Speaker
Description
Calculations, and more recently computations, have been intimately linked to physics since centuries. The 2024 Nobel prize in physics provides the latest illustration thereof. Modern physics really emerged when the use of mathematics was generalized to support the physical description of the universe. Computations nowadays have become a key tool of investigation in physics, both for modeling and for data analysis. Conversely, modern computers capabilities have been attained thanks to major physics-based developments, like the invention of the transistor. These relations between physics and computations are somewhat obvious and well documented.
The emergence of Artificial Intelligence (AI) both in science and all-day life might change the rules of the game. Fundamentally AI systems remain computer-based objects with all the above-mentioned links to physics. But AI now promises computations hardly conceivable only a few years ago. This might lead people to think that limits of AI will always be overcome, again and again. This would mean, on the long term, a strictly rational and deterministic viewpoint on the world and a latent hope to solve any “unsolvable” problem. Such a potential viewpoint may hold true in many sciences, not speaking of situations in all-day life issues.
However, it turns out that computations and AI suffer from intrinsic limitations, first due to technical issues in the representation of numbers they manipulate. These difficulties are well known, as well as strategies to overcome/control them, at least partially. More fundamentally, computational possibilities hit walls imposed by the laws of physics. The physical description of the world leads to address complex non-linear equations which allow chaotic behaviors. These chaotic features cannot be overcome, whatever accuracy is attained numerically, so that the practical description of the world is bound to integrate a chaotic component. Furthermore, quantum mechanics introduces a random component into the description of microscopic systems. This may have macroscopic consequences as for example a radioactive disintegration of a nucleus or the hit of a smartphone by a cosmic ray leading to an unexpected error in the system. All in all, to imagine a fully controlled, strictly rational and deterministic, access to the world by computational means is thus confronted with major physical impossibilities.
The aim of this contribution is to demonstrate and illustrate this fundamental impossibility. This is a key issue in our way we see and hope to understand the world. The premise of the analysis does not rely on vague arguments or on faith but on well-established, scientific, facts. It is thus important to keep in mind such limits whatever computations and AI might allow, both today and in the future. The point is not to dispute possible progress attained by AI, although caution should remain the rule, especially in terms of ethical and social issues. This latter aspect will of course also be discussed. But the major point is to identify the fundamental limits set by physics.
This concerns all of us. While physicists will in principle be aware of most of the aspects addressed here, it is clear that many scientists, users of AI in particular, are probably not. It is thus important that the scientific community, as a whole, becomes aware of these limits set by nature to AI and computations in general. More generally speaking, all educated people should integrate this aspect into their understanding of the world, as this very understanding is more and more mediated by AI’s and computations. The societal impact of such a realization is thus crucial.