Steven Waterhouse, co-founder of Nazare. How are you doing? Good. And you go by seven. to my friends, but you can do that. okay. I remember back in 2017, we talked about this a second ago, we were at a, um I led growth for another company at the time and we were at an ETH San Francisco event. And I remember when I first met you, you introduced yourself as Seven, so I figured I could call you the Seven. Perfect. I do more work in venture and talk to LP potentials and, you know, wear like slightly better clothes, or, well, you know, more dressed up a bit, then I tend to reduce myself less deeply. But yeah, partner calls me seven, and my team calls me seven, mostly seven. Well, you've been in this space for quite a while. Some might even call you an OG. So I want to do a quick intro really quick and then would love to hear your thoughts on a couple of things. Okay. So you started your PhD in neural networks, I understand from Cambridge, and then you went to NASA. And then you founded Orchid Labs with a pretty amazing team. And that's when we met back in 2017, 2018, something like that. and now you're leading Nazaré Ventures and want to definitely hear your hot takes on the industry, AI and crypto and what you're up to at Nazaré. Maybe if you could begin with sharing with us kind of like what motivated your kind of the transitions from being very focused on crypto generally and then privacy with Orchid and then now Now your thoughts around AI and starting a venture firm focused on, I believe the primary thesis is investing in AI projects, is that correct? Okay. Yeah, take us through your transitions. yeah, the old transitions. I've been in and out of venture since the early 2000s. I was an entrepreneur in residence at Foundation Capital in the early 2000s before I if I ended up starting an IP company and taking that public. And then again, of course, with Pantera Capital, got back into the venture game there. Then stepped away from it in 2017 to start Orchid. Stepped back in in 2021 and I helped my friend Richard Millhead with Fabric Ventures as an advisor and venture partner there. And then again, with Nazare and in between I've been angel investing and doing some interesting deals. was the first check in risk zero, the first check in speed router. So I've been successful with some angel things. I wanted to keep doing angel investing essentially like the feel of angel investing, to me that means I want to work with founders. want to, want to, you know, to the extent that let me like help them build, help them. construct their companies, raise their rounds, build something huge. And the fund gives me a platform to do that at scale in a way that you have to fight hard to do that as an angel investor. So that was the sort of idea of building Nazaré. As far as the direction, I kind of take a back on AI as a I used it in a number of companies along the way, but like at RPX, the IP company, was building models of patent risk and trying to value patents in order to, we bought patents to stop patent trolls from buying patents. Basically we kind of, we were good guys in the marketplace. And, but I'd mostly turn my back on a lot of AI because in the early 2000s, what I realized that most of AI was going to end up being used for quite a while to just mind humans to do e commerce advertising, kind of build an understanding the kind of the dark dystopian side of surveillance capitalism, right? And Orchid was a was the foil of that it was like, can we build something that gives you censorship resistance is free from surveillance, gives you the ability to step out of that equation. So obviously, I was quite, you know, like opposed to that kind of model. but AI wasn't the, just like any tool, AI wasn't the culprit there. was, it was the use of it. What I saw in the early days of, um, tools like stability AI, then of course the early chat GPT was the suddenly this was an enabling technology. was was a uh human enhancing technology, not a human mining technology. Um, and with, with all of the concerns people have around big AI and all these kinds of things, these tools are still useful. They're still like enhancing our ability to do things, our ability to be productive, to solve problems that we have. And that was fascinating to me. Now, I knew pretty early on that I wanted to do something in that space. the thing that really caught my imagination was not so much the applications, although I love many of applications. I'm definitely a power user. um And not so much the hardware because I like hardware, but I don't have a hardware angle in my investing um or background there. But it was really this idea of looking at the software infrastructure and the... the opportunity space that was being created there. Now, the thing that struck me was very similar to different pivot periods after the dot-com bust and also after the great financial crisis. You saw this move initially from very large... uh big servers in the end of the dot com. Like you had to have a Sun machine or a Paco machine in order to even get funded. I was like, where's your contract with Sun? Where's your contract with Oracle? And then within a year or two, that had moved a couple of years and moved to fully distributed server setups that Google was using and then Amazon was using and so on. Like this much larger distributed setup. that technology to do that was all software. Like there was this hardware, it was very simple hardware. A similar thing happened at the end of 2008, that period when people shifted from having their own server farms to having just putting everything on the cloud. And so I saw this opportunity of like, well, Nvidia is building this big iron. It feels a lot like the sun system that I was working on back in the early 2000s. What if there's a pivot, a change to a more distributed and perhaps a decentralized form of AI compute? And so our thesis is we look at the software that allows you to build and operate AI systems on the continuum from the biggest clusters, mega clusters, all the way through to the edge devices, like with everything in between. Now, maybe we can back up a bit just for the audience. um So when we think of AI, most people are familiar with the application, chat GPT, grok, et cetera, right? But if we go backwards, maybe walk us through like, what are the major steps to get to the application layer so that, know, at the consumer level so that... um You know, people have used the phrase, you know, algorithmic development, there's model development, know, data labeling, et cetera. Maybe take, maybe walk us through and then, and then given, given the, the AI stack, um, where are you specifically focused on and how does crypto play into that? If, if at all. So I'll get a little walkthrough of AI. So what we call AI today is primarily, almost universally, neural networks, which is a field that dates back to the 50s. But I'm not going to give you the full history story here, although it is interesting. There's a great book called The Alignment Problem. which deals very little with the actual alignment problem, but gives a very good overview of an historical analysis. if your readers, listeners are looking at something interesting to read, that's a good one. uh Essentially, these systems, they learn from data. so you need a whole bunch of data. You need a model with what we call weights, is essentially the variables that are learnt as it sees the data. And then once the model is built, you show it new data and it does some kind of prediction. So in the case of, say like image classification, which is a common thing that we do in AI, we'd say, okay, here's a bunch of pictures of cats. Here's a bunch of pictures of dogs. Now I show the picture of a cat. What is this? that cat or is it a dog? If it's a Shiba, it's not quite sure. gets confused. So that was sort of the way AI worked for very long time. And in order to do that, and so does, and the way to do that at scale in image labeling, that's what we're talking with data labeling. So we need, first of all, using a data set which is human labeled. So back in 2016, this was done using data labeling with Mechanical Turk on Amazon. And now that's done with a combination of machine and then human labeling. What you're using most of the time are one of two things in the chat, the chat box and sort of like the chat GPT and so on. You're either using an image model, which is to say, generate this thing, right? or you're using something you're just talking to and is giving you some kind of text output. Now, the text output and also some of the things that the image wants use a technology which is called a transformer. Now, this transformer was a new invention a few years ago, came out of Google. And it had a very unique approach to the network structure. It used a recurrent network, which means you wrapped some of the outputs into the inputs, but then add a special kind of what they call attention, which is like paid attention to certain parts of the previous inputs. And the fun thing about this, because it's recurrent, is that you can just take text and just feed it into it. And so the entire internet becomes the training set. You don't need to label anything. You just kind of just like, throw it through. the tricky thing that's happening with the transformer, especially on text, is all it's doing is predicting what the next best word is. And so I catch myself trying to remember that when I'm using it and thinking this thing is like, you know, even with my experience, I'm like, this thing feels so real. But all it's really doing is predicting the next next word. you know, and some art man and Ilya and some of the others talk about this and they say, well, maybe that's all intelligence is, is knowing what to say next. The new transformers have other tweaks. the reasoning models, the ones that are a bit slower, which we saw initially with DeepSeq, but O3 from Chetik and others have these characteristics. They're basically wrapping around the output. So they're doing something and then they're telling it again to think again. It's basically thinking for longer. That's a rough, simple description on AI as of today. What we're seeing more and more is just bigger and bigger models. There's a concept that was sort of running out of data, but now what we're doing is using the data that we have and the models that we have to create what's called reasoning data, which is synthetic. Essentially it's like, asking the model to solve a series of problems and then using those as training sets for the next model that is built from that. So we're kind of using models to build themselves. em That's the state of the art. I'm glad you walked us through kind of the history. um So in grad school, I studied computational linguistics. This was a while ago, right? um And uh my thesis was on um something called word sense disambiguation. And I created a program, a neural net in Perl. Can you believe that? mean, people used Perl back then. one example of this just for the audience is uh the word bank could have different senses. So bank is in like the financial institution or a bank the side of a river or bank when you shoot a basketball and you bank it off the side, right? And so how can a machine or how can a program know like what sense, in what sense are you using the word bank? And so I created this neural net using at the time it was uh pretty novel research. called analogical model of reasoning. And the predictability of this program was pretty high at the time. This was back in, don't, a while ago. I don't want to age myself, but I'm pretty old. no, this was pre-embedding. So I went to school in Chicago and studied under a guy named John Goldsmith who's kind of a big deal in the. computational linguistics space, or really linguistics, but at the time, computational linguistics was just kind of becoming, becoming, becoming vogue, I guess. um And computer scientists were interested, et cetera. And so, but anyway, it's really interesting to see what's happening now because now we have like all this data that's available to be trained on. You know, a lot of it has been labeled. the data makes sense. And, but what I'm noticing still is that the problem of word sense disambiguation like still exists. And so we, we still haven't solved some of the most basic things such as, you know, a machine can't really tell irony, for example, or sarcasm. and so like, if we were to apply the Turing test, like it's, you can still tell if it's, if it's a human or a machine. And so we're. despite all the advancements that we've made, there's still some very basic human things that a machine cannot do. It's like the Turing test is a spectrum now. It's almost like if you think you're talking to the machine, then you'll start doing different things. Yeah. I mean, there are cases where I'm using, you know, Grok or Chad GBT and especially with deep research, when you click on the deep research button, it's like, wow, it's really, it's reasoning, it's thinking. It really is. Now tell us your, what is your thesis around having been in the crypto space as an investor and as a founder, an operator, and now you're focused on the AI kind uh of a... you know, segment, I guess. I don't even know if that's the right word. Yeah, AI wave. But you don't quite see it as being completely separate. There's still there's there's some relationship there. And so tell us your thesis around AI plus crypto. And yeah, tell us that. scratched my head for a long time and had sleepless nights trying to find the answer to this question where people would say to me, what is the intersection? I was into AI, and I was like, oh, and I got asked this question so many times that in the end, I was like, yeah, I don't know. Like, I don't know what this section is exactly. I know these are two very powerful technologies, and I think there are ways that they can be used in combination. But I have a very extreme I almost kind of start you on crypto study, which is, think there are two product market fit happening. One of them more obvious than the other. The first one is just finance. Crypto is becoming finance. Tradfile is becoming, DeFi is becoming Tradfile. It's just like a merger. And we've seen that in everything from ETFs all the way through to JP Morgan using Uniswap tech and everything in between. So we'll see like real world assets. We'll see all these things, stable coins. It's just like a big thing. Now that regulatory environments has swung more positive, I think we'll just see a big race towards following that product market fit. And that's very refreshing in crypto when you've been struggling to find product market fit with many applications over the years. We have a little bit of product market fit and then it just kind of pulls back. And so we've had this thing where we've been trying to get crypto products into people's hands, but stablecoins just work. So I say, if you're in crypto, pay attention to stablecoins. em Secondly, if you're in crypto pivot AI, and I don't mean crypto AI, I mean, actually AI. So one of the things I work on with my companies and my investments that have crypto as part of the product set is be an AI company, service people in the AI business, service people who have needs, find your products, find your customers. And this is often quite foreign concept to many crypto founders who, you know, been trained in the last, I'll say five years or so where The primary thing in crypto is the token is a product. And so you build a company and your concept is like, how can I get as complicated and as nuanced with my technology and then get my marketing really good, get my KOLs in there and I got a token, that's the product. I'm like, no, if you're really lucky, you'll make bank on that kind of thing, back to your bank. But quite likely you'll miss the timing. And then you'd end up with this thing, which is like actually kind of a weight around the next thing. Like, well, what the hell am I doing? I don't have a product and I've got all this liability to deal with all these users, all these kind of not users, but traders. I'm saying like, let's go back to the old days. Let's find products that have product market fit. find customers. Let's focus on problems. And to the extent that crypto is helpful and beneficial in those situations, let's use it. But I'm trying to build AI companies and some of them have crypto. em Rails, some of them have backends, but I'm certainly not investing in some intersection. I'm not investing in a meme. So in our portfolio, there are things that are much more like very decentralized, almost kind of crypto native. For example, a merge from Stability AI is a company called Intelligent Internet. uh That one is a fully decentralized AI, computes, models, data. So, yeah. people. But then at the other extreme, have one of my investments, earliest investments is called Vast AI, which is a decentralized GPU platform without crypto. So you bring your GPU, you rent it out, you get paid in on Stripe, good old settlement. And, you know, they do the sort of auditing of the, and build a reputation model and so on. It's like an Airbnb or clearinghouse for the GPUs. Probably could ask crypto at some point, but that's just not the thing. And they make a lot of money doing that and service the AI market really well. We've also seen really interesting things where there's almost like a hybrid of these two. So you've got layer land, which is building benchmarking services for companies with building AI. So if you've got like a Tesla or Anthropic or something, building new models, how did they know? And how do they have an independent verifier of how good their model is on certain tasks? And so what LayerLens does is build benchmarks, these existing benchmarks, and then publish that information, either privately or publicly. And the way that they get their compute is they use an Eigen layer-powered GPU cloud. So they have a token there where people bring the GPUs and work that way. So you see how you can hybridize these things together, but you try and find the situations where does a product market fit? And we've had companies where we've invested in them thinking they were going to do a token and they pivoted back. And other companies where we've invested in the like, well, actually, we think we might add a token. We're like, we didn't expect it to, but sure, if you some help doing that, we're here to help. So the basic thesis is we're looking at decentralized and open source AI is one of the fundamental things we're looking at. And obviously in the case of decentralized, we may have crypto exposure in those situations. So maybe another way to think of it is rather than the intersection, you approach your investments with the AI lens first and to the extent that uh decentralization and crypto as an incentive mechanism makes sense, then that becomes part of it, but it's AI first. Is that a decent characterization of what you said? That's definitely it. And so for example, yeah, we're also like, we're not heavily focusing on like, you know, the kind of the finance side of AI either. Like some people looking very much at DeFi and trading bots and all these things. I think these things are interesting. I just think that I think there's so much opportunity within just the sort of the raw infrastructure opportunity with AI that we think we can really focus on that and do well. And we think it's very fascinating too. When you say raw infrastructure, tell us about, I guess, what are the opportunities right now? And I agree with you. I think there's a lot of crypto people that are looking at agents and how to make more money using AI, et cetera. But I think enterprises have a different view. I believe they're looking at how can AI help us save money, save money and save time. I'm not sure if we're at the stage yet of enterprises looking at how can we generate more revenue. Is that your experience and maybe how do you view kind of AI adoption in the enterprise space? Yeah, so on the agent side, it's very interesting. em I started looking at agents, like two years ago or so, but it was pre, so crypto agents, was like, it was before the infinite back rooms and all of that kind of fun stuff with virtuals. em And the area I was looking at was em in cooperative agent architectures and looking at that in the enterprise space. And I think if you look at the SaaS business model um and try to understand that, an example of this is if you were to build your own CRM today, what would that look like? And most likely it would look like a database, which you would structure in some way so that there was those nuances of knowledge embedded into it. Maybe you'd process it in some way. um, connected to a mail server, uh, have some kind of interactive component that you can update things on and then a language model. And so a lot of the, the kind of kind of grift that goes on with something like a Salesforce where they sell you this thing, which is kind of brittle, but very powerful, but in order to make it work. You need all their extra consultants and then someone else and someone else and their add-on, and you've got a tuner and working through so and so. You just want to talk to the data. And that's what Nelolens gives you. So you start thinking about a lot of SaaS businesses, and you can just flip them all around and say, let's have a data source and a model and an interface. And so I think a lot of SaaS will get replaced by agents on the enterprise side. Now, As far as like, I think that the test with AI will come as to how much it's disrupting the workforce. When we see a crop of graduates come out of the best schools and not get jobs in consulting firms, because the consulting firms are like, we're good. Or they get the jobs because they know how to use AI. And then a middle tier just gets eliminated or just doesn't get promoted because the consulting firm is very much up and out, right? So they just get out. They don't even have to fire them and just kind of listen. You're not going to make a partner. See those later. You're out. We're seeing it already in our investments where we see uh and this is to your point as to saving money. We see it where AI companies, know, AI software development companies we're investing in don't hire at the rate that we expect them to. And when we talk to them about it, they're like, well, we're just using more AI, and so we're spending more time to figure out how to use it to help us do what we're doing, whether that's writing code or operating systems. So I think that there will be a fairly swift, not even necessary reduction in team size, but just like a slowdown in hiring, which will be the same thing eventually. how that happens in the larger companies, we will see. But yeah, I think that these companies that we're looking at that are adopting AI, they will certainly become more profitable. Whether that leads to increasing revenue, we'll see. I don't know. It depends on the space. The problem with finance, find, and people saying, I'm going to use this AI model to make me trade better or lead to better results and so on, is that You're getting into the hedge fund world, right? You're getting into like competing with the two Sigmas and these people who have insane resources and models. And I love it if we built some kind of decentralized open source version of that, but that's kind of a whole game, a whole nother world. I've never really been into financial trading for financial trading sake. I'm much more of a product builder. want to build things that... that are not in that space, they're doing different things. But yeah, I think there's a lot of opportunity in that, not in the intersection, but as you said, AI first and then crypto second. Yeah. And I think the AI first product development approach is very interesting and I think forces you down a path of how can we make this actually useful, which is kind of the criticism against AI is that it's going to replace humans, right? But it's actually going to make humans more effective in what they do and allow them to focus on things that probably bring them more joy, probably bring them more satisfaction. Back to the agent space, I met with a project last week that's creating a they're in the agent kind of swarm space. so we talked about what swarms are. And apparently these agents, one of the big problems with AI right now is that it learns on static data for the most part. And then on the prediction space, like you described earlier, like predicts the next word, but it's based on a set corpus or data that's kind of, it's already existing. but what about data that's created as the machine thinks? And so we talked about that in this conversation that I had with this team. And they're trying to create a situation, infrastructure, allows agents, as it's thinking, it's creating data at the same time. And it's making decisions off that current reasoning. But that needs to be stored somewhere. And so that's one of the big problems with uh really replicating humans is as we're doing something, like for example, right now, I'm kind of talking out loud, that's not captured anywhere. um As I'm, you know, working in the garden, like all of my thoughts are in my thoughts, like it's in my mind. And so it needs to, that kind of information needs to end up somewhere in order for machines to learn from that. And so that was the conversation that we had and how that's actually a really hard problem. And the conversation also went to like maybe all the data that's currently available on the internet. It's been exhausted. And so what the next step then is like of the data that's really messy, it needs to be labeled. And so there's a lot of opportunities in labeling space. But also really it's like, how do we get all of the thinking, the thoughts, the reasoning that doesn't appear anywhere? How do we get that out so that these machines and these models can learn ah from that type of information that doesn't exist. Anyway, what are your thoughts on that? Because that's a very compelling problem, I think. I think that the... I think the concern of the lack of data is. It's valid, but there's also nuances. So synthetic data is not the terrible thing that people have blown it out to be. It turns out that hand-built synthetic data sets that are very effective, especially for chain of thought reasoning. Some of what you're describing is a little bit like, what is the human chain of thought uh when you're solving problems or ruminating on things, I guess. I mean, short of putting a wire in your head, I don't really get access to your thoughts, but I guess you could verbalize things and that would be interesting training data sets. And some people look at that kind of thing of like, do you, some companies looking at how to incentivize a collection of different training data. I don't know, I think also we'll see. I think we'll see the emergence of... Not necessarily like the biggest model you can think of that has all the data in all the world. Cause you can't get all the data in all the world because a lot of data is private. So maybe what you see instead is the emergence of, um, Models that I guess sort of act like agents. I mean, that's like, just ask them for their opinion. Maybe you have to pay for it. So you have to pay extra. So like, if you've got a subscription to the Bloomberg agent, you can talk to it. You can get that data out of it. But it doesn't expose all of its data all the time. And it's building data and contacts all the time. So can imagine this stitching together lots of different agents and then using other models to judge those outputs and form them into an adhesive whole without having to have a single model built on all data everywhere. So you can just imagine these mixtures built together. I think that's a reasonable model. that in that situation, do you see crypto as an incentive mechanism, maybe as a payment rail would make sense? Well, there are two outcomes here where either everybody's on the same rails. So that works here. Just everyone's on stripe, for example. Now, where wouldn't that work? Well, it wouldn't work where you can't get a It wouldn't work where you have different models which are in geographic regions where you don't have the same kind of pavement architecture. or it's kind of messy to use the same kind of architecture. And I think then you'll start to see crypto getting used. As far as incentive mechanisms, I think it's application dependent. I think that there's like a kind of competing force here happening, which is that I think the crypto rails will just get used all over the place. I think there are much easier mechanism for programming and for utility and for implementation. think we should, the thing we just talked about, probably should just use stable coins anyway. Because existing banking rails suck. So just use the crypto ones, but not because it necessarily needs to have crypto rails, but just because they're I think is the way I look at it. Because again, once you start thinking about it, you try and force yourself into this. The same as the intersection question, where you're just like, am I just forcing crypto onto this because I think it's cool? Or is it really the right thing? So that's why often I start these things out. We're just saying, well, if we all use the same rails, we could use Stripe, but that doesn't always work. So there'll be situations where that doesn't make sense. So identify the situations where that doesn't make sense. Cause that's going to be the default. Like, you know, you just program the thing you know how to use and yeah, it's a bit more expensive or maybe it's not. And yeah, it doesn't work sometimes and it's kind of annoying, but you know what? just works most of the time. So you just use the thing that works. How do you stop yourself from, I mean, you've been in the crypto space for a long time. How do you stop yourself from thinking, I forcing crypto onto this thing that doesn't even need it? Like, how do you, I guess, help reflect in a way that's honest and makes you really answer the question sincerely and whether or not the product actually needs crypto? Do you often find yourself thinking like that? Um. Yeah, so I have a simple way looking at most crypto, which is it's a way to route around centralized points of control. So that's kind of what it was developed for in Bitcoin, right? And so what I'm going look at is something I'm like, okay, what's the centralized point of control I'm trying to route around, right? What's going on here? So you look at decentralized GPUs, whether it's using crypto or not, I'm routing around Nvidia. I'm routing around CUDA, right? That's the centralized point of control. Why? Because it's expensive, kind of like a lock-in. Maybe we can do something different. Maybe we can build software that allows us to have decentralized GPUs, not just all centralized. That might be a more interesting world. Have more open source systems. This could be a bigger playground for developers. That would be interesting, right? My friend, Eric Voorhees, he built Venice, which has got lots of open source language models in there. What's the centralized point of control he's going after? He doesn't want censorship. He wants to be able to ask the Venice system anything and for it to give it, you know, what he calls an uncensored answer. Now, whether it's, you know, one person's bias is someone else's censorship. So like we kind of, you know, what exactly is not biased and not censored. But the point is, at least, you know, at least you can go back and use the model and say, kind of, know this thing like Hermes from Noose or, know, whichever the model, the dolphin model has come out. I know where that came from. I know where the answer's coming from. With Bitcoin, we're routing around centralized points of control, are obvious. uh ICOs, we're routing around people like me, like investors, or forcing us to play in a different game. And so I asked myself that question when I look at something with a crypto as part of the product is like, what am I routing around? What is the centralized point of control that's being overcome here? And could I build the same thing? without Twitter and would it matter? Would anyone care? All right. So. I think that's a good question that if you force yourself to ask that, especially as an investor and you're helping your founders, helping them reflect on that, it kind of gets you out of your head, especially if you've been steeped in either the crypto or the AI space. If in the AI space, then you can see crypto could help as a way to get around the centralized structures. focus on decentralization and really it's very supportive of kind of the agentic kind of mindset. um But if in the crypto space, an AI becomes a very different opportunity where that just gives, I think founders more breathing room in terms of what's available versus kind of forcing them down specific path. If we could go back to Nazaré and what you're up to there. So the AI infrastructure, kind of AI product focus is kind of top of mind. Tell us what kinds of projects you're interested in evaluating, investing in, anything that excites you right now. Yeah, I've described a few of the things. So we're looking a lot of things around, you know, what I call a continuum of compute, everything from edge devices through to mega-plasters. It's not either or, it's like, what does that continuum look like? How do you build software for it? How do you build AI applications? How do you execute those things? There's another area which we're very interested in. I just call this like, an algorithmic exploration. um And you can see some point examples of this already that are very successful. What I'm talking about is stepping outside of the big language models in terms of how they're architected from the algorithm perspective, the model design, and experimenting in that. So one example is Mixtral, which is a Paraspace company, uses a scaling technique called Mixers of Expert, which I did my PhD in back in the 90s. this whole story. They use mix of experts, the scaling tech. that scaling tech has now been used in many systems. It used in ChatGPT. It was then used in DeepSeq. DeepSeq developed another algorithmic innovation, which is sort of the chain of thought reasoning approach. And one of the things that happened with DeepSeq is that we've had this kind of catchy theme called matcha, which is make AI cheap again. And a little green hat with matcha. The the the sort of joke around that is that how did we get to billion dollar seed rounds? I mean, I know how we got there. But the way we got there is just by like, adding more compute, adding more data and just saying we've just got to go bigger and bigger and bigger to make these things better, which works. But at certain point, I don't know, someone starts putting tariffs on the world. People say, hang on a second, I don't have the budget I thought I had before I need to find something cheaper. And then we're back to 2000 or 2008. Like I talked about before. and people want to find something cheaper. Now what happened with DeepSeq, which is super exciting, is that suddenly people in labs all over the world said, hang on a second, I've got a new way to do a transformer. I've got a new way to make that model better. I know how to make it cheaper. Maybe I can get funding. And they call people like me. And so if you've got a new algorithm, I'm interested in talking. We're looking into that space as well. But the theme here is very much like, Trying to imagine the future software infrastructure that AI models are to get built on. We don't think it's a winner take all environment. We don't think that open AI or Anthropic are going to rule the world of models. We think there'll be lots of different approaches here. We also think that open source will continue to be a vibrant uh area for development. And we want to focus on the software infrastructure that allows these things to flourish. Yeah. So your article, Make AI Cheap Again, I don't know if you noticed this, but on the CoinDesk webpage, um above the fold, so your article's in the middle, and then above the fold, there's three ads circling your article. I haven't visited uh a website like CoinDesk in a while, and so was surprised to see literally so many ads above the fold. This is totally aside. And then as I scroll, there's two more ads under your article. It's crazy. um But that's how they, yeah, they gotta make some money. But it's an interesting article. So the title is, If You're In Crypto, Pivot AI Now. And maybe let's talk about that. I know we've kind of talked through some threads from the article, but. Maybe tell us the main thesis of the article and what your message is to crypto people. Well, this is a riff. I will give credit to Jason Calcanis who tweeted this, I think when crypto was crashing, FTX and everything was collapsing and the AI space was obviously taking off. em And it's important to understand that context is that when a lot of the founders and investors in AI were coming up, which was a couple of years ago, we were crashing. was 3AC. was FTX is the sort of final blow. Mm-hmm. know, the whole thing started with Terra Luna and that was like one of the worst years. I've been in this space since 2013 and I've been through Silk Road, Mt. Gox, Bitstam, Hax. Like I've been through it all, But at that moment, you know, I turned to my partner who's also in crypto and I was like, I don't mind, but like maybe this is it. Now I feel like actually it's going to be okay. And I sort of feel like crypto is going to make it. It's going to be. established it's going to be a thing. I feel like the babies leave in the house and the teenagers opt to college. I'm also finding that my excitement and the kind of the weird revolutionary characters that I love hanging out with in whatever field I'm working in, they're doing different things. Some of them are still in crypto, many of my friends still are. the whole gang every time I go to a crypto conference, but the weird new people are going and doing something else. And that thing is AI. um And that for me is a bonus. It's also a big attraction to being in a space as being on that risky, kind of weird, bleeding edge of things. um I certainly hope that we see some interesting fusion of these two worlds that I love over the next few years. But for me, you know, the focus is going to be, as you said, very much AI first. As far as pivoting, mean, it's a tongue in cheek thing. spirit of the article is what I talked about, which is I see us moving into very much like finance or like AI applications. And That doesn't mean that there is an opportunity because in the things that people aren't looking at or the things that aren't in vogue, that's often the best place to do something new. And I certainly expect to see, you know, I've been a big collector of NFTs and fascinated by the generative art movement, many friends in that area. em Seth from Bright Moments, a good friend, Kevin Abash, like we have, I'm deep into that crew. uh I think the digital art will continue to and become something that actually surprises people over the next few years. I think the gaming will surprise people. But I think those things will be largely eclipsed by, least in the short to medium term, the other two opportunities in finance and in AI. You know, as an aside, since you mentioned NFTs, this project that I talked with last week, uh their agents are actually NFTs, which gives us some really interesting, these agents have very interesting properties now of ownership because they are NFTs, which I thought was, you know, I think pretty cool. uh So Nazare, you're uh actively investing in projects and... Mm-hmm. and projects can learn more about Nazaré at nazare.io where they could submit their application to be evaluated if they want investment from Nazaré. Do you have a typical check size that you invest in? What range? I guess, uh yeah. we're usually investing in the the earlier stages of company and the pre seed to seed phase formation through to kind of first product. We invest between 150 through to 500 or a million depending on the size around and the team. Yeah, but we work alongside a lot of other investors, smaller and larger. We play well with others. We tend to be a value add for the deals we do, so we get to see some interesting things. be the first check, it sounds like. If we're not the first, then at least in the first round is the best for us. In AI infrastructure, and infrastructure like we've discussed is a very broad topic. Any specific categories within infrastructure that you are especially focused on right now? So we, a good friend of mine told me that the best way to describe a thesis is by what you don't do. So we don't do hardware and we don't do apps. So we kind of do all the squishy stuff in between. So everything else is fair game. Yeah. mean, we are often application driven and application focused in terms of the things we're investing in because we like to understand what the application is, we're not, we're not sort of, you know, particularly doing, in most cases, the customer for us is investments as developers of AI, um businesses using AI, not so much consumers in this iteration. Yeah. Any portfolio companies you want to, maybe you can highlight one portfolio company that you're particularly excited about. I mean, so many. m We've invested in nine companies over the last year and our full portfolio is on nathalie.io. I mentioned a few of them already with vast intelligent internet and layer lens. mean, probably just the fairest, the most recent one we just invested in is called Fairmass, which is building a fully homomorphic encryption platform, all open source. uh It may not be obvious why this is important for decentralized AI, but we think that it will be crucial in the future. And the team is fantastic. So that was the latest one. We just did. That's pretty cool. I've often thought about the intersection of ZK with AI and where the opportunities are, which kind of is in your space since you're an advisor to RISC0. By the way, I'm interviewing Rika, who is on the RISC0 team, on the boundless side. And yeah. team. We've looked, we've done a number of deals in ZK. LRAS is another deal we did in the ZK space, focusing on proving whether or not a particular AI model was executed. They have a platform for doing that. And then I've been doing ZK all the way back to Zcash. Pantera, we led the first institutional round in Zcash. I'm a big fan of that uh sub area, has been largely funded and propelled by crypto. It's fascinating how a lot of that research really got pushed forward by the need for scaling. And that need for scaling is now turning into privacy. And we're looking at another very interesting company right now. I can't mention yet, but We'll hopefully be closing a deal in that space as well in the ZK space. That's really cool. I met with Rob Viglione, who was the founder of ZenCache, which is around the same time as Zcache. This has been so fun because I've spoken with so many people with diverse backgrounds, but what's been really fun is speaking to the OGs who can take us back to that time and how it really was and how it felt. And at the time it was, everything was just so new and so uncertain. um But to see where ZK is now is like, it's with projects like RIS0 and Succinct and with the Ethereum Foundation talking about ZK proofs recently and potentially, you know, utilizing ZK proofs and on the execution layer. It's like, maybe like now is the time for ZK. It's really feeling that way. Whereas before it felt like a research project, but right now it feels, wow, this thing could actually be really useful to people and to human beings. That's quite exciting. I think it's ready for prime time. em Not everyone agrees with me, I think that's also the moment when there's this moment where there's a gap between what most people understand and what's possible. And then that usually flips pretty fast. Suddenly everyone understands that. Yeah. What's the saying? Things are really slow and then they go really fast. Or something like that. What's the saying? But it really certainly feels that way with having been in the privacy space for a while. You, and I have too with Oasis and other projects. It feels like there's a converging of... It now feels like the right time. And I'm excited for the intersection of not crypto specifically, but ZK or privacy and AI. think there's a clear need there um where it's actually helpful. one we were looking at recently is very much focused around this idea of independent crypto agents. If you imagine like model context protocol, which is essentially a language written by Anthropic to allow agents to talk to each other. And imagine an agent from uh IBM and an agent from Salesforce. And one of the masks, other one I questioned was like, well, so how do you know the integrity of that? answer, how do you know what it was built from? How do you know that? So like essentially, how do you trust each other? And so looking at, um, you know, understanding the, the, certain state of a model and what has changed in that state and being able to then prove the, difference between where you are today and where you are next. And so I think that we applications and middle of a here, but that's essentially like what we're looking at with how to apply ZK to these areas. That's really cool. That's very cool. Well, Sevan, this has been such a fun conversation. thanks for taking us through the AI stack, AI, the history of AI. And very excited for what you and the team are up to at Nazaré. And for any founders that are interested in learning more, go to nazaré.io. And in the show notes, I'll share information on Sevan, how you can get in touch with him. if you're looking for uh investment in your startup. Yeah, and like also just encourage people if they want to. I'm actually very approachable if you want to chat about what you're doing and just socialize. I'm usually actually pretty easy to get in touch with. So happy to advise. mean, not like advice, but just, you know, give some help. I'd like to see this space get better and people, it's hard to get going in these kinds of industries sometimes and I just need like a helpful hand. Yeah, and be sure to follow 7 at D7Tral, D-E-T-R-A-L on Twitter. Cool. Well, thank you so much, 7, and best of luck to you and the team and excited for what you guys are up to. Thanks Peter. All right, cheers.