Episode 8: Zero Revenue, Billion Dollar Valuations? Menlo Ventures’ Deedy Das on neolabs, Anthropic, and China

Deedy Das 00:40

Not all research problems are economically valuable. No investor can look me in the eye and tell me here's why this sort of zero revenue company is worth 1 billion versus 10 billion. I don't know how that math works. I don't think any of us know how that math works.

Yiwen Lu 00:56

Deedy happened to be one of the investors who had been following Neo Labs from the beginning, and he also offered sharp analysis on those companies' economics.

Deedy Das 01:04

So if you don't know your growth, then it's a little bit of gambling. Buy too less, I'm going to lose. If I too much, I'm going to be bankrupt. You never know how the big labs are going to change. They have this big compute cannon that can totally fire it in your direction and obliterate your chances of doing anything in a domain.

Yiwen Lu 01:19

Meanwhile, with Anthropic's IPO on the horizon, Deedy also offered unique perspectives on the culture of Anthropic, the compute challenge, and the Anthology Fund.

Deedy Das 01:29

From day one, they said the path to intelligence is through coding. People are paying a lot of money, like hundreds of millions of dollars, to get people out of Anthropic, and they don't leave.

Yiwen Lu 01:41

Besides Neo Labs and Anthropic, we also talked about Deedy's portfolio companies, including Goodfire, OpenRouter, and Pangram, as well as China and open-weight models.

Deedy Das 01:51

If you look at the list of Neo Labs, there's almost zero chance all of them work. But some of them might don't build for where the models are today. Build for where the models are tomorrow. So your product kind of shouldn't work today for it to really shine in the future.

Yiwen Lu 02:05

I often hear Deedy shared fun anecdotes from his interviews, and he did not disappoint us this time either. So here is my conversation with Deedy Das. Deedy, welcome to Valley101.

Deedy Das 02:19

Thank you so much for having me. Of course,

Yiwen Lu 02:21

I want to maybe start with yourself. You know, you have been at Menlo for over two years now. You came from Glean, and then before that, Google. Now you kind of look at founders who are raising a lot of money, and before they even had a product. So I wanted to hear more about, you know, what did you think that you understood about startups before that you know being an investor changed your mind about.

Deedy Das 02:46

I was you know always extremely interested in the idea of ever since I was in high school I would like think about the people in my class and I would just kind of make predictions like hey this person seems smart totally gonna make it. This person, I don't know. He seems smart, but it seems like it's gonna trip up somewhere. And I didn't think I'd ever be in venture. But when I was at Glean at the startup, and I saw it from the very beginning days in 2019, over four years, one of the things that you know, when you're embedded in the startup ecosystem, what I would do is I would look at other startups very deeply, and I would say, do I think this is going to work or do I think it's not? And that's where my fascination for venture sort of started. And I didn't think I would ever switch, but when I did switch, I think the big difference that I might have added versus many other venture capitalists was the fact that I take a very technological-driven view. I mean, my training is in computer science. My undergrad and master's was in like graphics, robotics, and computer science. And so when I look at companies, especially on the early stage, I'm evaluating a lot of things.

But primarily, the two big ones is you know the quality of the founder and how technically adept they are, which is a lot easier to do if you're technical too. But also just the strength of the like the technology. Like, have you really thought this problem through, or are you just pitching a good problem? It's so trivial for people these days to go on Claude, generate a slide deck, tell me how I should pitch this, and then pitch something. But how do you know if they they really know what they're talking about or if they just don't? So to answer your actual question, like what did I learn since I moved over to the investing side that I didn't know before? Is a lot of the way we judge companies, not just early stage, but every single stage, is about the founder. And I know this is a cliche thing to say, but like the way I phrase it is, like think about it. We have we have friends, and I'm sure you know people who want to be founders. It's so mainstream today, social network onwards, that everyone wants to start a company. But how many people want to start a company because this is their burning passion?

And they need to see an idea in the world versus people who want to start a company because you know being a CEO is awesome and being a founder of a company is a very high status fun thing to do. It's really there where the distinction lies because some of the best CEOs that we work with, you know, they you know they could make a billion dollars, they could make $10 billion. I don't think their mindset changes at all, and I think that's so critical for the venture business because you know what happens when a company gets to a billion dollars. Is that founder going to give up because they think they're successful, or is that founder just so like doggedly ambitious that they're going to go the extra mile and you know honestly sacrifice so many other things that a normal human wouldn't sacrifice to make sure that their vision of the world is successful, and thinking about figuring that out about a founder is is sort of like the hardest the hardest thing in venture. But I think that's really where what I learned is means more than even the idea or the approach. It's really like how determined is this person?

Yiwen Lu 06:02

Yeah. So when when you moved from Google to Glean, you spent a lot of time evaluating which startup to join as well. And now you're in a position as an investor. People came to you, ask you, you know, how what advice you would give them regarding what startups to join, and what types of founder would you advise them looking for? Like, what would you tell them?

Deedy Das 06:22

I this ask comes in all the time. So here's what I would say. The first thing is, I sort of never understood. I understand why people think this way, but a lot of people make I think the mistake of trying to find a company that's already a unicorn or some arbitrary valuation as a proxy for success, being on a VC on the other side, I can tell you that the valuation is very loosely tied to whether this company is going to be successful or not. So the first thing I would say is like, a don't look at the high valuation and think, hey, this is a company. Maybe this is a company I should join. I always recommend try to join a company as early as possible, as early as you think your risk tolerance lets you join. There's a couple of reasons for that. Number one, the earlier you have alpha, you can join an early company and get a meaningful amount of equity that might be a life-changing amount of wealth for you. That's not true typically when you join a company at a at a billion plus valuation, it's much harder. The bar for your economic guarantee is much harder. That's number one. Number two is you look at a typical like Series B, C company that a billion plus.

Not all of them, but they tend to be growing in employee count a lot faster. And so the reasons people want to join a startup is, hey, maybe I have more ownership, I have more growth, I can do more things. But they join these companies and find, oh my God, things are an absolute mess. I feel like there's 20 people coming into the office every month that I don't recognize. It can be very jarring for people joining, you know, a real growth stage company. Versus when you're at an early company, very often like things are flat until you sort of figure out how the product should work and whether you have product market fit. So pre product market fit companies, those are the best companies to join and the earliest possible. So that's the premise that I always encourage people to do. I don't think it's a it's a hard and fast rule, but I always encourage that you should start. Then the second question is how do you find them? And there's a couple of strategies I have. Strategy number one is: our job as VCs is not just to invest in companies; it is also to hire for our companies. And are we biased for our own companies all the time?

However, if you talk to enough VCs and you ask them, "Hey, can you give me some time? If you're a talented person, give me some time and just tell me what are some interesting companies at this stage, at this size in your portfolio that I may want to consider. And then what I would do is make an Excel sheet. This is I literally how I did it. Write down all the names, put the websites down, funding, who funded it, what stage, etc. When it was founded, who the founder is, founder's background, and then just ask yourself, okay, what do they do? Who is the competition? And just try to like role play as a VC for a little bit. The second big mistake that people make in the whole startup journey, the startup joining journey, is they don't spend enough time on the problem. Joining a startup is very different from finding another job. When you're finding another job, typical job, you're like, look, I know the 10 companies that I want to work for. I'm just going to apply to those 10 companies, and then I'm going to spend all my time on nailing the interview. This is very different in startup world.

In startup world, you want to spend three months, I would say, studying the market and understanding what you're looking at, and then the interviews are actually much quicker once you do that. So if, for example, if I had like $1 for every time my founder said, I got this candidate. He was strong. He didn't even understand what we do, and that's almost the most important thing for a company. What are they looking for?

Of course, smart, talented people, but they want somebody who comes in going like, Look, I've studied your business. I know where you are. Here's your competition. Here's your product. Here's some things I don't think you're considering, and here's how I think I can help you. That's the sort of pitch you want to make to an early stage company, which is compelling. And most people just show up like a job interview and say, "Hi, software engineer for five years. Please tell me if I can join." That's so uncompelling. So those are the kind of ways that I would look at joining a company, and that's how I advise people to do it. That's how I did it at Glean too.

Yiwen Lu 10:25

How often do you meet founders who, on the other hand, cannot really explain what they do?

Deedy Das 10:30

All the time.

Yiwen Lu 10:31

Yeah.

Deedy Das 10:31

All the time. We typically try not to invest in those companies. Yeah, makes sense. You can tell. Yeah. All the time.

Yiwen Lu 10:36

Yeah. I guess you also were very public about a lot of those advice on X, you had a huge presence there. Do you feel like publicly stating your opinions and being, you know, having such a strong opinion, does it make it harder to change your mind later?

Deedy Das 10:53

I think I use Twitter and social media as a way to put out my thoughts and analysis on things, a lot of people might read that, and then this happens all the time. And then they meet me, and they say, "I saw this tweet of yours. Is this something you really, really believe in?" And my answer is like, "No. This is the evidence I had at the time. This is the thought I had at the time. If the evidence changes, my thought will change. I don't live and die on these beliefs. I can't do this at my job. Like my job involves like ingesting new information, processing it, and then having like a new opinion based on the things I know. So, I think there's some downsides with social media. Obviously, you might get you might say something stupid. People might say, "Oh, your opinion is wrong." You might say something that you predict will happen in the future, and it doesn't happen, and you look really stupid. So I try to keep it as factual as possible.

When I do go into opinion territory, I take the risk of being wrong, but I would much rather use that channel as a way of being able to express how I feel about things, because the maxim that I have is if you're posting something and it can add a lot of value to at least some portion of people, you should just post it. So that's that's how I I run my social media. No ghostwriter. Everything is written by myself. No AI. Nothing. That's how I run it.

Yiwen Lu 12:16

Yeah, yeah. I learned a lot from your tweets, and I wanted to get into one of those things that you commented on and had analysis on it is Neo Labs. You were one of the first investors to popularize and codify the term Neo Labs. It has been, I think, like nine months ish since you published the ultimate list of AI Neo Labs. Do you feel like? How do you define Neo Labs now compared to then? Like, do you feel like your definition changed? Like, do you think the companies that we once thought of as Neo Labs has that changed as well? This

Deedy Das 12:46

is a fantastic question. The first thing I'll say is I can't take full credit for defining Neo Lab. I'm not sure who did it, but I'm pretty sure it wasn't me. Like, I think I've heard it somewhere, and then I'm like, that's a really nice term. Yeah. Okay. What is a Neo Lab? I have struggled with this question a lot since publishing that list. This is why I haven't published. I've published two of them, but I haven't been able to do it more regularly because my definition keeps changing. Here is my current best definition, and I think you could poke at it at the edges. Number one: Did you raise a significantly larger amount of money than usual. Again, variable. I don't have a strong amount, but say 50 million or more for your seed round total, or your early rounds total. So that's number one. Number two is the capital you raised primarily for large compute purchases, GPUs typically, right? So is is that is that the case too? Number three is your team largely made up of you know world class researchers in some way. Number four, and this is a controversial one, is typically you don't get to revenue very quickly.

And the things so that's my definition when I first published the list, and I think everything in that list, roughly was like abided by that definition. The things that I struggle with, one of my you know angel investments and a good close friend of mine runs a company called Cartesia, and I remember he sort of nagged at me at on that tweet, and he was like, "Well, just because we have revenue doesn't mean we're not a Neo Lab," and I'm like, oh, you know, like fair point. And so then it becomes a little difficult because now some of these many of these Neo Labs have started generating significant amount of revenue. Like, okay, well, then are they still Neo Labs? Are they functioning companies? Is there a ratio of revenue and evaluation that I that I care about? I don't know. But I think broadly that spec is still sort of useful to know. And the last thing, maybe I'll say is typically all Neo Labs are going after some interesting orthogonal idea in AI that the bigger labs are not going after. So that's how I would define a Neo Lab.

Yiwen Lu 14:54

Does it matter that much, you know, whether you call them a Neo Lab or not after they start to generate? Revenue like does the term matter at all? It

Deedy Das 15:02

matters to them. Yeah, I hear a lot of Neo Labs ones in my portfolio and not in my portfolio. They they they don't like being called Neo Labs. In fact, I'll tell you one such hot company recently. I walked in and I met this founder. I could probably say who it is. Sure, he wouldn't mind. It was Diogo from TypeSafe, and he's such a fun guy. And he was wearing this shirt called "Your Favorite Neo Labs Favorite Neo Lab." And I asked him a little bit about like, "Hey, like, do you do you consider yourself a Neo Lab?" And he's like, "No, some other people do, but we don't." And yeah, it's debatable. I think people think there is some sort of negative connotation, especially with the revenue comment, and then they see certain other Neo Labs and go like, "Yeah, I don't want to be like them." And so I think that's how people why people take some concern or issue with the the matter. But yeah, I think largely it's a useful definition for us to have.

Yiwen Lu 15:55

Yeah, yeah. A lot of those founders were formerly researchers. They came from an academic context, and you mentioned one of the I guess key components of a Neo Lab is compute. What is something that say you think would be better fit for a researcher to do versus for a company to do? Because a lot of the problems they're solving are like scientific breakthroughs, or at least you know things that we would often see in the academic context.

Deedy Das 16:15

The simple answer. Well, firstly, I'll just correct you quickly. Not all of them are academics. Some of them are like industry researchers. Many of them are academics, but could be both. I will say, why? What is the line between research and a company? Such a great question. We think about this time, this thing all the time in venture, especially because there's so many professors who tell us, "Hey, this is the research I'm doing at a lab. Give me X million dollars, and I want to do that research as a company, and I think the there's there's several parts of that answer. The simplest answer is some research just requires so much capital and compute that you cannot possibly do it in a lab academically. No one's going to pay you that amount of money. We know the cost of a B300 today. Like you know, even getting 1000 B300 is just not economical for any lab to get. So, part of it is just like I just need the compute to even feasibly do this research. And as we know, so many of the AI developments in the last few years have been as a result of scaling laws and scale. So, if you don't have scale in compute, it's typically it constrains the set of ideas that you can actually work on.

So that's one reason. The the second reason that is why people would start companies rather than doing it in a research lab is I just think companies with the intent of generating profit are far more focused and driven to achieving a goal than in academic lab, because the incentives are wholly different. If you ask any professor in academic lab, they will tell you, "Look, get the insight, get the idea, publish paper, NeurIPS, amazing. I am this is awesome." And then the novel idea is sort of done. Like, hey, do I need to do any more? Like, it's not that novel, and I don't know if I can get another paper published by just scaling that concept. So, you just typically in academia don't get research to translate into the real world, and in a company you do. So those are that's one of the reasons why you prefer to do it at a company than research. And you know, obviously, OpenAI and Anthropic and all of these labs are examples of companies that have done that.

Yiwen Lu 18:25

Yeah, yeah, that's really interesting. I guess one of the critique people have regarding academia these days is that a lot of those professors were going to labs and they're like starting their own companies and not pursuing the kind of research that you mentioned that that probably wouldn't be able to turn into real world applications. Is that a problem? Of course. Yeah. This is why I said it

Deedy Das 18:46

was a great question. Yeah. Not all research problems are economically valuable. Sometimes it takes. I mean, in some cases it takes 10 to 100 years to realize economic value from from research. However, some of it, some academic research is immediately economically valuable, or you know, investors would say we will bet on maybe this is economically very valuable, and so ultimately the investor bet when it comes to funding researchers is, do I see a short term you know economic incentive for this line of research, or no? And we can be wrong, and I think you know we'll inevitably be wrong. If you look at the list of Neo Labs, there's almost zero chance all of them work, but some of them might. And so it's just we try to thread the needle. At least I try to thread the needle between like, are you doing something that can't be economically valuable that is distinct enough from other things that also definitely needs capital to be done.

Yiwen Lu 19:46

Yeah, yeah. I want to get into some examples of that we were talking about Elorian before. I feel like that's a pretty good example of the thing that a Frontier Lab cannot do. They do visual reasoning models. What impressed you the most about their results?

Deedy Das 19:58

The you know I. Elorian, you know, is founded by this person, as you know, called Andrew Dai. He used to work for Jeff Dean back at Google, and there was a Harvard professor on board. There was one of the early xAI employees. There were some Apple senior AI veterans who designed the first LLMs for Apple. So phenomenal founding team there. I will say many Neo Labs have a very good founding team. The thing about Elorian that I found very interesting is their core bet was very simple. They said visual reasoning as a task, and visual reasoning. I'll give you an example. Like the example that I tweeted about, I thought was quite nice. Is everyone at this point has seen the sort of Tesla demo of the car and the thrusters and doing the backflip. If you were to put that video into a model and say, "Hey, explain the physics of this," that's a pretty hard problem to do from video. It's that's that's what we call a visual reasoning task. Elorian and Andrew's bet was most big model companies are just not focused on that task, and even if they do it a little bit, their models that do that are so much more expensive that it's very tricky to deploy this at scale.

If you need to do something like content moderation for TikTok, for example, that's a really hard thing to do if you only if you have a model that takes consumes so many tokens and costs like so much. You need to get the cost down. You need to get the model to be smaller, but you still need the model to be very visually like precise in their visual reasoning. That was the premise of the company, and his second premise was the model labs have different problems to worry about than visual reasoning. They are laser focused on code as a path to intelligence. So I think there's going to be a gap in the market where I can sort of fill that gap, and I would say you know that is a little bit of luck. You never know how the big labs are going to change. They have this big compute cannon that can totally fire in your direction and obliterate your chances of doing anything in a domain. So for now, I think he's been fortunate that he's found a little pocket that it is at the Pareto frontier of performance. It's as good as Astra on visual reasoning, but it's 20 times cheaper than Astra. So that's that's what it got out.

Yiwen Lu 22:10

That's a very good point. So these are things that using frontier labs could do, but they're just not doing right now. But for so for the Neo Labs that are for for math, for example, like OpenAI just released that 722 math proof solutions, and that could be a challenge to a lot of the Neo Labs focusing on math proofs and stuff like that. Yeah, is that does that mean that they're going to pivot? Like, what does it mean for those companies that are competing with the frontier labs at this point?

Deedy Das 22:36

I would say it's good. It's a phenomenal question. I the way the theory in my mind about startups is you cannot possibly go into startup building without acknowledging the risk that highly talented people at a bigger company might be able to absolutely overthrow what you

Speaker 2 22:55

do.

Deedy Das 22:56

It's sort of the classic like you know what if Google did this question that people used to ask startups when when Google was like sort of consuming everything in tech. Now the reality is the question to really answer is okay. Google or OpenAI or Anthropic could do a lot of things. What will they do? And I think that's where you know the value of startups sort of blossoms. I think that because all of these companies are so laser focused on the revenue right now, and a lot of that revenue is coming from coding, that a lot of these other things that they're doing is not necessarily in their roadmap. But math, you're right. Like you know, I mean, the results that OpenAI had in math were very impressive, and it just came as a side project of the main thing that they were doing. Yeah. And and does that mean the math companies that are trying to do this in as Neo Labs pivot? Not necessarily. It just means that they have to think about how they do things differently. You know, there's many other ways to think about novel math problems, can you think about getting to those solutions in a more data efficient way?

Can you think about getting to those solutions in a more compute efficient way? Can you think about better ways of formalizing those solutions? Can you think of ways of writing those reports that are actually readable by mathematicians, which today you know most mathematicians say these are very hard to read and understand. Those are all open questions. So that's where I lie on on that topic. Yeah,

Yiwen Lu 24:23

yeah, totally. And to build on that, you wrote about the economics of Neo Labs. We touched upon some of those points already: compute utilization, inference. But what are the economic challenges for for Neo Labs for this category broadly?

Deedy Das 24:36

Look, it's very simple. Like everyone is competing for one constrained resource. That's compute. The compute prices are going ballistic. They're going up and to the right, and the reason they're going up and to the right is because two major buyers, or you know, call it like maybe five major buyers, hyperscalers, the two big labs, are you know aggressively buying up compute, and there's high capex. Expenditure, etc., etc. Okay, what does that mean? If someone is buying compute at a certain price, they believe that they can convert sort of a price per GPU hour into more money for that GPU hour, and the the bigger companies have a lot more reason to believe that. Why? Because they have distribution, they have done this problem for a longer time. They've probably baked in a lot more efficiency into training these models. They know that they don't have to spend that much on additional marketing. They have distribution for these models. So there's a bunch of these things that stack up into why a bigger company has economies of scale when it comes to converting a price per GPU hour into revenue per GPU hour. Okay, what does that mean for everybody else?

If you don't have those advantages of economies of scale, you are still paying the same price per GPU hour, but now you have a much higher bar to clear to make this company economically viable, right? If you're paying, I don't know, a spot instance of B300 what is like $8–$9 per hour? Okay, how are you going to make $8–$9 per hour out of this? Even if you do basically take the most basic business, you do inference. You just take some GPUs, you think you know how to run a bunch of open source models, and you host them on OpenRouter to get some inference done. A lot of companies doing that. Even if you do that, you're going to have utilization problems first. You might get all the revenue, but you might get to only 50% utilization because you don't have demand baked in because you're a new company, and now you're losing money still. So that's why it becomes very difficult for new companies to compete. The second reason is, hey, and this happens to a bunch of Neo Labs, is maybe you're not very commercially savvy with these agreements, and you signed up for happened to a lot of people. You signed up for a three-year compute purchase for $100 million.

You put 15 million down, and now you don't have utilization because you ran this one big training job three months in. You're like, okay, well, I'm sitting on these GPUs. I don't know what to do with this. Okay, so you're paying interest and you're you're paying this aggressive cost. Your runway is going to run out. Your business may or may not make money at this point, and you might have to keep fundraising to survive. So, my point behind that tweet was: any startup is challenging. Neo labs, particularly, these are the challenges that you have. You raise all this money because you need to afford the compute, but that doesn't mean that it's easy to create to like create cash from compute. Yeah, it's very difficult. So that's the challenge that they have compared to the bigger players who have these economies of scale.

Yiwen Lu 27:35

I guess, given that backdrop, what would you make you invest in a Neo Lab now versus pass on one?

Deedy Das 27:41

It's it's it's a very multifaceted thing. Let me think of some some some some examples. Let me go back to the founder thing I said. Yeah. The first and foremost thing is, are these founders uniquely capable of doing something that most other people can't? I can give you some examples of that. One of the one of the Neo Labs I would say we invested in and led the series A for is a company called Goodfire. They are they work in mechanistic interpretability. They had this amazing launch about yet last week was really viral paper about how they were able to figure out that polysemanticity in neural networks doesn't exist. They try to understand the inner workings of a neural network. Now, when I looked at that company back in the day, you know, these were the best researchers I thought in the world in mechanistic mechanistic interpretability. Now, you know, if you ask me, do you think mechanistic interpretability will work and be economically viable? I'd be like, look, there's some chance. I don't know, obviously, I don't know the future, but if it works, these guys are gonna do it because they're the best at what they do. And so I had to be really honest about myself.

Like, you know, everyone wants to say I'm backing the best at what they do, but I genuinely think, you know, you had the people who started the OpenAI interpretability team, which is one of the first interpretability researchers ever. So just the credibility on that team was just phenomenal, and so that's one example of you got to start with are these founders uniquely uniquely capable of being the best at what they do. The second example, I mean, we already talked about Elorian. I'll skip that. The second example I can tell you is this company called Core Automation. If you look at Core Automation, you Jerry Tworek, the CEO and founder, led the discovery of reasoning models at OpenAI, amongst many other things that he did. Yeah. So when it comes to okay, is this somebody who understands AI the most deeply in the world? He's up there. So and then the whole team. I could go on and on talk about the team was phenomenally good at every single layer of the stack, so I'm like, look, do I love the valuation? No, no one does. Do I think this is going to work? There's a chance.

He had some interesting approaches to solving this this problem of continual learning and why he thought the big labs couldn't solve it. And more importantly, when he says. I trust him because he's worked. He led all of research at OpenAI. So those are two examples of companies with like leading top people. That's where it starts, and then they give you an idea where they're like, I think no one else is going to do this

Yiwen Lu 30:13

one. Yeah.

Deedy Das 30:13

So that's how I think about

Yiwen Lu 30:15

it. Just because we talk about Goodfire, I wanted to you know Interp is really fascinating. What are some ways that you envision interp becoming a business?

Deedy Das 30:24

There's there's there's plenty of ways. Firstly, there are many models outside of language models that are extremely hard to understand what they do. Genomics models, for example, that output DNA sequence tokens, those are very hard. You can't read a DNA sequence token really understand what's going on. So interpreting those models is very useful in figuring out, hey, what's what's actually going on. So they had a great paper where they published how they found a novel Alzheimer's biomarker by being able to interpret genomics models. Again, multimodal inference is another interesting one. Thinking about hey, can I? One of the things you can do in interpretability is something called steering, which is if you know how the weights work, you can also change the weights on the fly. And it's sort of if you do it right, and it's not that easy to do, especially across all models, you can figure out how to change model behavior at essentially a zero cost type of way or zero recurring cost type of way. You pay one-time cost to understand it, and then you pay zero recurring cost to change the behavior. That's very interesting in things like even speech.

Do I want to swap in and out a model, or do I want to take the same model and change behavior on the fly and react to how the speaker is speaking? Those are some examples. Then obviously the big labs when it comes to LLMs, and you know, Goodfire has some relationships. I can't say more with some of the big labs and talking about how can we interpret these models that are able to do superhuman tasks like math, like cybersecurity. Without that, RSI might be gated on that because models to be able to recursively improve models. How do you understand how the model is performing other than like these empirical benchmarks? The weights are sort of the the source of truth, and and and I think that model labs care a lot about this problem, especially having it third party and not being in house because there's some issues with that. So those are a bunch of these different paths that a company like Goodfire can go down, but it's definitely going to be a non-traditional type business. It's not like hey, everybody in the world goes and uses interpretability. I don't think that's going to happen.

It's really working with these major players that have models and figuring out how to get interpretability to to apply in these ways.

Yiwen Lu 32:37

Yeah, but that's really helpful because I feel like we usually associate interp with AI safety and alignment, but it sounds like there are a lot more than that.

Deedy Das 32:45

Look, it's extremely critical to have interpretability. We talk about this with the team all the time. Is yes, safety is great, but ultimately capability is also the most important thing in the world. So interpretability, you know, the analogy maybe that I would make is, if a neurosurgeon for your brain only told you how you were thinking, that would be somewhat valuable. But if a neurosurgeon could fix your brain when your brain is not working, that's very valuable. And similar with models, like there is some value to understanding. There's a lot more value to doing things with that understanding, and I think that, if everything goes right, could be fundamentally change how we think about developing smarter and smarter models. So that's really the bet.

Yiwen Lu 33:37

Yeah, yeah, totally. Going back to Neo Labs and exit options. World Labs, AMD just acquired World Labs. Does that frame? Does that? Is that? I don't know if that's the first Neo Lab acquisition, but does it provide like a framework of understanding exit options? I

Deedy Das 33:52

don't know if that's the first. It may be the most important one so far. You know, one thing that we talk about a lot in venture is what is the path for this company. How do we see our investment increase in size? It's ultimately our jobs, right? Like we all love technology, we all like interesting ideas and cool businesses, but ultimately our job is to return capital for our LPs. If you asked a venture capitalist 10 to 15 years ago how what their framework is for this, they will say, "Hey, this business has to make money. It has to keep growing. It has to have margin. It has to IPO, and it's going to take seven to 10 years." So if we don't see all the characteristics of that business, when we invest, probably not going to invest. So today, however, you know clearly that framework is very broken, which is why many investors who were very good in that era are kind of like lost for words today. They don't know how to think about things. I mean, what am I saying? Sometimes I don't know how to think about these things. Either how do you assess the this scenario, right? But like, like look at look let's look at the facts. I believe about nine out of 10 exits of all startups are M&A.

Now a lot of that M&A is at a value where you're not really making the kind of returns that you need in venture. But what's really changed in the last few years has been a lot of those M&A's are very good for the company and very good for the investor. These are significant gains and these are significant valuations. You talk about Hugging Face at 12.9. You talk about OpenRouter at you know you talk about World Labs at eight, Cursor at 60. I don't count the xAI one because that's just Elon buying his own thing. But all of these companies are significant outcomes. If you told somebody, hey, in seven years, if you invest in a Series A, your company will IPO at 8 billion, they'd be like, I'll take that every day. So what does that mean for in valuing companies? Well, obviously there then, and we made this mistake, I think, with World Labs. Clearly, right? Like, I think every company that's successful that we didn't invest in is a mistake. Let's start there. With World Labs, I'm like, look, is Fei-Fei awesome? 100%. There's no argument. I and I think I can speak for the firm. We spend so much time thinking about like, okay, the idea is cool.

Definitely, Fei-Fei is uniquely capable of doing this, but I really struggled with what are the applications, even if they get this right. How do you actually make money off of this? And that's why we didn't proceed further. Yet, you know, you look at an outcome like this, and you're like, man, that's incredible. It's really tough to underwrite. I think a lot of these things are, you know, you tell me if you find out a rule for which companies will get acquired and which companies won't. I think it's pretty vibes based. Sometimes, you know, you can conjure different reasons, but ultimately, I don't know what's going on in Lisa and AMD's mind when they bought Fei-Fei and World Labs, and I don't know which other Neo Labs might or might not be bought. Yeah. Frankly, so it's just good to know for investors to know. Hey, sometimes you don't have to understand everything about the business for it to be a very good outcome.

Yiwen Lu 37:10

Yeah. Yeah. It is very vibe based, and there's still no answer to the question you asked, which is what are the applications of this product? If I understand

Deedy Das 37:17

correctly, I don't think that business made any revenue, right? So like it's almost inconceivable to think of like a zero revenue company being bought for $8 billion. So how do you underwrite that? I mean I can just swing checks into random companies and go like, hey, maybe someone will buy you for 8 billion, but that's like not a very you know effective use of capital. So it's tricky. Those things are tricky.

Yiwen Lu 37:39

Yeah, makes sense. But then when four the Neo Labs that are raising their next round, you know, other than you know we talk about founders and that association between the decision to invest, but do they need to do anything to prove that they’re worth the valuation from the last round? Like what what's going on?

Deedy Das 37:57

Hey, look, if the acquisitions are vibes based, the valuations are pretty vibes-based too. That is the reality. I mean, no investor, and if they can, please come and tell me, can look me in the eye and tell me here's why this zero revenue company is worth 1 billion versus 2 billion versus 5 billion versus 10 billion. I don't know how that math works. I don't think any of us know how that math works. I can say that there's a couple of things that you know have clearly come out of this Neo Lab funding frenzy, which is increasingly common, and we can talk about this more if you want to. Is the idea of a tranched round structure, where you know I think Sequoia actually talked about this. Pat Grady actually talked about it's one of his podcasts. I think it's the first time I've seen someone publicly talk about this, which is the idea of if a company needs a lot of capital, but the investor needs to make their return.

What if the investor comes in as a capital partner at a much lower price, which you know a lot of people criticize because that price is never announced, and then you bring in more capital at the higher price, in which case you know you're like you know I kind of didn't pay the high price for it. My blended price is much lower, but someone still paid the high price for it. So it's kind of this idea of like hey, some investors care more about returns than others because nothing fundamentally changed in the business. So doesn't really answer your question, but one way people get to these absurd valuations is they forget that this is actually a very tranched round, and people have come in actually at many prices in the middle before they came in at that final price.

Yiwen Lu 39:30

I see, I see. Oh, that's interesting. Is that that's common among Neo Labs and I guess other? It's

Deedy Das 39:35

very common.

Yiwen Lu 39:35

Okay, got it. It's

Deedy Das 39:36

very common.

Yiwen Lu 39:36

Sounds good.

Deedy Das 39:37

I would say at least 50% of the Neo Labs have been funded that way.

Yiwen Lu 39:40

Yeah. Yeah. Very interesting.

Speaker 2 39:41

Yeah,

Yiwen Lu 39:42

cool. Let's move on to China. I would love to hear your thoughts on China. You know, you were also very early on in pointing that China's influence in AI is well beyond DeepSeek when R1 was released. That was early last year. DeepSeek just released V 4.1 4.1. Flash. There are a lot of good receptions about that as well. How are you understanding China's AI ecosystem now?

Deedy Das 40:06

You know, I'm just gonna say there's my there's like leaving all the geopolitical relations with America aside because then look, I'm fundamentally I'm a technologist. Is China's value in the AI ecosystem has been incredible. I mean, it's just I honestly think it's underrated the amount that China has done for even the the U.S. startup ecosystem. How many of the companies you I mean kind of name the top 30 companies outside the big labs? I mean, most of those companies would cease to exist without open source models, and all the open source models kind of came from China. I think China has figured out many many things. So let me let me start with some of the major ones. Getting open source models to be cheap, kind of obvious. They did that. Many many such models: GLM, Kimi, DeepSeek. All of these ones are incredibly big drivers of the ecosystem and what people expect on prices. So it helps the app layer think about how I can price my product cheaper. It helps the inference companies think think about hey I can host these models and and and make a business there. So that's been one big way.

Number two is some labs more than others, but have been extremely open, DeepSeek in particular, about sharing their research and methodology to a very deep level in terms of how do I rack my GPUs? How do I actually serve my inference? What are the tricks I'm using under the hood to make this model work in a more token efficient way than than others? You know, any open researcher or someone who likes open source or open science loves that because ultimately the American labs don't do that, so that's been incredible. And and the kind of third, maybe if I had to cut it off, there is I have a couple more actually. Third is I think the idea that making models is expensive, they completely dismissed. And I think this is particularly important to the seat I live in as as a VC. Is you know I'm looking at some of these these some Neo Labs, and I'm like, well, I know you're asking me for this amount of money, but like I know that in China they're training these models who are clearly, in some cases, better than yours, asking for way less money and able to do it. Like what's what's the difference?

And if it weren't for China, I think I couldn't even point to an example where it's possible, and so I think it's very important to sort of reset expectations on what the cost structure of AI even looks like. So I think that's phenomenal. And then you go down the stack, and I think it's just been incredible, right? Like you know, you you go to like sort of AI adjacent areas, and when it comes to robotics manufacturing, obviously, you know, I don't think America is even close today when it comes to, you know, I was reading about like more recently about like CXMT and like memory manufacturing. Like again, America really doesn't come close. So obviously at the model layer too, and then everything down from that, and hardware besides chips, I think Huawei still has some work to do. Their innovations have been pretty incredible. And then north of the model layer, maybe I just don't know enough, but about the Chinese AI app ecosystem, I hear some numbers and I hear some like apps exist, but hard to ascribe like core innovation. I think the biggest one is Manus for me. Yeah, I have a Manus story, which is when I first saw that product.

Sort of, I think it was about two years ago or something like that. I first saw that product, and it wasn't very popular then. And I sort of like played with it. I'm like, this was for me. I still maintain that was agents for the first time. I've never seen something like it because before we were still very much in like Q and A and reasoning mode, but there was nothing that could do a task for a long period of time. And I I'm playing with the Manus product, and I think the first query was maybe one of their query on their website is like, plan me an itinerary for my two week trip to Japan, and it was amazing.

And I'm like, this is it made a website. It had links. It had calendars. It had like here's what like a little pop up of a restaurant you should go to. Here's why you should go to the restaurant. Here's the reviews of the restaurant. Like everything just created this entire like website like for me to to consume, and I remember and it ran for 30 minutes to consume to make this, and I was so so impressed by the product quality. I I'd go so far as to say usually when I use products from China, I don't like maybe the the design taste or whatever is different, so I'm not always that impressed. Even like with the Kimi or DeepSeek core product, I'm like, okay, this is good, it's whatever. But Manus, I was like, this is slick, this is polished. And I remember I tweeted about it. They sort of blew up. I reached out to the team. I remember they came out here, and most of that team spoke very little English. Spoke a lot of like mostly Chinese. But his answers were amazing, like the way they think about product, the way they think about building.

It was like this is the difference I think sometimes between Chinese entrepreneurs and like American entrepreneurs is American entrepreneur a lot of like show, a lot of like grandiosity. Here's what I'm gonna do for the world, and you know I remember Red from Manus is like, look man, there's no moat. Very broken English. Like there's no moat. I'm gonna crush my competition because I work harder than my competition. So if they can copy me, great. By the time they copy me, I will be ahead. And I'm like, it's a phenomenal answer.

Yiwen Lu 45:32

Yeah.

Deedy Das 45:33

There's no like grand like I'm doing this because they're gonna do that. Like no, you do whatever you want. I'm gonna be. And I just love that energy.

Yiwen Lu 45:40

Yeah, that was a great answer. Actually, I want to hear your thoughts on that because a lot of times the first innovators were usually not rewarded. Like now, we see Muse, Instinct, Grok, Bot, like all the agent products, and and Manus is kind of like put into this place where they had to keep innovating to sustain their business, and this applies to a lot of the other companies too. How are you thinking about this in general?

Deedy Das 46:04

Well, I mean, Manus Journey is interesting because they sort of got acquired and they sort of got like not acquired. I'm not really sure. Yeah, I keep hearing different stories about who made money and who didn't make money. So I don't really know. It's hard to basically unconfound those two things. Like I don't know if Manus would be like very successful in personal agents if they hadn't been acquired. I think it's fair; they might have been very successful. When it comes to, I guess your question is like, you know, how do you think about incumbents versus competitors when it comes to these things? And muse and instinct. I think that the incumbent, the competitor, has a time horizon to capture critical mass to be able to compete with someone with a distribution advantage, and that's how it's always worked in startups, especially when it comes to consumer products. The cleanest example of this is why ChatGPT has a billion users.

Yiwen Lu 46:57

Yeah,

Deedy Das 46:57

you know, like did could Meta and Facebook and all of these other companies not try to do that. Of course, they were just late, and by the time they tried to do it, it's too late. ChatGPT already had the billion users, so clearly it's possible for a startup to compete in consumer and win. I mean, Claude has several 100 million users. ChatGPT has billion users. If you look at Instinct, I think the problem was just thinking about that particular company, and leaving Manus aside just because it got acquired. So I think it like it's a little bit different. With Instinct, if Instinct had three more months before Muse launched, could have been incredible. You know that would have been an interesting fight to see. The problem is when you, when when Muse launched, again I don't know the details about what's happening at Instinct, but I imagine it set them back a lot because now you're competing with somebody who's throwing their full force of the company at you, and it's really hard to compete as a startup. So, like one maybe a thought experiment is, if ChatGPT had launched, what it went did launch like late 2022, right?

If Gemini had launched in January 2023, then I think it would be difficult for ChatGPT to compete. But in this case, like Gemini, kind of like it was Bard. It was terrible. They did a gazillion things wrong, and like who knows even if the current app is good. And then like you know, they took like a year or so to like actually ship something reasonable. So yeah, I think a lot of it has to do with having the time window for a competitor to to scale.

Yiwen Lu 48:29

Yeah, yeah, makes sense. Going back to open models, where do you see founders are actually comfortable with using Chinese open-weight models? Like, do they draw the line at

Deedy Das 48:40

all? The reality. They draw the line at telling people they use it. But now, look, like I think all founders, all if you are, I mean, I don't like this word AI native, but if you are a company that is in the token flow and that sells intelligence in some form, then you have every incentive to use open-weight models because it's way cheaper, right? So some companies give their users the option to pick a model. I like that less a little bit. And then some companies say, "I'm going to give you a product, but under the hood, I'm going to use a bunch of models, and they may be Chinese models, and they may not be. Typically, when they do use Chinese models, they use the Fireworks, Modal, Basetens of the world to to host inference, which you know kind of alleviates the sort of I don't know if you if you think of like geopolitical risk of your data living in China, it sort of alleviates that. So they use it. That's typically how those companies use it. The main downside, I would say, practically of using Chinese models today, is competition. If like take two companies, I don't want to name them, but like this happens in several different industries.

Two companies offer an intelligence-based product. One is throwing the best frontier model at it for all tasks. One is like I want to cut costs. I'm going to maybe use I don't know Opus five five to do the orchestration, but I'm going to use four one flash to do the actual implementation. Okay, great. So the competition becomes now either I keep the same price and I will probably lose to this guy burning more money because his product is probably better quality overall. Or I lower my price and try to compete on price, and then try to undercut this person. And that's sort of the game that's being played. Like, are you competing on price? Are you competing on intelligence? How much can you burn? How much can you afford to burn? And and so those are sort of the pros and cons of trying to use Chinese models. If you do in some industries, if you have the inferior product, you will probably lose in the short run as well as long run. So they take the view of you know I'm going to eat bad margins for a while, and I'll think about this problem later.

And then in some other industries where it's less critical maybe to have the absolute best intelligence and cost is actually the decider, then they they undercut. Yeah. So a good example is like OpenCode is a good product that they offer multiple things, but one of the things they offer is a $10 a month plan for a Claude Code like agentic programming experience, and they primarily use open source models, and they're a big consumer of DeepSeek. And their whole offering is we're just cheaper. Are we going to be as good? Maybe not, but doesn't matter. Good for most things. And so those are sort of the ways that I see play out.

Yiwen Lu 51:35

Totally. Yeah, I guess also see parallel between that and OpenRouter. Like I guess that options is what benefited OpenRouter. When you invest in OpenRouter, is that like what did you see that other people missed?

Deedy Das 51:49

I there's a lot. I I love that company so much. I think the thing that people didn't see in OpenRouter is well a couple of things. One is you you have to believe at the time at the earlier rounds at least that this we were going to live in a world with multiple models. Most people didn't even believe that, which is you know to some extent maybe fair. They were like everyone's just going to use like one or two models and the rest of them are kind of whatever. No one cares. It's fine. The second thing is, I think like people struggle with venture people especially struggle with this this business model of it's essentially a marketplace business model. Like, what does OpenRouter do? The core product buys a bunch of inference capacity from a bunch of providers. Maybe they buy it on the spot.

Maybe they buy contracts, and those providers either host models or serve their own closed source models, and then OpenRouter gives their users an API to use all of these models in a unified API, be able to switch, be able to do all sorts of things like say I only want these kind of models, I only want those kind of models, and route me between them, and gives users that API, and then the business is taking a cut in the middle for that service. So some people just don't like that business model because they were like, well, if any big enterprise uses this, they'll probably churn, right? And some people said, well, you take a cut, but like overall the spend on OpenRouter is pretty small, so your cuts like pretty small, and you know maybe people just try different models for for in the beginning, and then they'll find one and they'll stick to it. It's not a recurring revenue business in the same way, right? So VCs really like to see Fortune 500 companies use me. They're paying me a million dollars each. They're super high highly recurring, and and they will love to invest in that stuff all day.

For this kind of stuff, there's all these risks, which is why people were a little bit more hesitant to to like this business. And the thing that I I loved about it is I'm like, look. My core thesis is, if you build a product that developers love, and that's one. And two, if it's if everyone thinks it's easy to build, but it's actually extremely hard to build, then it's a and you have a good business model. I think it works. I had built a version of OpenRouter at my previous company, Glean, because we had to basically have one API to route different customers to different models just based on what they chose and the kind of commits that they had on different things. When I was doing this, I started that project thinking this is a simple project. It's going to take me a week. And honestly, I was right. It took like a week to stand it up. What I didn't realize is every time a new model drops, it actually is very annoying to go and update again and again and again, and I saw a bunch of models drop at that time. Yeah. So when I left Glean and I found my venture job, OpenRouter was sort of my first deal. I was like, this is exactly the company I would have started. Yeah.

Because it's so obvious that everyone thinks it's easy to build, but no one knows how annoying it is to keep getting new models. To work correctly on your platform, so that's why I we decided to invest.

Yiwen Lu 55:05

Makes sense. Yeah. Why was Stripe the right buyer?

Deedy Das 55:09

I don't think a big lab would buy this because it's sort of a third-party aggregator. So they would buy it almost as a defensive acquisition, if anything, right? Like if if you're, you know, OpenAI, why would you want to buy the thing that serves other models? That's not you.

Yiwen Lu 55:26

Yeah.

Deedy Das 55:28

So okay, if it's not them, then who is who is a good buyer? So some, almost all the hyperscalers, a lot of them have affinity to their own models, so they're probably not good good good buyers too. It's really therefore like just by elimination, the only people left are like big-ish tech companies that have something to do with being sort of this gateway for AI. Now, of those big companies, why does Stripe like it? Like Stripe's mission is to be a part of the growth of the GDP of the internet, this was how they got started. Now, the the breakdown of the GDP of the internet fundamentally changed with AI. You have this enormous amount of spend on the internet that comes from tokens, and Stripe is sitting here thinking, well, well, technically they're a part of it because OpenRouter is powered by Stripe, yeah. So they are a part of it, and they also like kind of knew how big OpenRouter was getting, but they're like, you know, this is so in line with the thing that I'm trying to do at Stripe. I want to be the place where people go for independent financial transactions of vendors, right? Like think of it like Amazon for models. I hate that. I've never said that before.

But like you know, of course Stripe would be the payment gateway for a bunch of e-commerce stores. Why won't it be the payment gateway for a bunch of models, right? Totally makes sense for them to facilitate that transaction and take a small cut for that convenience. So that's why one of the reasons. The second reason is trust and safety, and I think people don't talk about this enough. But there are tons of token resellers. There are tons of fraudulent behavior in all of the AI apps of people creating new accounts, spinning them down, reusing services, trying not to pay their bill, and then reselling those tokens to other people, and I think you know it's very hard to build good trust and safety infrastructure for the first time. And Stripe, their business relies on good trust and safety infrastructure, so they knew how to do this really, really well. So those are the big two reasons why I thought Stripe was the right buyer. And to be honest, like if this is a more valued thing, but if if you ask anybody, like, what are some CEOs you respect the most? Patrick Collison is maybe like the most respected CEO in the Valley, just because of how purist of a technology technologist he is.

Yeah. So all founders, no founder I have met has regretted being bought by Stripe. In many other acquisitions, a lot of politics can happen. They might shut down your product. They might say one thing, but then you know reorg your whole team. Might fire a bunch of people. You know, there's all these things that can happen on acquisition. Patrick's reputation, 10 on 10, and so, and like lastly, Stripe is a great business. Yeah, totally. That's why. That's why that worked.

Yiwen Lu 58:19

Yeah, going back to China for a little bit, there are a lot of VCs going on China tours again after a couple years. Is that something that you guys are considering?

Deedy Das 58:28

Yes.

Yiwen Lu 58:29

Nice.

Deedy Das 58:29

We're gonna do it.

Yiwen Lu 58:30

Yeah. Yeah. Are you excited? I'm very excited. I'm very

Deedy Das 58:33

excited. I don't think people like I tell some of my colleagues this, but one of my side hobbies is, you know, my my, I'll go back home and I'll put on YouTube and I just watch random videos about Chinese tourism in different cities, and I just watch so much of it. I've seen like I think every video that covers Chongqing in the world, like every single one. I'm just like, oh, tell me, show me more, show me more, because I'm very impressed by by China. One of my best friend is a really good friend, at least, is Chinese, and she's a Chinese American. So you know, her family moved here. She was born and brought up here. And I remember catching up with her, and I was like, she says, when I, I'll be honest. She said, the last time I went to China, I kind of thought I was going to a third world country. I went this time, and I'm like, I think I'm flying back to a third world country in America. Like I can't imagine living here now. What is this? Like everything is terrible here. Like why is there homelessness? Why is there all this stuff? In China, everything was perfect. Everything was perfect. I can't. So he's she's like I was gonna go there for two weeks. I stayed for three months.

And she had no agenda and nothing to learn. She was just like I just want to travel and reconnect with my roots or whatever, and so I think it's pretty. You know, it's the country itself has changed, and yes, I mean, obviously, we want to go for to understand a lot about the economy and robotics and hardware. But I've learned a lot of. Everyone tells me that the country has changed so much. I want to go experience. That as well.

Yiwen Lu 1:00:01

So you had a funny story about a Chinese kid that you recently talked

Deedy Das 1:00:04

to. Yeah, that's right. So I was telling you that every every once in a while I'll meet some awesome kid and I'll just tweet about them anonymously. And every time I do this, people will be like, "This never happened." Like, and and I love because I'm like, yeah, that's how unbelievable it is. Because someone told me that, I would also say that. I meet this. There's this person on Twitter. I've been following him for a very long time. I had no idea how old he was or whatever, but I thought he posted some really insightful takes on AI and not not just takes like core technical depth. So he DMs me and says, "Hey man, I'm gonna be in San Francisco. Would love to catch up for like half an hour." And so usually what I do when it comes to these cold DMs, and it's kind of useful if anyone ever wants to cold DM anyone, but I'll go to their profile. I'm like, "Is it interesting or not?" And I look at his stuff. I'm like, "Oh my god, this guy's a genius!" So I'm like, "Of course I'm leaving. Come." So he comes over to this office. He's actually sitting right here, and I ask him. I say I'm not going to use his name to maintain his privacy. Like, tell me your life story.

And so he says, "Well, firstly, I'm like, how old are you? You're so you look pretty young." And he says, "I'm 17." Like, holy shit! Tell me, tell me a little bit about what's your journey been? So he says, you know, firstly, he says at 14, I barely spoke, basically didn't speak any English. I was in China, and I was sort of on the internet, and I learned about these new, you know, models that were becoming a big deal. So the first thing I did was everyone was talking about this paper called Transformers, and so I pulled it up, and I had no idea what was going on. So, and I could barely read English well. So I just copy pasted line by line into ChatGPT, and just said, "Can you explain in Chinese? Can you explain it?" And went again and again and again, and until I understood what was going on. And then I did more work on it, and then I read more papers, and then I found people on the internet who talk about it. And then slowly my English started getting better because I was reading all these papers in English. And then he says, I forget the exact age, maybe 15, maybe 16. He's like, I got amazing internship at one of Chinese, one of the leading Chinese AI labs.

That my understanding is very few people work there, so it's quite hard to get a job there. So he goes, does that internship when he's literally in high school for a year, crushes it, and everyone loves him. He's kind of full time, even though he's like he's an intern, and and you know he's doing very fundamental like core AI research work. So this is like a job that in the valley you get paid seven figures for. Says okay, this is great, but any kind of testament to his character, he's so he's such a growth mindset kind of individual. He says you know this is great, but I want to do more. He thinks that he needs to come to America to learn, and he wants to create more of a network and get to know more people, probably because what he sees online. So he comes to America, comes to one of the top 10 universities here, and enrolls in his undergrad program, and it lasts two weeks. And he says, "Yeah, this is kind of useless. Like these guys don't know anything, and they don't really want to learn anything." And I understand. He's right. So he's like, "Yeah, this is not it." So he drops out, comes to the Bay Area.

This is when we we met, and he says, "You know, I I just want to learn and work on AI. That's my only goal.

That's why I came to SF because the people were there, and I want to find different companies that I could work at." So we went through a list of companies that he was considering, and he said, "Yeah, I got offers from all of these five amazing, like, you know, I would say early stage Neo Lab type companies with amazing talent." And he's like, "Yeah, they all want to hire me, and this is what I just wanted to get your thoughts on which one should I join?" And I actually thought he was asking me like, "Hey, which one should I consider?" But he was actually asking like, "No, I got offers from all of them. Which one should I join? And so, and he's 17.

Yiwen Lu 1:04:06

Yeah.

Deedy Das 1:04:06

So the reason I tell that story is I just think it's so interesting how you know we often say things like, hey, you know, AI is changing the world, this that, but just to see in person, like a person who's 14 years old. At the very extreme, you know, sure, the median person might use ChatGPT to cheat on his homework, but at the very extreme, like he taught himself English, he taught himself everything he knows. He basically has one of the hardest to get jobs in the world, and he's 17. He never went to formal school, so I think that just is such an optimistic take on the value of technology. You can we love spending all day talking about ways this can go wrong, but here's all the ways it can go right, and I was just blown away by this guy.

Yiwen Lu 1:04:53

Yeah, impressive. Also, great story about how you can just reach out on X.

Deedy Das 1:04:57

Oh, 100%. Yeah, 100%. I'm gonna go and click on your site. I want to scroll down, scroll up, and be like, "Is this person interesting?" And then if it is, then it's a yes.

Yiwen Lu 1:05:07

Yeah, yeah, totally.

Deedy Das 1:05:08

And then sometimes I do it. Like I reach out to people on Twitter all the time just to ask them, like, "Hey, I saw this thing that you wrote, really cool, and I would love to meet you." So, yeah, I think it is at its best, used for like social, like actually like meeting people and things like that. My other great story that I had from, I don't think I've ever shared this publicly, but at the very early days of me using Twitter, one thing I would do is I would go and I'd read a research paper and I would try to use my tweet to sort of summarize what I had learned from the paper in like a very short form, so I did that for one of the papers. That was one of the initial papers about tool calling. So back in the day, it wasn't obvious to get GPT-3 or whatever to use another tool, and here's how you could make the output such that it would actually tell you here's how to call that tool and here's what input to provide it. So I posted that paper, and I remember I got reached out to by, you know, in my mind a hero of mine. It was Bret Taylor, who at the time was actually not at Sierra. He was the CEO, co-CEO of Salesforce, and he just said, "Saw this tweet.

Can you explain this to me?" And I remember I was at work that day, and I'm like, you know, cancel all my meetings, cancel everything at work. I actually have to reread the paper and make sure I get the best answer possible. So I did, and then eventually met him, and we had a great conversation. And he was telling me about Sierra, and he's going to start this company with his friend Clay. It was going to be a customer support AI company, and you know, sometimes I regret this, but I'm like, yeah, I kind of think that's kind of. But you know, I think you're gonna do a great job, and it was so. But it was really cool to see just that interaction came just from Twitter.

Yiwen Lu 1:06:48

Yeah, totally, totally. And I guess till this day, you still write your own tweets, like no AI, as you mentioned in the beginning.

Deedy Das 1:06:54

No ghostwriter, no AI. We're investors in this company called Pangram that is the best AI detector. Use Pangram, get the Chrome extension, scroll through my profile. You're not going to find a single thing that's written by AI. Yeah, I don't like use it as a just like a growth tool. I use it as a way to express my thoughts. And if I'm not writing my thoughts, then they're not. It's not really serving that purpose. So

Yiwen Lu 1:07:19

yeah, whether whether

Deedy Das 1:07:21

people like it or not, it's definitely for me.

Yiwen Lu 1:07:23

And Pangram is such an interesting company too. I I just saw Max replying to someone saying that they're not gonna add a humanizer like all the other companies are doing yeah to this product.

Deedy Das 1:07:33

It's very rare in our industry where we get to meet a true missionary. And when I met Max at the very early days of Pangram, it's one of the first questions I asked him. I'm like, Max, you have the best AI detector. You know that a lot of people want humanizers so they can cheat on their homework primarily, but some other things too. And have you considered making one? And he said, I'm never going to make a humanizer. And I'm like, I love this guy so much already.

Yiwen Lu 1:08:00

Yeah.

Deedy Das 1:08:00

I think you know, like people underestimate these things because they get like greedy about revenue in the short term. But people forget, like you know, like ultimately, I'm gonna take it a different direction. Ultimately, like I, you know, I watch a lot of anime and stuff, and like any any show about humans, people the biggest bottleneck to company growth is attracting people to your company, and nobody wakes up every day going, "I want to work at a company that wants to make more money." They want, "I want to work at a company that stands for something." You know, I want to work at a company that wants to do something in the world. So he doesn't do it for this reason. He's obviously very authentic about this. But one of the impacts of that is people join this company, and they are they're like I'm here for a mission. I think this AI slop in certain many cases shouldn't exist, and I don't want to contribute to the idea that some people can just you know fool other humans by posting stuff that's just not theirs at all. And it's such a great boon for hiring for a company like Pangram.

And you know, you were talking about this earlier, but you know, there's many similar connotations to even a company like Anthropic when it comes to the culture that they've built over time.

Yiwen Lu 1:09:15

Yeah, yeah. Let's talk about it from one startup story to another. You guys have this deep partnership with Anthropic invested in the. You also led Anthology Fund. Let's start with the team. What was your first impression when you met the team?

Deedy Das 1:09:28

So, for background, you know, before Anthropic was a consensus bet, because nowadays the joke is, if you ask any VC, they'll say they're investors in Anthropic. But we invested at their Series C when they had no revenue, then we led their Series D. Then we started the Anthology Fund exclusively in partnership with them, and we've sort of invested in every round since. I think that the thing that stood out about Anthropic, and a part of this, honestly, I will say, is hindsight, because it's hard for me to go back and say so much has happened. Since then, in a short amount of time, but similar to Pangram, which is funny because it's a tiny company, but similar to Pangram, they had a clear mission and direction, and they never wavered. When we all know this adventure, it's very very rare for a number two competitor to be able to to win a market, and to be very clear, they were number two to OpenAI. We all know that they came later. They came from OpenAI. So the question becomes: Okay, how are you possibly going to win? And the answer for many startups in many cases are focus, and you have to focus on the right thing.

From day one, they said the path to intelligence is through coding, and this was every single model that they released after Point was doing focus on coding. There was no voice, there was no image generation, there was no. They did some visual stuff because that's important to to general intelligence. But they ruthlessly focused on on just that core core experience. Even today, you start seeing these amazing I call emergent behavior of that focus, where like Opus five five for video generation is arguably the best video generator in the world, and they don't generate video; it generates like the scaffold around these informational videos. I think that was, I think, the most impressive thing about that team. It's that the ability to have a very clear goal and view of the world and focus on it. They were never. I know it sounds insane to say, but they were never driven by revenue. I I would argue they still are not driven by revenue. They believe in a way that the world should be, and they believe in a way that they can get the world there, and one of the reasons I look at this data a lot.

If you look at the retention curves of different companies, of employees at different companies, the highest retention company in the valley is Anthropic, and mind you, like people are paying a lot of money, like hundreds of millions of dollars, to get people out of Anthropic, and they don't leave. And so the question you have to ask is like, what does it take? Like, why? Why are people staying at this company? And I won't have all the answers. I don't work there. But from my the people I know there, from our partnership with them, is it's because like once you, once you're at a company for a reason that's not financial. Like I was saying earlier, like it's not just because the company wants to make more money or you want to make more money, then it's really hard to get you to leave. And I think they've done a very good job. They didn't do it for this reason, but of setting like the culture, and that's what's also very impressive about the company. It's led to a lot of the success, I think, of the company.

Yiwen Lu 1:12:45

Totally, totally. So, when you guys invested in it, you also mentioned earlier that ChatGPT was kind of the the leader in consumer adoptions. How were you assessing this decision to focus on the enterprise?

Deedy Das 1:12:58

It was very simple. It was like, hey, clearly the competitor, like the goal is to to generate to make extremely good and smart and safe

Speaker 3 1:13:06

AI.

Yiwen Lu 1:13:07

Yeah.

Deedy Das 1:13:08

Now to get to that goal, you definitely need to figure out a business model. Now your chief competitor has gone head first into consumer, and they're doing a very good job. Right? ChatGPT is you know arguably the best consumer product that we've seen in the last 10 years, I think. Okay, so what is left to do in AI? Where where does AI have value? Well, most consumers don't use AI for its intelligence, and I think that was something they picked up on then, right? Because if you think about how consumers use AI, there's this great report on how all of these, you know, breakdown of use cases. People use it like to chat with a friend, or like as a therapist, or as like, hey, like help me with my homework. Like, they're not using the utmost like top of the intelligence type queries all the time in ChatGPT. Anthropic's view was the value for the best intelligence is actually not in like the broad consumer; it's in the people who do work. Those are the people who actually have to think and do a job and do a task. And so, what would it look like to serve those people?

And and that was like sort of the very basic reason they were like well let's just do enterprise and then then the second question was how do we do enterprise okay then one of the big breakthroughs was hyperscaler distribution OpenAI made a fatal flaw where they were too tied to Azure they had to do it because they needed the compute and Microsoft was willing to give them a lot of it but okay you left a hole you left two or three holes, actually. So, the then GCP and Amazon, who are clearly big investors in Anthropic, decided they will distribute Anthropic APIs through their services. So immediately they had a big enterprise distribution advantage for their APIs, and and then you know you have get more data as a result. If we make more models, as a result, you also see the behavior of people. Like, what are they using it for? What are they not using it for? And that makes that flywheel really help them win an enterprise. And it turned out that market was actually a lot bigger than consumer. Which this is maybe just my opinion. Which, if you asked anybody at the time, I don't think that was a. I would have even agreed with that.

It was not easy to predict the enterprise market would be bigger than consumer,

Yiwen Lu 1:15:28

because

Deedy Das 1:15:28

if you look at every other software field, it's just not the case, right? Like Facebook makes, Google make a lot of money from consumer, and not as much from enterprise. Like you, if you like look at the search and YouTube business versus the cloud business, it's it's very different. So I think it was totally not obvious that that coding thing would happen, agentic behavior would happen, huge usage of token would happen, and that it would it would just be a more phenomenal business for the enterprise than even for consumers.

Yiwen Lu 1:15:58

Yeah, yeah, makes sense. This is really interesting. And compute is that still a constraint? Well, I mean, I'm sure this is an industry-wide problem. Is that still a constraint for Frontier Labs for for this topic?

Deedy Das 1:16:11

I would say that it's a concern for everybody because nobody knows how much they really need. Right? There's a couple of different factors to know how much you need. Let me start with theory number one. Most economic value will accrue to someone who has the best frontier model. That's sort of what we see in the data. That's what most of the big labs believe. Okay, so how do you get a frontier model? To be to get a frontier model, of course, you need a bunch of data that's specific to the things that you want to improve on, but you also and some algorithm improvements, etc., etc. But you compute. I think that's probably the biggest bottleneck. The other two are more solvable. That's the biggest bottleneck. Okay, but how much compute can you get and how much can you afford? That's a very non-easy answer because if I make the amount of risk I'm willing to take, and Dario goes into this in one of the main podcasts with Dwarkesh, is if I make X amount of money, I can probably only take debt for some amount related to X in compute, but I also don't know how my growth trajectory looks in the next two months or three months or six months. No one knows.

Every single lab has undershot or overshot their growth trajectory. It's impossible to predict this. So if you don't know your growth, then it's a little bit of gambling, right? That you're like, ah, if I make this, then I can afford that. If I if I if I don't make it, then I'm going to be underwater. And if I buy more, then I'm going to also be underwater. Buy too less, I'm going to lose. Buy too much, I'm going to be bankrupt. So you're sort of treading this like very fine line of predicting how much compute you need, but you want to maximize that because the first point, because ultimately you think more compute will let you get better models. There's also this idea of it will also let you serve more models at once. But so both of those. But that's how that's ultimately the source of the compute war, and it's obviously a constrained resource. Power is a constraint. Enough data centers can't go up fast enough. Chips being made as a constraint. All of these things are being resolved right now.

Yiwen Lu 1:18:27

Yeah, yeah, makes sense. Cool. I want to talk about Anthology Fund. You guys, it has been also like two years to close to three years since you guys started that. And yeah, you started that like shortly after you joined Menlo, and you guys wanted to invest in companies earlier. And how how is Anthology Fund different from say like a corporate venture fund that could have been housed under Anthropic and also what you guys like typically do?

Deedy Das 1:18:54

Yes, the Anthology Fund started as $100 million fund in partnership with Anthropic. When we were working with them, it's almost you know comical to go back to that time where they were a much smaller company, and they said the discussion we had with them is you could run it as a corporate fund and then lose focus, and most corporate VC funds don't work very well because they're not incentivized correctly to make great investment decisions, or you could partner with us and then we could run it and with you, so we chose to do that, and it's been, it's just been amazing for a bunch of different reasons. I think it helped. Hopefully, I can speak for them. Hopefully, helped Anthropic really understand what kind of developers and companies we're building at the really forefront of use cases, right? Because what one thing that tends to happen with labs is a lot of your revenue might come from one place, but they're not necessarily the biggest innovators. And you really want to design your product for the innovators, because these are like the ones who pay you a lot are laggards, and they will follow the. The thing that your innovators want.

So, the Anthology Fund helped them get very close to a lot of companies that were developing really at the cutting edge of applications and things like that. So, that was one big benefit for them. For us, I think what it really helped us do is typically, you know, Menlo as a fund has been around for 50 years, and we like to write lead checks into companies, and usually do you know a partner will do two to three investments a year. So we're highly focused and we go go deep. Now in AI, if you did that, you would learn very slowly because you can only learn so much when a company pitches you. You only really learn when you actually see the company operate. And the Anthology Fund let us invest somewhat small checks in so many companies that we got a sense of like okay this is working this is not working this is interesting this is not interesting this is a bad idea this is a great idea and it helped us hone our knowledge of like where is what is interesting and what is not in startups so those were I think two of the main good use cases for for anthology. What are some

Yiwen Lu 1:21:04

interesting ideas that came out of that? Goodfire was also like part of.

Deedy Das 1:21:08

So the two cleanest examples I would say is OpenRouter and Goodfire. We invested in OpenRouter at the seed and Goodfire at the seed, and we eventually from the Anthology Fund, and we eventually led both of their next rounds, and they both ended up being. I mean, OpenRouter definitely a success economically, and Goodfire has been on a great run. There's many other companies like that too. There's another one we're closing right now that was also an Anthology check that is going to be a big company. I can go on and on. If you talk about these two companies, I can speak for for myself here, but one of the things is a very tricky going back to the first thing I said in this conversation is startups are a lot about investing in founders. However, in our day-to-day jobs, what tends to happen is someone a founder will come to me and say, "Hey, I'm raising my Series A. I'm raising $30 million. When can I pitch you? They'll come in. They'll do a pitch. They'll run me through some slides. They'll say all these great things about their company, and then I'll ask them when do you need to make a decision by. And they'll say two weeks.

Now, in two weeks, and they're doing this for a lot of firms, obviously. In two weeks, how am I going to understand this guy or a girl? I don't know. Are you good? Are you bad? I know what you said about yourself, but it's the same thing as you know. This is a very classic venture saying, but it's the same thing as it's like it's a shotgun wedding. You meet a meet a if you meet a guy for the first time, you're going to know in day one whether he's a good guy. No, I'm not going to know if a girl is a good girl. I don't know. It takes time. So what? Coming back to the OpenRouter and Goodfire thing, we got the small check in early or smaller check in early. We didn't lead their first seed round, so we were invested in them, and then we spent a lot of time really getting to know the founders. And you know, to me, it's quite surprising. It's actually pretty interesting that takes three or four meetings when they're not when a founder is not pitching for you to be like they're awesome. Like it's almost like you know a hum a person reveals their true nature when they're not you know on show. If you're pitching and you're selling, then obviously I don't know who you are.

But if I'm just hanging out with you, and no, there's no no one listening, no one caring. That's when you'll like sort of know this guy is actually a good guy versus not a good guy. It's very similar, I think, for founders where we we saw them through their hard periods, through how they were thinking about company building, what their focus was, and it just became so obvious. Like these guys are awesome. We need to we need to back them. So gets helps you get to know the founder. Yeah.

Yiwen Lu 1:23:40

Does this closeness to Anthropic help the founders in any ways? What can they access?

Deedy Das 1:23:45

You know, I would say not as much as most people. You know, maybe the the strongest ask that people think that is not true is no, they don't get premier access to models, they get maybe a week in advance of a release. Like there's an early access program that's even beyond Anthology, so they get they get that. But the biggest help for the companies has been closeness to Anthropic. As you can imagine, for OpenRouter, for example, or for any other company, you your traffic might spike today, and you might be like, oh, I need quota now. I need to know if I can like meet my customer demand in a matter of like 24 hours. If I don't know anyone at this company, it's very hard to get a human on the line to say, hey, I need this today. And with the program, I think those kinds of requests have been a lot easier. Like I'm doing a launch, I need to bump my quota up by 10x, or I just launched. I need to bump my quota up. Help me sort of figure out this demand and this capacity. Can you do it? So those kinds of things have really helped companies. It's also helped companies to sort of understand, hey, Anthropic, where are you going as a company?

They've always been very honest and consistent about it, but that's also helped a lot. Companies think about where they're going. Like some of the advice Anthropic always gives to startups is, don't build for where the models are today. Build for where the models are tomorrow. So your product kind of shouldn't work today for it to really shine in the future. So it's like small strategic things like this about how to work with models. About you know we do a lot of talks with Anthropic leadership and and people on the team, which allow these companies to sort of be the the first to understand. Hey, how do I do an agent architecture? What does that look like? How do I do tool calling? What kind of tools make sense? Which models should I be using? Why use Sonnet versus Opus and the and vice versa? So it's like that kind of like stuff that really helps those companies more than like, oh my God, you're gonna get Fable 5.5 like four months before everybody else. Like that's not the purpose of the of the program. It's these other things.

Yiwen Lu 1:25:49

We're close to the end, so I'm just gonna ask some big picture questions. Go ahead.

Deedy Das 1:25:53

We'll rapid fire.

Yiwen Lu 1:25:54

I was in well, I was in SF for the past two weeks. If there's any takeaway that I got one takeaway is that everyone is selling data and RL environments. I feel like you guys don't talk that much about it, and and I I mean I have very complicated feeling towards RL environments. I feel like researchers could spend the time doing other more interesting things. How are you feeling about the AI data industry?

Deedy Das 1:26:18

I always tell people there's a website called Pavlov'sList.com. It's a beautiful website because it just is a is a list that one of my friends Chris runs. That is a list of all of the data and RL environment companies. We are investors in some of them, but that was before they became data and RL environment companies. They were because we just liked the founder a lot, and this is what they decided to take the business. It happens a lot in our industry. When it comes to those businesses, a lot of them here's like the classic set of risks. They're non-recurring revenue. They seem very services-y, but they make a ton of money. You know, I've seen companies go from zero to 100 million in three months, four months. So they make a lot of money. Some of them obviously make more than that, and but they have a lot of customer concentration. So it's typically like you know, 10 customers. Some bigger ones have like 50, 100. But typically, you know, we all know who the main buyers are, and it's not that many. So I look at those businesses, and then you know the margins vary all over.

Like some of them are very low margin, some of them are slightly higher margin, and some of them are a little bit higher. But you know that's sort of the construct of all of those businesses. And again, last thing is very little differentiation across across those businesses. But again, some would argue some are a little bit more. If you look at that slew of businesses, the main issue I have is is a why are you asking for a valuation on your revenue multiple? You are not a recurring 80% margin software business, and you have no path to do that. So you know, don't like I don't like it when people are like, hey, I'm at 10 million run rate, and therefore give me a 500 million valuation, like you would have given a software company. I'm like, well, you're at 10 million run rate from four contracts that you made last month that are non-recurring, and so maybe I don't want to give you a 500 million valuation, right? Like, so I think the valuation and the venture story is a little bit unclear for data businesses, which is why I tend not to love investing in them. Even though we have some checks in the budget, like I said.

On the other hand, what I always tell those founders is like, dude, just go pay yourself a lot. What's wrong with you? Like, why are you raising venture money?

Yiwen Lu 1:28:34

Yeah. If

Deedy Das 1:28:35

you're making $100 million off of these contracts, go back, go pay the team, pocket the cash, that's great. Why do you care about raising money? You don't even raise money to scale your business, right? Because like the theory of venture capital is I I raise external money because I want to scale my business because I don't I don't have the money. But if you're coming to me and saying I have the money, but I just want to raise just 'cause why are you raising? Use your money, grow your business, hire more people. It's such a great cash flow business. That's how I feel about it. And look, like I will admit, you might have a freak acquisition here and there. We had two: Scale and Mechanize. I don't think I don't know if Mechanize is fully public, but both of them are fantastic acquisitions. So that can always happen, and then I will be wrong in those cases. I might even be wrong. Maybe one of these companies' IPOs. I just don't think the IPO prices are anywhere near where the private prices are, especially for this business. But again, might be wrong there. That's that's kind of what I view it.

Yiwen Lu 1:29:30

Yeah, yeah. What's your biggest bet now?

Deedy Das 1:29:32

In in my portfolio?

Yiwen Lu 1:29:35

In your I guess in general, industry wise, vertical wise.

Deedy Das 1:29:40

Oh man, that's a hard question. A oh man, so hard. There's so many things.

Yiwen Lu 1:29:46

There are, there are, and also

Deedy Das 1:29:48

things change so much. So you know, the reason it's so tough answering this question is every two weeks I get new information. I'm like, well, what I was thinking is really is really not the case. So and then and then. Then the answer just ends up sounding super generic.