The Exchanges

Q Are, are a lot of these companies building tools like that, or is all the focus just on like, let's one shot the next Oscar film?

A Yeah, it's a great question. There's a decent amount of vertical focus. I think also, like, it's very hard for this, for startups, like, if you're not a Google to, or an OpenAI or whoever to play in the game of, like, let's train the largest one-shot best video model. Um, so a lot of them are focusing on verticals, like, um, Luma, which is a company we've invested in, did this really cool tool where you can upload, like, a, a nine by 16, like, iPhone style video, and you can just say extend, and it just Basically like out paints around the existing video and makes it look like, so you can just change the dimensions of your video really fast.

AI assessment note: “There's a decent amount of vertical focus. I think also, like, it's very hard”

Q to corporations that want to vend in talking avatars. For example, are you seeing a lot of that? Um, or like, I guess how, how, how, uh, rigorous or, uh, stringent are founders about like, Hey, we are trying to build a consumer app or we're just building a cool technology and it might land as a consumer product, but it also might land as a B to B play.

A Um, most people are not at all rigorous and often don't even know at the beginning. So what we actually saw with the first generation of AI video was it was only researchers making these like magical models and they had no idea what the use cases were going to be. They all just put like a text prompt box in front of the model and then you got an output and like a big company and an individual were using the exact same interface. Now I think we're starting to see more at like what we call the app layer, which is essentially like How do you productize this? How do you create workflow? And, and therefore, how do you go into two specific verticals? Um, maybe an example of that is all of the like standalone video ad, um, creation products. So things like Creatify or captions where, or Hey Jen has a product for this too, where, um, you can literally just like paste in a link to your Amazon or Shopify store. It will pull all of the info about your product, your logo, your brand, and it will generate a talking head avatar. Like holding your product and describing it. And that's something that like a VO three or cling, the general video model companies won't do today.

AI assessment note: “most people are not at all rigorous and often don't even know at the beginning.”

Q think we're going to see a lot more of this. Uh, and we, we've seen some of this to date, but do you expect that to be kind of like a new, a new, Like, how, how bullish are you on, on sort of, like, entirely AI creators getting, um, getting real adoption, following, turn, followings turning into real businesses? I'm sure you've, you've followed a bunch of them already.

A Yeah. I'm, I'm personally super excited about it because it kind of separates, um, the content from the character. Like now, you know, before AI, if you were on Instagram, like you were both the character and the person coming out with the content. And so you had to like look in a way and present yourself in a way and talk in a way that was like interesting to the Instagram algorithm. And now it's like anyone with a good idea can create a compelling character. Um, and so I think Some of those are human characters. I've already seen way too many examples in my Reels feed of OnlyFans models who promote themselves with AI avatars of themselves now, which works shockingly well. There's some photorealistic human influencers, but honestly, some of the more interesting ones are things that could never be influencers before AI. So there's one called, like, Raccoon Stole My iPhone, and it's a, it's an AI raccoon influencer. There's, like, AI capybara influencers. There's, like, mystical creatures, like, all of these things that just come out of people's imagination.

AI assessment note: “I'm, I'm personally super excited about it because it kind of separates”

Q this like behemoth kind of quietly hiding in a discord server. Still, uh, I saw some examples of video. It looks fantastic. Seemed like they hadn't added audio yet, but how, how do they fit into the whole, the whole piece? Because it seemed like early on, they, they developed a really great, um, feedback loop for the data that maybe wasn't happening with some of the other model providers.

A Yeah. So the mid journey model came out this morning, um, really conveniently, like 10 minutes after I put out my market map without the journey video because it's not yet available. Um, I was just playing around with it too. It's, it's really cool. They do image to video. Um, and they, and so they don't do text to video, which is actually sort of easier. They can start with their, the super high quality images that they generate on the platform and then animate those. I think they have like a low motion and a high motion setting. Um, From what I've tested so far, um, it's, it's better as sort of like a low motion scenery environment light interaction type thing. Like you have a photo of a person and you can then animate sort of rain and wind and them walking slowly. And it's not as good at like what I call physics heavy world model type things, like two cars running into each other and exploding. That sort of thing requires, um, Uh, a very, very large and costly, usually like text to video model that is more difficult to train. Um, whereas mid journey, I mean, I, I have no idea how they did it. It's a great model. They could have taken one of the open source image to video models and fine tune it on all their own data.

AI assessment note: “They do image to video. Um, and they, and so they don't do text”

Q And I'm sure there's reposts of, American videos over there. So it's not like even has a unique flavor probably can generalize pretty well, right?

A Totally though. Yeah. Like I was one of the very early users of all of the Chinese video models when you still had to access them on Chinese apps with Chinese phone numbers. And they were definitely very good at things that were more, um, China oriented than the U S models. Um, oh, okay. The other, the other thing that's important to mention on data is, Um, in video in particular, it's not just the volume of data, it's also the quality of data and the quality of data labeling, because essentially you can't just feed a video into a video model and assume it can understand what's going on and, and pull out the relevant info. You have to have really sort of dense labels is what we call them, or super detailed captions about like, this is this style shot, shot from this sort of camera, the camera is coming from this angle, This is the sort of character. This is how the character is interacting with the background, and that quality data is what drives quality in the video models. Um, and China has really benefited there because there are so many more PhDs than there are here, and it's much cheaper for these companies to hire them, um, to do these, these dense, uh, labels for the video data.

AI assessment note: “they were definitely very good at things that were more, um, China oriented”

Q Are you seeing any infrastructure players trying to do, like, do anything on, like, content verification side and, like, trying to create some sort of, um, mechanism to, to prove whether something was, like, authentic, you know, actually shot on an iPhone, right? Right. You know, proving through the metadata and some type of, like, public, um, setting. Is there, is there any pitches, uh, from, from that side?

A Yeah. So, most, Largely, honestly, today that has come in two places. One is the model companies themselves will often watermark the content in some way, like the VO three generations has a little VO three, 11 labs, it's just the audio. They actually have a site where you can upload any audio and it will tell you if it was generated with 11 labs or not. That's cool. Um, which, which is pretty cool. The other, um, place we've seen development there is for like prominent individuals, um, like, you know, celebrities or someone who's There's like value behind their brands and who potentially even might want to monetize it in the age of AI. Like if you're an actor and you suddenly don't have to, you know, film, go fly back to LA when you're filming a movie in Australia to tape like five ads for some cell phone brand, and you can have your AI avatar generated to do it instead. And it looks just as good. Like you might actually want to, you know, have some licensing company that owns your AI licensing rights, whether it's your traditional talent agency or not. Um, who can manage that for you?

AI assessment note: “Largely, honestly, today that has come in two places.”