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:Then I 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 Chris Keevil, founder and CEO of Versos AI, a company helping AI labs find the needles in the haystack data they need to train their models. Chris works at the center of a new market. It's where the frontier AI models are paying for video most of us would never think to value. Ten thousand hours of hands tying shoes, ice melting, micro dramas from Asia, and a coffee shop in China.
Rob Kelly:AI wants it all. We talk about why some AI training clips sell for 30¢ a minute, while others go for a dollar per second, and why some AI companies are now doing whole corpus deals to buy entire content libraries in bulk. Chris also explains why the demand for training data will never run dry as AI moves from understanding words to understanding the physical world, motion, gravity, culture, behavior, and eventually even industrial systems, like oil refineries and hydroelectric dams. We also discussed some gold mines of AI data, such as security camera archives, niche media libraries, and even the video sitting on your phone. And finally, Chris shares why he's a cautious optimist on AI, despite believing it could massively reshape jobs, business, and society.
Rob Kelly:Please enjoy my conversation with Chris Kevil. What was the moment you realized, oh, AI companies don't just need video and data, they need it prepared a certain way?
Chris Keevil:About a year ago, I guess it was probably March 2025, one of the studio partners that we were working with on our early version of the platform told me that something strange had happened over Christmas. And I said, well, what happened? And he said, well, I got approached by one of the large AI labs who wanted to license my entire library for data training.
Rob Kelly:Wow. Which AI company was it?
Chris Keevil:Well, here's the funny thing about this business. We're sworn to secrecy on all the deals we do. They are so hypersensitive to disclosing any of that. I'll just say one of the companies that you read about all the time is being, you know, the most valuable companies in the world. One of those.
Rob Kelly:And what kind of studio company was this?
Chris Keevil:It was a video production company. They had 10,000 of video assets in their archives and they were able to license it to the training model.
Rob Kelly:Gotcha. So they get basically a cold call out of the blue of a type of business partnership that they had never done before, an AI company that they recognize that wanted access to their data.
Chris Keevil:That's right. Christmas twenty twenty four, like early twenty twenty five, this whole new and emerging industry was just getting started.
Rob Kelly:I'm just curious how important is that line of business now for that production company?
Chris Keevil:Well, for some of these companies, it's become very important. We partner with Curiosity Stream, who is NASDAQ traded. They're a video streaming company.
Rob Kelly:Had the CEO Clint on the podcast?
Chris Keevil:Oh, yeah. Of course you did. Yeah. Yeah. You had Clint.
Chris Keevil:Well, then Clint would have told you how important it is because he has shared on his earnings calls that he sees this as being as big as his subscription revenue.
Rob Kelly:Yeah. Amazing.
Chris Keevil:Yeah. It
Rob Kelly:is. How many companies were you the founder of before Versus AI?
Chris Keevil:Three companies. Rob, I've been founder of three other companies in, I'll say, Media Tech over the past number of years. I've lost count of the years, but I can keep count of the companies.
Rob Kelly:So you've got the supply chain. Let's call it the rights owners or the content companies. Yep. You got on the other end the demand coming from AI companies, the OpenAI's of the world, Anthropix, and Google, and so forth. And then in the middle, you're enriching these assets.
Rob Kelly:What do you call your industry?
Chris Keevil:We would be data annotation or data enrichment. A lot of these things don't have good labels yet because it's so new.
Rob Kelly:Speaking of Curiosity Stream, you mentioned last time we chatted, they may have the best premium video library for AI training in the world. Why? What makes their library so trainable and valuable?
Chris Keevil:A lot of it is unscripted. So a lot of their library is documentary educational and it's produced at the highest quality. So it's top quality production, multi cameras, well lit, well performed, full production quality. But the essence of the content I think relates to that as well. And I couldn't say for sure, but I would say it's because the world models that we're training are looking to train on the known world.
Chris Keevil:And you know, the known world includes dramatic television and film, but it's more accurately captured in unscripted documentary. So that kind of content is the chronicling of the known world as we would know it through film. The world models can't go out into the world. The world has to come to the models and it comes in the form of video and a good format for that video is unscripted documentary.
Rob Kelly:So factual content unscripted gets an advantage over scripted just because it's real life?
Chris Keevil:I necessarily say it's an advantage exactly. It's just got a very broad use case because there's also a market for training on scripted full episodic drama. We're sourcing and delivering short films, what are called micro dramas from Asia. These are shot on an iPhone, linear format, three minute or four minute episodes of what you would think of as an American soap opera style story. It's almost a TikTok video, which is user generated is somewhere in between that and a daytime soap opera.
Chris Keevil:It's not good quality performance and acting. It's just low cost, high volume. There's another extreme on another end of the spectrum that's in high demand. Part of what's interesting there is the cultural diversity that comes from that kind of content. So they're looking for multi language, multicultural content all over the world from every corner of the earth in order to train the models.
Rob Kelly:And why would an AI model want a microdrama from some other corner of the world? What would be a use case for it?
Chris Keevil:The training purposes are multifaceted, but you know, here would be an example. If your model was training on generative video, so if you're creating a generative video model to create film and it's generating video from a prompt and someone's looking for a scene in their video of a coffee shop in China, it needs to have been trained on what does that mean. What is a coffee shop in China? It's different than a US Starbucks or a Canadian Tim Hortons. I'm Canadian.
Chris Keevil:So there's and there those are those are three different cultural experiences. And the model wants to understand them all.
Rob Kelly:So in that case, the model might want to take the micro drama data or content for someone else who wants to maybe create micro dramas themselves, but also someone could create a full length feature film and they need that coffee shop in Yep. In China kind of realistic footage.
Chris Keevil:Right. And here's another example. The the models want to understand things like character development and a story arc and how that develops through stories. If you think about a TV series, each episode has a beginning, a middle and an end. Each season has a beginning, a middle and an end.
Chris Keevil:And the series in total has its own arc. Training the model to understand those arcs across entire series of TV content, you need to feed all that content in. They wanna understand how characters develop through a show as well. And the only way you can do that is to feed the beast.
Rob Kelly:How will the AI model know it was a good story as opposed to some crappy micro drama?
Chris Keevil:You know, I I I don't know. Now you're into the data science and the weights applied inside these models. But if my data scientist cofounder was on the call, he would try to explain it and I bet to say, I sure still wouldn't understand it. It's I don't know.
Rob Kelly:Got it. I love it. What can you do that a content or data company can't do themselves so easily?
Chris Keevil:Well, there's quite a bit. So in many cases, the library owner has little or no engineering capacity. In the simplest form, it's the movement of this data into a cloud is in and of itself nontrivial. So there's an engineering job to move the data and get it to the cloud. More importantly in our platform though, is the transformation and the data enrichment on those assets.
Chris Keevil:So the trainer doesn't just want a video file. They want a video file with extensive metadata attached. Now the metadata that we attach or enrich the file with is important to the trainer. It's things like describing what's in the file or describing what's in a scene within a video. Our processing would identify characters in a TV series, name them, and then track them through the entire series, say 200 episodes.
Chris Keevil:And then timestamp every time that character acts in a scene, we would create a clip of just that character and we would transcribe their audio, provide it in text, and then we would describe what's happening in the scene. And we've done that through an entire series which could be 10,000 scenes. It's data intensive. So 10,000 scenes across maybe a 100 characters in the series is a lot of individual diarized scenes. So the complexity in doing that is in many ways unique to our platform.
Rob Kelly:Are the AI companies enriching the data themselves? In other words, will they gladly just take millions of hours of video from some content owner and then do this annotation themselves?
Chris Keevil:They have the ability to, but they would prefer that that burden is pushed down into the supply chain. They would rather acquire training ready data sets in many cases, not in all cases. There are examples where they either don't need a lot of data enrichment or they can do it themselves. The challenge with the model builders is their data science teams are very high skilled and very high priced. And if I've got a team of data scientists and I'm in a race to build the next model, I really want my data scientists focused mainly on model building.
Chris Keevil:And so the dataset preparation is you could argue from their perspective, less strategic. It's more of the plumbing work. So we're doing that plumbing work so that they don't have to and they can stay focused on their model building.
Rob Kelly:Makes sense. Can you talk about what's getting trained? In other words, to most people, you've got ChatGPT and then you've got some video models out there using media. But what when the models take in this nicely enriched data and content, is the output mostly just LLMs, or what else are they using it for?
Chris Keevil:Well, vision models, they're building models for autonomous driving. They're building early stage models in human robotics and industrial robotics. So the video source to train a robot, you could think of as being ten thousand hours of a close-up of hands tying shoes, which speaks back to your earlier question, why don't they just want large format video files? Well, inside large format video files are a few seconds of hands tying shoes, but they want five thousand hours of hands tying shoes. You've got to search across thousands and thousands of hours of video to find a thousand hours of hands tying shoes.
Chris Keevil:That's part of this processing and enrichment is finding those needles in the haystack for the very specific scenes that they wanna train on.
Rob Kelly:What are the AI companies looking for exactly?
Chris Keevil:They're looking for, in some cases, long format video assets. And that's where a lot of the market has been over the first, I'll say year of its existence is the shipping of large video files. That technically isn't hard to do. Where it's moving to though and increasingly they're looking for shorter segment assets or scenes described within a larger file and having each of those assets enriched with extensive metadata that is explaining what's in the video. So it's giving structure and context to video, which is by its nature unstructured data.
Rob Kelly:In terms of the use cases for what AI model companies are looking for, can you first just give the common requests or the the table stakes for what AI wants these days? Just a couple examples.
Chris Keevil:Sure. So the base would be a scripted TV program, full episodes with minimal metadata attached. They wanna know just file name, format and length. So that's the most basic use case. You move up from there.
Chris Keevil:It's they want full files, but they want those video files understood. By that, I mean, they want to provide metadata that describes what is in the show. Then from there, they might say, all right, I'm not just looking for a description of the file. I'm looking for you to break up the file into its contextual scenes. And then within those scenes for each scene, I want you to identify the character, identify the actor name, describe the scene, transcribe the scene, and then maybe even clip the scene and create a new derivative file from that.
Chris Keevil:So if you think about a single shot, I'll make up a scene here. Okay. So a single shot is a dragon is flying through the air and landing on a mountain and breathing fire on a castle and the castle burns down. That could be one shot if you thought about a camera motion in a show. But contextually, that's three or four things.
Chris Keevil:That's a dragon flying, that's a dragon landing, that's a dragon breathing fire and burning a castle. Those are three contextual scenes. In some cases, they want it segmented by camera shot, and in some cases, they want it segmented contextually. So it's it's that level of specificity that they're looking for.
Rob Kelly:How does the actual request get submitted? In other words, are the AI model companies just sending an email over to the content owner, the rights holder?
Chris Keevil:Yeah. Email and phone conversations. Yeah. There there's relationships that are built between the large buyers and these large library owners and and aggregators.
Rob Kelly:There's not a system or some sort of database where, you know No. OpenAI could fill out a field saying we need these. You know, we need this.
Chris Keevil:Not yet. No. No. Alright. That'll come.
Chris Keevil:That'll come, but it's it's still very early in this market.
Rob Kelly:And last time you and I chatted, we talked about how I had just heard an anonymous tip that an AI model company this is not through Curiosity Stream or you, just to be clear, but an AI model company needed two hundred thousand hours of video of people doing chores. That was just a fun cocktail barbecue conversation Yeah. For me and my friends and family. Yeah. Give me some of those.
Chris Keevil:Hands tying shoes. A close-up of hands handling fabric was one that was interesting. Yeah. So those are a couple interesting ones recently.
Rob Kelly:And in those cases, what's your guess as to the use of hands tying shoes and hands handling fabric?
Chris Keevil:It could be a robotic use case because they're trying to understand the mechanic within the hand. Eventually we would anticipate and this may have already started, although we haven't seen it, there will be the inclusion of sensor data from the hand. So it won't just be a video of the hand tying the shoe, it'll be the hand with tension sensors showing, well, how tight is the grip to tie that shoe? I mean, the tying of a shoelace is a complicated activity. It probably activates hundreds of muscles and nerve ends.
Chris Keevil:Right? And to train a machine to do that is highly complex.
Rob Kelly:So just playing video of hands tying a shoes is certainly valuable. But you're saying that the data from sensors attached to one's hands that would then show the kind of force being used and the muscles being used, that would take the quality of the data up to another level.
Chris Keevil:That's right. That's right. And and you can anticipate or I would anticipate that they'll want close-up of hands doing lots of things like auto repair or surgery. They'll want a close-up of hands to train the machines to do these highly technical, high motor skill activities. The other thing that where I think this is likely to go is out of the popular known world into the industrial known world.
Chris Keevil:And the industrial known world doesn't have video of it. So there's not a lot of video of the working of a hydraulic on a hydroelectric dam or pumps in an oil refinery. Right? How does the inside of large machines, how do those operate? Well, those videos don't exist, right?
Chris Keevil:We don't film that in film and entertainment. And so we could anticipate that video is gonna be commissioned specifically to train on in the future. We had a request fairly recently for training on the inside of buildings to create from a two d video to create a three d mesh out of those videos.
Rob Kelly:So training meaning just video of inside of buildings with no one in it? Just a video of how it looks?
Chris Keevil:Walk around an apartment or walk around an office. Yep. Look underneath the furniture and around the chairs.
Rob Kelly:And what's the use of that, do you think?
Chris Keevil:Don't know. It could be video game development. Could be scene development in three d for a game development platform.
Rob Kelly:K. Let's get back to more of these odd edge cases. They're just ones that are unusual. Hands tying shoes type of stuff. So last time we chatted, you mentioned AI models needing video of people cooking an omelet, making a bed, doing laundry.
Rob Kelly:Yeah. What's that all about?
Chris Keevil:Yeah. Chores, it would be even less task oriented. Videos of a hand crushing a can or a hand crushing a water bottle to show tensile strength and motor skill activity and you know, how the muscles act. Fine motor skill activity, yeah, like making a bed or washing your face. Rob, it's literally anything you can think of that you perform in the real world.
Chris Keevil:Like, I see a guitar in the background. I haven't had this request before, but it would not surprise me at all if somebody was looking for a thousand hours close-up of hands playing guitar or musicians on stage. It is literally everything that exists needs to be trained on, which is what's exciting about the opportunity because it likely never runs dry.
Rob Kelly:Now how about here's another use case you brought up last time we chatted. Dashcam and self driving data covering edge cases, like diverse weather, night driving, rare road events. Is it just all the self driving cars?
Chris Keevil:Yeah. Well, it would be a lot of autonomous vehicles. You know, it wouldn't just be self driving cars. Right? So all self driving vehicles.
Chris Keevil:Think of it as autonomous activity. So the the car is effectively a robot, but what other machines are gonna do the work of people then the the models need to be trained on that activity. So viewing vehicles in wet weather conditions on the road from a traffic camera is important data. Closed circuit TV and drone video footage is important because you need to understand it from various vantage points. Depth and dimension is really important.
Chris Keevil:So depth and dimension of a car driving down the road different than the depth and dimension of a person walking through a room or a cat walking across a table. Right? All these things need to be understood both in two d and increasingly in in three dimensions.
Rob Kelly:So if I've got you right then, it's interesting. Even the case of needing self driving data with edge cases and diverse weather and night driving actually is not just for a self driving or autonomous car. That training could then be used for a robot taking out the garbage at nighttime in the rain. Am I getting it right in that case?
Chris Keevil:Yeah. You're right. You're right. And so what does it look like when a person slips and falls? That needs to be understood to train the robot not to slip and fall as he takes the garbage out to the corner.
Rob Kelly:Amazing. What about first person wearable camera footage of folks walking through public spaces? That was one you mentioned to me.
Chris Keevil:Right. So egocentric cameras. So think Google Glass or a GoPro on your head.
Rob Kelly:And this is the term that folks use sort of in the industry now is egocentric data. It basically means wearable video camera is kind of the easiest way to describe it.
Chris Keevil:Yeah. But it wouldn't have to be wearable. It could also be me walking through the space with my phone. Right? That's also first person or or ego centric.
Chris Keevil:And so we wanna capture that and then convert that two d flat file into a three d mesh so that you can then look on the other side of the table.
Rob Kelly:How about, you mentioned last time, footage of objects on scales?
Chris Keevil:There's there was an ask for video of object on scales and verify the weight. And, for the life of me, I can't imagine what they were looking for that information for except to say they're trying to understand the physics of the universe. Right? Like, what do things weigh? How does weight feel to the visual eye?
Chris Keevil:How does weight get communicated through video?
Rob Kelly:And is that related also to you mentioned AI needing videos of melting ice.
Chris Keevil:Yeah. They would refer to that as transforming objects. So ice melting or water freezing or water boiling and turning to steam.
Rob Kelly:One thing that surprised me about this is how specific it is. So if you're on the the research team, the engineers over at the AI model companies
Chris Keevil:Mhmm.
Rob Kelly:How are they even getting to that granular part yet? Are they getting a request demand for it, or are they breaking down how to understand the world? The more of these use specific use cases, the better. Is is that how it's working?
Chris Keevil:They need to understand everything. So they'll need to understand a boat cutting through the water. Then they'll want a diversity in hull shape to understand how water moves off the hull differently. You said it probably best, Rob. They need to understand the known world.
Chris Keevil:And so all the things that we take for granted that are happening all the time around us, I'm just looking out the window and there's wind rippling on water. They'll need to understand that. So it all needs to be trained on. And I think I said once before, you can't take the model out into the world to do it. You've got to bring that data in the form of video or some rich media asset could be a three d object.
Chris Keevil:Like I referred to a mesh. You've got to bring that as the best representative of the world in the model to train on in volumes. Right? They don't need one hour of wind effect on water. They need thousands of hours of wind effect on water to train on.
Rob Kelly:Do they always give the volume when they make a request?
Chris Keevil:Yep. They give volume because just like everyone else, they have budget to train and the video asset marketplace for training is sold by the minute. So they'll say, I need x thousand clips or x thousand hours or minutes of video under the following specification.
Rob Kelly:Now in the case of someone like a Curiosity Stream or these micro drama companies, are the models now sitting on all of the content and then coming to you for help finding needles in the haystack, or are they sitting on none of it and just coming to you for the needles in the haystack, or is it somewhere in between?
Chris Keevil:They're not generally sitting on anything. They have specific requests and they go out to the marketplace, to the supply chain with a specification of what they're looking for. So they're looking for ten thousand hours of team sport videos, but they want that team sport video with annotation on play by play. They could be looking for scenes of falling objects to understand gravity. So they're looking to understand the physics of the world.
Chris Keevil:So all those very specific sources are usually in the gathering mode, Especially when they're fine tuning models, they're looking for more specific content types in specific formats, processed and annotated in specific ways.
Rob Kelly:What's the name of the job or position at the AI model companies that you most often work with? What type of team are they on?
Chris Keevil:Yeah. There's training teams. Each of these large hyperscalers have multiple training teams. Those are the engineers. And the engineers go to sometimes called a partner manager, which is really a purchasing person, someone who is sourcing training data.
Rob Kelly:Anyone else? Is that in the partnership manager then comes to a CuriosityStream or a Versos?
Chris Keevil:Yep. Yeah. Exactly. They would go to a CuriosityStream. We don't provide the data to the training models.
Chris Keevil:We provide it on behalf of the library owner like CuriosityStream. So they would they would go to a CuriosityStream sales representative and and look for the content. And then on the Versos platform is where that content is sitting and then processed and managed on behalf of Curiosity Stream.
Rob Kelly:In that case, you called him a sales representative from the AI company. Same as a partnership manager?
Chris Keevil:No. The the salesperson is the library owner. The library owner has a salesperson. Yeah. Yeah.
Chris Keevil:They're the seller.
Rob Kelly:Gotcha. The partnership manager would go to the sales rep at the content owner.
Chris Keevil:Correct.
Rob Kelly:Okay. Gotcha.
Chris Keevil:Yeah.
Rob Kelly:And is there anyone else any other departments involved? In other words, what if an AI model company is creating a new product? That's very maybe a vertical. You know what I mean? Yeah.
Rob Kelly:Is that change how the flow works of who calls who?
Chris Keevil:No. It's it all comes down to the training team. It's the data scientists inside these companies that are the are the users of the data and the builders of the models.
Rob Kelly:Okay. And the training teams are usually engineers Yeah. And data scientists. Are those one and the same, interchangeable, or different?
Chris Keevil:Yeah. Probably interchangeable. I mean, I think you can think of them as data scientists more than, you know, what is a conventional notion of an engineer. But, yeah, they're data scientists.
Rob Kelly:Okay. So in that case, the the AI model company knows it has a good partnership with someone like your OsteStream. They're calling to see if they got this new use case, and then Curiosity Stream would come to you to help find some needles in the haystack if it's
Chris Keevil:Exactly.
Rob Kelly:Super hard to find. Do they come to you for most requests?
Chris Keevil:Yep.
Rob Kelly:Yeah. They just outsource that part because as you mentioned, it's super hard.
Chris Keevil:Yeah. Wouldn't even say it's outsourcing. We kinda operate with our library owner partners as a you know, it's one team. We are the data processing platform managers on their behalf. They don't have this technology themselves in house.
Chris Keevil:That's what we've built in a proprietary way.
Rob Kelly:Can a person at Curiosity Stream use Versos, like, the platform themselves and not even bother your team?
Chris Keevil:They could. Yep. For a simpler job. Yeah. A lot of these jobs these days are new for the first time.
Chris Keevil:It seems like
Rob Kelly:Yeah.
Chris Keevil:Each job that comes in has a flavor of its own.
Rob Kelly:So in the future as as sort of the industry starts to mature, it might be that you guys are hardly ever talking except for odd edge cases.
Chris Keevil:Well, we still like people but Yeah. Yeah. The the idea is the idea is to automate and remove as much of the friction in the supply chain as possible. That's our goal is to make it easier end to end to source process and feed a dataset into a model.
Rob Kelly:How many AI companies have you met with so far directly? Like, the big AI model companies.
Chris Keevil:Half a dozen.
Rob Kelly:And what are the meetings like?
Chris Keevil:Engineering meetings. These are discussions around the specification of a dataset build.
Rob Kelly:So this is when a Curiosity Stream or a content owner company maybe can't answer as technical a question about how to get a needle in a haystack kind of data. Would that be why they meet with you and and not them?
Chris Keevil:Yeah. I mean, we would meet with them together, but it would be around the engineering of a dataset through our processing pipeline to make sure the data structure was accurate, to make sure we're capturing the right amount of data, to make sure that the machine annotation that we do is measured for accuracy. Yeah. So those conversations are it's all around the engineering specifications.
Rob Kelly:These are usually the engineers from the AI companies, not the partnership managers you mentioned earlier?
Chris Keevil:It would be partnership managers that would sometimes bring engineers or data scientists from the model building teams together.
Rob Kelly:What's the just rough price range these days?
Chris Keevil:Oh, it's a wild wide range. So from 30¢ a minute to $50 a clip of a thirty second video. So wide range.
Rob Kelly:What's the highest price data? Both the amount, but also the type of data needed.
Chris Keevil:The drivers of the value are how specific is the ask, how specific is the metadata that needs to be attached to that ask, the quality of the production in the source data, the video quality production, and the scarcity of it.
Rob Kelly:I'm just curious. Why not if you're AI, why not just go to Curiosity Stream and say, we want your two point five million hours of video. Just send it all over and maybe OpenAI or the models work with you on chunking it down. But instead of doing all these little one offs, why not just try to grab it all and pay for it kinda one time and then just have it? You brought up scarcity, and then that's it.
Rob Kelly:They own it.
Chris Keevil:In some cases, they have. In some cases, they've done what are known as whole corpus deals. Right? So I want I wanna license your entire library of two million hours of content. Just send it into my s three bucket and I'll take it from there.
Chris Keevil:Those deals have been done, but it puts a lot of burden on the model trainer to then sift through and process all of that video data. And their talented data scientists are busy building models and they would be better served by getting datasets that are already prepared. So that's one reason. The other reason is it all doesn't exist in any one library. So they have to go to multiple sources or they tell the library owners like Curiosity Stream, we need five thousand hours of talking head video of broadcast news.
Chris Keevil:Well, Curiosity Stream may or may not have that. They go out to their sources and then they would acquire and aggregate that content.
Rob Kelly:Can you give an example of a whole corpus deal?
Chris Keevil:I know of whole corpus deals, but I wouldn't be comfortable sharing it.
Rob Kelly:Are they just exclusives? I mean, is that another way to say exclusive deal?
Chris Keevil:No. They're not even exclusive. Right? It's not as if they couldn't sell it again. They're not looking for exclusivity.
Chris Keevil:It's the point you made earlier, Rob. Sometimes they just want volume. They need a bulk load of video to start doing some foundational training.
Rob Kelly:So in the same way that they reached out to that production company that you mentioned, the studio early on, and just said, we want your whole thing?
Chris Keevil:Yeah. We want everything you've got.
Rob Kelly:So that's a example of a whole corpus deal.
Chris Keevil:That's right. That's right.
Rob Kelly:And are the AI companies still going direct out there doing these deals or do they now go through brokers you mentioned earlier, someone in between?
Chris Keevil:Increasingly, they're going through brokers. They'll go to a large library owner directly. But we would think of a large library owner as somebody with a million hours or one hundred thousand hours. But a lot of volume sits in smaller collections of ten thousand hours or less. I'm dealing with a library owner right now that has six hundred hours of high quality MMA fight scenes.
Chris Keevil:And they have people coming to them directly looking to find that particular source of content. And that's not a large library, but it's High quality. High quality. It's premium.
Rob Kelly:Yeah. I mean, specific, I guess, is probably the better. Yeah.
Chris Keevil:And it's high quality shot production.
Rob Kelly:What are examples of content or data industries that are sitting on a gold mine of data, but they don't know it?
Chris Keevil:Well, there would be, in some cases, security cam data is a training gold mine, but a privacy nightmare.
Rob Kelly:And we know it needs to improve because there's just a Wall Street Journal article on how
Chris Keevil:Yeah.
Rob Kelly:You know, Ring cameras and others are giving warnings that you've got a bear in your backyard and it's like a golden retriever.
Chris Keevil:Right. Right. That's right. You don't want your personal drone security firing device to go and get that poor golden retriever when it is mistaken for a bear.
Rob Kelly:Okay. So security cam data that if you could get past the privacy?
Chris Keevil:Yep. Or if you could get past the transport. When AI meets security cam data, you've also got this challenge of getting at the data to process it. Because data is hard to move through public internet, right? It's just a big heavy dataset.
Chris Keevil:So if you were thinking about a use case for AI intelligence applied to video in a security environment, it would be stadium cameras scanning the crowd, looking for signals of say violence, right? Some guy's reaching into his pocket and he pulls out a gun or he pulls out a knife. Processing that in real time is a challenge because the processing has to happen, what they say on the edge. The processing has to happen in the stadium environment. You can't do that through central control because you can't move the video fast enough to be able to capture, well, that's a guy pulling a gun out of his pocket in order to get there in time.
Chris Keevil:So you asked about untapped sources of video for training. Independent studios all over the world have archived video footage that sits on tapes and drives in storage containers. Getting at that and moving it is difficult, but there are people that are helping folks move through that. I would say in time, not in the too distant future, the greatest source of video content is going to be user generated content. And it's going to come from the libraries that we all carry around in our pockets.
Chris Keevil:So they aren't huge libraries, but they're very, very specific. They're very culturally diverse. They're very geographically diverse. And people's phones and the video libraries that they've stored in the cloud are huge volume of video assets to train on there. And I think that's where it goes next.
Rob Kelly:How will folks get that data if they want to opt into that from their phones to AI? How would that work?
Chris Keevil:It could be as simple as helping an individual monetize their library by micro payments into their bank account. Every time they take a video and release it to the training cloud, if such a thing were to exist, the training cloud is the place that captures all of these videos and makes micro payments to the individuals in order to capture the known world.
Rob Kelly:So in the future, you might be sitting there with your own smartphone. You take a video and perhaps something pops up saying, would you like to train AI? You can make, you know, 12¢ by uploading this video to the cloud.
Chris Keevil:Exactly. Or or maybe it says you you kinda gave it after the fact, but what if what if this notional training cloud that we just invented sends a message out to opted in subscribers that says, I'm looking for videos of yellow butterflies. I need 10,000. Go get it. And the training cloud sends the shooters out to capture the content in order to train the model.
Chris Keevil:And that could happen, you can imagine, at scale with large numbers of users in near real time.
Rob Kelly:And just so folks don't think it's a completely wild idea, already GoPro allows the users to opt in to allowing any footage that they record to be uploaded and they share in a revenue share that is coming from training AI.
Chris Keevil:That's right. Yeah. GoPro is doing a version of it today already.
Rob Kelly:That's right. Are there data factories in the world set up for all this training data? Meaning, folks recording hands tying shoes with sensors and having wearables on and egocentric data. Are there giant factories out there already doing all this?
Chris Keevil:Not yet, but there will be.
Rob Kelly:I saw one video on YouTube of, like, a factory in India, I think, where they all had, like, a GoPro on just recording themselves. You know what I mean? Like, in a factory?
Chris Keevil:Yeah.
Rob Kelly:It was unclear what it was for. But
Chris Keevil:That's right. Yeah.
Rob Kelly:But these are gonna happen. It's inevitable.
Chris Keevil:Yeah. These are happening. Yeah.
Rob Kelly:We've talked a lot about LLMs, large language models. Can you talk about world models and what that means to data licensing?
Chris Keevil:You know, the language models like ChatGPT, the original ChatGPT large language model was trained on text and language. And there's a lot of reasoning and logic embedded in language and text, but there's more dense reasoning and logic embedded inside of video. And the next generation models in the race for artificial general intelligence, you know, the notion of near conscious intelligence, requires training on the known world. It requires training on denser data sets. It requires training on video.
Rob Kelly:And what are the types of things we'll be able to do with a world model just as a user that we wouldn't be able to do with an LLM, a ChatGPT?
Chris Keevil:So the world models are what are gonna be driving autonomous vehicles. They're gonna be driving the robotics revolution. They're going to be smarter and understanding three d depth and dimension within space. Really moving from the thinking world to the activity world is maybe the best way to describe it.
Rob Kelly:Got you. More physics oriented.
Chris Keevil:Mhmm.
Rob Kelly:Do you expect the new devices that come out? You've got Meta's smart glasses already out. You've got OpenAI working on something with Apple's old head designer, Joni Ives, both AI devices. Are these gonna be data collection devices?
Chris Keevil:Sure. I mean, everything everything will be a data collection device.
Rob Kelly:Do you envision AI companies acquiring content companies just to have the data in house and have it be real scarce, only they own it?
Chris Keevil:It could happen. It could happen. What I'm told by the data scientists is that there's little competitive advantage to the model builder in owning the training data. The competitive advantage is how they train on that data.
Rob Kelly:Do you consider yourself an AI optimist, pessimist, or some other descriptor on the spectrum?
Chris Keevil:I'm an AI cautious optimist. I'm an optimist in general. So I think I have faith in mankind. I can certainly see the threat of a fully autonomous AI driven world or military system or industrial system. I think that the displacement of jobs, particularly the jobs that people find themselves sitting in front of a computer all day is going to be dramatic.
Chris Keevil:And I think it's gonna be highly disruptive in society and it's not gonna be solvable anytime soon. I think it could in fact drive the need to a new societal economic system where the distribution of the benefits of AI are managed more broadly. What I mean there is I don't think anybody wants a world where, you know, there's 30 infinitely rich people and everyone else is starving and poor, not even the infinitely rich people would want that world.
Rob Kelly:What do you tell someone who's worried about AI replacing their job?
Chris Keevil:There are a lot of new jobs being created as a result of AI. You know, in software development, my company, to build what we have built in a pre AI environment would have taken 200 people and instead we're doing it with 10 people. So does that mean that 190 jobs were replaced by AI or does it mean that 10 jobs were created from AI? And I think it's the latter, not the former. So there will be a lot of wealth creation and a lot of invention as a result of AI.
Chris Keevil:I think that you're not going to see that necessarily happening in very large companies. I think you're gonna see it happening in very small companies.
Rob Kelly:What else are you telling young kids in particular, young folks in your life about changes to make in this new world of AI?
Chris Keevil:My kids are grown and they wouldn't listen to me anyway. But I would probably say on the one hand, it shouldn't make much difference. If your passion is sport or music or social work, you may not be directly impacted by AI. But if your passion is commercial passion or engineering or software development, then you really wanna be an AI first professional in that field. Whether you're on the business side of things or the development side of things, you would need to be kind of an AI native person.
Rob Kelly:If AI does a lot of the work in your life and you had endless time suddenly, what would you do with all your newfound time?
Chris Keevil:I'd find a new vocation. I don't think I would go seeking leisure. I think I would find new work. Because I just think purpose in life comes from your work and your love. It doesn't come from your leisure.
Chris Keevil:Leisure is a don't get me wrong, it's a wonderful side effect and reward for hard work but it doesn't bring for me anyway, and I would argue for most people, it doesn't bring a whole lot of purpose in their life. So we need to work. We're working animals.
Rob Kelly:It sounds to me like you'd go create another company.
Chris Keevil:I would probably create another company.
Rob Kelly:And have you or will you create an AI avatar of yourself for family, friends, or others both while you're here and also when you've passed?
Chris Keevil:No. How come? I don't think people would want that much of me and they certainly wouldn't want that much of me three generations from now. No, I don't think that's the way it's supposed to be. I think the people get as much of me as they want and they'll have as much of me in their memories when I'm gone.
Chris Keevil:I wouldn't want to be an avatar that hung around here forever.
Rob Kelly:Is there anyone you wish you had an avatar of who's no longer around?
Chris Keevil:No. No. I've lost people close to me, and they're still close to me. I know when they're around. I can sense it and feel it just fine.
Chris Keevil:I wouldn't need to see them.
Rob Kelly:Hey. Thanks so much for investing the time today, Chris.
Chris Keevil:Hey. Thanks a lot, Rob. I really appreciate it. I really appreciate the call. I like that last question the best.
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 see can 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: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:And if you have a moment, an honest rating helps me make this better for you. You 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. See you next time.