated with private industry in one way or another.

10:13And I think this in large part shapes their research agenda as well. So it's almost like, hey, you have all these professors and researchers go after the same research directions because this is what the private sector says is important. This is what people are willing to fund. And I think that's not necessarily the best dynamic we want to see in academia. Now tell me about your trajectory because you started as an academic. Tell us about your PhD. Like what was your dissertation on? Yeah, so I did a PhD specifically focusing on economics and AI. So a little bit of like econ, a little bit of like CS, AI.

10:47And specifically, I was really interested in game theory and AI alignment. Since, you know, I got into AI back in the very early days of LLMs back in 2017. What was the world for LLMs like then? Like Paint us a picture. Well, this is when, you know, the initial LLMs were coming out. And at the time, the hottest thing wasn't the LLM. People were talking about LSTMs. People were talking about other types of - What's an LSTM? So it's another model architecture. You don't need to know what it is. This is like ancient history. This is an ancient LLM. It's a fossil. Yeah, and you know, the hot models back in the day were computer vision models, some reinforcement learning models.

11:23And at the time, right, I became very involved in the AI safety community and actually ended up working for the Future of Humanity Institute at Oxford University for a summer. And then I became really immersed in the world of AI safety and became really interested in what AI safety means. And that inspired me to want to do a PhD related to this area. And specifically, I was really interested in how do you create incentives for AI, which is why I went into econ and specifically focused on game theory, because game theory is all about like, you know, how actors interact with each other and also how, you know, you create incentives and structures for actors to behave.

11:58And that's an area that I was very, very curious about when I was a PhD student. And what about the transition to open AI? How did that happen? It's a very interesting story. I actually left my PhD after the first year to join open AI full time. And then I was one of these stories where you want to compute? Well, in large part, yes, because I thought that academia is like just so much behind real industry and really wanted to have an impact on the AI industry and especially in AI safety and that's what inspired me to like leave academia and join an industry lab and then OpenAI and DeepMind were two of my top choices at the time.

12:34Ended up going with OpenAI. It really pissed off my mom because she was like oh you're leaving a really good academic program to go to this you know you're a dropout now little like non-profit in the Mission District in San Francisco. She was very pissed about it at the time. because this is 2017 no this is 2021 2021 okay but chat gbd still hasn't come out yeah oh so okay so you started opening what were you working on there what did you learn there yeah i mostly focused on benchmarking like large language models like benchmarking evals was like the main thing that i was working on specifically i was looking at like um chinese language models at the time so like you know large language models that dealt with like um you know like Chinese language, not just English language.

13:17So that was the main thing that I was looking on, looking at. Now there's a world of people who are founders who are like, oh, I did six months at OpenAI, so now I can go found a company or I can go be a VC. But that wave hadn't started yet at all back then. What you were doing is probably pretty unique. Did people look at you like you were crazy? I think so. At the time, right, like I had a lot of friends who asked me, why did you leave you know why did you specifically like leave the research world for venture I think that was the question people were asking me the most and not like hey why did you leave open AI because a lot of my friends who left open AI ended up working for anthropic because that was the early days of anthropic as well and um I would say like my research fit would be a good fit for anthropic just because there's a lot of overlap in what the two labs are doing but I think I made a career decision that's very different than a lot of my friends who ended up like working for another big lab but I just decided to like do something else all together if you had to distill it down why didn't you want to work for another big lab I think fundamentally I'm pretty entrepreneurial I wanted to like really own something and really have the ability to shape an institution and when I was at open AI I really felt that like hey this is a really amazing organization but I'm still sort of a cog in a wheel I don't really want to say that because it's a research lab and it was a small company it's not like Google or Meta or any of the you know 30 ,000 person companies but But it was getting there now.

14:42It is getting there. It is getting there, which is why, you know, I probably wouldn't have liked myself if I stayed there for another three years. What has most surprised you about OpenAI's trajectory since you've left? I think the biggest change, and I've actually written about this. I wrote a whole article about this, which is OpenAI is transitioning from a research lab to a platform company. OpenAI wants to be the everything platform for the U.S. It almost wants to be WeChat for the U.S. And it's one of the few companies in Silicon Valley that has that kind of ambition because the vast majority of Silicon Valley companies wants to do one thing and one thing very well.

15:19And OpenAI is going off after consumer and shopping and APIs and enterprise. And it's going after like all these different directions. And I think it's going to face a lot of competition in a lot of these directions. But at the same time, it's probably going to become the platform company of this decade. I was going to say the other thing too, there's so much talk around open AI. What do you think people most misunderstand about working there? Well, I think when people think about open AI, they think about it as open AI today, right? And I think the biggest thing is that like it's very different working there in the early days versus working there today.

15:58In the early days, people were taking like a huge pay cut just to go work for open AI because this is like, you know, nonprofit. It was very idealistic. There's a lot of focus on safety. And today it's a lot about, hey, you know, my total comp is like three million at OpenAI. Right. So it's a very different culture today compared with like three, four years ago and before Chagiput went viral. But I think this kind of like transition happens like in every single organization. If you want to go from like niche to mainstream, you always attract a larger following. So I think when people are saying like, wow, did you get a huge comp when you're working at OpenAI?

16:32I was like, no, no, it's it's a lot lower than Google. go at in the early days. You're sitting there, you're like, I was not being paid millions of dollars, I promise. Tell me a little bit about your strategy. You describe it as research-driven. What does that actually mean? So the inspiration for the research-driven strategy comes from my experience working as an AI researcher. So basically what I found is that there's a massive disconnect between what researchers are seeing and what investors are seeing. What we're talking about today here at fortune ai or at all these business conferences is at least three to five years behind what researchers are thinking about and it's crazy you don't look me look at me like that but it is crazy um but we are so behind the technical frontier and that's the gap that i really want to bridge and i really believe that with ai there needs to be a new generation of founders there needs to be a new generation of vcs and this is the first time in the last like 20 years that VCs have to underwrite not just the market, but also the technology.

17:40Like SaaS companies were built on a very stable tech stack that doesn't really move, right? Your infrastructure provider doesn't threaten to absorb you. But now you have all these LLMs that might build your company into a feature in ChadGPT. And this is the first time in the last 20 years that VCs have to underwrite not just the market, but also the technology. And we think that because of that, you're going to have a new generation of VCs as well. And this generation of VCs will have to be as technical as the founders themselves. So what are researchers talking about that we're not talking about right now?

18:15They're going into all sorts of details about scaling, about data, about reinforcement learning, about multimodal, all those really, really nuanced and very technical questions that we're not really talking about as business observers. What is one of the things that you think every business observer here who's going to be watching this podcast needs to understand about the frontier of where AI is at right now? I think the biggest thing that business observers should know is that AI progress isn't linear. A lot of business observers just assume that AI progress is going to happen like in a very linear way.

18:53But in reality, AI progress happens in lumps. You have a new model architecture. You have a new training run. you have a new way of doing things and suddenly you get like a big bump right it happens like you know like periodically and that's why it doesn't really make sense for you to say like hey why is AI progress slowing down or why is it like you know like happening so quickly because I think business observers will say two things about AI progress they'll either say like oh it's happening so quickly that you know like AI today is like so different from AI two years ago or they'll say like hey it's happening so slowly that we haven't seen any progress since GPT-4 So, you know, I think it's neither of those two extremes.

19:30It's somewhere in between. And the reality is AI progress happens in little lumps. Now, you've previously talked about the role your Asian American identity has played in your career as an AI researcher and now as an investor. Can you tell us about that? Yeah, absolutely. You have kind of a contrarian take. Yes, I have a very contrarian take. I'm actually Chinese American and you can tell by my last name. I actually think there's a huge advantage in being Asian American and specifically in being Chinese American if you want to be in the AI industry. And I know a lot of people are like, oh, there's discrimination.

20:06There's like all the political stuff that's happening. But I think the advantage comes from the fact that AI is very, very heavily Chinese. You go to any AI conference, you go to New York, you go to like, you know, AAAI, you go to any of these AI conferences. I would say like 60 percent of people have Chinese last names. I don't want to assume people's nationalities, but like 60 % of people have Chinese last names. And I always joke that if I just go in, you know, with my Chinese name, I would just blend in perfectly. And I would say that being, you know, like just being Chinese or being Chinese American in AI is a huge advantage.

20:44It's because like you will have a network of AI researchers like within your circle, like pretty easily. and my hot take here is actually that the ai race between the u.s and china is really a race between chinese and china and you know chinese and chinese americans in the u.s right so the tiger moms are doing something right here that's such a hot take but also sort of to this point what do you think americans in particular misunderstand about ai in china i think people always think that there's like a six to nine month gap or even like longer than that between Chinese models and the frontier US models and people always think that that gap is going to be there for the long run right because oh we have more resources we have more GPUs.

21:28They're assuming a steady state. We have yes exactly we have just you know more but I think the reality is that the Chinese labs are growing really really quickly because they iterate really quickly and also because they have a lot of unique data as well like I know a lot of Chinese labs are paying mathematicians from Qinghua and Peking University to do data labeling for them and the cost for doing that kind of high-end data labeling is a lot cheaper in China than here in the U.S. because you know how can you pay someone say$30 an hour and do mathematical labeling for you right that's just on you know that's unimaginable if you're in the U.S.

22:02it's a lot more expensive than that and then also there's a lot of distillation happening in the Chinese labs as well. People don't like to talk about it, but I know for a fact that like almost all the Chinese frontier model labs are to some degree distilling American AI models. And it's very, very hard for the American companies to stop them. And that's another reason why the gap between the US and China is like decreasing pretty quickly. And I would say like, don't assume a steady state. I would actually predict that in 2026, the feature velocity coming from open source models is going to exceed closed source models.

22:40And you're going to see models in China ship new features and then OpenAI and Anthropic are catching up and copying them. Sort of for our audience, actually, what is distillation? So distillation basically means that you use OpenAI's chatbot or Anthropic's chatbot, right? You ask ChatGPT a bunch of questions and you say like, hey, you know, what's the, I don't know, like what's the distance between Earth and the moon? And then you get all these answers and use those answers to train a smaller model and that way you kind of absorb all the knowledge from the bigger model remember when deep seek came out you were one of many people i called being like what's going on here um one of the things i found really interesting about that moment was the absolute shock do you think we're in for another deep seek moment i actually think there's a lot of new model that have come out that exceeded the performance of deep sea but because we had the deep seek shock already we're not as shocked about it yet.

23:36I think the next DeepSeq moment will be the moment where you see an open source Chinese model exceed the performance of the state-of-the-art American model because the DeepSeq moment was DeepSeq overtaking OpenAI's O1 model, but O1 was already like a, you know, last generation model for OpenAI. So there was still like a nine, six to nine month gap between DeepSeq and frontier OpenAI models. But I think the next moment is going to be like Quinn or DeepSeq or whatever model in China really exceeding, say, D55 or Claude, you know, Clop 4.0, I don't even know, 4.3 or 4.0 whatever. And then people are going to be really, really shocked when that happens.

24:16What does that look like when that day comes? Do you think it changes anything we do on the ground here? Do you think it speeds up that pace we were talking about? What do you think happens? Let's play out the scenario for a second. I think what's going to happen from a foundation model company perspective is that they're really going to rethink their whole API business. And that's actually one of my predictions for 2026, just because the gap between open source and closed source is closing so quickly that the foundation model companies are going to rethink whether the API business is a good business to be in in the long run.

24:46Just because if companies are able to use a model that's like almost as intelligent as like Claude or ChatGPT, like what's the point of like calling the OpenAI API, paying like a 10x, 20x premium for a closed source model? And I think that day is happening. It's going to happen like very, very soon. And then it's very hard for the foundation model companies to perpetually maintain a lead over open source. And then from a policy perspective, I actually think that's going to be when the U.S. really tightens export controls and tries all sorts of ways to, you know, limit China's AI progress. And this could mean banning GPU exports, banning all sorts of GPU manufacturing equipment, and anything related to computer AI.

25:30So it could facilitate a business model shift and a broader geopolitical shift, essentially. It could be that big of a deal. I think so. I think so. I think it's going to be a real shocker when it happens. Now, let's talk about the AI bubble. You think there's a bubble, right? We've talked about this. The question is, how big is the bubble? I think before we talk about the AI bubble, I would like to think about what is an AI bubble. And essentially an AI bubble is when there's a gap between value and valuation. Now, a lot of people say that there is an AI bubble, but I think the big issue is no one tells me exactly how you should value AI companies because AI companies aren't supposed to be valued at zero.

26:10So when you say that there's a bubble, it's very intellectually lazy for you to say, oh, there's a bubble, but you don't really tell me how how much you should value the AI companies now what would the valuation be like if it wasn't a bubble yes exactly and one of the concepts that we introduce and one of our research pieces is called zero value threshold and basically what this means is that when when companies when AI companies cross a certain threshold their enterprise value goes to zero for foundation model companies this basically means that if they're overtaken by the you know the best open source model that their whole business goes down right we've seen this happen to a lot of second tier foundation model companies over the last couple of years where they're not as good as the open source models and therefore they go to zero.

26:51We've seen this happen to application layer companies as well when their whole business gets absorbed by OpenAI or Anthropic or any of the foundation model companies. So that's where the zero value threshold kicks in for AI businesses. And what this means is that because there's a zero value threshold, AI businesses should not be valued as expensively as SaaS. it should actually be valued cheaper than SAS. The multiple should be lower for SAS because there's a zero value threshold. Now, aside from that, there's a couple of other factors that also means a point to the direction that AI should be valued lower than SAS.