00:37:20 · Eric Seufert
[00:37:20] S2: Here, you gotta look at ByteDance. What ByteDance is doing is incredible on this front. Right? So they put out this paper a month ago. Mhmm. I did a summary on Twitter and and LinkedIn, but, what they're doing with TikTok shop is these real like, you know, basically, photorealistic three d avatars that are selling stuff. Right? Like, so infomercials. And they've, you know, they've built custom models to build those ads. And those are all ads. A lot like, I mean, not all of it, but, like, if you go on TikTok and you're looking at the TikTok shop stuff, a lot of it is AI generated. Mhmm. And so what they they in this paper that I summarized, like, they invested in, you know, in in, essentially, like, fine tuning this model to make sure that there was no collision with the hands. What they were finding is that, like so when you get in that uncanny valley Mhmm. Situation where people can tell it's AI, then they turn off. Like, there's lot of research that's been done in this. If people know that it's AI, they penalize the ad. But when they don't know it's AI, the AI ads outperform the human creative ads. It's it's really fascinating. But so what they found was, like, when you saw the collision between someone holding something in their hand and the object, then people the the the click to rates dropped, the conversion rates dropped. But so what they did was they fixed that. So they built this whole, like, visual interpretability model that just focused on that with it with with an expert, like, in the model. And so it just addressed the hands. Mhmm. And so, like, if you but that's use case specific. You needed a general purpose model that's gonna build photorealistic video. We're probably pretty far off of that for, like, all ads of all types. But I think with something like, you know, okay. Well, we need human photorealistic kind of, like, infomercial style. I think you could get to that point now. And maybe we were there now, and maybe ByteDance is there now. But I think it it takes a lot of investment. Right? I mean, they had something that if I remember correctly from the paper, it's been a while since I saw, like, twelve thousand hours of live human product interactions. I mean, it takes a lot of data to do that. Right? And so, you know, it's just it's whatever you wanna invest in.
00:39:03 · Eric Seufert, Thibault Sottiaux, John Coogan, Jordi Hays
I think if YouTube wants to do that for a general purpose photorealistic video ad tool, we're probably pretty far [00:39:08] S4: off. [00:39:08] S1: Mhmm. [00:39:09] S0: How do you think that the market will react if Muse 1.1 is like a very much like a base hit on the API, where they they get some you know, big companies move over some workloads, but it's not, you know, this runaway hit. And and I say that because so many models that have been good, not great, have a little demand just because there's a lot of demand for AI, but but they don't sort of, like, have these sort of breakout, revenue charts.
00:39:41 · Eric Seufert
[00:39:41] S2: Well, you know, they clearly are gonna be very aggressive on the pricing. I mean, they talked about that today. Right? And so my sense is, like, what you're gonna start seeing is that people don't need to operate at the frontier, and you have a lot of use cases that work just fine with and and and, basically, are just good enough, right, with some legacy model. And so it just comes down to then, okay. Well, that's commodity, and so are you priced like a commodity? Yeah. Like like, if if think about, like and and also, like, I think we're gonna see a lot more, like, people relenting from needing to be on the on the frontier when they've built stuff using a model that then gets upgraded. And the whole idea there is, like, well, am I gonna upgrade this tool because I have to adapt it to the new model? Right? Like, if I'm if I if I you know, because essentially, like, the model name like, if you if you use, like, Vertex AI. Right? You you just got a model name as a variable. Like, you're sending this, you know, system prompt to to, you know, Google, but, like, it's just a variable. You could swap that out in thirty seconds. But the the fact of the matter is you're sampling from a new distribution if you do that, and it's gonna change the output. It's gonna qualitatively change the output, and it might change it in quantitative ways that, like, with retention and engagement. And so the thing is, like, if okay. Well, now we're talking about this big cycle. It's a new product development cycle because I have to adapt this product that I built to this new model and the output that it provides. Right? And maybe it was working perfectly. It was working exactly as I expected it to before. Mhmm. Now I've got to invest a bunch of hours, much of engineering time in adapting it to this new model. So even if it's just a swap of a variable name, it's still a whole lot of testing, QA, determining, like, how that impacts long term retention. They're gonna do AB tests. So my sense is, like, you're gonna see a lot more people just saying, no. This works fine. This is perfect. And and getting more getting, like, more robust output let's say the token price is exactly the same. Getting more robust output wouldn't benefit me. Why am I gonna invest the resources into adapting to the new model?
00:41:18 · Jordi Hays, John Coogan, Eric Seufert
[00:41:18] S0: Mhmm. Yeah. But isn't that so so if they're going after workloads that are running on old models that are working fine And it's a lot of work to switch over. [00:41:30] S1: It's not a lot of work to switch over. Not It's it's it's some risks, but it's not a work. [00:41:34] S0: To switch, but then, again, you have to go through this process of, like, QA ing and running it through your own benchmarks and all this stuff. Like, the question is, like, how I I I just don't know. Like, I I'm I'm thinking about a scenario where, like, a year from now, we're sitting here and, like, Meta has a $3,000,000,000 AI API business. Mhmm. And the analysts that you're talking to are, like, that like, to me, they're like, that doesn't get you to 40% year over year revenue growth on the business So [00:42:05] S2: it, [00:42:05] S0: like, doesn't solve at least what those analysts in particular were talking about. And, like, 0 to 3,000,000,000 on, like, a new business line Crazy. Would be crazy and is like, you know, only been done a handful of times throughout history over the last few years. Right? Yeah. So Yeah. I'm just saying like, we're we're there's such big numbers now that there's possibility where you have like this incredible breakout revenue growth, but it doesn't actually move the needle enough that that the market still says, like, hey, we're we're not super confident about about about, like, the next CapEx cycle. Right? This twenty twenty seven numbers that are coming up.
00:42:42 · Eric Seufert, Jordi Hays, John Coogan
[00:42:42] S2: Well, that's and that's the problem with this whole business line in the first place. Like, I think it's a mistake. I think it's a capitulation. I think you're gonna get much more value out of that compute if you apply it to your own core business, which is advertising. My sense is you get better growth. But, like, the problem is the investors don't buy that right now. Like, they've got a narrative issue. It's not a it's not a productivity or a competency issue. It's a narrative issue. And, like, the problem is, like, you know, and, you know, you you you cited Ben's brilliant essay from yesterday or the day before about, like, what stocks should say. He should say that. Like, he Ben is totally right. He should say that. He should come out and say, look. We like, senator, we run ads. Yeah. Senator, we run ads. When he said that, I was at f eight. It was, like, the next month. Every Facebook employee is wearing a shirt that said, senator, we run ads. They know what business they are in. Zuck seems to be confused about it. Like, I don't understand why they should come [00:43:29] S0: Here's here's a big question that'll be interesting. This year, they're gonna spend I don't know how much they're gonna spend on external models. Like, I would say I would I would expect them to spend, like, maybe, like, $10,000,000,000. [00:43:38] S1: 10,000,000,000. Right? Yeah.
00:43:38 · John Coogan, Jordi Hays
[00:43:38] S1: 10,000,000,000. Right? Yeah. [00:43:39] S0: Like, something in the range of $10,000,000,000 from from Google, Anthropic, and OpenAI. If next year they can say, we're not spending money on any external models, then then that could help them with the narrative issue of saying, like, hey, like, we're we're invest we're basically getting instead of having to give this money to other businesses, we're just using our own infrastructure. It's a lot more efficient. We can, like, token max. We can use way more tokens. We can do way more workloads. And so that's potentially it's potentially setting but but the question is, can they actually move all the workloads that are on these other models to their own models?