Dear Readers,

In this exclusive interview, Kim Isenberg, Editor-in-Chief of Superintelligence, sits down with Zeev Farbman, CEO and Co-founder of Lightricks and LTX Studio, to discuss the future of AI-generated video. They dive deep into Lightricks' strategic pivot, the decision to open-source their foundation models, and the broader implications of open versus closed AI ecosystems for creators and developers worldwide.

All the best,

Kim Isenberg

Kim Isenberg: Today we're talking about Lightricks and LTX. Many people know Lightricks from Facetune, but the company is now pushing hard into AI video and multimodal models. A few weeks ago, Lightricks open-sourced LTX-Video (LTX-2), an audio-video foundation model with open weights and training code. The company says it is optimized for Nvidia GPUs and can generate synchronized audio-video with a focus on local deployment, control, and privacy.

Today, Reuters reported that Lightricks is splitting its consumer app business from its AI video platform, LTX, as investor interest in AI growth increases. This is a great moment to talk not only about the model but also about the bigger strategy behind it.

Today I'm talking with CEO and Co-founder Zeev Farbman.

Kim Isenberg: Zeev, it's a pleasure to have you here.

Zeev Farbman: It's a pleasure being here, Kim. Thank you for having me.

Kim Isenberg: Please take us back to the beginning. What was the spark or insight that led you to start this company?

Zeev Farbman: Imagine the year was 2012, and the word "selfie" just became part of the Oxford Dictionary. My co-founders and I were finishing our PhDs at Hebrew University, and most of our research was somewhere on the border between computer graphics and image processing. I spent some time collaborating with Adobe's research departments, and I realized that at least back then, Adobe didn't really see mobile as a nascent platform for creative tools. The assumption was that mobile is great for social media or gaming, but less so for creative tools.

For us, it felt like a great opportunity to try and build cool creative tools on mobile. We actually thought about Adobe as a role model for what we were trying to create. Initially, when people asked what we were doing, we said, "Listen, we are trying to become the Adobe of mobile." We started with Facetune, which was supposed to be a mobile Photoshop that was really easy to use. It was supposed to be a small first project where we were learning how to work with iOS and figure out how to leverage mobile GPUs. It turned out to be a way bigger opportunity than we imagined.

Kim Isenberg: The switch to AI happened in 2022. You started with Facetune, and now you are building open, production-grade AI models like LTX-2. Could you tell us what is fundamentally broken about today's creative AI tools or workflows, and why did you feel a new foundation was needed?

Zeev Farbman: That is a great question. Our push into AI comes from seeing a certain failure in the business model that plugged the world of AI at the moment. In 2022, we saw diffusion models like Stable Diffusion 1.5, DALL-E 2, and LLMs like GPT-3. It was clear we were going to experience a paradigm shift in both how we create pixels and how we interact with machines at large.

Until then, we really focused on creative tools for non-professionals. But at that point in time, it felt like this would be a completely new generation of tools, and we could start building professional tools from the ground up for the world of AI. Initially, we thought about ourselves as a product company. We naively thought the world of AI was going to be open, and we would be able to license the technology and build products on top of it. That's how we started. We wanted to build a pre-production tool for professionals, which became LTX Studio over time.

Initially, we partnered with Stability AI. We licensed their models in order to fine-tune them for our needs. The video model Stability AI offered back then was SVD (Stable Video Diffusion). At some point, OpenAI started to tease Sora. They didn't release it, but they released this amazing marketing material that was mind-blowing. What we realized is that another step function in the quality of the tech had arrived. OpenAI with Sora moved from a UNet architecture—which we used back then for diffusion models like SVD—to basically using a Transformer for video creation.

AI stocks. Metals. Crypto.

Surprise, surprise; gold crashed 16%. Silver plunged 34%. Bitcoin dropped to 1 year lows.

All supposedly "uncorrelated" assets moving in lockstep largely because of overleveraged margin.

JPM strategists warn that the same leverage is still a risk.

Those markets may be recovering now, but cascading liquidations could trigger quickly across several asset classes simultaneously.

So much for diversifying away risk, right?

But get this–

70,819 everyday investors have allocated $1.3 billion fractionally across 500+ exclusive investments. 

Not real estate or PE… Blue-chip art. Sounds crazy, right?

Now it’s easy to invest in art featuring legends like Banksy, Basquiat, and Picasso, thanks to Masterworks.

They do the heavy lifting from acquisition to sale, so you can diversify with the strategy typically limited to the ultra-wealthy.

(Past sales delivered net returns like 14.6%, 17.6%, and 17.8% on works held longer than a year.)*

Our subscribers skip the waitlist

_{*Investing involves risk.  Past performance is not indicative of future returns. Important Reg A disclosures: masterworks.com/cd}

Once the picture became clear and everyone realized Transformers are great not only for LLMs but for diffusion models as well, I flew to London to talk to Emad, the CEO of Stability AI.

We had a really good conversation, and he outlined his plans for what Stability was going to do. But I think a week after I landed back, Emad stepped down or was fired, and we started to see key people leaving the company—like the guys who formed Black Forest Labs.

We realized these guys were experiencing a setback, but the world doesn't stop. So we were building an entire product layer on very shaky ground. We started talking to closed model providers to see if it was possible to do licensing. That's when we realized the stark business realities of foundation models. None of these guys wanted to give us the weights of the model, and the API pricing just didn't make sense for the kind of products we wanted to build. We also wanted to deploy these models on mobile for our consumer base in our apps like Facetune, Videoleap, and Photoleap.

We realized there was literally a gap in the market. The vast majority of Western companies want to serve their models through API. I completely understand why—it's an extremely lucrative business model. You get access to all the data, you don't expose the weights, you can constantly raise the price, and you control the whole thing. But it is very problematic for a certain category of builders who want to fine-tune the model for their needs, who want control over IP and their specific data, and who need predictability. With API pricing, you're always on shaky ground.

My big suspicion is that some of the biggest and most powerful companies in the West are trying to create the most lucrative business model the world ever saw. They hold access to this "philosopher's stone" of technology, the Promethean fire. You can build everything on top, and they will be very nice to you initially when you're trying to build—the price will be subsidized. But it's not going to be like that forever, because at some point they have to justify their valuation.

We saw these things play out in the world of tech before. In the 90s, Microsoft leveraged its monopolistic position in operating systems. But then other companies around open technologies, like Red Hat around Linux, were able to build a viable and successful business. Maybe not as successful as Microsoft, but certainly significant.

So what we decided to do basically is almost initially solve our own problem. We wanted to create models that are open, efficient, and IP-safe for the relevant players. That's how we got into the world of models.

Kim Isenberg: Let's talk about the recent announcement. Reuters reported that Lightricks is splitting its consumer app business from its AI video platform LTX. Is this primarily an operational decision or ultimately a capital market valuation decision?

Zeev Farbman: It is both. Imagine that at the moment, we are this combined entity that has a very profitable existing mobile business and a very unprofitable "moonshot." It's something that is very hard for investors to wrap their heads around. People who invest in existing businesses look for a certain type of investment strategy and expectations of return over a certain period of time. When we're talking about AI moonshots, we're talking about a completely different crowd. It's kind of hard to attract people who are simultaneously interested in both pieces. The decision ultimately allows both the profitable business and the moonshot to be judged on their own merits, which makes sense for investors on both sides of the aisle.

Kim Isenberg: Our goal at Superintelligence is to make AI understandable to everyone. Could you walk us through what LTX-2 actually does and why it matters?

Zeev Farbman: I'm going to do a little bit of hand-waving here to try and tell a grand narrative. When people think about the world of AI at the moment, the vast majority think about models based on Transformer architectures.

We can say we have two big buckets. We have LLMs (Large Language Models), where the Transformer basically takes words (tokens) and outputs the next token. The other bucket is models creating images, videos, and sounds. Here, the output token doesn't represent a word, but a patch of pixels or audio. When we're talking about a multimodal Transformer, it's something that holds all these modalities together. It has tokens representing pixels over time, audio, and text, and these models know how to work with all these tokens together.

We can think about any interesting frontier model at the moment falling into one of these two buckets. There are a ton of experiments with different architectures. For example, in the world of LLMs, some people are trying to take ideas from diffusion models and figure out how we can diffuse text. In the world of diffusion models, people are trying to figure out how, instead of generating an entire batch of frames, we can work auto-regressively, like LLMs—generating a frame, and then another frame, and conditioning the generation on previous frames.

But if we are looking at the world of AI at large, this "shadow" of the Transformer covers most of it. We are seeing two parallel trends that almost run in opposite directions. On one hand, we are constantly trying to increase the capabilities of the models—more parameters, more compute, more data. But there are clearly still some diminishing returns, and some problems are just really hard. For example, the context window of LLMs. We are somewhere around hundreds of thousands, maybe a million tokens. Under the hood, it's a quadratic problem, and we don't have tools to easily extend it without a ton of approximation techniques.

The other trend we are seeing is that we're trying to optimize things—trying to deploy things on edge, reduce latency, reduce costs. With LTX, one of the big technical bets we are pushing is trying to create an extremely compressive latent space. When you're creating this very compressed latent space, you need fewer tokens when you're training the model. It reduces your training costs and allows faster experimentation. And on the inference side, obviously, the whole thing runs faster.

The big challenge when you're creating this extremely compressive latent space is that the reconstruction can be pretty cool, but you're losing "diffusibility"—it's hard for the diffusion model to adapt to it. That took a while to figure out. But once you do, it allows you to basically reduce the cost of training and run it on consumer-grade GPUs, like a 3090 graphical card that is more than five years old.

Kim Isenberg: You are very successful with open-sourcing LTX. What have been the most important milestones regarding developer adoption?

Zeev Farbman: We are seeing this gap in the market. That gap was, to some degree, served by some Chinese companies for a while. Notably Alibaba's Qwen model and Tencent Hunyuan. For a while, they kind of filled this gap. But then what happened is that for some reason, Alibaba and Tencent's teams decided not to release the newer versions. Tencent is lagging behind a little bit. So I think builders like us are facing a situation where they literally have no foundational tech to build upon. It's really annoying.

We want to gain the trust of developers. We want to show them that we are committed to building an ecosystem around it, and it's very easy to use the model across different layers of the software stack. For example, we released an integration as part of the MCP (Model Context Protocol), showing how you can integrate it as a rendering engine inside existing architectures like Blender. So you basically have a plugin, you're working on your project, and then you're using LTX as a renderer. We want to show that you can build an entire non-linear editor that runs on top of it.

If we look at computer graphics, we saw a rapid advancement in real-time computer graphics since the early 90s, when John Carmack programmed Wolfenstein and DOOM, up until the early 2000s with programmable shaders. Since then, if you look at a game like Crysis that came out in 2007, the last twenty years feel like a really incremental improvement.

With these new pieces, we will be able to create a completely lifelike environment. When you connect it to technologies like Augmented Reality and Virtual Reality, you can start imagining a metaverse, but for real. You'll have these persistent, multi-participant environments where big-brain LLMs are orchestrating the whole thing, and diffusion models run locally as rendering engines.

There is another huge application people are focused on: using these multimodal models as world models. They try to create a physical simulation using them. Usually, they take data from a specific domain—like a warehouse, recording what's going on with robots or drones—and use this data to fine-tune the model so it excels at predicting, given state T, what the system is going to look like at state T+1. This is a very cool thing because it allows you to basically reduce the cost of training robots in a software simulation that approximates the real world.

Kim Isenberg: This all sounds like a huge success story, but were there moments where you were totally wrong about something?

Zeev Farbman: Listen, to tell you the truth, the jury is still out on how much place there is for open models in the West. If you look at the Chinese ecosystem, it looks very, very different from what's going on in the West. There is a bigger understanding of how AI can be infrastructure and how things should be built there. In the West, it's still an open question. We need to prove that there is actually a viable business model here.

I think a lot of investors are going to say, "Okay guys, it's really cool that you are able to create these models at a fraction of the cost, but who is going to pay for it?" And then I'm going to tell them the story, and some of them will believe it and some of them won't. So the jury is still out. I don't feel like we are out of the woods and we're [not yet] able to prove that the alternative exists.

The conversation around AI to me is really funny. For a while now, I'm tracking what Yann LeCun is saying, right? And for a researcher, it completely makes sense. There is nothing too controversial there. He's pointing at certain things—he has his own thesis—but no one talks like that. Everyone wants to have this singularity in two years. I don't know what to do with it. If you know how to achieve singularity, achieve singularity. If you don't, maybe go back to doing your research.

I call bullshit on most of it. You can't predict the exact shape of the technological landscape until you know certain answers that you don't have at the moment. Some people are very adamant and determined to show they have an exact view of where the value is going to be, and they're already building data centers in space. To me, it's just really funny. I think people should realize that assigning different parameters to all these questions is going to determine a business value. If you pretend you know the answer to technological questions we don't have the answer to yet, it's going to leave you naked one night, you know.

Kim Isenberg: Thank you so much for attending in this interview Zeev. See you around!

Why it matters: The open vs. closed AI model debate isn't just philosophical — it has direct consequences for every developer and company building on AI infrastructure today. If Lightricks can prove a viable business model around open video foundation models, it creates a meaningful alternative to API dependency and could shift how the entire creative AI ecosystem is built.

If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.

This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.

Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.