Mathematician and winner of the prestigious Fields Medal, former MP for Essonne and ex-president of the French Parliamentary Office for Science and Technology (OPECST)In 2018, Cédric Villani published a remarkable report entitled "Donner un sens à l'intelligence artificielle" ("Giving meaning to artificial intelligence"). This issue, which may have appeared to be a preoccupation of specialists only a short time ago, has made a spectacular irruption into offices and homes in 2023...
365: What'changed since the publication, in 2018, of your report on the'artificial intelligence ?
Cédric Villani: From one sector to another, the impact of AI has not been in line with forecasts. In healthcare and gaming, it has made much more progress than expected, while in the ecological transition and the automotive sector, it has disappointed. But it was completely predictable that the progress of these experimental technologies would not be ... predictable! Even more impressive was the arrival, in 2023, of "consumer" generative artificial intelligence: with Chat GPT, everyone could suddenly make it their own, which is a minor revolution!
365: In scientific terms, can we talk about'a conceptual break ?
CV: Not at all, in fact! It's been ten years since Yann Le Cun and others brilliantly reintroduced neural networks to statistically generate texts, so their use is not new... And it's been five years since Google engineers published the principle of transformers, the key "building block" of ChatGPT networks. But we didn't expect such a performance so quickly! As recently as the summer of 2022, a famous INRIA research director showed me how to comically trap the best conversational agents... And yet, three months later, Chat GPT foiled all these traps. What a shock!
365: How can we explain this spectacular but unexpected breakthrough? ?
CV: In science, sometimes experiments come after theory - radio waves, for example. In AI, it's the other way round: theorists "run" after experimenters. It may be decades before algorithmics can explain the efficiency of these generative AIs... It's a narcissistic wound for theorists - some of my mathematician colleagues still think that AI isn't science... At the very least, it poses an exciting challenge!
365: It's a new situation ?
CV: The injury dates back to at least 2012, when Le Cun, Hinton and Bengio crushed their competitors by perfecting old techniques that everyone else had disdained or abandoned... Today, the theoretical challenge is matched by an extraordinary performance visible to the general public: inspired by the corpus of the Web, the AI establishes relationships, occurrences and figures of speech. In just a few seconds, it can produce a summary of a book, or a general policy speech so realistic that it seems to have been written by a pro. A simple technical evolution, then, but one with enormous practical effects!
365: How can we explain the machine's unexpected progress in terms of "creativity"? ?
CV: There are certainly things we haven't yet understood, from a theoretical point of view, about the structure of knowledge: basically, our great orators and writers are more predictable than we thought, which is disturbing... Be careful, though, if AI is having a major exploratory impact in the physical sciences, it still hasn't created a single innovative theory. AI conceiving general relativity is a long way off!
365: Who do you think will control the'AI and its power?
CV: Billions are being poured in by companies and investors worried about missing out on Eldorado, but we still don't know what will come of it. A crucial question for public authorities is how to ensure that at least one of the new generative networks is open access, given the breakthroughs of Microsoft/OpenAI and then Google. Yann Le Cun, for his part, anticipates that, in any case, open models will eventually be much more effective, because they will have more data to learn from, and will not be biased by the gazelle-like prudishness of the big private players, who censor themselves to preserve their reputations. It's a possibility...
365 : Chat-GPT opens a series of moral questions about learning, from "cheating cheating from "cheating" at school to changing jobs in the workplace, where the'initiative may seem less necessary.
CV: The arrival of the calculator in schools gave rise to the same fears. I'm not worried. As with the calculator, student assessment will be divided between tests carried out "at home" with technology, and those "on the table" where the student's reasoning will be checked by himself. There are more difficult challenges to be met: in the workplace, for example, even beyond the "thankless tasks" that can be easily replaced, we are seeing a growing subordination of professional practices to machine guidance. Yes, there will be upheavals...
365: With this conceptual quirk of'an inverted situation where the machine teaches man, becoming the master and man the pupil !
CV: With AI, we become, de facto, pupils of the pupil... We're back to the old problem of technology enslaving as much as it liberates, a problem that began with Socrates and was re-explored by Jacques Ellul, Ivan Illich, André Gorz... Historical examples sketch out a contrasting reality, far from black and white: technology brings us more capabilities but more dependence, more wealth but more needs, more speed of execution but more tasks. In research, digital technology has brought us fabulous comfort, and it would be absurd to do without it; but it has also clogged up the system beyond what is bearable. On a personal level, we need to find the right compromise between the Socratic ideal - understanding by doing - and the immoderate use of AI for knowledge. To stay strong, we need to master the tool and maintain our personal intelligence at the same time! And find ways to develop our memory and "inner" abilities to compensate for the increasing externalization of these faculties, including by embarking on a seemingly useless but rewarding path. For the past few years, I've been learning poems - excellent for memory, style and reflection...
365 : Isn't Chat-GPT a risk factor for'increase in inequality, because not everyone can'not everyone has access to it, but also because of its differentiated impact on the'employment ?
CV: In general, a new tool increases inequalities because the best-informed, most affluent group has easier access to it than the others. But within the same use, the same "tribe", it reduces them. An excellent programmer told me that his productivity increased by a factor of 4 with Chat-GPT, but that it also reduced the gap between excellent and average programmers.
365: How should a company manage the'impact on employees of the'introduction of Chat-GPT?
CV: Inevitably, there is some reticence about such a potentially disruptive effect on internal habits and power relations! But when it comes to AI, there can be no know-how without experimentation: this is crucial for both the individual and the company. The most important thing is to establish a framework of trust, through human relations, to install the idea that, thanks to AI, we're all going to work more efficiently for a better common future. 
Interview by Charlotte Bourgeois-Cleary (Partner) and Philippe Manière (Chairman, co-founder)
