More than 150m2 on the ground and 45 tons on the scale. A monster of cables, hard drives and graphics processors. Performance of 36.8 million billion operations per second. The Jean Zay supercomputer, named after the former Minister of National Education, has been highly prized by French researchers in artificial intelligence (AI) since its inauguration three years ago on the Saclay plateau, near Paris.
“It’s a tool that many countries don’t have, at the cutting edge of technology and very easy to access,” says François Yvon, researcher at Limsi, the Computer Science Laboratory for Mechanics and Life Sciences. CNRS engineer. We rent time slots, from a few days to several months, to train data sets. Applications are currently granted within 24 to 48 hours only, thanks to an accelerated procedure. “At the speed at which AI is progressing, we really cannot afford to wait six months”, slips the expert. Welcome to the world of high performance computing (HPC), the basis of many applications that are exploding in the eyes of the world today. The ChatGPT star in the lead.
In the shadow of this new giant, France has its say. At the height of his art, Jean Zay gave birth last summer to a great linguistic model (or LLM) called Bloom, with a power comparable to that of GPT-3, the brain of the famous virtual assistant. It doesn’t matter if OpenAI has just unveiled its fourth generation GPT-4, which is even more efficient. Bloom symbolizes French know-how in the field. Another way to approach AI, too. Open-source (whose code and sources can be consulted), trained on scientific articles and sports reports, Bloom is finally more multilingual than its competitor. The icing on the cake: 1,000 researchers from various backgrounds participated in its design, around the start-up HuggingFace. What confirm all the potential of this nugget, founded by three Frenchiesand which shines internationally.
It all started for her, like ChatGPT, with a rather special chatbot, since the interlocutor was represented in the form of a dog. “He was called Talking Dogsand it looked a bit like a Tamagotchi (Editor’s note: a Japanese virtual pet, created in the 90s)”, laughs Julien Chaumond, its co-founder. It was in 2019. An eternity on the scale spatiotemporal of AI. HuggingFace is today considered the “GitHub” of generative AI. A cap that could not be more serious. The start-up’s library, freely accessible on its website, opens access to more than 150,000 different models of machine learning – intended for translation, text explanation, information extraction from a video – as well as about 25,000 data sets such as text, images or videos. Two essential ingredients to bring an AI to life. There is literally nothing more to do than help yourself; more than 10,000 companies are already doing it, like Pfizer or Samsung. “Our goal remains that anyone can get their hands dirty and practice,” says Julien Chaumond. HuggingFace is valued at more than two billion dollars and monetizes in parallel computing power as well as its precious advice, powered by around forty sizes in artificial intelligence.
A developed ecosystem
It is therefore not uncommon for a young French AI shoot to rely, in one way or another, on Jean Zay or the HuggingFace website. “The open model of Stable Diffusion, published in their library, served as a launching pad for us”, confirms Matthieu Rouif, head of PhotoRoom. This French start-up walks on the flowerbeds of the famous Photoshop software, by cropping photos and generating, thanks to artificial intelligence, new backgrounds. The app is aimed primarily at merchants and SMEs who want to avoid an expensive photo shoot for their products. She raised $19 million in the fall and produces no less than three million images a day.
Laurent Daudet, from LightOn, trained his very first artificial intelligence model called “Pagnol” (like Marcel) on the supercomputer. The entrepreneur prides himself on being able to compete directly with OpenAI by training specialized generative AIs for business needs. Among them: summaries of videos, documents, the sorting of customer comments… “Many are looking for tools that are both efficient and personalized”, explains the entrepreneur, who also draws on the tricolor fiber: “companies can have data sensitive that they do not want to put in the hands of just anyone.” Some mainstream AI, like that of Bing (always inspired by OpenAI), have indeed already shown that they do not always know how to hold their tongue.
The demand is there. “This is the first subject of our customers at the moment: how can I use AI to increase the performance of my company?”, says Stéphane Roder, head of the consulting firm AI Builders, well established within of the CAC40. In line with LightOn, Dust intends to exploit the blind spots of models such as those of OpenAI to answer this question. “Large Language Models (LLMs), such as those that underpin ChatGPT, are very powerful, but they also offer sometimes surprising and even incorrect answers. However, for a number of years, engineers working with these models have known that we can use them to obtain more precise content. This is called orchestration”, explains Gabriel Hubert, one of Dust’s bosses, associated with a former OpenAI employee. According to New York Times, the American fund Sequoia Capital is leading the seed round. Dust, still in the testing phase of its product, could be valued between 30 and 40 million dollars.
France, already mature on generative AI? The findings come as no surprise to many. “Take a tour of the AI pioneers: Yann Le Cun (Meta), François Chollet (Google) or Luc Julia (ex-Apple, now Renault). The Gafams, by entrusting their teams or their flagship products to of the French, were the first to highlight our excellence in research”, recalls Matthieu Rouif, founder of PhotoRoom. Google and Meta have long established well-endowed research laboratories in France. The country’s appetite for mathematics – the true language of AI – probably has something to do with it. The Ecole des Mines, Centrale, Polytechnique, the MVA master’s degree at Paris-Saclay: all of these courses are reputed to be among the best in the world for learning mathematics and learning about artificial intelligence. “Students are running data learning algorithms quite early in their school curriculum”, observes Maxime Piatte, young leader of Veesual. Politically, the awareness also took place early. As early as 2018, the “AI for humanity” plan led by former majority deputy Cédric Villani had highlighted the need to support the sector, to provide it with a large-scale supercomputer firmly anchored in French soil. “Artificial intelligence often seems like a promise, but make no mistake, this revolution will not happen in 50 or 60 years: it is happening right now,” said the mathematician, winner of the Fields Medal, who became famous for her pretty ascots.
A glass ceiling?
The train is moving and France has other arguments to put forward. Imki uses generative AI to create technical drawings for designers. “We have the best art schools in our country, such as the Gobelins, or the National School of Fine Arts. The most important museums, I am thinking of the Louvre or the Decorative Arts. We have a large database of data in the art world. These are incredible assets for training AI”, says its founder, Frédéric Rose. Aqemia, on the other hand, takes advantage of French skills in the physical sciences and health. “We have developed a generative AI to invent drugs”, explains its creator, Maximilien Levesque, who recruits with a vengeance in French research centers at the forefront of the sector.
“This coupling of generative AI and theoretical physics allows us to find active molecules against targets thought to be unassailable”, he continues, like pancreatic cancer, a pest if it is not diagnosed very early. On the proposal of the machine, its researchers then chain the analyzes, then if the molecule is really promising, tests it on animals. The time saved, compared to a traditional drug research circuit, is enormous. Two years, with a ladle, estimates Maximilien, former CNRS and Normale Sup. Sanofi and Janssen are working with the start-up, which recently raised 30 million euros. “We have no other ambition than to become the world leaders in this field”, assures Maximilien Levesque.
It remains for this young ecosystem to consolidate its dikes. In tech, for the past twenty years, disenchantment generally occurs in three stages. The best talents fly away, as was the case in the older generation, towards the big companies of Big Tech, whose means remain disproportionate. Then, investments decline. “Jean Zay is no longer powerful enough for us”, regrets Laurent Daudet, who is turning to high-performance solutions on the cloud, the quest for computing power being the subject of fierce global competition. Finally, the nuggets, after a promising start, are bought up or go under foreign ownership. “Beyond 100 million euros, it becomes very complicated to obtain financing in France”, comments the former deputy Jean-Michel Mis, specialist in digital issues.
Even on a smaller scale, competition rages between banks and investment funds. Among the start-ups interviewed by L’Express, almost all of them, despite the delicate economic context, receive phone calls from across the Atlantic. A bill, carried by LR deputies, calls for the creation of a sovereign fund to support promising companies. “France is still disarmed on the side of capital tools. The 30 billion available to the Public Investment Bank (BPI) are on the scale for the development of start-ups, SMEs, at best ETI (…) It there is a ceiling to its ability to intervene”, note the parliamentarians. It would be a shame to miss the (very) high-speed train of generative AI. Especially after contributing so much to his incredible progress.