Does AI Belong to Physics?

To the 2024 Nobel Prize Committee in Physics,

“Foundational discoveries and inventions that enable machine learning with artificial neural networks” can never be considered a contribution to physics itself. It does not make sense to award machine learning and artificial neural networks a physics prize.

In general, AI is a tool that can be used to study physics, but it can never be part of physics.

From Plato to Einstein, natural philosophers and physicists have shared the belief that the physical world is comprehensible and can be explained through a series of fundamental laws, often called first principles. For Platonists, the primary goal of physics is to uncover the fundamental constituents of matter, discover the universal laws of the Universe, and reduce the phenomena of the nature world to these laws. Empiricists, on the other hand, view physics as the pursuit of understanding the causes behind matter’s interactions in space and time, with the key requirement that these causes be falsifiable.

AI has no direct connection to the natural world. It is a human-made tool designed to simulate certain aspects of reality, drawing on limited knowledge from fields like linear mathematics, physics, neuroscience, and others. It is more akin to an engineering effort in bionics, but irrelevant to the first principles that govern the natural world.

First of all, the term “AI” (Artificial Intelligence) is vague and continuously evolving. Many people may associate AI with robots or machines that exhibit human-like intelligence. However, just as ancient alchemists failed to understand the fundamental nature of matter and thus could not transform stone into gold, modern human beings are far from comprehending the essence of human consciousness. Creating a robot with true human consciousness is well beyond our capabilities. The Francis monk Roger Bacon’s legendary Brazen Head was a myth eight centuries ago, and it remains a myth today.

Modern AI originated from scholars’ attempts to automatically prove mathematical theorems. They wanted to build a machine capable of solving mathematical problems. Scholars later shifted their focus to developing a huge “if-else” system, known as the Expert System, aimed at answering questions and making decisions like human experts. This plan ultimately failed, mainly because it required too many “if-else” rules, making the system overly complex and beyond the capability of nowadays’ human technology.

Meanwhile, other methods became feasible thanks to advancements in computational power, the most significant being artificial neural networks (ANNs), which were inspired by biological neural networks in animal brains, with the initial goal of mimicking the brain’s ability to receive and process signals. With the development of GPUs and enhanced computational capabilities, ANNs have been applied to a wide range of fields — from defeating human Go players to analyzing protein structures, interpreting signals from physical experiments, and even powering the latest large language models and generative AI models.

Now, let us examine the science behind artificial neural networks.

Have ANNs improved our understanding of animal brains and uncovered physical laws within them?

No.

People do not yet know how to effectively model animal brains, so researchers turned to statistical mechanics to develop methods for building and optimizing these networks. Jim Fan has compared machine learning to statistical mechanics, highlighting parallels such as:

• Machine learning -> statistical mechanics
• Loss function -> energy functional
• Optimizing the model -> minimizing free energy
• Trained model -> equilibrium distribution

On the surface, machine learning appears similar to statistical mechanics. But here is the key difference: statistical physics is governed by the universal laws of thermodynamics, while machine learning/ANNs lack any fundamental laws behind it.

Simple question: Do ANNs provide foundational laws similar to the three laws of thermodynamics?

No.

Do ANNs propose fundamental principles that govern the Universe?

No.

Do ANNs offer basic theories that can be falsified through experimentation?

No.

Thus, ANNs are not part of physics, nor do they constitute science in the traditional sense.

So, what do artificial neural networks do?

They are merely tools, fundamentally no different from least-squares fitting methods or other mathematical models.

Let us see an example:

Statistics is not physics, but statistical mechanics is. Pioneered by Maxwell and Boltzmann, statistical mechanics used statistical methods to establish a connection between microscopic particles and macroscopic phenomena. Its core theory is the principle of maximum entropy. The most important aspect of statistical mechanics is not the use of statistics as a tool, but rather its foundation on the assumption of the existence of molecules and atoms. This theory is falsifiable according to scientific standards. If atoms do not exist, German physicist Ostwald’s energetics theory would have won. The experiments and explanation of Brownian motion in the early 20th century provided strong evidence for the existence of atoms, leading to the success of statistical mechanics and cementing its place in mainstream physics.

What is the foundation behind ANNs? Biological neural networks? Not exactly. ANNs are an oversimplified and far more energy-consuming beta 0.0 version of brain neural networks. They are a form of bionics, with no fundamental connection to physics.

Let us give another example:

Calculus is pioneered by Newton and Leibniz and later developed into a rigorous mathematical branch by Cauchy, Weierstrass, and others. While calculus itself is math, not physics, its certain application — such as using it to discover that celestial bodies follow elliptical orbits obeying the inverse square law of gravitation — is physics. If we were to observe a celestial body moving in a triangular orbit, Newton’s law of universal gravitation would be disproven. In contrast, ANNs do not contain any fundamental laws, like the law of gravitation, that can be tested or falsified.

In short, artificial neural networks and the vaguely defined term “AI” are not physics, nor are they traditional science. They are tools, inspired by bionics, but fundamentally no different from least-squares methods and other statistic methods.

At last, let me add one final point:

If there were to be an award for ANNs, it should not go to John Hopfield and Geoffrey Hinton, but rather to the Japanese scholar Shun’ichi Amari. Neural networks trace their origins back to the Ising model proposed in 1925, and it was Amari’s pioneering work in 1972 that first extended this model to establish the foundation for neural networks. John Hopfield almost replicated a similar approach ten years later. Anyone with a shred of scientific integrity should refer to the Hopfield network by its rightful name: the “Amari-Hopfield network.” As for Geoffrey Hinton’s famous “Boltzmann machine,” its roots lie in the earlier work of Soviet scientists Ivakhnenko and Lapa. The complete omission of recognition for Amari, Ivakhnenko and Lapa in the 2024 Nobel Prize in Physics reflects a glaring oversight by the Nobel Physics Committee.

It brings to mind an old saying: “Physics is over, offer it a prayer.”