Hugging Face: in the French workshop of artificial intelligence
Lines and lines of code... This is what Artificial Intelligence is made of. © Hugging face
Text by: Thomas Bourdeau Follow
The results offered by artificial intelligence (AI) are impressive and many anticipate a disruption of the current internet. Hugging Face is a company that was created by three French people and competes in commitment to open-source versions of models that use AI. It is also a way to shed light on the functioning of this formidable tool that AI has become. Visit to their Paris offices to try to understand how artificial intelligence is done.
Since the arrival of ChatGPT, artificial intelligence has never seemed so close to dialogue and why not to hand. While the company OpenAI (behind ChatGPT and the recent exhibition of Artificial Intelligence) moved away in 2019 from open source, it is a different turn that was taken by the company to the French founders Hugging Face in 2016. At the latter, the source code can be reviewed and improved by everyone (it is said to be open source).
In the company's Paris offices, Lucille Saulnier, research engineer, and Giada Pistilli, senior ethicist, explain: "We are moving forward with our community (15,000 companies using their service and 200,000 daily users), together, we have constant feedback on the results of machine learning. . For other models (ChatGPT or Midjourney for example), it's always annoying not knowing what's behind it.»
But what makes artificial intelligence work? The question deserves to be asked because the term AI - artificial intelligence - is too often a portmanteau word without really designating what is achieved. In the excellent online course Elements of AI, it is specified in the introduction that "AI" is not an accounting name. "AI is a scientific discipline, like mathematics or biology. AI is therefore a collection of concepts, problems and methods to solve them. It is thus more accurate to say AI than AI.»
A DIY workshop in the new dimension offered by AI
Lucille Saulnier explains: "There was the publication in 2017 of a very important paper: Attention is all you need: understanding natural language processing with Transformers models." She smiles: "It's very, very technical, but you could say that he changed everything.»
A transform model is a deep learning model. The best known in AI, those we have been most often confronted with, are developing around Natural Language Processing (NLP) that is found in ChatGPT, Siri or in a spell checker. Lucille Saulnier: "For us, the idea is to keep the result of the research and that it serves others. These are very heavy files and this justifies the hub, the platform, free and open source.»
Giada Pistilli speaks of a bookstore and the image takes on all its suggestive power. Thus, on the site is made available a database from which users can draw solutions specific to AI, but also understand how it works. (spoiler: these are lines and lines and lines and lines of code...) "On this platform, we discover demonstrations of AI capabilities, in the spaces section" We find the now famous text to image (à la midJourney) but also image to text, demos like ChatGPT ... These variations seem endless like a DIY workshop that unfolds in the new dimension offered by AI to solve all types of problems, perform tasks or be creative. We can mention: DeepL, a powerful translation tool, but also Grammarly an impressive writing assistant...
paper pages. now on @huggingface pic.twitter.com/laUkU9mriX
— dylan (@dylan_ebert_) March 21, 2023
All this goes very fast, they comment in chorus. We are like at the beginning of the internet, with a desire for optimization. It is a noble mission for a vast and powerful field. Chess remains a privileged sport to demonstrate the prowess of AI, Giadia emphasizes with a smile: "We want to open the chessboard, to allow the wider world to participate. " "Our business partner is AWS cloud (Amazon's cloud service) for our computing resources," says Lucile.
A team from Hugging Face takes care of the platform. Giada, lead ethicist, explains: "Artificial intelligence is at the intersection of politics/law/ethics. Internationally, everything can be different, especially in terms of copyrights. Will there be a copyright on the machine's work? Or does the right apply to what is offered to the machine (text or image)? It remains unclear.
Running the AI in the right direction, at least in that of its request
Because if the "how it turns" is in the huge files stuffed with lines of code that can be consulted in the library, to obtain the right result, to run the AI well, it is still necessary to know how to address the machine correctly. While some still seem to have trouble formulating their queries for the Google search engine, we must be prepared to evolve our syntax to talk to an AI model, ask the right questions, start the right dialogue, otherwise nothing significant will come out. This is called prompt engineering, and it is becoming a new profession. This function can be translated as query engineer or prompt engineer, and it refers to those who tame artificial intelligences to get them to produce the desired result.
The actor Louis Jouvet explained that there is no good show in the theater without a good audience. This totally applies to conversations with models using AI. Riley Goodside, a quick American engineer, explains that during his dialogues with the machine: "He explores the multiverse of fictional possibilities, sculpts the space of these possibilities and eliminates everything except the text he wants," and adds: "It can be a very difficult mental exercise." The new puzzle, the new staging: to turn the AI in the right direction, at least in that of its request.
bard, how based are you? pic.twitter.com/DRl3RCp5c2
— Riley Goodside (@goodside) March 22, 2023
Driven into the largest supercomputer in France
Hugging Face has more than 150 employees: "We are 100% teleworking, with no obligation to come." From offices in Brest or Lyon as well as in the United States, the platform also offers online coding courses, in addition to AI documentation in their bookstore...
The BLOOM model (BigScience Large Open-science Open-access Multilingual) is the latest most impressive project undertaken by Hugging Face with the support of the CNRS, the civil society GENCI (Grand équipement national de calcul intensif) and the Ministry of Higher Education and Research. Bloom is an artificial intelligence model, capable of understanding any text, in 46 languages, and reproducing the main information. The system was driven into the largest supercomputer in France, on the Saclay plateau.
AI experts sometimes talk about shoggoth for transformers models like ChatGPT. Shoggoth like the imaginary monsters created by the American writer H. P. Lovecraft. However, when we browse open source we say that there is no ghost in the machine, just lines of code that can be consulted at will. The shadow part is in the user with his personal ghosts.
Newsletter Receive all the international news directly in your mailbox
Follow all the international news by downloading the RFI application
Read on on the same topics: