# Physics in artificial intelligence Author: Luis Eduardo Paniagua Martín Original publication: 2025-10-29 Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-109-el-motor-oculto-de English URL: https://elcontemplador.github.io/estrategia-english/essays/109/ Status: Published translation This is a translation of the original Spanish essay published on 29 October 2025. Its claims, examples and forecasts retain that historical context. English publication: 2026-09-29 > Artificial intelligence rests firmly on physics, from the foundations of chip design and computing architectures to the mathematical models that inspire its algorithms. > > For this reason, **estrategIA** welcomes the **[Colegio Oficial de Físicos (COFIS)](https://www.cofis.es/)**, Spain's professional body for physicists, as a regular contributor, thanks to an agreement with [Institución Educativa ALEPH](https://institucioneducativaaleph.com/). The participation of this body, a leading institution in Spanish physics, strengthens the rigour and scientific perspective we want to bring to this space for reflection on AI's impact on politics, government and society. > > The collaboration begins in this issue with an article by [Luis Eduardo Paniagua Martín](https://www.linkedin.com/in/luis-paniagua-824a891a7?utm_source=share&utm_campaign=share_via&utm_content=profile&utm_medium=ios_app), COFIS representative for Extremadura and a data scientist at Babelgroup. It will continue with future contributions from the Colegio Oficial de Físicos to broaden this conversation between science and society. The Nobel Prize in Physics last year, 2024, was awarded to John J. Hopfield and Geoffrey Hinton “*For foundational discoveries and inventions that enable machine learning with artificial neural networks*”. Specifically, Hopfield invented a type of neural network inspired by physical models of systems composed of and designed with many small parts. Hinton, in turn, used tools from statistical physics to create what is known as the “Boltzmann machine”, which can recognise a subset of characteristic elements within a complete dataset—in other words, it can recognise patterns. But why is physics so closely related to artificial intelligence? How important is physics within this field? With a Nobel Prize—the highest recognition in the world of research, and one that caused some controversy within the scientific community—physics has driven the development of artificial intelligence not only through the immensely important work of mathematical development, but also by developing all the hardware responsible for running the multitude of algorithms inspired by physical processes and concepts, among others. Starting with the creation of an idea written in the language of mathematics, the many existing algorithms are subsequently translated into programming code and, finally, into the on-and-off bits through which all a server's systems are coordinated by the processor's clock frequency: the computer is executing the idea. Besides allowing us to delve deeply into a particular scientific field, physics connects processes across different areas of knowledge, helping communication and understanding in the spaces where they meet. It is the translator that understands mathematical language and needs to translate it into software code. It is the translator that understands the language of software and needs to translate it into machine language, integrated circuits and transistors—into hardware. ## From the idea to the machine The first multilayer perceptron, the first neural network that could learn from examples, was created in 1958. Earlier, in 1943, scientists Warren McCulloch and Walter Pitts created the first concept of a neuron. This simplified mathematical model laid the conceptual foundations for modern neural networks. Yet until the arrival of GPUs—graphics processing units—in the 2010s, artificial intelligence was a science separate from mathematics, still entirely at the research frontier. Today, artificial intelligence is known to everyone, especially generative artificial intelligence. But fields within it, such as *machine learning*, have been established in technological processes for many years, thanks to the development of algorithms and their communication with servers of modest power. Imagine a collection of financial orders of every kind, containing information on the equity market, the corporate world and fixed income—government debt—as well as commodities, derivatives and other complex products, together with different order types, currencies, brokers and much more. Is it possible to extract patterns that help detect potentially fraudulent or anomalous behaviour? *Machine learning*, a subfield of artificial intelligence, offers many algorithms for doing this that are not necessarily based on neural networks. *Deep learning*, another subfield of artificial intelligence, also offers a set of neural networks designed specifically for this task. Yet without the ability to validate the result, how can we be sure we are heading in the right direction? Which of the two algorithms is correct? Can we quantify the algorithms' results and compare them? One of the most important concepts in physics helps shed light on this: entropy. ## Entropy as a connecting concept Entropy is a measure of uncertainty or of average objective information content. To understand the concept of information, we need to understand that the data least likely to occur within a dataset are those that provide the most information. For example, if we treat the words in a text as a dataset, very common Spanish words such as “la”, “el” and “a”—the two forms of “the”, and “to”—provide little information, while less frequent words such as “ordenador”, “algoritmo” and “ratón”—“computer”, “algorithm” and “mouse”—provide more. Deleting the word “a” from a text will probably not affect comprehension, and the text will still be understood; the same is not true if we delete “ordenador” from that original text. In general, when all the data in a system or dataset are equally probable, they all provide irrelevant information and entropy is at its maximum. Entropy is the amount of “noise” or “disorder” in a system, not just in a dataset. By applying entropy as a measure of the information obtained from both algorithms' results, we can therefore quantify which is providing more information and which less, identify the point at which exactly the same information is obtained, and better understand the algorithms themselves. An algorithm that reduces the entropy of the resulting subset of data compared with the remainder classified as non-anomalous—that is, the purely random data content—can be considered more effective at identifying atypical patterns, because more information is obtained on average. Physics acts as a connector between two independent algorithms. ## What might physics contribute next? Years ago, when the first concept of neural networks began to emerge in the offices of mathematics departments, the infrastructure capable of running what would later become an algorithmic revolution did not exist. Today, the world stands at the frontier of knowledge, with the largest infrastructure humankind has ever built, and yet generative artificial intelligence models seem to have reached a computational limit. Might physics have something to contribute? If the limit is computational, could we create new methods capable of revolutionising even the field of artificial intelligence? Perhaps the development of quantum computing has something to say. [Luis Eduardo Paniagua Martín](https://www.linkedin.com/in/luis-paniagua-824a891a7?utm_source=share&utm_campaign=share_via&utm_content=profile&utm_medium=ios_app) Senior data scientist at Babelgroup Representative for Extremadura, Colegio Oficial de Físicos ## An illustrated explanation *Since we know that most of you, like us, do not have advanced knowledge of physics, and that some of these interesting reflections may seem rather complex, we asked Gemini 2.5 Pro to help us generate an infographic from the article, taking advantage of the improvements Google has introduced in this field in recent weeks. Here it is:* *Archive note: the four original infographic panels below were generated with Gemini 2.5 Pro for the October 2025 newsletter. Their explanations are translated in full as the original editorial aid; they are not a new scientific validation. The Nobel citation above was already in English in the source.* [![First infographic panel on the 2024 Nobel Prize and physics as a translator between mathematics, software and hardware.](https://elcontemplador.github.io/estrategia-english/assets/images/c010f3aa-a7ec-4b4e-9013-0f8c3def7ab7_1203x1553.png)](https://substackcdn.com/image/fetch/$s_!ZGGw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc010f3aa-a7ec-4b4e-9013-0f8c3def7ab7_1203x1553.png) **Panel 1, translated: “Physics in your AI? More than you imagine! For estrategIA readers.”** **A Nobel Prize in Physics… for AI?** Yes! In 2024, Hopfield and Hinton won the Nobel Prize in Physics. Their achievement? They used ideas from physics, such as complex systems and statistics, to create the first neural networks that could learn. **Physics is the universal “translator”.** Physics connects the abstract idea with the real machine. It acts as the indispensable bridge: Idea (mathematics) → Code (software) → Reality (hardware). Physics is the engine that allows software (code) to run on hardware (chips, transistors). [![Second infographic panel explaining the gap between early artificial-neuron ideas and the computing power needed to apply them.](https://elcontemplador.github.io/estrategia-english/assets/images/074d6840-b2b2-4427-bf4c-10977fe25c20_1225x668.png)](https://substackcdn.com/image/fetch/$s_!lroU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F074d6840-b2b2-4427-bf4c-10977fe25c20_1225x668.png) **Panel 2, translated: “The idea existed, but the muscle was missing.”** The first ideas of artificial neurons date from 1943. But AI did not “explode” until the 2010s. Why? Because that was when physics gave us GPUs (graphics units): the immensely powerful hardware needed to train the AIs we use today. [![Third infographic panel illustrating entropy through a before-and-after diagram of fraudulent and ordinary data.](https://elcontemplador.github.io/estrategia-english/assets/images/7ab2d6dc-8501-4ec4-b871-09fc64f0765d_1088x1147.png)](https://substackcdn.com/image/fetch/$s_!9P_f!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7ab2d6dc-8501-4ec4-b871-09fc64f0765d_1088x1147.png) **Panel 3, translated: “How do we measure and compare AIs?”** Imagine looking for fraudulent transactions. At first, all the data are in chaos. This is where a key physical concept comes in: **entropy**. **Entropy = level of “disorder” or “noise”.** | Before AI (high entropy) | After AI (low entropy) | |---|---| | Labels reading “Data” and “FRAUD!” are mixed together. | “Anomalous patterns (fraud)” contains three “FRAUD!” labels. A separate group is headed “Normal data (noise)” and reads “Data, Data, Data, Data…” | | Everything is mixed together. Finding patterns is difficult. | AI “orders” the chaos and reduces disorder. | A good algorithm reduces entropy: it “orders” the chaos, groups the patterns—the valuable information—and separates them from the noise. [![Final infographic panel asking whether quantum computing could address a computational limit for AI.](https://elcontemplador.github.io/estrategia-english/assets/images/e5455ced-8b0f-4db7-aff6-d6636562ec23_1113x682.png)](https://substackcdn.com/image/fetch/$s_!FM2T!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5455ced-8b0f-4db7-aff6-d6636562ec23_1113x682.png) **Panel 4, translated: “What now? The computational limit.”** Today's AIs seem to be reaching a computational limit. If hardware is the limit, could physics once again have something to contribute? Perhaps quantum computing has something to say… *Infographic based on the article “Physics in artificial intelligence”. estrategIA.*