Transcription
In general, there is still a lot of demand for image models, and there seems to be some kind of convergence on quality. But then, each model really has its own differentiation.
With video, we're earlier in the competition. There's still a lot of leapfrogging happening. There's just so much more to build, and there's just, like, you know, we haven't hit like a quality bar where there's just like marginal improvements. We're not there yet. So over there, it's like more fierce competition, and it's, it's very hard to predict what's going to happen next month. It's like that's we're operating at the scales of like weeks at this point.
I remember when Sora came out, even in our team, people were like, "Oh my god, OpenAI is like so far ahead that no one's going to be able to catch up." And then like Luma released their model, Runway released their model, Cling released their model, Minimax released. And every release, like, if you're not the best, you're not releasing, generally that's how it works. You can never like say, "Oh, this is the model," and then this is not going to have any competition for like even a month, right? Like even like two weeks is like, I think that we are operating at like weeks.
As I said, we started about 4 years ago. The origin story is, I I used to work at Coinbase, and they had a lot of infra issues with respect to machine learning, and fraud was a big problem there. So I kind of grew up in that environment where like we're just constantly fighting with fraud using machine learning models. And the initial idea had a lot to do with building these pipelines for for companies to be able to like train these models. But about a year and a half into us starting the company, ChatGPT happened, DALL-E happened, the whole world of machine learning and AI changed. So we sort of adopted as as things were developing.
We're going to definitely dig into the 2021 wind shift happening on on the multimedia side. But before we go there, how did you meet and recruit Batuhan, the foul guy now? He is the the mascot on Twitter of of F. Yeah, we're both from Turkey. And like I first saw Batuhan online. I was pretty curious about his work on on Python. He's a pretty big Python contributor. I just DM'd him on Twitter, and we had a call. And we had just started the company, and he was in, I think, Poland. Yes. uh in a in a dorm room, I think. And I was pitching F like, "Hey, you know, we're doing a lot of Python stuff, would you come join us?" And initially, it was like, very much an intro call. And I think at the time, you know, he was working on something else, it wasn't going to work out. Few months later, we actually raised from you guys. So this time I was like, "Okay, we have some funding, great investors." I'll go pitch it again. And this time we got on a call again, and he happened to be leaving. And then believe it or not, when I said, "Hey, like, we just raised from," and it's not announced yet, "we're going to recruit a lot of people, like you're going to be one of the first people to join." That's when he was really convinced. I'm really glad to be a very small contributor in this Helen acquisition.
What were you doing in Poland? I went there for a university, but I dropped out in like the third week or something. Was preparing to leave Poland, go back to Turkey, continue working, you know, at developer tools. I want to explore like different options. Like I come from a developer tools background. If you consider compilers and programming languages, developer tools. I just wanted to work on those a bits. And I was like leaving a data company. So I wanted to do more like deep developer tools. But then seeing Burkay, seeing your cam, the excitement around the product, I think the traction was just starting to come up for the product. And seeing that, seeing what I can bring to the table, convinced, "Okay, these guys are the best people I've ever seen in my life, vibes-wise. I want to work with them for the next decade." Let's just let me just join. Awesome. So you and Gorkham Vive sniped for fun? Yes. Yes. And you also got into Python at a very young age. Maybe tell the story of how you got to. I I first started doing operating systems, even before compilers. And one component of building like your own operating system, which is like very basic, you know, a super basic kernel, and then like some sort of graphical user interface, is writing your own shell language. So I started building that. And the the thing that I enjoyed the most of all that operating system journey, from like writing a custom bootloader to, you know, writing a custom file system, was this like shell language. So I started doing more on programming languages. And my go-to language was Python. So it was natural for me to just like go jump in, see, understand how the fundamentals of Python worked. And from that point onwards, I was able to go start doing small patches, contribute, understand like, you know, what the community wanted. Like it was also like my introduction to being like an open source contributor. And I I started contributing. I think a year and a half later, I became one of the committers and the maintainers for the parser, core compiler, and some of the parts of the interpreter. That was one of the, you know, most favorable and highest learning experience that I've ever.
Going back to the pivot into the direction of generative media in 2021, all the rage was about large language models. Yes. But when you and Gorham saw this shift, you picked generative media, which is image, video, audio models. Why did you make that choice? The way we got into this, I I remember this like so vividly. This was like November. It was a Thanksgiving member on a call. You were in a cabin somewhere upstate. That's right. I I was living in New York for like one year. I I was in upstate New York. We had a Python runtime at that time. And the idea was, we think the next big thing is going to be workloads running in a Python cloud. We just thought Python is the next biggest thing. And we got a lot of inspiration from Snowflake's success with SQL. We thought like Python is the next big thing. And this is pre-ChatGPT or pre-boom of AI models in general. We're just like sitting there and we're thinking about what's the Python workload that's going to be really, really big. The more we we saw the excitement around language models, image models, the less excited we felt about the old school enterprise workloads, right? And and that was one thing that we were like, "Okay, it really feels like the big opportunity is in these new kinds of models," and especially the scale, how large of a scale you have to run these models. That basically brought us to LLMs or image models.
This November, we're just sitting there trying to run Stable Diffusion 1.5. We just kept asking ourselves, "Why the heck is it so damn slow?" I remember SD1.5 taking 19 seconds, maybe 10 plus seconds to like run. We had a lot of patience back then to come up. We had Batuhan at the time, like we were like five or six people. And I mean, he he has like amazing experience in like compilers and, you know, performance engineering. We just went super deep into SD1.5 and just like optimized the hell out of it. That was like really the start of it. And it it didn't really have anything to do with, "We think this is going to be massive." It really started with like a technical curiosity, which which is, I think, like really interesting. And and, you know, we did that extremely well. And then that became our like wedge into tackling some of these use cases. And and people started like associating us with speed and performance. And from there on, we made more directional bets and like business decision decisions that that kind of brought us to like the overall media story, which is less to do with tech and so much more to do with the market.
People tend to forget that back at those days, there was a very massive GPU crunch because it was like even before the AI investments and Stable Diffusion was popping up. ChatGPT was going crazy. I remember like we even tried to run Stable Diffusion on Hugging Face, and there was like a 3,000% queue. Like Hugging Face had this queue system and like just suddenly submit per prompt. So it was also like coming from like a necessity where the GPU resource was very scarce, you know, like you you had to maximize the efficiency competing with the language model uh workloads. Exactly, where you were essentially trying to get allocation to that little amount of GPUs that's given to you as a startup, right? Like Google had zero incentive to give us uh GPUs back at the days. I remember our CUDA was eight GPUs total, like all of our system. And we were trying to make do with that. And I think we were able to extract an insane amount of value just by optimizing these programs, going to the first principles, trying to understand why the slow, where are the bottlenecks, and how can we resolve those.
Since you mentioned you have to get a lot of compute and iterations out of just eight GPUs, where did you even start with optimization? For me, like this was my first time entry into machine learning performance space. As Burkay mentioned, I come from like a traditional compilers background, and everything is like performance engineering. Like you just take a program, profile it under different conditions, figure out where's the bottlenecks, and try to understand why isn't this operating at the maximum achievable speed for your CPU, for your clock speed, whatever. I did the same math with GPUs. This is the GPU, this is the horsepower, this is the maximum achievable flops that I could get out of the GPU, and this is how much flops this workload needs. What is the difference there? And then you start noticing patterns where simple calculation might be getting blocked and not being able to utilize the fullage the GPU. So you start sharding these workloads, uh threading them in in a way that's much more efficient. You figure out common patterns within the execution layer that you can group together. So like figuring out these small bottlenecks and then picking up from there, one one optimization by optimization. But it's like extremely satisfying work, just going it like going down on speed. There was a lot of work around the quality as well that I can talk more doing this optimization space. Just like, and it was December, I had nothing else to do. So when you combine these two, you you get these like our optimization sets that we did for Stable Diffusion, you know, just much more efficient kernels for the workloads that we were targeting, and making sure this is like a pluggable system. So when the new model comes, we could just apply the same set of optimizations.
Most efficient kernels is the best Christmas gift. It is the best Christmas gift. Yeah. Back then, a lot of the inference platforms and Python clouds, if that was even a thing, were agnostic of different types of workloads. But today, we're certainly seeing F.AI emerging as the leader of generative media workflows, and as the company, the first company that's uh coining and defining that term. Where are we in these two paths of diffusion models or generative media versus language models? The point when it kind of clicked for us was actually, I think Llama 2 was a a pretty pretty big turning point. When Llama 2 got released, there was this big rush towards inference platforms in general. You know, "We're going to host that, and we're going to host the image models, and we're going to host everything. Everything that runs on GPU, we're going to host it." One of the reasons, I remember this also very vividly, like one one of the the key reason we didn't want to do it. It was going to take a lot of time to optimize. We were just putting so much effort on image models. So we just didn't have the resources. I think at the time, we're probably still like six, seven people. And then the other reason was, these image models, there's so much demand, and the use cases are starting to appear, and it's looking like its own thing. Why don't we just focus here? Build something that's more clearly targeted towards its own users. That was kind of like the first moment where we said, "There's something here that is is kind of special about these models. Maybe they should be its own market. Like maybe we should be a specific platform for just images." And we had the idea that if we can do really good images, we can probably do really good video. We can probably do really good audio and like full experiences. Just imagine like where where this can go. And I think we had like a very early on, very like strong conviction around how the space is going to evolve and that it's going to evolve into its own market. And that that's when we said, "Okay, like we're going to just focus on here. We're going to beat the drum of generative media. We want to be the torchbearer of the space, and we're just going to be super focused here."
A few months after that, I think Sora got announced, and that was like definitely the right bet. We were just so happy that, you know, someone proved ourselves right, and we got really excited and like we doubled down even more on our messaging. I remember like going and changing the website copy, like, "This is it, you know, this is what we're going to push." And so we can talk about the whole arc. What are the types of things people were building back then, and like how that's changing? A lot has happened, I think, even like in this in this short period of time, definitely. And you you put out a prediction that 2025 will be the tipping point of AI video. Of course, announced it quite a bit earlier than that, but now we are seeing a plethora of video models. What's behind that prediction? There was a very big push from Chinese labs in particular that initially when they were first pushing these models out, it wasn't very clear like what was happening. And it took some time for especially like in the West, like for people to realize like what's going on. So this, we we kind of caught on this fairly early, like sort of was like February. And then in September timeframe, I think we first saw like Minimax come out, and it was like kind of foreign to to to everyone on. And like people would just see things on Twitter and Instagram and get very like surprised about it. And we thought, "It is starting, and it's going to just go blow up from here." That was kind of the the bet we were making, uh, that, you know, we're going to have more models, we're going to have more capable models. And here we are with Google's V3 coming up. It sounds like your 2025 bet is correct again. Yes. I mean, definitely what we see internally is, every time there's a big shift in capabilities of models, the adoption and the use cases just, it's like a step function. It just keeps growing. We're just, I think, scratching the surface. Audio is cool, but like, there is actually really interesting things that like the V model is doing, for example. It's very good at comedic timing, for example. That's something like people say, like, and if you look at a lot of the like, you know, shorts people create, it's like really freaking good. So there's things that you just experience with these models that like it's hard to kind of point out what they are, but these things are getting better. And these are the things that like consumers and and enterprises are looking for in these models. So these capabilities are going to keep growing, keep getting better this year and next year.
And one big learning from even our side on the modality bifurcation is, given these are different workflows and different workflows too, like people are using image, video models more as a chain workflow where you process image, remove background, you improve the resolution, you add some more color palettes on top of it, and then turn it into potentially more produced image or even video. It's a pretty complex and bespoke custom workflow that requires not just the inference engine itself, but also how you built this workflow on top of it. The whole Comfy community are focused on that, and that's just very different from language models. It it is like language models are very generalizable with just like an example, you know, we can zero-shot, one-shot most of the stuff that you can get. But especially with the initial image models, the the quality was not there to one-shot your generations, right? Like editing or like upscaling, you know, rem background, whatever. It didn't generalize. It was just a text-to-image model. And Comfy UI, thanks to Comfy UI, people start discovering that you can just chain multiple stuff. You can chain text-image model and then do use the same thing as image-to-image to refine the image, change the image a bit, or upscale the latents to get like a higher resolution image. So these like workflows started getting out. And as soon as we started seeing that, we we start thinking, "How can we make this more efficient?" Because like one one thing, this this worked well with consumer, but it didn't scale up for as APIs. It didn't like, you know, there was like different dependencies, and every part of the stack is going to be optimized differently. So we even built our own workflows product, just seeing the adoption of these multiple different models getting chained together or the same model getting used differently. And then people are expecting to put a prompt then and get a result back, you know, for their own use case. I think that's really great for us to see this like happening, because that means more creative use cases can appear, even from the least capable models. And that's part of the reason why, even though there are a lot of attention on these diffusion transformer models as well, that can get the end results through through just prompting with a lot of controllability, you still need this more fine-grain control on the image model itself. It just adds more capabilities into the overall pipeline.
We also saw fine-tuning being like a major component to this, where if I have to guess, I think there's been 1,000 times more fine-tunings in the image space than the language space. People also fine-tune language models, but not this much, because this was the only way you could provide context. Now, there there's like GPT Image, One High Dreams, Edit Model, Black Forest Labs is a context model that gets additional context. But we still see people fine-tuning for particular tasks, you know, like virtual try-on or characters, products, to get the highest amount of consistency out of it.
Talking about the fine-tuning workflows and demand, we have seen a lot of difference in just like people focusing on more large-scale pre-training or large-scale post-training workflows, where a lot of developers on F.AI are more doing these LoRA training and building their, you know, unique styles. Is that deliberate from the F.AI side to focus more on that type of workloads versus going all the way back to large pre-training, and why did you make that decision? As a generative media cloud ourselves, I think we we are trying to target both use cases, but the one that has the highest volume in terms of being able to use as an API, we saw the fine-tuning was the most effective solution because it costs less. There's so much opportunity for other people to use it versus like there's much less amount of people who can do the post-training themselves. But we now have like post-training solutions and distillation solutions for these large models that people can come, give their data, and get something out of it. And then we have been doing this one, like we we even did one of these examples with Freepik. We trained like a we trained an open-source model in collaboration with them called Flight, and that was trained fully on commercial data using our proprietary training and data processing stack.
Talking about the open and closed source models, who is winning in both the image camp and also the video camp, and how does F.AI prepare for both sides of workloads? Yeah. So, the winner is literally changing month-to-month. That's that's what we're seeing. It's like a fierce competition. It's very difficult to predict what's going to be the biggest model this month. In general, there is still a lot of demand for image models, and there seems to be some kind of convergence on quality, but then each model really has its own differentiation. Just to give some examples, like Imagen 3, Imagen 4, very good at like character consistency. Flux, amazing at having an ecosystem of different tooling, you know, so that like you can just kind of use an existing workflow someone's built, you know, or a fine-tune. Each model really brings its own capabilities. And I think foundation model companies also know this, right? Like they need to have their own edge also. So they they kind of play into their strengths, and they they all have their one thing that they're like really very good at.
With video, we're earlier in the competition. There's still a lot of leapfrogging happening. There's just so much more to build, and there's just like, you know, we haven't hit like a quality bar where there's just like marginal improvements. We're not there yet. So over there, it's like more fierce competition, and it's it's very hard to predict what's going to happen next month. It's like that's we're operating at the scales of like weeks at this point. Our leaderboard is is essentially the traffic that models are getting, and that just keeps changing all the time. I remember when Sora came out, even in our team, people were like, "Oh my god, OpenAI is like so far ahead that no one's going to be able to catch up." And then like Luma released their model, Runway released their model, Cling released their model, Minimax released. And every release, like, if you're not the best, you're not releasing, generally that's how it works. So like everyone who's releasing is on the top. And then like other players start entering, like open-source Genmo released their model, Tencent released Hunan, Alivo released one. The competition is like very fierce. And just seeing like V3 got released, and two weeks later, I remember seeing like ByteDance C-Dance model just like leapfrog them in the arena. You can never like say, "Oh, this is the model," and then this is not going to have any competition for like even a month, right? Like even like two weeks is like, I think that we are operating at like two weeks.
As I said, do you stay pretty agnostic for all the model uniqueness and quality, or do you also poke around whenever a new model launch to figure out if this model is uniquely good at this thing and just like generate a bunch of image and videos to know how to let's say serve the model to your developers? I think we built such a team that people are interested, even if they weren't working at F.AI, on which model excels at what. Like they they know everything. And like luckily, we partner with these companies, so we get to play with them before the release so that we can advise our customers, because they are the ones that are asking us, "Oh guys, what is the best model for doing product photoshops or doing, you know, virtual try-on?" And we are the ones that is advising them. And like getting playing with these models, having a feel, even like six-month or nine-month-old models might excel at something, and people are like very sticking with it because it's it's like really good at logo generation, let's say. And that's that's a phenomenon that we're seeing. Like the best model of today might be best for generic tasks, but there's very specialized tasks that people are using even like older models for that.
On this point, more and more F.AI starts to become this two-sided marketplace where on one side, you have a lot of this developer attention, customers, and users that have unique use cases that are coming to you for the advice and also for the API endpoints on these uh image video models. On the other side, because you have this developer attention, which is awesome, and more and more signing up every every day, every week, the model players wants to, you know, expose their models to the end users as well. Is this strategy sort of deliberate from from day one, or it happened given this sort of booming of of diffusion models or or creative models, and what's the strategy now? I would actually date this back to another moment where I think Cling was being released or it was released, and it wasn't clear like where it's coming from, how does one run it. And I remember there was this guessing game of like, "What is this model? Where is it from?" Yes. Yeah. And why is it so good? Exactly. Like like Yeah. So, so it was a mystery to a lot of people. And like I think we've really helped kind of bring those models to like the market here. And and, you know, it wasn't necessarily deliberate, it just it was just something like we just felt very strong urge because we just saw such demand from from our customers. So, you know, being being a generative media platform, we just had to do it. And that sort of created a very interesting, like you said, like marketplace dynamics where after doing that a few times, and I think after seeing how successful these models got, a lot of other companies actually wanted to come list things on F.AI. So that that sort of created a good flywheel effect for us where developers are at F.AI because there's like these cool models, and then, you know, model listers want to list on F.AI because there's all the all the developers, enterprises that we work with. So so luckily, that turned into it its own business model for us, essentially. And and yeah, it's been it's been very successful.
Another interesting angle there is that because of our success in like hosting open-source models and like running this infrastructure extremely well, we also help some of these companies optimize their workloads, prepare them for launch day, and help them benefit the infrastructure that we've built, capacity planning, and also anticipating what's the the peak uh and spiky workloads, and also how to load balance. People generally overlook that aspect of the business because we don't really talk about our infrastructure that much because we think it's table stakes, right? Like every like if you want to do great inference engineering, you need to have like great infrastructure. We have an amazing infrastructure team. We're managing tens of thousands of GPUs at at certain peak points. And to do that, like we're essentially building a distributed supercomputer, getting chunks of compute from like different vendors, making sure all the workloads can be orchestrated and can scale up, can scale down, have access to super-fast distributed file system to load the model weights. There's so many different problems there that we don't talk about, but it's like it's a challenge that every foundation model company is trying to build in-house. And we are essentially trying to help them. You don't have to build this yourself. We built it. You can deploy your workloads here. You can optimize your inference while you deployed. And then when you need to scale, we're going to be there for you.
Let's talk about that aspect. Speed is literally running through everything about the company. How you catch up with every new model launches, how you run every model, open source especially, optimize it to, you know, the fastest speed on the market possible. And also how you spin up and down these uh different workloads, different models, given this massive infrastructure. It's just permeating through the whole business. How have you gained the knowledge of what are the the critical infrastructure pieces that you need to really build from ground up versus let's say leveraging maybe one of the three cloud providers, and where which layer are you have you strategically placed bets on that needs to be owned in-house as a secret sauce? From the start, we were like, "Okay, when you remember the 8 GPU CUDA's that I mentioned, we were like, 'Okay, we're never going to be able to get capacity allocation from hyperscalers as a seed-stage startup back then.'" And we we decided, "Okay, we're going to build a multicloud system with our own orchestration on top of it." We tried like, we tried using Kubernetes, other solutions, but we found them to be too slow for cold starts because till to to that date, like no one cared enough to start a workload and shut it down like after a request finished and start another one in like less than a second. You know, when you were trying to do, when we were trying to do multi-cloud Kubernetes, we were seeing like five-second delays to orchestrate a single container just to for the to start. And like five seconds might be acceptable for like web workloads, but not for our case where five seconds of GPU time was very precious. So we started building our own orchestration system to be multicloud. And when you when you go multicloud, another big component is you need to be able to have access to the same data. And for our system to have the same level of developer comfortability as like other products, we needed file system to be accessible because like, you know, using S3 doesn't have the same level of comfort. So we built our own distributed file system storage using like uh existing solutions, building stuff on top of us, building multi-layered caching, so we could cache at the data center, within the nodes, within nodes memory, because like these nodes come with like two terabytes of RAM, so like building all these solutions, it's just like seeing it's it's an endless performance engineering work, just going faster and faster. There's no limit till like everything's at the light speed, you know, everything's at the theoretical maximum. We still have a long way to go, but we have come a long way. And this is one of the parts part of the infrastructure that, you know, our infrastructure team, which is amazing, is handling and helping serving our own, you know, applied ML and inference engineering parts of the team. They're like literally the service provider for those.
Which optimization, if you can recollect or recall, has delivered the largest performance gain for infrastructure? I guess distributed file system caching was one of the biggest points, you know, just making making sure that if I load the same model weights in the same data center, I can just read from my peers which I'm connected the 100 gigabit. And if I use it from the same node, I can read from the NVMe. This was like a very big turning point versus not using like any caching and going to Google Cloud Storage or S3 every time you want to load model weights. This shaved off like a single amount of time. And like this is still not like widely adopted because people don't really have the same issues when they're using a single cloud. If you're just using GKE or Amazon's like EKS, you're just using a single cluster, and then you can have your storage there. But like this is these are the new challenges that we were like building for that enable like that required these sort of extremely well-thought distributed caching systems.
And do you fundamentally believe speed is a mode? However, you optimized for the model to run really fast, and also how fast you can I guess optimize to the best performance point possible in that time. When I look for open source, if I evaluate our inference engine from a year ago to open source today, it's falling behind because open source is catching up. I don't see inference engineering speed as like that mode. It's always focus and being one step ahead. Always being at the peak, right? Always putting, if open source has a great idea, adopt it and then build your stuff on top of it. It's it's like if you're just saying, "Oh, I built this great thing. Here's the good set of kernels," it's going to get outdated because Nvidia has like 50 people working on these sort of stuff. I don't know, Meta has like 100 people working on this stuff. So for us, it's always this is our focus: running these diffusion models, diffusion transformers extremely fast. And whatever we can do, we are just going to be try to one step ahead, have this focus, and then always be at the edge.
And how do you structure the team to keep up with what's happening, I guess both at the big labs, but also at the, there are probably at this point a dozen or two dozens of generative media companies are developing different styles and types of models. What is cutting edge? What is the best performance price point for whether it's image generation or video generation? Like it sort of is a constant, you know, moving target. And and for you to provide the best price quality for your end user and customers, you kind of have to stay very educated about the whole market. So how have you I guess structured the team to to do that? I think pretty early on, we put a lot of effort in engineering, obviously. Like I think until we were like 28 people or so, we were all engineers. And I think like the 28th hire was a non-engineer. So like our core is like very engineering heavy. But around that time, like the business was also picking up, and we thought that we also have to have a very strong business team, like go-to-market team. So these were very very important to us. And I think in F.AI's culture, we are results-driven. We are really revenue-driven. That basically meant that like we have to be just as good at go-to-market as as engineering. So, you know, we have a small go-to-market team. We're currently 40 people. We have a smallish go-to-market team. It's like six people. Up until that point, you know, we were doing founder-led sales, three, four, like all of us, four of us in addition to that, six people. And then the rest of the team is like pretty much engineering. And within engineering, I mean, Batuhan can talk more towards it, but we're responsible for optimizing the whole journey of a request, essentially. So there's infra engineers, performance engineers, there's product engineers, right? So we can deliver this whole technology in like a really amazing usable way. And then we have a massive applied ML engineering team. And that's that's like, you know, one of our, I think, secret sauces, I believe. The team is like 10, 11 people. Love anyone. More than half the engineering team is applied ML. And these are the folks who are literally like obsessed about, you know, this market and this space, getting all these models, putting them in production, um, being like the experts for our customers to answer any questions they have. And within this team, even so, we have we have some folks that like sit outside of this team that focus on performance. But within this team, we we also have like performance engineers, some that are like more customer-facing, some that are more like post-training, fine-tuning uh focused. So it's it's really a mix. And and that that team is like, you know, bulk of our engineering team.
Let's talk about go-to-market enterprise because it's so it's so fascinating to me how obsessed of these 28 engineers are at the business problems, because it's very rare to find when you're already so deep into the infrastructure and performance engineering at the same time solving really bespoke customer problems, right? I've not met any other engineering teams that are so customer-focused, customer-obsessed, and also market-driven too. Even for the founders yourselves, being just very forward in having the sales conversations with customers, leaning into enterprise sales, it is just a very interesting and unique culture. Do you think that's also a deliberate decision, or it's more being pulled by this customer-driven uh development uh fashion? How did it come about? It's very deliberate. It's just like, I think it's part of our culture. We tend to hire people that are very ambitious. I mean, in order to win, you have to win the market, like you have to win these customers and have them work with you, right? So like it really means that, you know, we have to not only like build amazing technology, but like we have to also like grow the business. We heard a lot about enterprise sales, right? And like from including you, in fact, like you introduced us to our first like, you know, sales advisor. I think with technical founders, engineering general, there is this like massive skepticism about sales. Famously like engineers hate to like talk to sales people or be marketed to, you know, all this all this stuff. And yeah, I think I think it was, "Hey, like, let's listen to, you know, what people are saying. They've seen thousands of companies. There must be some merit to that." And initially, I think even we felt a little bit uncomfortable, but it didn't take that long. I think once we saw the results, we were just like, "Okay, like this is something we really need to be doing." So I think we're just, you know, Gami Batuhan, like we're just very open to learning and and adapting and and do whatever it takes to like really win. And and, you know, that's that's kind of what what actually ended up happening there. And and like we We got a very good hang of it very fast. And and we had some really amazing people like supporting us. And and yeah, and, you know, we got so into it, I think we like couldn't let it go. And we did like founder sale, founder net sales for so long. And then eventually we said, "Okay, like, you know, we can actually hire a team. We can scale this. Turn this into like a, you know, machine." And and, you know, that's that's kind of what we've been doing. And it's been very successful. We we are like really customer-centric when consider a company. I think we might be one of the company like we might in the in the top leaderboard of Slack connects across all Slack users. You know, we have Slack connects with almost all of our customers. The channels are open. So engineers are inside the channels. Like every channel might have like an average three, four different engineers from applied ML team, product. We are like very centric. And that's the sort of profile that we try to hire with sales as well. People who are going to listen to these customers, people who are here to serve the customers, not sell stuff. I think that arises from the fact that we're coming from an engineering background. We don't want stuff to be sold to us. We want people to serve us, help us, grow with us, partner with us. And that's the profile that we have been hiring. I think that's a really great point. Is sales today, especially when you're selling infrastructure as well as dev tooling, it's very different from let's say 5, 10 years ago when you're building this beautiful deck and talking about a lot of the performance and trades of what you get from a third-party vendor versus building in-house. It's very much about how you can learn with us even faster, and how we can be embedded with your problem so we can serve that demand as well, just because the speed of iteration that nobody has a perfect knowledge, but you have a very unique insights about how generative models work, and you can serve the customers best when you're so entrenched with them. And I have introduced a lot of sales advisors or sales leaders to companies, but I think you guys have embraced it perfectly, and really it shows from the company culture. Yeah, we want our sales people to be advocates for the customers so that we don't drop the ball on their future requests, on their needs. We always are there to serve them. Making that the culture thing worked really well for us.
Last question. If you put yourself a year or two ahead in 2026 or 2027, if generative video do not take off or as big as it is, what would be the reason that caused the slop? I think it's impossible that it doesn't take off. It's here. It's all over my feed. I I don't know about your Instagram feed, but like I couldn't tell which ones are real, which ones are not these days. It's it's it's one of those things I think like it's it's too late now, you know, it's like, you know, cat's out of the box. I think the question is more like, how is it going to be distributed across like different industries? Where are the biggest opportunities? And like how can we actually go invest in those areas? And what are the areas that you see opportunities but not enough developers are paying attention or figured out that these could be use cases or these could be interesting products to build on top of F.AI that you feel is underrated or underdiscovered now? Yeah, we have like a god's view on like all the things people are building, which is very very fun. And and we see lots of very interesting things when we talk about this among ourselves. I think what we all agree is that we're very excited about the net new use cases, recreational, you know, image generation, like people just like having fun with it, and that becoming like now video, that becoming maybe like short playable games or whatever. Like those kinds of things are going to start happening as these models get bigger. We're most excited about these net new use cases, which is not to say like everything else is small, but it really feels like these technologies are so powerful that like we're going to get a lot of net new things being built on top. Yeah, I'm for, for one example, very excited about the the real-time generated ads that's placed in videos too, because I that's always been the dream of how you can adapt the advertising in a video in a show to display certain products, but now it seems very much possible given what's happening in the generative media world. And also just all the IKEA rooms unpacking videos are so fascinating. They're great. I mean, like cat Olympics. Yep. Exactly. Yeah, those are very Those are amazing. Yes, even the interviewers themselves, they're going to be AI generated. That's true. True. But it won't be this fun of a conversation. Thank you so much, Batuhan Burkay. This is awesome. Really great to have you guys. Thank you for hosting us. Of course. Yeah.