A (0:00)

Welcome to the podcast. I'm your host, Jaden Schaefer. Today on the show I want to talk about foul AI. They've just recently raised $140 million in a series D and they've just come out with a brand new model. This was kind of their year end surprise. What's incredible here is it's built on top of Flux 2, which is an open source model from Black Forest Labs. I've talked about them a lot. They were kind of the original image model that powered Grok over on X. What FAL has done that is making, you know, it's surprising a lot of people is that they've been able to build something on top of that model to create Images that are 10 times cheaper and 6 times more efficient. I want to cover this today because I think this is the direction that all of the AI companies, all the large AI hyperscalers, everyone is going to be moving in the direction of. And so I want to break down what they, what they're doing, how they're doing and what some of the innovation is. Before we get into that, I wanted to say a big thank you to today's sponsor, which is delve.com if compliance is something that's slowing down your deals, Delve is an incredible resource. They help with SOC2, HIPAA, GDPR compliance. Busywork can definitely kill momentum inside of your organization. Delve uses AI agents to automate compliance. They do this end to end. They collect evidence, they fill out security questionnaires and they customize controls to your actual business so you can get compliant in days, not months. Something else that I think is awesome is that you get one on one slack support from real security experts who respond quickly. There's over a thousand fast growing companies right now that are using Delve to help them close deals faster and stay compliant as they scale. If this is something that would be interesting to you, go check out delve.com I'll leave a link in the description to go check out Delve. All right, let's talk about what's going on with fal. So I think the, you know, obviously they've just raised a whole bunch of money. $140 million. What's interesting here, they've just unveiled a flex to deb dev turbo. This is a faster, cheaper and more efficient version of the open weight model which was originally released by Black Forest Labs. So this new model, it's already outperforming a lot of large competitors on public benchmarks. It's available on hugging faces today, although there is a really important caveat. It is distributed under a custom non commercial license which was originally created by Black forest labs. So Flex 2 Turbo is not a full standalone image model. I'll just definitely put that out there in a traditional sense. Instead it's what's called a Lora adapter. This is a lightweight optimization layer that essentially attaches to the original Flex 2 base model and then when they attach it, it dramatically improves the performance. So the result is that you get these really high quality image generations you and they're all delivered in a fraction of the time and they're at significantly lower cost. So this is an incredible innovation if they can apply this to Flux 2, which is kind of this open source project. So that's why, you know, they're able to even work on this. But if they can apply it there, there's so many companies that could take the same technology. OpenAI Claude, who doesn't really have an image generation model right now, there's a lot of other players that could also take the same strategy and we could see much cheaper, much faster images, which I actually think will make a huge difference. I don't know if any of you struggle with this. I definitely, definitely do. Basically every time I use ChatGPT to create an image, like, yes, I know it's a magical AI machine that can make incredible any image I want. And so like I should just be grateful for what I have. But it really is sort of annoying to sit there and wait for like two minutes for my image to be generated. If you could make this thing six times faster and 10 times cheaper, I think that would be an incredible innovation. So I'm hoping that OpenAI can take some of this technology as well. I think one of the most important things that the model's open weight. So for engineering teams that are trying to, you know, grab a good software solution. Of course there's so many closed APIs right now. And this new Turbo shows how this kind of targeted optimization of open models can actually get some really big gains in speed, efficiency and cost control. I think what's interesting here to me is kind of foul's bigger bet, which is this infrastructure over models approach. They're really positioning themselves not as a model company, but as an AI media infrastructure. So they're kind of serving as this centralized hub for real time generative media. They're offering developers access to both open and proprietary models that they have image, they have video, they have audio, they have 3D generation. According to a recent press release that they did more than 2 million developers are now using their platform. They also operate a usage based pricing on their product. So essentially they're charging per token or per asset. FAL is actually someone that I think we originally looked at. I'm not, I'm not sure if we're using, we might be using one or two things from them. On AI Box, which is essentially my product where you can, you know, you use an AI and you can build tools without knowing any code. You just describe what tool you want it to build and it can chain AI models together. What we were originally using FAL for or what we looked at it for, I think what we're primarily using is together AI. But VAL does something very similar which is essentially have all the open source models, they host them. You can use an API and you pay them just like you would pay OpenAI for using their model, for using a lot of open source models. So foul or together. A lot of these players are doing a really good job in this space. But I think what's even more impressive is beyond just kind of, you know, offering, beyond just running the open source models and offering an API subscription, which is very useful for developers like, don't get me wrong, but when you start building software on top that optimizes the models that are there that nobody else has, like this, you know, improved version of Flux, I think that's when these companies can become really, really valuable. They said apparently on their whole platform in the past year they've become one of the fastest growing backend providers for API generated media. They said they're serving billions of assets each month. A lot of that has drawn from their investments that they've raised. They've had money From Sequoia, Nvidia, NVentures, Kleiner Perkins, Andreessen Horowitz. And their customers are, you know, ranging from solo developers building creative tools to a lot of enterprise teams running large scale personalized media pipelines across retail, entertainment and internal design workflows. So right now Flex 2 Dev Turbo is the latest addition to their stack that they've built. It's one of the most developer friendly image models currently available on the open weight ecosystem, I think. But what does it really do differently? I think it is a distilled version of the original Flex 2 dev model, which was released last month by Black Forest Labs, which is a startup founded by former Stability AI engineers. The base model was positioned as kind of this open alternative to offerings like Google's Nano Banana Pro, which is part of Gemini's image lineup, and OpenAI's GPT image 1.5. So when the original Flex 2 required roughly 50 inference steps to produce a high, you know, a really high quality image, Turbo achieves a really comparable output in just eight steps. So going from 50 steps to eight steps, this is a massive improvement. That acceleration I think is enabled by customized DMD2 distillation techniques. So I think the speed has not come at the expense of quality in evaluations which are conducted by artificial analysis. My favorite, you know, one of my favorite tools for determining the best AI model for something. Turbo now holds the highest ELO score among all of the open weight image models. It has a rating of 1,166 which is passing the competitor, some other competitors like Alibaba, a bunch of other players on the YUP benchmark which is factored in the latency price and the user ratings. It has Elo score of 1024. In you know, 66 seconds it can create an image for.008 dollars per image. It's basically the lowest price that is currently on the leaderboards. And I think that is a really like, that makes a big difference because if you're looking at like what AI model to power image generation, like I think a lot of interesting tools and use cases that we don't think about a lot. For example, one of them is Suno AI, which is a music generator. Every single time it creates a song for you, it also creates like an album cover for you. They're all the same kind of like style of Suno. But you think like you, you think of Suno as like this music generation company, but they're also an image generation company because they are making just as many images as they're making songs and that is millions and millions and millions, perhaps a day. And so if you can get you know, one of these models that can drastically, you know, make it six times cheaper or 10 times faster to create an image, then I think a company like Suno is very interested in that technology and running with it as well. So and Suno's like one example, there's probably like a hundred of these where they, they have to generate a ton of AI generated images and being able to make this a lot cheaper does make a huge impact on their company. One thing that I'll say is despite you being you being able to access this license, it's not licensed for commercial deploy. So this model is governed by the Flux non Commercial License V2O which essentially allows personal use research, internal evaluation, but doesn't let you use this for any revenue generated applications. Without a separate agreement. So you can make your separate agreement and then you'll be able to use this for inside of your organization. Essentially. I'm really excited to see if we see similar technology rolled out in other big players and we see a speed up in the overall image generation space. I would be a massive fan. Thank you so much for tuning into the podcast today. If you enjoyed the episode, make sure to leave a rating and review wherever you get your podcasts. As always, make sure to go check out Delve.com the sponsor of today's episode, and go check out AI Box AI, my own startup that lets you build AI tools without knowing how to code. Just describe what you want to build and we'll build it for you. Thanks so much and we'll catch you in the next episode.