Moonvalley was built for change: to change the way generative AI is created; to change the way media is made; to change how the world thinks about AI and art. The company’s generative AI model, Marey, is the first state-of-the-art video generation model to be trained on 100 percent licensed content—something nobody was sure would be possible. Marey is an astounding achievement, but it’s just the first manifestation of Moonvalley’s pioneering mindset and purpose. Here, co-founders Naeem, Mateusz, John, and Mik discuss Moonvalley’s groundbreaking technology, its innovative company structure, and the impact of investing in transformation.  

What does it mean to say that a model is “clean”? Is this what people mean when they talk about ethical AI?

Naeem: From our perspective, the spirit of a clean model is simply that you can’t create technology that produces art by ripping off other people’s art. We collected all of the visual data we used to train Marey directly from the original creators, with their permission. We licensed every piece of content, and of course, we compensated them. 

I’ve never liked the framing of “ethical models,” because I think it’s up to the media and entertainment industry, and the professionals creating the work, to decide what ethical means for them. It’s our job to build what the community wants. There’s still an open question in terms of what level of “clean” is acceptable. You could go to an absolute extreme: Has every single bit of machine learning that’s ever been used in any part of the pipeline been trained on indemnified data? That’s a very deep rabbit hole. But the industry has been loud and clear about the importance of protecting creators’ rights, so that’s been our North Star. 

Why is it so challenging to build a model on licensed content?

Naeem: The biggest challenge is that you just have way less data to work with than companies that scrape the internet to train models—a smaller volume of data, and also less diversity in that data. To get a model to the same quality as one trained on more data, you need to build an architecture that’s substantially better.

Mateusz: On the other hand, the quality of licensed data is often higher than the quality of  random YouTube clips. And if you have high-quality data, you need less of it to train high-performance models. Ultimately, that means that once this technology becomes more accessible, it might actually end up being more cost effective. 

Why do you think more companies haven’t tried to crack the problem of building a clean model?

Mik: A company has to be willing to make a big investment and take on a lot of risk. The cost of training a big model is already very high, and acquiring rights to training data is yet another financial factor on top of that. And because nobody had done it before, we couldn’t be certain it would work. It’s not like we could train one clean and one non-clean model and see how they compare. Training a big model is a one-time thing—a YOLO thing, as we like to say at Moonvalley.

What are the ramifications of proving that it’s possible to train a model on fully licensed content?

Mateusz: We’re setting a new precedent in this space, which could change a lot in terms of building the legal framework around generative AI. Until now, everyone could say that training models on scraped data is “fair use” of copyrighted content, because this technology is very important and the only way to build it is to scrape the internet. But we’re showing that there’s actually another way. This could impact the legal perspective on LLMs as well as visual models.

Naeem: It won’t necessarily make it easier for others to achieve, but it will start to happen more. It’s like the first time someone ran a mile in less than four minutes—everyone thought it couldn’t be done, and then all of a sudden a bunch of people were doing it. When you set the precedent, other people will raise the bar. And industry professionals have something concrete they can advocate for as an alternative to the status quo. 

Moonvalley is building products for everyday use, but you’ve built an incredibly pedigreed team of AI researchers. Is your long-term focus more about research or products?

Naeem: We are a research company at the heart of everything, with a primary focus on innovating foundational AI models. But that doesn’t make us any less of a product company. We believe the most compelling products of the future will come from AI-native companies in which the research org and the traditional product orgs aren’t necessarily delineated. Good research inherently results in good products, and the best research isn’t done just for the sake of research, but for the sake of expanding human potential.

Mateusz: Research is necessary for anyone in this space because there are still fundamental problems to be solved, especially around getting models to genuinely understand the physics of motion in the real world, and giving users fine-grained control over the outputs. My view is that the key to solving these problems is seeing how people interact with the technology, and that’s where the product comes in. In the end, the ultimate measure of any research output is its capacity to transform society by advancing human knowledge and increasing productivity.

John: Another important point to add is that we tend to talk about research as the deeply technical stuff, but if you look at Asteria and the groundbreaking things they do with film production, that’s very much R&D. Our DNA, across the board, is about innovation applied in real-world contexts.

Mik: For many researchers, it’s really great to see their work impact the world, which is fairly uncommon. My career started with me being a math geek, studying pure maths. At some point, I realized that doing research for the sake of research, as Naeem said, doesn’t bring that much joy, at least not to me. I think what brings many researchers to Moonvalley is the fact that this research can actually change people’s lives. And that is really, really satisfying.

There’s obviously plenty of controversy around the use of AI in entertainment. What do you say to people who think generative AI supplants human creativity?

Mateusz: I see these tools as a way to augment our intelligence. If you’re using a notebook, it augments your memory. How far could you go if you couldn’t write down your thoughts in that notebook? Probably not far. We want generative AI to be a tool that augments our creative intelligence in the same way—it makes us able to do more.

Naeem: The biggest misconception is this idea that you’re just going to type a few words on a screen and generate a movie, and that’s it. I think that’s an incredibly reductionist approach to craft. In reality, what you see when you work with filmmakers is that the overall process doesn’t change when you introduce generative AI. You’re doing the same things, it’s just that parts of it become easier. And some parts become more complicated because you have more capabilities. But you still have actors, you still have writers. In our software, the first thing you do is sketch a storyboard. Understanding this as “hybrid filmmaking” is what’s really compelling, and that’s what engages filmmakers.

What will people be able to do with Moonvalley’s tech that they couldn’t do without it?

Mateusz: Weirdly enough, I sometimes underestimate the power of technology because, as a scientist, I see all the imperfections. AI-generated video can still have strange artifacts: a third leg appears; there are more hands than there should be; there’s an extra head. What is really striking to me is that artists take advantage of those imperfections to express themselves in new ways. I have seen—I don’t know what to call them, maybe “special effects”—that I have never seen before. A filmmaker might use that extra head to make an effect where two heads morph into one, and they do it in such a skillful way that you don’t see it as an artifact. It becomes part of the creation. I believe that this technology will be used by people in unexpected ways that could fundamentally transform the field. We can’t even imagine today how people will use this technology in the future.