At first glance, Stability AI’s Emad Mostaque doesn’t seem like the kind to go up against two of the biggest companies in the world — Microsoft and Google. But that’s exactly what he’s doing in the generative AI arms race.
OpenAI’s release of chatbot ChatGPT and image maker DALL-E have catalysed the biggest VC hype eruption in recent memory, capped off with Microsoft’s near-unprecedented proposed $10bn investment into the company.
But Mostaque — with all the swagger of a former hedge fund manager — doesn’t seem phased.
“$10bn may seem like a lot — it’s actually still just the start of this,” he says casually.
Stability AI is no slouch when it comes to fundraising either. In October the two-year-old London-based company raised $101m from elite investors like Lightspeed and Coatue to scale its generative AI technology, which has been made famous by its open source image generator Stable Diffusion.
And when asked about how he plans to differentiate his company’s generative AI offering from the giants of Microsoft and Google, Mostaque doesn’t pull punches.
OpenAI “fundamentally wrong”
His criticism of OpenAI centres around the Silicon Valley-based company’s more “gated” approach to its AI software. That means it charges for services like DALL-E, and blocks users from creating images based on certain words associated with hate, violence or pornography.
But it’s not just words being blocked. People in Ukraine also currently can’t use ChatGPT, along with people in China, Russia, Afghanistan, Belarus, Venezuela and Iran.
“I think that’s fundamentally wrong,” says Mostaque. “They banned all Ukrainians from using it. If they were the only image model provider, that would mean an entire country that’s being subjugated would not be able to be able to access these creative tools.”
In response, an OpenAI spokesperson told Sifted: “While we would like to make our technology available everywhere, conditions in certain countries make it difficult or impossible for us to do so in a way that is consistent with our mission. We are currently working to increase the number of locations where we can provide safe and beneficial access to our tools.”
“Everyone in India and Thailand and other countries will use this technology in the next 10 years, and they will most likely use our models”
StabilityAI, meanwhile, has come under criticism of its own for having too few controls, after some people allegedly used Stable Diffusion to create porn and deepfakes.
Mostaque doesn’t seem too phased — in an interview with TechCrunch he responded to the criticism saying “a percentage of people are simply unpleasant and weird, but that’s humanity”.
Instead he’s more concerned with getting the potential benefits of generative AI into as many hands as possible.
“Everyone in India and Thailand and other countries will use this technology in the next 10 years, and they will most likely use our models, as opposed to these models by private companies that are gated,” he tells Sifted.
Turning a profit
While some open source evangelists worry that StabilityAI taking on hefty VC investment could put greater commercial pressure on the company to get users to pay for its products, Mostaque is quick to reassure that his backers are behind his vision.
“I did the whole round in like six days and I made sure that they were fully supportive and we maintained our independence. So there’s no commercialisation pressure or anything like that,” he says.
He adds that the VCs on his cap table, like Lightspeed and Coatue, have extensive experience in backing open-source tech companies like machine learning platform Weights & Biases and analytics and monitoring solution Grafana.
Like many other open-source tech companies, Stability AI will keep its core services like Stable Diffusion free while charging users who want a more scalable and personalised service via an API — essentially allowing companies to integrate the tech into their own backend.
“That is a really good, tried and tested business model,” Mostaque argues.
He adds that Stability AI will be focusing on “media, education and gaming” as it continues to develop its generative AI tech, which will soon include a ChatGPT-like chatbot, as well as video and audio makers.
“In movies we’re already seeing people that we’re working with saving millions of dollars on production using our technology,” says Mostaque. “We are even working on fully AI-generated movies with certain well-known directors.”
Stability AI has “lots of partnerships to come” but has already announced a collaboration with Bollywood producer Eros.
“In movies we’re already seeing people that we’re working with saving millions of dollars on production using our technology”
“We have an exclusive licence for them to transform all of their Bollywood assets into intelligent, brand new assets via avatars, or creating whole new movies from that dataset,” he explains.
Mostaque says that when Stability AI does release its ChatGPT competitor, it will be optimised for education.
“It’ll be an intelligent AI assistant for education that’s open source and available to the world, starting in Africa,” he says.
But while Mostaque might get you to believe that generative AI is going to change the world for the better, not everyone is so excited. Last week a class-action lawsuit was launched against Stability AI and other AI startups for alleged misappropriation of copyrighted art in the AI’s datasets.
The plaintiffs say that Stable Diffusion has copied potentially billions of artworks without the artists’ consent, which allow the creation of “an essentially unlimited number of infringing images” that “will inflict permanent damage on the market for art and artists”.
Grassroots artists’ campaigns are beginning to pop up, with designers saying that clients are already ending work contracts, preferring to use AI-generated images than pay for human labour.
Mostaque seems calm about the lawsuit. “We haven’t received anything in the mail yet, so — you know — we’ll see how that goes.”
He adds that the legal question centres around the question of “fair use”, and that he believes generative AI is “transformative”, meaning that the technology changes the nature of the material, and therefore doesn’t infringe copyright.
The reality is that there’s little agreement on whether generative AI image makers infringe copyright, but an upcoming US Supreme Court case involving Prince and Andy Warhol will give a clearer idea of the courts’ legal interpretation of the area.
Mostaque says that, in the next version of Stable Diffusion, artists will be able to opt out of being in the dataset — currently Stable Diffusion’s dataset is based on a wide scrape of the internet that includes images from artists.
“It is something that I think is reasonable to do,” he argues. “We offer it, nobody else does that I know of. And we’re also super public about the datasets that we use, which again, nobody else is.”
This move won’t cover StabilityAI for the alleged misappropriation of copyrighted images that plaintiffs say has already happened though. That decision will be a matter for the courts.
Tim Smith is senior reporter at Sifted. He tweets from @timmpsmith