Decentralized Data Foundry for AI | Rowan Stone, CEO of Sapien

Rowan Stone is the CEO of Sapien, a decentralized data foundry where AI models can access verified human expertise worldwide. They've raised raised a $10.5M round led by Variant. He's also the co-creator of Coinbase's layer 2 network called Base.

Rowan's favorite book: Outlive (Author: Peter Attia)

(00:01) Introduction
(01:09) The Flaws in Centralized Data Models
(04:10) Mechanism of Knowledge Transfer and Expert Incentives
(07:08) Supply, Demand, and Market Dynamics for Training Data
(10:22) Chain of Thought Reasoning and 3D/4D Data Use Cases
(12:22) Building the MVP: What Worked and What Didn’t
(15:17) Acquiring the First Five Customers
(17:59) What They Got Right and What They’d Change
(20:15) How to Scale from Early Customers: Advice to Founders
(22:02) Data Infrastructure Opportunities in 2025
(25:57) Designing AI-Native Databases
(28:04) Biggest Startup Challenge: Messaging and Clarity
(30:22) Future of Data Collection Mechanisms (2 to 5 Years Out)
(32:07) Autonomous Vehicles and Demand for 4D Data
(35:33) Emerging AI Use Cases: Memory, Wearables, and Robotics
(36:19) Rapid Fire Round

--------
Where to find Rowan Stone: 

LinkedIn: https://www.linkedin.com/in/rowan-stone/

--------
Where to find Prateek Joshi: 

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
X: https://x.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.481)
Roman, thank you so much for joining me today.

Rowan Stone (00:05.166)
Thanks for having me. Great to be here.

Prateek Joshi (00:08.359)
Let's start with the basics. Decentralized data foundry. Before we start, let's talk about data foundry. So for people who may not know, can you explain what a data foundry is?

Rowan Stone (00:24.136)
Absolutely. And so before I do that, maybe it's better to take a little step back. We are an AI data company. And what that means is we source or structure data for companies that are building AI models. And when we say a data foundry, we literally mean a system or a platform that enables people all around the world to, in a structured way, either provide expertise and knowledge that they may hold in their heads.

or help us structure information that's gathered by customers, essentially forging net new information, which we then use to train AI models, or rather, the customers then use to train AI models.

Prateek Joshi (01:09.415)
In the traditional centralized world, can you talk about what gaps or shortcomings you saw that led you to this?

Rowan Stone (01:20.29)
Yes. And so there's really two main reasons that we believe this is a better model to create the data that's needed for AI. The first one is the quality of data. so AI models are very much garbage in, garbage out. If you feed it crappy data, you will get a crappy model. The model will hallucinate. It will say things that aren't true. It'll generally just not understand the world around it because you're giving it inaccurate inputs. And so therefore you're

make it inaccurate outputs. One of the things that's problematic is in the centralized model, whereby you're typically operating in like a hub and spoke, where you may have decentralized people that are providing inputs or providing information, but typically quality control is done in one place. And with quality control done in the Philippines or Bangladesh or any other country in the world, it really doesn't matter. The problem you encounter is that if you're

Creating truth via a thousand 20 year old Filipino men, that is probably going to work really well if you're building a model for that market. Like if your intended audience is literally 20 year old Filipino men, great. Like the source of truth is aligned. The problem comes when you're trying to build something that's globally relevant or something that gets actual ground truth. The difficulty being you're going to have huge biases in your data, whether they are age,

geographic, religion-based, any number of different slants will appear in the data because you're relying upon one specific subset of the population to give you your source of truth. And so our view is that very simply, every single person on earth, whether they are a podcast host, a VC, a musician, artist, lawyer, doctor, engineer, doesn't matter. All of us have information in our heads that's super valuable to AI.

really what we're trying to do here is build a system or a platform that structures that ability to transfer knowledge so that people can actually monetize what they have in their heads. But importantly, the companies that are building these new vertically specific models can access the data they need. so circling back, because I'm going on a tangent here, the first big part is the quality of the data. And the second part is that it's 2025. We can do this in a smarter way.

Rowan Stone (03:47.874)
and can involve more people and we can do it in a more ethically fair way. And so the second part of this is that it's just a better model. It doesn't make sense to build five offices in five different countries and fill them with thousands of people when in actual fact, all you need to do is align incentives between a network and have people participate from wherever they are.

Prateek Joshi (04:10.581)
Now, I would love to dive into the mechanism of the knowledge transfer because, obviously, MD, like a doctor in Romania, amazing knowledge in the head, but because they're so rich and they're happy with it, like, I think getting them to transfer knowledge onto like a training dataset would be difficult, versus somebody who really needs the money, they're willing to put in the time to do that.

Can you talk about the mechanism in which the knowledge transfer happens and also in that spectrum of great expert, but they don't want to do it versus they really want to do it, but they're not experts. So how do you balance this spectrum?

Rowan Stone (04:57.558)
Yeah, great question. And so I'm going to give you two different answers here. There's like a now in terms of how our system works today, and there's a future in terms of how it will work very soon, but not quite yet. And so the real answer to what you're saying is just simply supply and demand will dictate the price and price will slowly increase for this, for the kind of more expert level knowledge until there's someone that's willing to do the work. And that's

the reality. These models, if they want this knowledge, they need to pay for it. And so if the $300 an hour incentive for a doctor to do cancer screening on radiology images is not enough, then they're going to need to increase their bid and increase their bid until somebody says, you know what, I'm going to do that piece of work. Because the vast majority of people want to earn more income. And so this is an opportunity for them to, between patients and the doctor analogy, to have that extra income.

And so the way it works today is that we are actually doing the pricing, which is really not what want to be doing. We are going out to market and we're finding either from within our network or outside, depending on how specialized the task actually is, we're finding the people that can do it. And we are pricing it up and then we're quoting the customer and we're doing like a really manual enterprise sales mission with a team of people. And we're doing a good job. We've secured 27 customers, including some Fortune 100s and

fairly ridiculous companies that frankly we don't have any right to be doing business with, but here we are. And it just doesn't scale. Like we are not going to grow a massive team of salespeople in every continent. That's not the right way to build this business. Instead, we are going to flip to an open marketplace and we are going to build the system that connects supply and demand intelligently and we will take a flat fee. And so in this world, supply and demand will literally dictate pricing. We won't be involved.

The customer will have an idea in their head as to what they need and how much they're willing to pay for it. And if they don't get the work, they need to raise their bid. If they get too much people replying, then clearly they've had their price too

Prateek Joshi (07:08.479)
I love the markets are going to market. Somebody tweeted this recently, like, market's going to market, meaning the supply and demand, have to figure out how to meet somewhere where people demanding the service will get what they need, and they have to pay up. Now, in the world of training data, where are we in terms of just the training data?

what OpenAI or Anthrobic, they're scraping the entire world, they're building their foundation model. But in terms of just the training data needed by big companies, can you just paint a picture of where a Walmart or some big company is in terms of the training data sets?

Rowan Stone (07:56.504)
This is really hard to answer because every single company has a different perspective on what good looks like and are building a vastly different model. so maybe I can just talk specifically about two of our kind of most commonly requested data types. The first one is 3D and 4D data. And so think of this as the information that a car like a Tesla, I'm going to use Tesla as an example, we do not do work with Tesla.

They actually do their own work now. They realize the importance of data and brought this in-house. But it's a great example. Everybody is familiar. Elon Musk and Tesla have been working on full self-driving capabilities for a long time. Their cars have a plethora of sensors around the periphery, and they collect literally petabytes of data between their whole fleet. They're constantly encountering situations that they don't understand. Weird roadworks or people jumping out in front of them or just random

world things. And so they need human assistance and input to A, help them figure out like what are they actually looking at and B, to figure out how to deal with that particular scenario in order to keep the passengers safe and everybody around the car safe. And so that's one of the most common types of work that we do. It's 3D and 4D data. We don't work with Tesla, but we do work with the likes of Toyota and Amazon Zooks who do really cool robotaxis across Las Vegas.

GAC and a few others. This type of data also lends itself really nicely to robotics, either manufacturing robotics, because they have very similar sensors and they're looking at space and time in the same way, or humanoid robotics, which will slowly start to see an increasing numbers out in the wild. And these are essentially just a self-driving car with legs instead of wheels. From a data perspective, it's almost identical, except rather than annotating and helping them understand

road positioning and things that may be hazardous around the road, it's within a house or doing chores, working on a construction site, working in a factory, whatever the case may be. So 3D, 4D data, big part of our everyday work. The other thing that's very commonly requested is net new information. We talked a little bit about kind of expertise. Typically that's called chain of thought reasoning. And so if we used a doctor example from earlier, this could be

Rowan Stone (10:22.498)
I believe this particular part of this radiography image is an issue. That's like your bounding box type labeling task, very simplistic, however, completely invaluable. It's not very valuable to a model because what they really need is the understanding and the chain of thought that led the person to make that choice. And so it could be

This particular part is shadowed and has a ground glass occlusion around it. And so therefore I think it is X, Y or Z. And this is the real value to an AI model to help them properly understand. So data collection is the other huge part and it can be everything from voice. I'm from Scotland. I don't sound like someone from the US or I don't sound like somebody even from England. And so when you are using Google Home or Alexa or Siri,

If you have anything other than a plain vanilla, perhaps West Coast US accent, you're probably used to having a bit of a hard time, or at least you were. Things are getting better now. But things are getting better because these companies are asking for huge amounts of voice from people like us candidly, who have different accents and are able to help use that information in lots of different languages, be them English, Spanish, Portuguese, whatever, to make their voice assistants more useful.

We also have image capture, video capture, and a whole bunch of others. But stepping back again, it's 3D, 4D data and data collection. That's kind of two of the biggest asks from our customers.

Prateek Joshi (11:57.333)
Now let's go to the launch of Sapien. I'm going to take you back because many of the listeners are either early stage builders or people who are thinking about building something. So what did the MVP look like and also how did you decide what features would go into this before you just ship it?

Rowan Stone (12:22.808)
So MVP for us was a very simplistic labeling interface where we didn't have any modality other than here's an image, tell me what you see or draw a box around something like draw a box around the AirPods type stuff. Computer vision can do everything that our MVP did back then automatically without any input from any people. And so you'll hear people saying that this type of work is very low margin.

I would go a step further. This type of work is literally worthless at this point. Computer vision model will do it for us. We do not need this. Really the puck is moving up to much more sophisticated types of data or nuanced understanding of context. And so the MVP was really simplistic. It was like two colors. Web app user was able to log in just with an email and they were tagging. And so they were like,

Here's an elephant in a field. we actually did a, bizarrely, it was valuable at the time, but we did a project for a university in Canada and they were building a model to recognize different types of animals. And so we had literally like, what is this type thing? And that was one of the first tasks live in the platform. But again, no advanced functionality, no gamification, design was terrible. Just a very quick and dirty

Proof of concept to see A, are people even remotely interested in doing this? And B, if they are, can we get anywhere near an acceptable standard of data out of this? And we learned a ton. We quite quickly learned that people will try and game anything. I don't think this is particularly new, but we kind of reaffirmed the understanding that we're going to need to be very careful with how we frame things. And we're going to be very careful in terms of the bot protections and different layers of

of kind of checks and balances for quality. And then we also learned really quickly that it's difficult to get good quality data when you're enabling literally anyone anywhere to participate. And so those two learnings really started to feed the roadmap. And then learnings from our customers, particularly in terms of like, this is completely nonsense. We don't need this. We would love some of this started steering us down a different path. And so today we're building a very different business. We're no longer building

Rowan Stone (14:50.046)
was essentially decentralized scale AI. We're now building a protocol ultimately. And if we're not speaking to people that are in the kind of tech or crypto space, then it's probably better to say a system. But we are building a way, a framework to allow people of any type and background to transfer the context of the knowledge that they have to models that need it.

Prateek Joshi (15:17.205)
And you talked about the early customers. So can you go into how you acquired your first five customers? Like what are all the things you had to do to get to them, talk to them, convince them, and obviously convince them to not only use the product, but pay you money.

Rowan Stone (15:38.508)
Yep, it's beg, borrow and steal is basically the short answer here. We knocked on a bunch of doors. We had a lot of conversations. We learned a lot in terms of what people actually needed and wanted. And we used that to scope and hone what we were building to get it to a point where it actually could fit their needs. We did a ton of things that don't scale as you should as a startup. And I think the important thing when you're doing these things that don't scale is to kind of understand

a future path that brings you towards a system that you can actually operationalize. But the really short version is we knocked on tons of doors and we went to places that are underserved. And so if we think about the current data landscape, probably the biggest incumbent provider is a company called Scale AI. They're awesome. They've built an epic business, co-founded by Alex and Lucy Guil. Lucy's actually one of our advisors. She's been super helpful as we build out the system.

But they are in a position where they're doing work for like the US government and the Pentagon and a bunch of other different kind of government level agencies. And this is really useful for us to know because any company that is doing work with government level agencies in the US is not going to be doing work with places like China. And so China is a huge technology hub with some of the most talented builders in the world.

And there's a ton of really cool models coming out of that part of the world. Scaly Eye is not there. Maybe we should be there. So we sent our chief operations officer, Henry, who is fluent, which is very useful. And he was basically banished to go live in China for a couple of months. And he came back with our first few large customers, customers like Baidu and Alibaba. And that gave us a little bit of a footing and allowed us to learn a ton. And since then, we've just been knocking doors and

applying those learnings.

Prateek Joshi (17:37.237)
Now, looking back, the 0 to 10 customers, you already got past that. So in that journey, 0 to 10 customers, what's one thing that you think you got really right? And also, what's one thing that if you had to go back, you do very different?

Rowan Stone (17:59.542)
I think if I was to go back in time from a different perspective first, I'm quite a negative person so I can always start there.

I think I would have pivoted sooner. I think we had line of sight that what we were building probably wasn't going to be super useful long-term, but you kind of get stuck with blinkers on and you convince yourself in like an internal circle jerk that actually you have the solution. Even though you're hearing from customers that perhaps it's not exactly what needs to be in market.

I think if I was to go back, we would have started building what we're building now faster. We wouldn't have wasted six to eight months trying to go down a path that didn't appear to be that solid. The things that we did right, I think I just talked about a second ago, we went where we knew there weren't a ton of competition or competitors. That enabled us to really just punch up our weight.

we were a tiny team with not very much resources and we were winning business from some of the largest companies in the world. And so that wouldn't have been possible had we went to a super saturated market. However, now that we've managed to win some very large customers and proven that we can deliver good quality data, we can now go back to these saturated markets with real case studies and real references from large household names and say, we can do this. Like we have the ability to get the data that you need. You want to build a model?

where your partner. And so I think that was smart.

Prateek Joshi (19:37.961)
That's great. Now, once you've got a first handful of customers and you know, hey, we are onto something, this is kind of working. And after that, it's just, in any startup, it's just a sequence of experiments with the hope that something clicks before you die. So if you were to advise a young founder who's in that, we got our first five customers, kind of something is working. What experiments do you?

recommend they should run to quickly figure out how to double, triple, or quadruple the number of customers.

Rowan Stone (20:15.946)
I would probably actually recommend rather than doing a bunch of experiments, go and just spend a bunch of time with those five customers, learn about every single pain point that they're facing and try and figure out where you can add any additional value to make what you're already providing them more useful. And that really is going to dictate your early roadmap. It's also going to give you a really good flavor of what your sixth, seventh and eighth customers are likely to want.

because they're probably at a similar boat. We always like to think that every company is vastly different, but ultimately they're all facing similar problems. Right now, if we think about the AI data space or just enterprise businesses in general, they're being squeezed by AI. They're worried that they're going to be made irrelevant because AI can all of a sudden do a big chunk of what they've been historically making money doing. And so the response is we need our own AI model to augment what we do and make us

more efficient, faster, better, stronger, whatever, or provide our customers with more value. So we can give them a new product to play with or help them deploy code faster or write emails faster, whatever the product may be. And so being able to recognize the pain point and address that pain point, provide a solution is for me more valuable than going and doing a bunch of A-B testing on new product features. I kind of feel like product should typically be

steered directly by demand rather than blindly experimenting. I feel like blindly experimenting is good in marketing land, but products having learnt the opposite is better steered by real market demand.

Prateek Joshi (22:02.069)
That's actually a wonderful advice. Just kind of, if something is working, like double down, meaning spend time, figure out what's working, and then be even more useful to them. Now, if you had to take like a 5,000 foot, like 10,000 foot view into data infrastructure, meaning you spend time with big companies, you know where they're willing to spend money, and more importantly, you know where they wouldn't spend money.

Data infrastructure opportunities. If you had to pick a couple that are interesting in 2025, what would that be?

Rowan Stone (22:46.286)
I think probably, and I'm going to be selfish and talk from our perspective again, the biggest opportunity I can see is standardization. And what I mean by that, my experience, at least in recent time, is particularly in the crypto space. So I'm going to use crypto analogies and hopefully they make sense to some people that are listening to this. But in the crypto space, relatively nascent technology

not been built for very long. We're like 10 years in, 11 years in, something like this. And there are perverse incentives.

Rowan Stone (23:35.778)
up with a really clunky, like almost inoperable world. And if you're a user who's trying to navigate this world, none of your tools speak to each other and nothing works together. Now, from my perspective, the AI space is going exactly the same direction whereby everyone wants to build their own model. Nobody is sharing data really. There's a couple of open source things, but the vast majority is proprietary, both in algorithm

actual data and the compute, like nobody wants to share their compute to train their competitors' models. And so where we are is that we have 500 different models that don't talk to each other, and we have 500 different types of data set that aren't compatible with each other. And the net result is that we're going to be very slow to develop anything new, much like crypto, because we're all doing our own thing, we're not working together.

And so I think one of the biggest opportunities that we're currently thinking through is how do we create glue between all of these disparate worlds? And how do we create a shared technical standard that can at least solve some of the interoperability problems by allowing datasets to be shared between companies and ultimately expanding what we're building now from being come and self-serve and get the data you need.

Two, instead being, if you're a data company, come and plug in and source the people that you need or source the data that you need or basically just join the network and build alongside us rather than the usual, let's compete with each other. And so I think creating a shared technical standard for different modalities of training data is super important and something that we're actively thinking through and working with some partners on.

Prateek Joshi (25:57.149)
Okay, now thinking about the database layer for a second, if you were to design a brand new database that's completely AI native and you don't care about any other use case, what does that look like? Like what should a database be able to do well to serve this AI native wave we are in?

Rowan Stone (26:25.88)
It's a great question. It's not something I've ever thought about, to be honest. So I'm probably not going to give you a particularly good answer. I feel like if we took our CTO on here, he'd probably have a really good answer for you. And so for me,

The most intelligent answer I could give you is that it needs to be readable by AI and agents. And I think that's a topic that people are starting to get their head around in that even just the way we advertise our businesses is going to need to change pretty dramatically in a world where most people are actually searching things with ChatGPT before searching things with Google.

And I think a database would need to be positioned and framed and made accessible in the same types of ways that we will index and make accessible our websites or product portfolios or whatever to the AI world. And so that's a commercially answered answer. But yeah, if you want a better technical answer, we can get our CTO on here and he can give you.

Prateek Joshi (27:27.615)
No, think you kind of hit it on the head, is many of the databases that we've used or we're using, they're designed many years ago. They weren't designed, they weren't even thinking about the LLM era. And that's why I always wonder, for some reason, there's no single global standard database. There are so many of them, and they all serve their own vertical use cases. So no, that's great.

Coming back to company building, if you had to pick the biggest challenge that you had to overcome in building Sapien, what would that be from day zero to today?

Rowan Stone (28:04.272)
Rowan Stone (28:15.354)
Probably.

So once we figured out what we actually were building, we went through a couple of iterations of potential product. We essentially did the shotgun, like smash a bunch of stuff against the wall, whatever sticks, let's double down on that. And it became very clear that the path we're on now is the right path. We have good traction from customers. We have good traction from users. But I think probably the thing that we've done

The least good job on so far, going to be nice to ourselves, is explaining storytelling, messaging, imagery. Our website just now is a mess. I really hope that everybody listening to this just doesn't look and pretends it doesn't exist. We're actively working to fix these things, but I think it's such an important part of the puzzle. If you can't coherently explain why you exist and why people should care, you have zero chance of people caring. And we are very lucky that

what we've built is in such high demand that even though we've done a pretty shitty job of explaining, we've still managed to get good traction. And then the other reason that we're very lucky is that people all around the world typically like to earn extra money. And so because we're offering people to be able to earn extra money, again, even though the messaging perhaps isn't perfect and needs some work, we're seeing good traction. And so I think the thing that we need to double down, triple down on

our stage today to get to that next level is just massively upgrade everything that people can see. Get to the point where within 30 seconds, if you're an enterprise, you're going to understand the value we can provide to you and your business when you're building your model. And if you're a user, a person, someone looking to make money within 30 seconds, you can quickly decide, A, do I have the skills, time, desire to do this? And B, if I do, what is in it for me?

Rowan Stone (30:08.559)
And I think those two things are what's gonna help us get to the next stage of business. And it's definitely what I think we've done the least good job of.

Prateek Joshi (30:22.101)
No, as always, think startups are resource constrained, time constrained. There's always like 17 fires you have to put out. So think it's always good to get that perspective. Now, if you had to look at the future of this data collection mechanism, meaning the way in which we source the data, we collect it, we label it, we curate it, and we make it usable for AI models, so the whole cycle.

Where will we be in two years and maybe five years?

Rowan Stone (31:30.648)
Okay. And so where will we be in the next two years, three years, five years? Can you just, are you meaning specifically from a data perspective or do you mean like as an AI industry full stop?

Prateek Joshi (31:47.553)
the latter. Basically, the industry as a whole, the data industry as a whole. So where will the industry be in two, three, and five years? Like what would be the way in which industry will function?

Rowan Stone (31:53.157)
Mmm.

Rowan Stone (32:07.214)
Okay, I think it's actually a really fun question. I think within the next two to three years, we are going to see a fairly dramatic shift in how we move around. And what I mean by that is that I think we're going to have a meaningful percentage of cars driving themselves. And that means there will be a huge amount of 3D, 4D data that needs to be interpreted.

and that needs to be sculpted and structured and then sent back for inference and training so that we don't kill a bunch of people and that we have cars that drive and operate safely and can do millions of miles every year in lots of different places. Unfortunately, this is not one of those things that you can teach a car to drive an SF and then take it to San Diego. think Waymo have quite famously just found this out for themselves. They got their Jaguar F-Pace or whatever they are, the white cars to drive really well.

really nicely at SF. They took them to San Diego, I believe. Doesn't work. Doesn't work at all. And so they're going to have to start again, basically, to teach them to drive there. And so think the next two to three years, we're going to see most cars having this functionality, at least most new cars. And a meaningful percentage of miles driven will be done autonomously, which means there'll be a lot more demand for 3D, 4D data. That's at least the bet that we're taking. And we're doubling down on that area of the market.

I think the other thing that's going to be net new is that we're going to see robotics change. Manufacturing robotics previously has been static arms, assembling components, welding, gluing, spraying, things like this. And I think we're going to start seeing essentially the same stuff, but on legs or wheels that can move around. And again, to us will generate huge quantities of demand for 3D, 4D information.

You may be listening to this and thinking, no, we're going to have a year of demand and then we'll be able to extrapolate everything from there after. I would push back on that and say that randomness is going to create demand almost in perpetuity because you will always have a weird scenario that we haven't yet encountered. And the safest way to handle that random event is to do nothing until you have clarity and confirmation what the right thing to do is.

Rowan Stone (34:32.58)
Particularly if you're a 10 ton robot with people walking around, or you are a two ton vehicle driving through a city. The safest thing is always to do nothing. And so I think that will continue. And then I think looking all the way out into the future, we're probably going to have most AI living within some sort of different device. Today we are talking to our phones via text. Some people may be talking to their phone via voice.

and we are on our laptop augmenting our coding, augmenting our communications. I think it's not going to be long before we have a different way of using AI. We're already seeing companies like Granola who are listening to essentially all of your meetings in the background and then giving you actually really nicely structured and have decent notes. I have no dealings with Granola professionally. I don't know the team, but here's a free plug for them. Like if you're not using Granola, you probably should be. It's really good.

Prateek Joshi (35:31.379)
Yeah.

Rowan Stone (35:33.232)
I think things like that will become much more common. We're seeing things like pins that will just kind of listen to your daily life and then give you an oversight of what's happening or help you with basically augmenting memory. All of these things will generate different types of data and need different types of structuring to help make them useful. And ultimately, I think we're going to end up with AI essentially living within some sort of robot, whether that's a humanoid or a

avatar style thing that's kind of walking around your house looking like a dog, cat, rabbit, whatever. I think that's going to be the next big step. And that will create a whole new type of demand for data and information, which will look in our view, very similar to your 3D, 4D type stuff.

Prateek Joshi (36:19.145)
Right, amazing. With that, we'll add the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready? All right, question number one. What's your favorite book?

Rowan Stone (36:30.384)
Let's do it.

Rowan Stone (36:35.344)
most recently Outlive by Peter Attia.

Prateek Joshi (36:40.959)
That's an amazing book. Yeah, it's really good. read. 

Prateek Joshi (37:29.909)
What has been an important but overlooked AI trend in the last 12 months?

Rowan Stone (37:41.42)
I think likely overlooked is a little bit tricky. think AI in general is very talked about, so there's not much that's overlooked. But I think what I just spoke about is probably the most overlooked at my view, that the way we interact with AI is going to change dramatically and it's going to change very quickly.

Prateek Joshi (37:58.705)
what's the one thing about decentralization that most people don't get?

Rowan Stone (38:08.343)
It's a buzzword. People are always saying, decentralize everything, decentralize this, decentralize that. To us, to me, that makes no sense. There's reason to decentralize certain things. It adds massive value in certain ways. A lot of the time though, if you decentralize something, you make it worse. And so don't decentralize everything. Decentralize the things that actually benefit from being decentralized.

Prateek Joshi (38:32.103)
What separates great AI products from the merely good ones?

Rowan Stone (38:37.848)
understanding context and high quality data.

Prateek Joshi (38:41.769)
What have you changed your mind on recently?

Rowan Stone (38:47.312)
I don't know that is humanoid robotics. I am a massive Star Trek geek. And so for me, humanoid robotics is just data. And I've always thought about it as nothing but a good thing. However, in recent months, I've been thinking about it more and more as actually potentially quite a sinister and perhaps not good thing, but we will see.

Prateek Joshi (38:50.463)
for better or worse.

Prateek Joshi (39:08.979)
What's your wildest AI prediction for the next 12 months?

Rowan Stone (39:14.32)
wildest for the next 12 months, AGI. I think we are so close.

Prateek Joshi (39:21.141)
All right, final question. What's your number on advice to founders who are starting out today?

Rowan Stone (39:28.314)
Do it. Literally just start. Take the first two steps. The whole path may not be clear. You will not know until you start walking. So start walking.

Prateek Joshi (39:38.483)
Yeah, it's funny how many founders, I think it boils down to this single essence. Like there's no magic, big, amazing thing. It's just, you just do it. That's just pretty much it. The thing is like, yeah, that's about it. So that's amazing. No, Robin, this has been a brilliant episode. Love the learnings, the hard insights, the pivot. And I think that the journey of just like founding and building a company.

It's amazing. So thank you so much for coming onto the show and sharing your knowledge.

Rowan Stone (40:13.496)
Thanks for having me. Really appreciate it. It's been fun.