00:06:50 · Burkay Gür
[00:06:50] S2: There was also, like, a very interesting point sometime last year when Veo 2 came out. I think at the time, like, Kling was probably one of the best video models and there was also, like, Minimax. And the prices were, like, somewhat kind of going lower. I would say, like, they started at, let's say, 40¢ per five second video and then they sort of settled around there, maybe like 30¢ for a five second, like, clip. And then Veo 2 was, very surprisingly, it was like $5 per five second video. So it jumped like, you know, more than 10x. And we were like, Okay, like, who the heck is going to pay all this money, right? Like, this is just another five second video. Turns out, like, there is essentially infinite demand. So people started using that model and then even more expensive models came around and people started using those models. And what was really happening under the hood was we actually saw more and more professional use cases start happening for ads, like real people who are in marketing and ads and actually movie production, they started using these models because the quality went so much higher and there was just more demand that other models weren't able to actually capture. So, it's like the usage is going up and then the capabilities are going up and then everything's getting more expensive over time. So, essentially, yes, like people ended up spending so much more money. And in terms of progress, it's still happening. Like we're nowhere near, you know, with LLMs, I think people talk a lot about hitting a wall or like diminishing returns. I think in this side of the market, we're nowhere near actually hitting diminishing returns. I think there's just so much new stuff happening with CDAS V2 and V2.5. It's just like the quality is going up, prices are going up, and people still want to use all these models.
00:08:56 · Molly O'Shea, Gorkem Yurtseven, Burkay Gür
[00:08:56] S0: For sure. I mean, do you guys remember Sourcery? Yeah. [00:08:59] S1: Feels like it was a lifetime ago. No. [00:09:02] S0: I mean, that was a really cool moment, but clearly, too expensive. [00:09:06] S2: Yeah, it was too expensive. Like, in fact, we saw we also hosted Sourcery on our platform, and, like, we saw huge growth in the, like, last three months before they decided to, like, take it down. We were like, What's happening? Like, there's, like, infinite volume here, essentially. But, yeah, it turns out, you know, the rumor is like there wasn't enough compute for everything else. And it's also expensive to run. So, yeah. But I think it had a great future. Maybe it'll come back sometime. [00:09:35] S0: I mean, it's really interesting because earlier today we were down at AppLovin' and this will probably get released later than that, but I was talking to Adam, the CEO there, and Gio, their CTO, and we were kind of talking about how they've created their algorithms for all the ads and what works best, and the best ads are actually sixty seconds long, and I was like, Okay, well, are you thinking about using a Higgsfield or any of these other companies out there and offering that or acquiring any of these kinds of like generative ad platforms. Cause some of them just really, I mean, most of these cases are ads and like, it's just really not there yet. It's like the best they can do is like five seconds. So it's like not every brand can do In terms of where you guys came from and where it's going, advertising is going to be the biggest use case, you think?
00:10:22 · Gorkem Yurtseven
[00:10:22] S1: I like ads, like as a business, like internet runs on it. And also like the content these models are able to make is perfect for these UGC style ads because the content is usually throw away. If it doesn't work, make a new one, like it works great. But also, like, the quality, like the lighting, nothing has to be perfect. It needs to be almost looks... Needs to look homemade, so it's like easier to do. And, you know, the best ad may be sixty seconds, but you don't have to do a continuous sixty second generation. You can do fast small cuts, which works maybe better in certain cases. And then what you can do is you get a lot of data back, right? I don't think anyone is particularly doing a good job at it, but if because you can post these ads and see how people are interacting with it, you can have like continuously improving or you can do small changes to do what you are posting and just show more of the things that are performing well, and this can all happen automatically, right? So this is I think it works great with video models. No one has done it perfectly, but we would love to be part of this automated ad story. I think it's going to be a great business.
00:11:42 · Molly O'Shea, Gorkem Yurtseven, Burkay Gür
[00:11:42] S0: It was really funny. I was going to mention this in the interview, but then I realized it was probably kind of inappropriate, but example you just mentioned of reinforced learning models on themselves and just kind of optimizing over and over again. There was, and I don't know if you guys see this in your Instagram feed, but I do this because I make clips and I make like stuff for the internet. And they're like, one of the best hacks that I had was showing my armpits. Like, I got so many more views from showing my armpits. [00:12:12] S1: As the hook, right? [00:12:14] S0: Literally. Literally. It's like I got 10x views just from, like, having my arms up and like, and I [00:12:19] S2: was like, what [00:12:20] S0: the hell? Yeah. [00:12:24] S1: Imagine, like, you can have yourself in 100 different positions. Let's go. [00:12:28] S0: Yeah, I mean, that's like a thing. It's like, the thumbnail is always testing. I mean, for like YouTube, for us, it's like, you gotta test thumbnail, you gotta test the title, you gotta test like the copy, all of that kind of stuff because it works the algorithms differently.
00:12:28 · Molly O'Shea, Burkay Gür
[00:12:28] S0: Yeah, I mean, that's like a thing. It's like, the thumbnail is always testing. I mean, for like YouTube, for us, it's like, you gotta test thumbnail, you gotta test the title, you gotta test like the copy, all of that kind of stuff because it works the algorithms differently. [00:12:42] S2: Exactly. I think there's, like, really sophisticated users out there that are doing it, like, at, like, smaller scale, maybe they're, like, a creator, but also, like, large companies is doing it right now. I don't think it's, like, super spread out yet, but, like, it's out there for sure. And, like, we work with a few companies that actually do it, but I think it's gonna get even more available, I would say, because it's not very straightforward, I think, to do the whole reinforcement learning and all technical stuff behind it. It's not fully there yet, but I think in a few years, this is going to become the norm. [00:13:19] S0: So let's talk about the engine behind fall, and then we can kind of get into how your culture is keeping up with the growth because it's crazy. So you were saying you started off with compute, you started off as a compute company, you found your way into this area space. What was that evolution like for you and like how have you completely embraced this new category?
00:13:42 · Burkay Gür
[00:13:42] S2: Right, yeah. So we started in compute but like a pre AI era, let's call it, right? So we started in 2021 and the whole thesis was like, okay, there's, you know, this previous data infrastructure wave led to some amazing companies like Snowflake, Databricks and what do they have in common? They own compute, you know. So, in order to build the next generation of like the data infrastructure company, we also have to focus on compute. That was the initial thesis and we thought the next big thing is going to be Python. So, basically like what Snowflake did and what Databricks did to a certain extent was like enabling SQL to run on, you know, in the cloud. And they own the compute and they let you scale the compute elastically. So, we thought Python is going be the next thing because AI is the next big thing. So, we ended up working with some really awesome teams like trying to bring their data science and machine learning, old terms, like workloads into the cloud. So, that was like our initial starting point. And then, soon after we started, some insane things happened in the market. I think from a technology point of view, like we had a massive breakthrough with ChatGPT and in our space, stable diffusion and DALL E. So, we kind of had a choice at that point. Like, we can either stick to our like, you know, data infrastructure, you know, initial idea or we could go into this like new world of AI. And there was language models and at the time, I think people were very excited about a bunch of open source models. If you remember, like Llama came out shortly after and there was this like huge price race at the bottom and we just thought, okay, we're like five, six people right now. We don't have that much money in the bank. We're a pretty early company and we can't really go and sell tokens at a negative margin. And let's do something more interesting. Trend still holds. Where the trend still holds is let's focus on image models because like there's something interesting happening there as well. There was no sight of video models. Like, no one even talked about it, I think, back then. But we thought that like we could take image models and make them run a lot faster in the cloud on the same infrastructure that we built. And so, did that. So, we had a very technical team. We thought like we could do a really good job running these models faster.