When Prem Akkaraju took over as Stability AI’s chief executive in June 2024, he inherited a business in turmoil. The UK-based company behind Stable Diffusion, the popular open source image-generating AI model, was in debt, had just lost its co-founder and CEO Emad Mostaque, and was fighting a series of lawsuits over whether it breached copyright law by scraping the internet for images.
On his appointment Akkaraju joined a group of prominent investors, including Coatue Management, Lightspeed Venture Partners, Napster founder and former president of Facebook Sean Parker and former Google CEO Eric Schmid, to inject fresh funds into the company. He came from the film industry, where he led Wētā FX, an Oscar-winning visual effects studio behind films such as Avatar and the Lord of the Rings. He also has an executive producer credit for the 2021 Oscar-nominated film The White Tiger.
In this conversation with the Financial Times’ AI correspondent Melissa Heikkilä, he explains how his company approaches data and copyright, and how he sees the AI sector working together with the creative industries and artists instead of against them.
Melissa Heikkilä: You took the helm of Stability AI last year during a very tumultuous time for the company. Stability AI’s co-founder and CEO Emad Mostaque had just left, lots of other high-profile people had left, they had a long litany of lawsuits and dwindling cash reserves. What made you want to take the job?
Prem Akkaraju: For the exact reason that it’s all coming to fruition now too, because of the technology and the team. Nobody has the monopoly on AI research or machine learning. There’s dozens and dozens of really impressive, important companies out there, if not hundreds. [Sean Parker and I] were pretty convinced [despite] the departures that happened, that there was still incredible talent.
And before we even did the deal, we went out and met every single one of our researchers and applied researchers. We met over 80 people. And not only were we convinced that these were some of the most talented researchers in the world, they had two other things that you can’t see on a resume, which is extreme passion and extreme loyalty.
MH: You come from the movie world, and now you’re in the AI world. What has surprised you the most about this transition?
PA: I’ve always been at that point of the merger of art and science, so it’s really like bringing technologies in to have artists be able to realise their storytelling and their dream. And . . . that’s the difference between Stability and other companies . . . we’re truly putting the artist at the centre, and then building the technology around him or her to tell their story. And instead of a CGI, now it’s a generative AI and GPU-based.
MH: Interesting that you said that you want to put the artist at the centre, because I’m sure a lot of people in the artist community wouldn’t say that. In fact, they’re extremely angry and upset at Stability and other AI companies. What’s your message to them?
PA: What’s interesting is that clearly not all of them are angry and upset. I think we have one artist, named James Cameron, who’s an investor and board member of the company, who’s been very tech forward. There are ones who have touched technology, who have been able to really fulfil their visual dreams and create these visually arresting moments and scenes and shots in film and television, who really have embraced technology.
But change is scary. Change is hard. Every time there’s been this type of sea change in the entertainment business, it’s always been met with great apprehension and fear. Going back to 1900, when the movies went from silent to sound, these talkies that nobody really wanted, they were meant for Broadway. In fact, the industry thought it would be destroyed if you added sound to movies. It’s such a crazy thought.
Same when it went from black and white to colour, they felt like it would lose its cinematic quality, and these movie stars wouldn’t look like movie stars anymore if they looked too real in colour. And, of course, again in the early 2000s, when it was a shift from shooting on film to digital, it was also the same revolt. You’re going to lose the depth, you’re going to lose the warmth of film. There’s a transition, and then when the transition happens, everybody uses it. When they finally break that feeling and understand that it’s actually an empowerment, not something that is actually a deterrent, it moves pretty quickly.
MH: And what do you say to the criticism in the copyright debate that this technological development is basically possible because of theft?
PA: I think that if you speak really honestly with artists, they’ve been inspired their entire life by other work, whether that be art, whether it be photography, whether it be other films, whether it be other artists, other directors. In a way, it’s very much like the culmination of those types of things is really happening again.
Now, we do put the artist in the centre, therefore, we’re going to abide by, and we do abide by, all the existing laws and we continuously will be doing so. And we think actually it’s one of the greatest opportunities for artists, for many reasons.
One is they can monetise. If you think about it, it’s all about input and output. On the input side, we’re using free-to-use data. We have some bespoke license deals as well. I think that’s a great opportunity for artists to monetise their wares.
But then also, think about the different empowerments that they’re going to have on the output. Be able to very quickly iterate on their vision and be able to monetise that even further. I think it’s going to be ultimately an empowering tool for artists.
I’ll cite one conversation I had with an artist [who] said to me; ‘Listen, I don’t want AI to make more art for me. I want AI to do my dishes and clean my house, so I can then work on the art.’ This is a very important statement. What I said back was, there’s a lot of dishes and house cleaning in making the art. It’s not just the creative process. It’s really this gigantic workflow of non-creative work that goes into the ultimate product. AI is really good for that part.
Many, many people talk to me and ask me what are all the things that are going to change in filmmaking with AI? Well, I spend just as much time thinking about what’s going to change as what’s not going to change. And that artistry is deeply human. I don’t think it’s going to change that there’s going to be a director and a camera and an actor in front of the camera. That physicality of film production is super important to the creative process.
What’s not important to the creative process is taking three months to relive a scene or doing paint and rotoscope or camera match, or these very, very laborious, non-creative workflows. Think about AI in the film and TV business as aiming towards non-creative workflows so creative people can be more creative.
And those people who are doing those workflows aren’t doing it because they really love it. They’re doing it because they really want to be creative, they want to be storytellers, they want to be producers, or actors, or writers or directors. And this technology will allow them to move up the chain, if you will.
MH: Do you think there’s a case for a Spotify model for AI training data, so that artists could get compensation? What are your thoughts on that?
PA: I think that’d be a really great idea. I think that a marketplace for people to opt into and then upload their art, I think that’s going to happen. Actually, something we’re working on, where artists can actually have a marketplace or a portal where they can say, ‘hey, you could train on this,’ and then that actually gets licensed and used by us and others, and they get compensated for it. I think it’s really smart.
MH: There’s still a lot of bad blood in the artist community over copyright. Are you open to being closer to them or having a conversation or engaging with them?
PA: That’s an industry-wide conversation. This happened, we’ve seen this movie before. It’s called the music business. It happened with streaming music versus ownership. It happened with sampling. Now, that actually has been sorted. I think we’re going to get there a lot faster with the artist community and AI and its use, but that’s a gigantic, industry-wide debate.
MH: The artist community has also started fighting back by deploying tools such as Glaze and Nightshade, which mask images from AI scraping and add poison to datasets that make models break, respectively. How much have they impacted your models and your work?
PA: I don’t think that they’ve had that profound of an impact on what we do. We have a lot of different workflows and customised workflows on top of our model that provide a lot of the functionality that people are really looking for, and we’re adding to them every single day.
When you think about Nightshade and you think about these other [tools for] protecting data, that’s part of what we’re working on. You mentioned Spotify for images. [Another] great solution would be a Shazam for images. You can have this as a fingerprinting type of technology. And I think that’s going to take a lot of opt in from both sides. And I think that’s technically where we’re headed with a potential solution.
The question is, then, what’s the economics involved? Because we’re not duplicating or replicating everything. Like in the music business, you’re actually using the actual sample. Here, you’re not. The AI is essentially inspired by billions of images at one time, and definitely not duplicating or replicating anything. In fact, it has to be novel by definition. What the economics are going to be is one thing, but should they be compensated? Of course.
MH: AI training data is a very contentious issue. A few years ago, researchers at Stanford found child sexual abuse material in the LAION dataset, that you have used to train models on. And you’re facing lots of lawsuits over the copyright of material used in AI training data. Have all these cases made you rethink the way you use training data?
PA: No, not really. What we’re using is free-to-use data, as well as some bespoke license deals. I think that the way we’re doing it is the right way. And we strive every single day to have a clean and sanitised training dataset.
MH: Generative AI images come with massive risks, such as non-consensual deepfake pornography, child sexual abuse material or misinformation. How are you thinking about these risks in your models?
PA: That’s why we have proactive features to prevent those types of misuse. When you’re actually entering those types of prompts on our API, we’re immediately flagged, and those people are actually banned. That’s a very proactive safety workflow on top of our model that we do.
MH: I want to go back to your film background. You talked a little bit about how the moviemaking process or the creative process will change. You were an executive producer on the film, The White Tiger. That came out in 2021, and obviously, that was before this generative AI boom. But if you had to do the film again now, how would you have done it with AI?
PA: That’s a great question. No one’s ever asked me. That movie . . . [is] an adult drama, which means that a lot of it is in camera. There’s not a tremendous amount of visual effects, for instance, in that movie. We could have explored India a lot more, because we were very limited in the budget, so we had to shoot a lot of the exteriors there. And I think that you would have been able to do a lot of really cool exteriors.
I think you could have had more optionality on lighting and textures and things like that. But it’s important to say that the acting and the heart of that movie was all in camera, which I think a lot of movies are going to still preserve. That actor in front of the camera and the director pulling that performance out from human to human, I don’t see AI really replicating that. I don’t think AI has a role in that. I think AI has a role in a lot of the other workflows, to the left and right of it.
Editing and colour grading and all those types of really nerdy, boring workflows, is where we would have been able to use that.
It would’ve been faster and cheaper to iterate more. I think what also people overlook is the value of failing quickly, using AI in filmmaking and creative process. Because for that one shot, you could be labouring over hundreds and hundreds of different takes. Doing that without AI takes so much time and money. I think that failing quickly — and getting to what the vision of the director is quicker — is a huge benefit that I don’t think AI gets a lot of credit for.
MH: And where do you see this technology going next?
PA: We’re looking to upscale the technology into a professional, enterprise-grade format [for] real artists that are making high-end feature films, television series, as well as gaming. WPP invested in our company, so we are using it for marketing and advertising pipeline. We’re working very closely with them on bringing AI into that process. Not only just making the actual advert, but also being able to customise those adverts into different demographics in real time.
In advertising, one of our direct clients is MercadoLibre, which is an online retail site in Latin America. We were able to increase their impressions in click-throughs dramatically, double-digit increases, just by using some AI imagery in their product images, and to really bring to life the product in a different way.
MH: When Stability first started, the company really highlighted openness and open source. Is that still the direction you want to go, or has that changed?
PA: We just released another open-source model. We’re going to continue to release open-source models. The community is very important to us. We will release open-source models, if you’re using it for non-commercial work. If you’re using it for commercial work, like big companies who are paying us now, then they do a licence agreement with us, where they pay for the API.
MH: If you had to do a post mortem of your time at Stability, what would you say? How has it gone? And where do you want to take the company next?
PA: If you look at the press, you’re absolutely right, there was a lot of negativity. But that’s what the definition of an opportunity is, it’s me seeing something that you don’t. And we saw that within Stability. And because we knew the core team that was existing there, plus the technology and our developer community was enormously valuable, everything I’ve done in the last nine months to clean things up, we have a totally clean balance sheet, everything is spick and span.
Now the base is clean, we have zero debt. We have amazing investors. We’re well funded . . . Our revenues are spiking. We’ve got amazing blue-chip, notable clients all around the world.
We did a deal with [semiconductor designer] Arm. We’re the only AI company in the world that actually was able to generate rich media audio content on the mobile phone on a CPU, so on an Arm 9 chip. I feel very proud. I think you’re going to see that perhaps nine months ago [we] wouldn’t have been able to be partnered with a $150bn technology company like that, now it is possible.
MH: What excites you right now?
PA: I’ll be honest, every single model we make, I have that feeling. I feel like there’s a breakthrough, even just with our own models, every about two to three weeks or four weeks. The rapid change in our output, whether that be higher resolution or we actually find a problem that we’re really trying to fix. And we’re finding problems that maybe sound really nerdy again and boring to you, but they’re really exciting to me, like rig removal.
We have a credible model now that, let’s say you and I are being filmed right now in a shot, there’s a boom mic that’s sticking out that we didn’t notice. We have a model now that can actually see that shouldn’t belong in the composition of the shot and paint itself out automatically.
That would normally take a month or two months, even if someone then detected it. Now, we have ways of actually doing that on its own. I love stuff like that.
This conversation has been edited for length and clarity.