Founded by three French people, this platform sees itself as the toolbox of the artificial intelligence revolution. It already hosts thousands of models and the essential components of the most popular AIs.
“During the gold rush, it was not the gold diggers who got richer, but the pickaxe sellers.” Who will benefit from the AI rush that obsesses the tech industry today, from American giants like Google and Microsoft to endless startups? Among the best-placed pickaxe sellers, there are three Frenchmen.
Julien Chaumond, Clément Delangue and Thomas Wolf are today at the head of one of the essential cogs in this growing ecosystem: the Hugging Face platform. A name still little known to the general public, and which was even less so six years ago when the company had to go into exile in the United States to hope to raise funds. “AI seemed so far away at the time, it was impossible to raise funds in France by investing in such a long term”, explains Tech&Co Thomas Wolf.
To say that the situation has changed would be an understatement. Each week brings its share of new AIs with amazing abilities, from automatic text writing to image creation through the reproduction of voices more real than life. And Hugging Face has managed to place itself at the center of this revolution.
The AI toolbox
Hugging Face is a do-it-all AI platform that’s part toolkit, part library, part sandbox. On the site, anyone can design their own AI, train it with their own datasets or already existing ones, and publish it for everyone to use for free or comment on. Like a GitHub (which allows anyone to store lines of code and make them available to everyone), Hugging Face hosts nearly 150,000 models designed by its users.
Until recently, the platform mainly hosted programs specialized in language processing, such as Google’s Bert model. But 2022 has been there. Among the most popular today are “text-to-image” models like Stable Diffusion, or its predecessor Craiyon, whose graphic style made it the favorite meme generator on social networks.
“We had to add servers urgently to keep up with the demands,” laughs Thomas Wolf with Tech&Co.
But Hugging Face is not (only) an “AI App Store”. The platform also hosts certain parts essential for the construction of the main current models. The Transformers library, which allows language models like ChatGPT to understand a sentence in context; or Diffusers, which allow Stable Diffusion to generate images from a cluster of randomly colored pixels; these libraries are maintained by the Hugging Face teams, updated and available to everyone. All this in open source.
Reconciling AI and ethics
Because Hugging Face also cultivates an image of an ethical and open company, unlike tech giants like Google or OpenAI, which very often keep the functioning of their models or training data secret. “There is a lot of knowledge that is private and not shared around these models, we think it is not something healthy for the field in the long term”, believes Thomas Wolf.
“AI should be a kind of common good, not just the property of a few American Big Techs,” says Thomas Wolf for Tech&Co.
This open policy is reflected in the price of the platform: it is free for individual users, with paid accounts only for large companies that have greater needs in terms of computing power. In May 2022, the platform had around 1,000 of these paid accounts, for companies like Renault or Bloomberg. Enough to raise revenue while remaining open to the most creative researchers and users.
But the development of these AIs also fuels concerns. Aren’t there risks in making so many powerful models available to anyone? Hugging Face claims to be aware of these dangers, and constantly thinks about how to reconcile AI and ethics.
“We know that we manipulate powerful systems that can have negative repercussions on society,” insists Tech&Co Giada Pistilli, chief ethicist at Hugging Face.
Discretion around turnover
The company already employs big names in AI ethics research, like Margaret Mitchell, formerly at Google. Among its 160 employees, it has recently had a team specifically dedicated to these issues, responsible for designing ethical charters to guide its developers through complex moral questions.
“If I have a model who can create erotic art, how much do I allow her to go as far as pornographic? How much do I have to allow or block what the community asks for?” asks Giada Pistilli for Tech&Co.
These theoretical reflections must then be put into practice, and evolve over time and changes in the law. What Hugging Face applies to, while regularly publishing articles detailing their methodology for spreading good practices. A protective and empathetic image summed up by the name of the company and its logo, a hug emoji.
Are ethics and open source compatible with profitability? The company refuses to give its figures publicly. The cost of the computing resources required to process so many models remains very high, but Hugging Face seeks to reduce them, in particular through the privileged partnership with Amazon Web Services that it entered into in 2021.
Point evokes revenues of 15 million dollars in 2022, and optimistic prospects. “The machine learning community will be multiplied by 50 by 2030 because software developers will get started,” promises co-founder Julien Chaumond in the weekly. For pickaxe sellers, the future looks promising.
