The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
00:24:30 · Alex Mashrabov
[00:24:30] S0: Yeah, look, I think let's just look at the outcomes which we have today. Out of all the incumbents in The United States, when I look at Open Router Data, the only company which is relevant is Google out of all the incumbents. When I look in China, where probably obsession with benchmarks probably is less, We have Tencent, Xiaomi, Alibaba, three incumbents being completely relevant, and, obviously, like, Biden's obviously trying to catch up as well. What's your takeaway from that? I just do believe that there there is just... Obviously, in the tech bubble, there is a strong obsession over the benchmarks, which do not necessarily represent the reality. But I can talk specifics specifically in the... For video. A lot of benchmarks today for video is really text to video, which does not represent actual workflows at all. The way to think about video models today, it's just modern rendering engine. Think about this as like Unreal Engine or Unity, but just different types of inputs. And it's virtually impossible to really define a visual output and direct the execution just through text. If you just go into our open source projects like this movie, which I mentioned, average prompt length is over 3,000 words. That's the first thing. And look, all these benchmarks which we are talking about, they are not as comprehensive in terms of the details of prompts and people who are labeling, they obviously read 3,000 word long prompts. But also on average, there are at least 10 image references for every scene. The reason why it's important is because it's important to define how the characters look like, how the background looks like, how actually characters are located to each other in the scene and so on. And so that's why prompting and just the workflow is so complex. Benchmarks just don't represent that.
00:26:18 · Harry Stebbings, Alex Mashrabov
[00:26:18] S1: Going back to the model selection, why did we decide we're gonna do our own and then why walk it back? [00:26:24] S0: So it's true that with VFX and camera control, we got very, very quickly from maybe 1,000,000 to 20,000,000 in ARR within maybe the first three months. Then we released our own image model, which is really good at aesthetic photoshoots and product consistency. This is what allowed us to scale then from 20 to a 100,000,000. [00:26:50] S1: So help me understand, Alex, why did you decide that you were gonna do your own models and why did you abandon them? [00:26:57] S0: We still do them whenever we see specific use case like these photoshoots, but as soon as this is what our customers want. So it's all driven based on the customer feedback, not just by ambition to conquer the vaults and build the best model in the vaults. [00:27:13] S1: Do you think every company will have their own models? Like we're seeing Harvey, we're seeing Cognition, we're seeing McCore, Ramp build their own models. And we'll see every company have their own models with their own data, or we actually all use a series of providers?
00:27:13 · Harry Stebbings, Alex Mashrabov
[00:27:13] S1: Do you think every company will have their own models? Like we're seeing Harvey, we're seeing Cognition, we're seeing McCore, Ramp build their own models. And we'll see every company have their own models with their own data, or we actually all use a series of providers? [00:27:30] S0: So first of all, whenever, to be honest, whenever someone says we build our own models, very likely what they mean is something what's happened with Coursera. We we do remember. Right? A lot of companies, they actually take open weights model and just post training on own data. And post training can happen in two ways. Most important is whenever you have customer data around decisions they make, like sequence of decisions, and you can teach the model to actually learn how to compress these 10 steps into one step, this type of reinforcement learning is the most valuable. And I think increasingly more and more companies will have to do that. We see this in the market as well. So most of the companies in the world today, most of the businesses, they don't necessarily need Astra specifically. They don't necessarily need the newest Fable model. That's why OpenRouter reports that share of open source models went from below 30 to over 60 within this year. [00:28:28] S1: What do you think share of open models will be in two years' time? [00:28:32] S0: Look, I do believe that's just because the capitalism works. I mean, OpenAI Anthropics still are gonna have more than 50% of the markets. [00:28:39] S1: In terms of the dollars generally.
00:28:39 · Harry Stebbings, Alex Mashrabov
[00:28:39] S1: In terms of the dollars generally. [00:28:40] S0: In terms of the dollars, right. And especially because for coding still remains to be very prolific use case where coders are always jumping to the recent model. But for our markets, we're seeing completely different dynamics. What's actually happening in social media marketing as companies start to print hundreds of ad creatives a week, they wanna have maybe cheapest, more steerable models because PhD level intelligence is not necessarily needed to make viral social media video. And that's where we actually have seen that we get 80% plus margin whenever we run open source models, like post trained open source models, but it can be way more cost efficient for our end customer compared to the proprietary models. [00:29:26] S1: What's the comparison on margins between open versus closed fee?
00:29:26 · Harry Stebbings, Alex Mashrabov
[00:29:26] S1: What's the comparison on margins between open versus closed fee? [00:29:31] S0: The margin on own models and open weights models is over 80%, and then it almost doesn't matter. And for closed source models, it's probably between 2030%. And then what becomes important is can we actually steer the traffic? What makes me excited about Higgs Field is that Agencik grows so quickly. And actually for us, as companies start to actually create those agentic workflows to make more ads, we choose which model we can use. So we choose what model to use in over than 40% cases. In a way, model routing becomes a core feature of the business, though? Yeah. We call it tokenomics essentially. Right? As there is certain amounts of work customers wanna do, how can we optimize number of tokens which require it, and how we can pick the most efficient tokens for them. There are actually two incumbents in The United States who figured out models. It's not just Google. It's also NVIDIA. Why do you think that is? What I'm constantly seeing is that there is diversions of models. So there are these state of the art models which have to be really good in computer use, like Astra or in coding, but they can be prohibitively expensive. And we were chatting about that. It's like on average at Hixfield, person on the team spends over $10,000 a month on various models. And remember, we are split across United States and Asia across the So United how much do you spend on models per month? So internal usage of models a month is over 4,000,000. [00:31:06] S1: Wow. How many people do you have?
00:31:06 · Harry Stebbings, Alex Mashrabov
[00:31:06] S1: Wow. How many people do you have? [00:31:09] S0: We have close to 400 people and just wanna make sure that the math adds up. Yes. It's it's definitely over $10,000 per person. [00:31:17] S1: How has that changed over time? [00:31:19] S0: That's the best question of the whole show, by the way. That's the best question. What actually started to happen is the creative team started to do vibe coding. Like, this month, I was... I I just caught a guy who spent over 30 ks in a week on Astra model. Because he was frankly frustrated that some asset organization workflow, and as you said, basically auto editing, is still not very good in production. And he said, oh, I'm just gonna do this myself, and just went, like, five nights straight on Astra. And it works? We learned a lot. I wouldn't say it was production ready, but we learned a lot. [00:31:57] S1: 30,000 in a week. [00:31:59] S0: Yeah. Yeah. Many people spend over 10,000 in a week. Do you mind? Yeah. My finance team will probably say... I don't know if if if, you know, if you ask any of them, but they will probably say that I'm, like, being too stubborn, too relentless to control this spend because sometimes I feel it goes like... It really goes out of control. Like, 30 k in a week is quite a lot. But we learned... So this was actually a net positive experience. [00:32:22] S1: Okay. So the internal spend, 4,000,000, about 10,000 per head. What will that be in twelve months' time, do you reckon?