Subscribe
Share
Share
Embed
AI is the biggest technology shift of our lifetime. This show is about how to profit from it together.
Each week I talk with the founders and CEOs closest to AI and Content, the ones figuring this out in real time.
I’m also building an AI content business myself and share the lessons I learn along the way.
WHAT WE COVER
The Titans -- How companies like OpenAI, Anthropic, Google, Meta, and xAI are moving, and why their decisions matter.
The Incumbents -- How content giants like Disney, News Corp, Universal Music Group, and Reddit are responding to AI, and what it means for creators and publishers.
The Playbook -- Real lessons on AI business models, content strategy, creativity, IP licensing, distribution, and getting paid.
Family & Our Future -- Every episode ends with me asking my guest what AI means for our jobs, our families, and the next generation.
ABOUT YOUR HOST
Rob Kelly has interviewed Steve Jobs and Bill Gates, helped pioneer early web content licensing, and built multiple companies with more than $100 million in total sales. His work has appeared on CNBC, CNN, TIME, and Entrepreneur.
Thanks! -Rob
I'm Rob Kelly, this is Media and the Machine, a show about the biggest technology shift of our lifetime and how to profit from it. Each week, I talk with the founders and CEOs closest to AI and content, the ones figuring this out in real time. I'm also building an AI content business myself and share lessons of what I learned along the way. You know, life's funny. I began my career lucky enough to interview leaders like Steve Jobs and Bill Gates.
Rob Kelly:I Then went on to be a three time founder and CEO, driving a $100,000,000 plus in revenue and some failures too. And now I'm back at the table, interviewing this new world's current and future leaders. This isn't only a business story, it's a human one. So every episode ends with me asking my guest what AI means for our jobs, our families, and the next generation. We'll figure this out together from the inside.
Rob Kelly:Welcome to Media and the Machine. My guest today is Marty Pesis cofounder and CEO of Troveo. Troveo helps content and data owners license their video, audio, text, and other data to companies training top AI models, models, like OpenAI, Google Gemini, Anthropic, and AI unicorn startups too. Marty says he may have the largest AI data corpus in the space. We open the interview with Marty's very first AI data deal, request so strange I had to stop him and make sure I was hearing it right.
Rob Kelly:We get into how Tesla cars are really training machines on wheels. We talk about the future of data collection. Picture factories with workers wearing GoPros, wrist cameras, shoulder cameras, sensor gloves, and narrating what they're doing so robots can learn. We cover the pricing of data, including what type of content AI models will pay a thousand dollars per hour for, and the one library of content he'd want if he could have any. He also shares what he learned from Troveo investor Alexis Ohanian, the cofounder of Reddit, and also from mister beast who bought Marty's last company.
Rob Kelly:And we end as usual on AI and families, including Marty's view that because of AI, college is not even on his radar now for his young kids. Please enjoy my conversation with Marty Pessis. What was the first AI revenue you made for an owner of content?
Marty Pesis:That's a great question. The first AI revenue that we made as a company, we were working with a company building a speech model where it was detecting different barks from dogs and trying to turn that into English words. And so what we actually did was we went out and sourced a big audio data set.
Rob Kelly:So the AI model company wanted to turn barks into the equivalent, like, human words?
Marty Pesis:Exactly. So the idea would be if a dog's owner comes home and the dog barks a certain way, the human could actually hear the dog speak a certain word. So it was a dog collar that was trying to personify dogs. So if you've ever seen, like, dog days on Disney, that kind of idea where the dog wears a collar that allows it to speak, the collar would had a little speakerphone on it. This was the concept.
Marty Pesis:So we went out and we sourced a ton of dog content. We actually created about half of it and we sourced about half of it. That was our very first project when I started the company.
Rob Kelly:Was it just a database of barks that you got, or were you also getting the barks and the guesses of what the barks were in human language?
Marty Pesis:So a little bit of both. It was mostly just different barking sounds, but we had to have at least a label on it. Right? So in addition to barking, by the way, there's other dog noises that we were going after as well. So even, you know, weird ones like vomiting and crying and, you know, all all sorts of, like, urgent bark, scared bark.
Marty Pesis:We kinda had to highlight these different ones. But, yeah, that was our very first deal.
Rob Kelly:But I just got into Hobag. So did the dog collar get launched?
Marty Pesis:We supplied a lot of data for it, but I never saw the dog collar actually go live. I think that would be a pretty big viral launch video, I would say.
Rob Kelly:Yeah. I would buy that for my best friends who have dogs. I wanna hear it. How large is your library of data currently?
Marty Pesis:We've got about eight million hours of video, just over four million hours of audio. We've got billions of words of text, hundreds of thousands of hours of gameplay data, and hundreds of petabytes of data overall in our library across these kind of six key different types.
Rob Kelly:Where does that rank in the world in terms of a dataset for AI to use?
Marty Pesis:We're confident that our video library and our overall corpus of data is the largest in the space.
Rob Kelly:Gotcha. Could you walk through just the AI ecosystem in your world?
Marty Pesis:Yeah. Absolutely. I'd say there's really the model makers, ones that are training models, are tuning models, the Frontier Labs and Mag seven and the big tech companies, that's one bucket. Then there's the startups. The startups in this space, obviously, lot of them are very well funded.
Marty Pesis:Right? Seeing a $1,000,000,000 seed round is not super frequent, but you see it in the space. Right? And a lot of these startups, they hit, they're scaling very quickly and data is a huge part of it. So that's who we're often licensing content to.
Marty Pesis:That content then goes into the training runs. And that's really kind of the demand side of our ecosystem and our business. And then, of course, we have the supply side of our ecosystem, and that's the content owners, the data owners. These could be video production companies. These could be businesses.
Marty Pesis:These could be game studios, animators, independent filmmakers, huge wide range of of different content owners and data owners.
Rob Kelly:Would you call yourself an aggregator? Because you're not creating the content from scratch as you pointed out in every case.
Marty Pesis:We call ourselves a data provider, whether that means collecting and sourcing data, whether that means enriching and annotating data. We're helping our customers get the data they need to accelerate their model development.
Rob Kelly:And you mentioned the data enrichment. So if I'm sitting on a thousand hours of video, it's not as simple as me just opening that feed up to an AI company, right?
Marty Pesis:Correct. If we're talking about video, for example, identifying what you're seeing in the video, different movements and maybe different camera shots. You can also add a lot of context to the video, right? So actually understanding the human reasoning and the decision making. So if somebody is running down the street and they take a right turn, the context layer basically provides the answer to why did that person turn right instead of turning left.
Marty Pesis:And this all kind of falls under the category of data enrichment or data preparation.
Rob Kelly:How many creators of content contribute to the entire library?
Marty Pesis:The latest number was just over 8,500 actually. Content owners and data owners, AI companies only need so much of each kind of type and category of content. So it's very important that we're always thinking about diversity. It's like when we were talking about the dog content. If somebody needed dog audio content and all the barking dogs were from golden retrievers, that would not be a very diverse dataset.
Marty Pesis:Right?
Rob Kelly:I want to go through some AI use cases. And can we just start with, like, kind of the common requests from AI companies?
Marty Pesis:Yeah. I would say the common and most frequent requests we get, especially on the video side, would be things like trying to ultimately create content that looks cinematic. Everyone wants their model to be able to produce cinematic, super high quality, very crisp and clean videos. So that's something that we see very often. Talking head, like in the avatar space, is really prominent as well.
Marty Pesis:Everyone wants to be able to produce human faces that are very real, that are synced perfectly to audio. Another one that comes to mind as well is videos with a lot of interaction. Right? So sports is a good example of this. There's a lot of interactions that happen inside of a video.
Marty Pesis:Right? So people catching a ball, two people colliding, a ball being dribbled, let's say. And the reason for that is that it can really help teach physics and getting down to those very specific details. So if we use the ball bouncing, for example, the model needs to know that when the ball bounces, it doesn't just move up and down. It actually hits the floor, compresses, and then bounces back up.
Marty Pesis:And there's a lot of examples in sports content specifically that are very similar to this. Same as when two people collide, what actually happens when those two people collide versus just two objects coming together. So anything where there's, like, a lot of interactions is very prominent as well, and we get a lot of requests around that.
Rob Kelly:Now can you give me the more wild than usual use cases that AI companies are asking for these days? The type of data or content they're extra hungry for?
Marty Pesis:So an example would be like a humanoid robot that would potentially be doing chores around the house and be doing kind of household to dos. Let's use one specific example around, like, cracking an egg. Right? The data required to train a model and a robot on how to crack an egg, it gets really complex, right? It's not just a video of an egg being cracked over and over again.
Marty Pesis:They actually need to capture like sensor data, let's call it. They have these gloves now that allow a model to understand how much force is being applied both from holding the egg itself to not crack and when you actually, you know, kind of slam it down on the counter. If that same robot is folding clothes, for example, it needs to be able to detect, is this a shirt? Is it a sweatshirt? Is it pants?
Marty Pesis:As you get deeper and deeper into it, the nuances around the data needs is what really stands out to me and gets me kind of nerding out about it. And you can go this deep in really every single data type and every category, but that's ultimately what makes and breaks the dataset and ultimately leads to better performance on the AI side.
Rob Kelly:Oh, by the way, just on the cracking the egg, I got to know this last part. So are there people out there right now wearing gloves with sensors cracking eggs?
Marty Pesis:Certainly. This is happening in, yeah, in a major way in in the industries and
Rob Kelly:Are you getting the data? Like, you have this data in your dataset already or are getting it?
Marty Pesis:Yeah. Definitely. Think about manufacturing lines at a big warehouse. They're wearing a GoPro on their head, video cameras maybe on their shoulders, on their wrists. They're also wearing sensor gloves.
Marty Pesis:In an industrial setting like that, you're trying to capture data from every different dimension and every different angle. So video data, audio data, sensory data. A lot of times people at manufacturing shops are also voice overing what they're doing to help teach these models, ultimately robotics models, helping to teach them what they're doing as they're going. Yes. Capturing this data across all the different dimensions is hugely important.
Rob Kelly:Right now, who's willing to put on these sensor gloves and crack eggs for you? What type of person or company is even doing this? Is this just overseas, outside The US?
Marty Pesis:It's all the above. I mean, I think as the world is starting to understand the importance of data and how much of an impact it's making on these models. There are more and more people thinking about capturing data, but all of it is really driven by the demands of the AI companies, of course. Right?
Rob Kelly:And you mentioned factories earlier. So do you picture, like, going forward that if you walk into a what we would call just a regular factory today of someone making whatever, that they're wearing some sort of GoPro, that they are wearing a glove with sensors, and they are possibly even speaking out loud doing voice over about what they're doing, all to help train an AI. Is that the future you're picturing?
Marty Pesis:Absolutely. Absolutely. And it doesn't just extend it's not just factory workers and people in the manufacturing space. It's really happening across all companies. I'm sure you saw this week or last week Meta announced that they're gonna be screen recording and capturing data from work of all their employees.
Marty Pesis:I think we're gonna see that become a lot more common.
Rob Kelly:Streaking people out.
Marty Pesis:Yep. The Tesla dashcam dataset is another great example of this. Tesla's got a massive fleet on the road. All those cars are capturing data from multiple different perspectives that's ultimately helping to train the full self driving products and and features that they offer.
Rob Kelly:Yeah. Can we just talk about that for a moment? These are really data collection machines, aren't they? The Tesla cars?
Marty Pesis:Absolutely. And if you ever read some of Elon's docs, he always has these manifestos and these master plans that are kind of three steps, very simple usually. You know, put out a Roadster, a very affordable car. Ultimately, that kind of helps build the dataset that can then go train the models that make these cars self driving. Right?
Marty Pesis:Think about a Tesla. They have multiple external cameras set up around their car. They've got a primary one in the front, primary one in the rear. They have at least two on the sides, those pillar cameras that provide that three sixty view. So they're capturing the car moving throughout the real world and capturing all these different scenarios, people walking on the sidewalk, people walking in the bike lane, bikers biking in the bike lane.
Marty Pesis:You know, they're capturing every different scenario that happens on the road. That data is all taken, cleaned, structured, organized, labeled, annotated, etcetera, and put into the training pipeline to be able to create the autonomous driving, the features that we all love in our Teslas. Right? So the full self driving features.
Rob Kelly:Right. So it makes the Tesla better.
Marty Pesis:Yep.
Rob Kelly:So Tesla's data would be better than Waymo's data, like their equivalent?
Marty Pesis:There are a lot less Waymo's on the street than there are Tesla's.
Rob Kelly:And so Tesla now pretty much overnight, if they complied with regulations, can build a self driving fleet of cars like a Waymo because of all this data they've been collecting for years. Fair?
Marty Pesis:Yeah. Absolutely. This is all part of the master plan. Put out a car that's affordable. Elon's a ton of data.
Marty Pesis:Absolutely.
Rob Kelly:And then but not only that. And, the data that the Teslas are collecting can also be used for any other robotics and other parts of Elon Musk's empires, including training robots to do what? What else could they do besides self driving cars?
Marty Pesis:Yeah. One example of this would be walking down a sidewalk. Right? So this very similar data that you need to have a robot walk down a sidewalk. You know, it it would look a lot like a car driving down that same exact street.
Marty Pesis:So just capturing the square root
Rob Kelly:data the robot doesn't get hit by a Tesla.
Marty Pesis:Absolutely. Yeah. Or any other thing intruding it. Right? Or understanding what a stop sign is or, you know, all these things.
Marty Pesis:So, yeah, there's definitely multiple applications for the data that Teslas collect via their fleet of cars. And, yeah, that can certainly be reused.
Rob Kelly:So Chris over at Versos and I were talking about how that type of dashcam data could help a robot figure out how to take the garbage out on a dark rainy night. Right? The same dashcam type data, and then the use cases are just endless. Right?
Marty Pesis:Yeah. Absolutely. And some of this data can be captured when your car's parked as well. Right? So then they become static cameras basically that are observing the world around them.
Marty Pesis:But again, it always just comes back to what data are you collecting. And ultimately, in our world, in Troveo's world, it's data is the differentiator. So And the companies that have a really unique way to collect data and aggregate data and source data are winning in a major way. And so that's why that's what we're trying to help companies do too as well. Unlike scarce, unique proprietary data that they wouldn't otherwise have.
Marty Pesis:Really quick, too. I think video understanding is another, you know, really interesting use case that we're kind of, you know, starting to talk about as well. We've worked on a lot of deals around surveillance and CCTV footage, where people are basically trying to make better use and understand what's happening in all the surveillance video they create, and ultimately even predict what's going to come. Right? So if you want to better understand what are the chances that the footage or the content that you're seeing in surveillance footage might be real risky and drive to some sort of outcome that you're trying to avoid.
Marty Pesis:You need millions and millions of hours of surveillance footage, and then you start to capture these events and what happens right before to be able to start to predict when these things are going to happen. Right? So, for example, if you have a a video looking out on an intersection, it's like what happens in the seconds right before a car crash? But this would all be just another example under a kind of, like, video understanding and how people are making use of this different type of data.
Rob Kelly:If you could have one library of data or content that you don't already have, what would it be?
Marty Pesis:We have a lot of this, but I think one thing that's very valuable right now in the space is multi camera video. Let's say you're watching a sports game, there could be 18 or 20 different cameras watching the same game, all from a different perspective. Some are fixed, some are moving, but having those multiple angles of the same subject is extremely valuable right now. World model companies need this. Video generation, video understanding.
Marty Pesis:So, that's one area that we're really focusing on specifically in the video space. The other big area that we're focusing on right now is business data, enterprise data, and ultimately decision traces. So, understanding how real world businesses run and operate and make decisions, that's been a huge focus of ours over the last quarter or two. And I think we're seeing a lot of heat and momentum picking up in that category.
Rob Kelly:What's an example of the business data, these decision traces?
Marty Pesis:You can kind of think about any function inside of a business. So let's use sales as an example. So if you think about how a sale happens at any company, really, you receive an opportunity. Let's say this opportunity comes inbound. It may happen through a form capture.
Marty Pesis:Right? So, an inbound lead comes in, a salesperson picks up that opportunity. They probably have a conversation in person with their team to talk about the opportunity, and then they jump on a phone call with the customer. They get all the information they need, and then they jump into Google Slides, put a proposal together in a pitch deck, then they go back and present it to the customer back on Zoom to present it. Then they come back, get feedback, maybe they chat on Slack about the opportunity, and ultimately then it moves into legal.
Marty Pesis:So this entire workflow, right, what I'm trying to highlight is that it actually touches a lot of different surfaces, a lot of different platforms. And the better that you can capture that entire workflow end to end, that's ultimately what the agents need to be able to reproduce work that looks a lot like that. Cloud Cowork and all these kind of agentic tools, they're all built on these captured workflows. And so each different function of a company has these workflows, they all need to be captured a little bit differently. They all use multiple systems of record in different platforms, and these are the decision traces.
Marty Pesis:The real crux of the value of all this is how humans reason and make the calls that they do. So why did that person decide to put this slide into the presentation and not this one? Or why did they decide to put a slides deck together in the first place? So the decisions and the reasoning is really what we're after on the decision trace data and the business and the enterprise data.
Rob Kelly:So you've paid out $20,000,000 to content data owners so far. Is that since the founding of Trovio?
Marty Pesis:Yes. That's since the founding of Trovio. So in just under two years, we've paid out over $20,000,000 to content owners, which is a number we're very proud of and very excited across that figure.
Rob Kelly:How many content owners have actually been paid something?
Marty Pesis:Over over a thousand that have been paid something. We have multiple people who have made over $1,000,000 on Trovio. And we hear these stories. It's really exciting. They reach out and they tell us what they've done with some of these new revenue streams.
Marty Pesis:We've had independent filmmakers and other people that are sitting on valuable data reach out and tell us they've purchased homes, launched new business lines. So that is one of the most rewarding parts of this whole experience for me as founder of the company.
Rob Kelly:Can you maybe describe just the top 10 or so revenue makers of the content owners?
Marty Pesis:It's a pretty diverse mix. I would say it could be video production companies, studios, independent filmmakers, podcast networks, and businesses. Businesses are the other ones, and they're moving up the leaderboard right now.
Rob Kelly:Is there anyone who's now making pretty much all their income from this? In other words, that's their business?
Marty Pesis:I think people are seeing a larger percentage of it, but that's something that we really take pride in at Trovio is that everyone that we work with creates this content for their own purpose. They're really not creating the content exclusively for Trovio. They're a video production company that are making movies and TV shows for their own core business. And that really spans across our entire network. It's very rare that somebody is creating content specifically for Trovio.
Marty Pesis:The vast majority of our of our licensers and the content owners have their own core purpose that they're building content for. But, again, yeah, I would go back to the multi camera video. I think that's the probably the most important and most actionable thing that somebody can do that is already creating a lot of video content today to increase the value of the data they're producing.
Rob Kelly:And you've got, at least last time I talked to someone who feeds some content data to Troveo, I think they called themselves a referral partner. Am I getting the right kind of phrase?
Marty Pesis:Yeah. We have a big affiliate network at Troveo, and that's really, I think, early on how a lot of our momentum was picked up. And what actually happened is the first maybe 20 or 30 people that we signed when we first got started, they sent us data, we helped them monetize it, and then they came back to us and said, my entire network is people like me. And so we created a program that unlocked our network and really helped us scale immensely by servicing them. We've got over 500 partners now that are active, and they're helping us source and procure data in many different categories.
Marty Pesis:But, yes, that's been a huge factor in our growth on the supply side.
Rob Kelly:Can you just give a rough pricing range of what people will pay for video for AI training?
Marty Pesis:We've seen data go for as little as $10 an hour, and we've done deals and data in the thousands of dollars per hour as well. It totally depends on the metadata it's paired with, the uniqueness and scarcity of the data. If you can provide context with the data itself, that's something that increases the value significantly.
Rob Kelly:What kind of video could make a thousand dollars per hour? Just give me one example.
Marty Pesis:It's mostly about what it's paired with. So let's say, like, multi camera video around the same subject, super high quality video, super high quality audio that has multiple audio channels can increase the value pretty significantly. I would say, in addition to that, the way that you actually prepare the data can increase the value pretty significantly. So how you clip it, how you prepare it, how you annotate it. And then the final piece really is this context layer.
Marty Pesis:I think that's how you really make the value of your data exponentially greater. If you can explain the why and the reasoning and provide that level of context, that's where we're seeing the data value skyrocket, and that's what the market is asking for primarily these days.
Rob Kelly:What's the normal revenue share for again, a range is fine, but just for a content data owner, how much do they get to share when you strike a deal with an AI company?
Marty Pesis:The rough range is somewhere between $50.50 and $75.25.
Rob Kelly:On the high end 75 going to the content data owner?
Marty Pesis:Yes. Correct. Yep.
Rob Kelly:Okay. So fair to say if you've done the 20,000,000 in payouts since your founding that you're doing something in the order of 40,000,000 in sales since your founding?
Marty Pesis:We've done across our entire company, across data licensing, data prep, and, you know, we do other services to AI companies too. But, yes, we've been north of that figure. The 20,000,000 in payouts to licensers is kind of the number that we share publicly.
Rob Kelly:Okay. How much of your revenue is coming from the pre training versus the post training?
Marty Pesis:I would say overall revenue is pretty even, but what we're seeing in the market now is that there's a lot more post training deals just because there are a lot more people doing post training and fine tuning. It's basically this larger trend in the space going from quantity to quality to specificity. If you think about that being like the timeline of model training in any category, right now we're entering specificity, which is where all this precise granular data starts to matter a lot more. So revenue wise, I would say it's pretty equal. But in terms of like number of deals, we're seeing a lot more in the fine tuning and post training space at the moment.
Rob Kelly:Is there something you could do quality of data wise that is very hard for another content data owner to do?
Marty Pesis:I think the context layer is where things really separate and really stand out. Because we work directly with the creators of the data, we're actually able to capture a lot of the context layer as well. And so the more context that we receive with the data that we are collecting, the more valuable it is. And I think that's one thing that has really helped Trovio stand out.
Rob Kelly:Walk me through the process of AI data content licensing deal. Who calls who in a typical deal?
Marty Pesis:Usually, it comes from the research team. They're working on a very specific problem. Right? So
Rob Kelly:This is the research team at an AI company.
Marty Pesis:Yep. Exactly. You know, they have a very specific need, and then usually, they're working with their team, oftentimes a product manager on their side, to help define what data is needed to help solve that problem. Once that data is defined and very clearly outlined, they'll bring that to us and ask for a sample of the dataset. Right?
Marty Pesis:So they'll come to us, let's say, and say, hey, we need a hundred thousand hours of baseball content. We'll go back to our library, put together a sample dataset that we then send to them, get feedback on, and do a few passes to kind of tighten the scope and make sure we're delivering the exact data they need, and then ultimately move into a data licensing deal. So it usually starts with a need on the AI side. They come to us with a brief. We answer that brief, work with them, go back and forth on a few rounds of feedback, and then ultimately put together a data set that is hyper specific to their need.
Marty Pesis:What we've seen over the last year or two is that the needs are getting more and more precise and more and more granular. And that's a reflection of post training and fine tuning. There's definitely a higher bar these days in terms of being able to really understand what data we have and get down to those, like, really granular kind of precise data needs to ultimately deliver what the client is looking for.
Rob Kelly:And when you mentioned the research team at an AI company, these are engineers. Right?
Marty Pesis:Engineers and researchers, they can often play similar roles or different roles, but ultimately, they're the users and ultimate decision makers around what data is valuable. And researchers are constantly testing and experimenting as researchers do to ultimately see what data provides the model uplift they're looking for. And so when we provide the samples, oftentimes they're testing it and evaluating it against their own models to see if it's actually improving and uplifting the model performance on the output side.
Rob Kelly:If the researcher is not always an engineer, else would you call them?
Marty Pesis:They're mostly engineers. They can be data scientists. A lot of the research work is running experiments just to see where model uplift happens. Right? So it's like the engineers are putting the model architecture together, the then researchers are testing all sorts of different data and kinda trying new things to see if they can ultimately get model lift.
Rob Kelly:So in the case of the video generation model for the baseball, you've got it this a large AI model company, a product manager who wants to get some data for video generation for baseball. Where did they start that process? Did they have a customer at the AI company asking them to do this, Or are they in a vertical at the large AI company and they wanna make sure baseball is represented? How does that work?
Marty Pesis:Yeah. I think the latter is more accurate. If you're building a generalized video generation model, you really do need to account for every different type of use case and every prompt that somebody could ultimately put in, right? So, people's mind go to the edge cases, right? So, they want to be able to put in, I want to see a video of a horse playing baseball at Yankee Stadium hitting a home run that goes all the way to space, right?
Marty Pesis:And just their mind really races, right, and goes somewhere crazy. And so, if you're building a generalized video generation model, you need to account for all those crazy scenarios and situations. Then, obviously, as part of the whole development process, you're constantly testing and evaluating the model and identifying where there are deficiencies. So, maybe the model is confusing baseball and cricket, and they need to make sure those are very clearly differentiated. So then you might need a lot more baseball content and a lot more cricket content to make sure that if a user prompts a baseball prompt, they don't get a cricket response.
Marty Pesis:Right? Yes. In the case of more generalized video generation, there's just a lot of coverage required. That's why a lot of data is needed in those situations.
Rob Kelly:So at the Googles of the world or OpenAI's Anthropix, is there likely a product manager who is covering the vertical of sports, let's say?
Marty Pesis:Yes. And I think it also depends on where they are in the model development. Right? So the training stage is where it's more important to have massive volumes of data and coverage across really every different topic. Then as you get later and later into the stages of development, you're doing more post training and fine tuning.
Marty Pesis:And that's where you're identifying specific deficiencies and plugging holes. But think about it as, yeah, pre training is very broad, very large scale data sets that are very diverse. Then you start to move into the post training world where fine tuning becomes more important and much more precise, granular, hyper specific data becomes more and more important as you're trying to solve very specific individual problems.
Rob Kelly:You also made me think about new AI devices. So things like smart glasses from Meta, and then OpenAI is working on this. They bought Joni Ives company famous for designing Apple's products and are rumored to be coming out with an AI device. Are these for sure 100% going to be data collection devices too, whatever comes out?
Marty Pesis:Yes. Absolutely. It's a key part of it. If you look at the terms of service across a lot of platforms these days, they've updated their terms of service to make sure that they can collect the data and use it for AI training and AI tuning purposes. So I think we saw it across the platforms.
Marty Pesis:We saw we saw this happen in a wave. Right? It was all the big social platforms kind of one after another. The data is just so valuable these days that, yeah, all these companies have to be thinking about it.
Rob Kelly:And I think you're probably referring to I mean, one of them's Google with YouTube and changing their terms and conditions to make it okay for Google to use the videos uploaded to YouTube for AI training. Right? So arguably one of the most valuable collections of data in the world?
Marty Pesis:Yeah. Absolutely. YouTube is certainly the largest video platform out there, I think, by orders of magnitude. And then, yes, the ability to train videos on YouTube has obviously been pretty hotly contested subject for years, and I think these legal cases will take years to sort out and create clarity on. But the YouTube data specifically is uniquely valuable just given its video at a scale that you really can't find anywhere else on the Internet.
Rob Kelly:You're trying to get, you know, the Vimeos and the Daily Motions of the world and get their video data into your library?
Marty Pesis:A lot of what we focus on actually is non public video. So over 65% of our library is non public, meaning it's in archives, it's behind paywalls. So, you know, a simple example of this is if you're filming a TV show that ends up being twenty two minutes long on on TV, you're probably capturing between fifty and two hundred times more content than the final episode.
Rob Kelly:So you want the b roll footage?
Marty Pesis:Yeah. Absolutely. We'll take the b roll footage, and then this applies to every single data type as well.
Rob Kelly:Mister Beast, so my 10 year old son's a big fan. I'd be remiss if I didn't ask you to tell me the story of how you sold the company to Mister Beast.
Marty Pesis:We were building a hiring platform called Vouch. The whole idea was we were helping creators, big YouTubers, build teams, and no one could do it at MrBeast's scale. And so when he would post a job, they would get hundreds and hundreds of thousands of people to apply. But ultimately, I decided I wanted to pivot into this business and wanted to focus on Trovio. And so, MrBeast and Beast Industries was a customer at the time.
Marty Pesis:They had about 40 open roles and they were launching a lot of other businesses as we've seen them do over the last few years. And it just made sense for them to have a hiring platform internally. And so we were able to sell them our hiring platform, which helped them just unlock a lot more efficient recruiting and ability to find and retain great talent.
Rob Kelly:What's one thing about Jimmy Donaldson, aka mister beast, that most people wouldn't know that you learned?
Marty Pesis:Unbelievably hard worker and very, very intentional about everything he's doing. What he has built is not by accident. He is incredibly meticulous about everything, so detail oriented, and they've really got the production business, especially for YouTube, down to a real science.
Rob Kelly:Is there an example of just how hardworking and meticulous he is personally, Jimmy?
Marty Pesis:I would say when I went through and kinda learned about their entire workflow, it just it blew me away that he's so involved in every step of the process, and every step of the process is so in-depth and complex. So and I don't think there's any video that goes out that he didn't play a huge part in every single step along the way. And just his presence and his involvement really blew me away.
Rob Kelly:So Alexis, the cofounder of Reddit's invested in Troveo.
Marty Pesis:Yeah. Alexis has always just been a huge supporter of data being the differentiator and around figuring out a way to make these deals work out for both parties.
Rob Kelly:Can you just tell me something you've learned from Alexis?
Marty Pesis:I think Alexis' message to me is always just to keep going and have a bias for action. Every time I bring an idea to him, he asks me some questions to validate the idea, but really his message is just very supportive always. He's like, if you have a hunch, you know, that's why I bet on you. And I'd say go for it. I think the very first time I brought this idea to him, he said, I love it.
Marty Pesis:You better move fast. I think we should wrap this call up and get started right away. So, yeah, he's always just about taking action and moving quickly.
Rob Kelly:Do you represent any data from Reddit?
Marty Pesis:No. We do not represent any data from Reddit. We represent a lot of data from other portfolio companies at seven seven six, Alexis' fund. But no, we do not represent Reddit data.
Rob Kelly:I can tell you're an AI optimist. I mean, you're bullish on it. But tell me what could go wrong in an AI world that's got you worried.
Marty Pesis:I think it's just the unknown. We obviously have never experienced life with tools this powerful, and we kinda just you don't know until you know. Right? So I never imagined having some of the capabilities that I do today. And so what does that look like as AI advances is the big question.
Marty Pesis:The other the other kind of big one is, of course, kind of consolidation of power. I think it's really important that there's a lot of competition and that there's many companies that are controlling this incredibly powerful technology. But it's kind of classic fear of the unknown, I guess, I would say.
Rob Kelly:What are you telling kids and younger folks in your life? I think you've got children. Right? Yeah. What are you telling them to help them prepare for this world of AI?
Marty Pesis:I think in this world of AI, having a real domain expertise and real domain knowledge is super important. The tools can automate a lot of the steps. But to really understand things at a granular level and to really understand the nuances of what make things work, I think that's ultimately what's becoming most powerful right now. So, overall, I would say on the AI side, stay cutting edge, stay up to date, do the best you can to learn and understand the latest with these tools. But outside of that, that's where I'm seeing people shine the most is when they really have a deep understanding around one subject and can amplify that with AI tools today.
Rob Kelly:How does AI change your thinking about your own kids going to traditional school and college? Does it change it at all?
Marty Pesis:Yeah. I'm certainly questioning it. I I definitely question it and and wonder. Right now, we we I use AI with my kids a lot. But I think the two things as my kids get older that I think about is, number one, I wanna make sure that they remain extremely curious and just want to learn and dive into all these different topics.
Marty Pesis:And then resiliency is the other one that's super important to me. You know, it reminds me when I was growing up of learning multiplication tables. Obviously, like every other kid, I said to our teacher, why do I need to learn this? I have a calculator. And that has just been blown extended to, like, every single topic.
Marty Pesis:And so I think those two things are what really stand out to me.
Rob Kelly:I've got a 10 year old. How old are your kids?
Marty Pesis:My kids are younger. They're two and four.
Rob Kelly:I'm just curious. Selfish question. If you had to put a percentage on it, what are the odds that they're going to college? Whereas in the old days, probably almost a 100%. Right?
Marty Pesis:Yeah. Incredibly low in my eyes. I think college is not on my radar at all for my kids, candidly, right now. There's a huge social component to it, of course. Right?
Marty Pesis:I met some of my closest friends now at college going living away from home and kind of figuring it all out on your own is hugely important, and we're, as a world, gonna need to figure out how we check that box for kids.
Rob Kelly:But what would replace the rest of college, the non social part? In your view, what would our kids be doing, let's say, if they don't go to, quote, college after high school?
Marty Pesis:I think beginning to to work and be productive in industries. It just feels like a lot of these skills and things I learned in school, like, lot of it was more around the discipline of learning and understanding kind of how to learn, not necessarily applying exactly what I learned. I was an econ major, and I don't think I'm really, like, applying average cost curves to, like, my day to day, but it did teach me how to learn and how to study and how to have discipline and take the test seriously and things like that. When I think about college today, I think about being part of a society and a community and also learning those skills of how to learn, how to study, and that discipline. I got a lot of that in sports.
Marty Pesis:It's kind of similar. Right? Like, I I don't play a lot of hockey, but I learned a lot of things playing hockey that I apply today in terms of discipline and things like that. So, yeah, it's a tough question, but college is really not on my radar for my kids at all right now. It's more about how do we check those boxes in other ways.
Marty Pesis:I just I think the world has changed enough that we need to make those adjustments right now. How you thinking about
Rob Kelly:it? Yeah. It's interesting. I hadn't even thought about it until you just mentioned it in this way that maybe our kids are going to be working, again, quotes, like, a job earlier than our generation and our parents' generation because of AI. Like, they'll be diving in at an earlier point, like, right after high school or before high school.
Marty Pesis:It's just really intriguing that they're that the the path to launching a business and trying things and starting things has just become so much easier, and there's so much less friction around starting a business and spinning up products and building products, things like this. It's just yeah. There's you know, I think the age of entrepreneurship will just continue to be unlocked. And, yeah, I totally see young kids and high schoolers, whatever it may be, being valuable members of society. From a commercial standpoint, it seems like a no brainer.
Rob Kelly:Will you create an AI avatar of yourself for those around you, family, friends, while you're alive or when you've passed?
Marty Pesis:It already starts with building an AI brain inside of my company. I have a chief of staff that is fully agentic. It listens to all of my meetings. It captures content across our whole company and proactively recommends things. It gives me ideas.
Marty Pesis:Probably my number one brainstorming partner at the company is my AI brain and the agent that we've built to do that. And so I think it's already taking steps to that.
Rob Kelly:Are they one in the same? Yeah. Is your chief of staff and your this is one in the same. Okay.
Marty Pesis:Yeah. We yeah. We've I basically, yeah, spun up an agent that that can help do a lot of the work that a chief of staff would do. But I think that's already a step in that direction. Right?
Rob Kelly:And how about an AI avatar for your family and friends for once you're gone?
Marty Pesis:For, like, a longevity and the idea for me to live on post life, I guess, post humanistic? I really haven't thought much about that. I think I'm pretty focused on staying in the present. I have two young kids, two and four, and I feel like they're growing up quicker than I'm comfortable with. I wanna stay on the cutting edge and obviously have a deep understanding of what's going into AI right now, but I haven't thought a lot about what happens to my AI avatar after I die, candidly.
Rob Kelly:Do you wish you had one, an avatar of anyone in your family, like parents or grandparents?
Marty Pesis:Not really. That's not something that I feel like I need in my life right now.
Rob Kelly:More of a in the moment guy.
Marty Pesis:Yeah. It just goes back to fear of the unknown to an extent. Right? Like, if that becomes common and normal, then it would feel weird to not have it. But no, I I'm I think there's a lot of work to just stay up to speed and in the moment right now.
Marty Pesis:So I think these things are gonna come quicker than we think. But I I do think it's important to stay present, especially with really young kids. I don't wanna I don't wanna give my head too far away from these moments.
Rob Kelly:Hey. Thanks, Marty, for all the time today.
Marty Pesis:Appreciate it. Great conversation. We'll we'll be in touch.
Rob Kelly:Well, this is Media and the Machine. A few things about you and me. If you wanna hear about the next new episode, make sure you hit follow on the show in your podcast app. If you wanna go a little deeper, head to mediaandthemachine.com and subscribe. When you share your email with me, you can see handcrafted transcripts, read the essays in my newsletter, and be the first to hear about who the guest is on the next show.
Rob Kelly:You can also email me directly from there. Maybe you wanna recommend a guest. I'll give you a shout out if you do. I love paying it forward. From time to time, I also open up office hours and host small meetups for subscribers, just to meet, talk, and build things together.
Rob Kelly:And if you're creating something of your own or thinking about it, I'd love to help. Maybe you've got a podcast in you. Finally, I don't have a marketing budget for this show. So if it's finding you and others, it's because someone like you passed it along. I'm genuinely grateful.
Rob Kelly:If you have a moment, an honest rating helps me make this better for you. You can just go to the show page and click one of the stars. And I'd rather you give me a low rating than no rating at all. I mean it. It pushes me to get better.
Rob Kelly:Thanks again, and see you next time.