1. Beyond the Shadow of Language: AI as the New Observer of Nature

05/2026

For centuries, science has been bound by a profound limitation: the bottleneck of human language.

When we talk about Artificial Intelligence for Science (AI4S), we often think of it as a tool to process data faster or write code more efficiently. But in my view, the true essence of AI4S is far more radical. It is about empowering artificial intelligence to bypass the barrier of human language entirely, allowing it to directly observe, experiment, synthesize, and reason with the raw fabric of reality.

The philosopher Ludwig Wittgenstein famously noted that language is merely a tool to describe the world, not the world itself. "The limits of my language mean the limits of my world." When an AI generates text, its meaning and logic are subjectively judged by humans. It is simply a shadow embellishing another shadow.

However, the molecules, proteins, and materials we aim to discover in the lab are not abstract linguistic symbols. They are nature itself. Their existence and interactions are not governed by grammar or syntax, but strictly by the Schrödinger equation.

To truly understand nature, we must ask: What happens when the observer is no longer human?

The Physics of Observation: Human vs. AI

To understand this paradigm shift, we can look through the lens of quantum mechanics. In quantum theory, the act of measurement is what collapses an infinite wave of probabilities into a definitive, observable reality.

The Standard Quantum Framework

In traditional physics, measuring the energy of a quantum system is expressed as:

• $\widehat{H}:$ the Hamiltonian operator, representing the energy measurement.
• $\psi$: the intrinsic, uncollapsed state of the quantum system.
• $E$: the energy eigenvalue yielded by the measurement.

The Human Observation Analogy

If we view humanity as a collective operator measuring the wave function of nature, the resulting eigenvalue we obtain is what we call human language.

• $\widehat{H}_{\text{human}}:$ the human observer acting as the measurement operator.
• $\psi$: the intrinsic, unobserved state of nature.
• $L$: the resulting eigenvalue, collapsing into human constructs such as English, Chinese, or classical mathematical symbols.

Language is an evolutionary compression algorithm. It is highly effective for human survival and storytelling, but it is a profoundly lossy, low-dimensional projection of a high-dimensional, complex universe.

The AI Measurement Paradigm

This raises a fascinating, paradigm-shifting question. If an AI system, unburdened by evolutionary biology or linguistic syntax, acts as the operator observing nature, what will its eigenvalue be?

• $\widehat{A}$: the artificial intelligence, such as a neural operator or foundational model.
• $\psi$: the intrinsic state of nature.
• $X$: the novel eigenvalue generated by the machine.

Uncovering the "Alien" Eigenvalue

Unlike human language, this new eigenvalue does not manifest as a linear string of characters or words. It manifests as a sprawling, high-dimensional topology. It is encoded in latent manifolds, multi-dimensional feature spaces, and billions of network weights.

When a machine observes a folding protein or a quantum material, it does not translate that phenomenon into a noun or a verb. It translates it into a continuous mathematical geometry where distance equals structural similarity, and trajectory equals chemical reaction.

This means that a neural architecture and its specific combination of parameters might inherently encode scientific truth in a purer form than human language ever could.

The future of AI4S is not just about using machines to read scientific papers so they can summarize them for us. It is about allowing machines to read the universe directly, capturing the wave function of nature in a language we are only just beginning to learn how to translate.

2. From Observation to Integration: The Four Paradigms of Scientific Discovery

09/2025

Science is defined by how we observe the universe. Over centuries, our methods of observation and discovery have evolved through distinct paradigms, each fundamentally changing the speed and scale of human progress.

The First Paradigm: Empirical Science

The foundation of scientific discovery lies in directly observing, recording, and describing natural phenomena. Scientists formulate hypotheses, test them, and develop laws based strictly on physical validation.

Concrete examples: Galileo rolling spheres down inclines to measure motion, and Mendel meticulously breeding pea plants to deduce the laws of inheritance. This is the classic hypothetico-deductive method.

The Second Paradigm: Theoretical Science

Knowledge advances through abstract reasoning and mathematical synthesis, often extending far beyond what can be seen in a lab. The goal is to derive fundamental theories that explain observed phenomena and predict entirely new ones using limited experimental data or purely thought experiments.

Concrete examples: Maxwell unifying electricity and magnetism into a set of elegant equations, and Einstein deriving general relativity. These theories capture the underlying mathematical structure of physical laws.

The Third Paradigm: Computational Science

As physical equations became too complex to solve by hand, science shifted toward creating virtual models. By translating physical systems into algorithms, we can simulate phenomena that are too small, too fast, or too dangerous to observe directly.

Concrete examples: Density Functional Theory calculations solving quantum mechanical states of many-body systems, and Molecular Dynamics simulating the atomic movements of materials over time.

The Fourth Paradigm: AI for Science (AI4S)

The current frontier integrates artificial intelligence, machine learning, and high-performance computing with traditional methods. AI does not just crunch numbers; it identifies hidden topological patterns in massive datasets, proposes hypotheses, and bridges the gap between pure statistics and physical laws.

Concrete examples: DeepMind's AlphaFold predicting protein structures, symbolic regression uncovering new physical laws from raw data, and neural networks predicting complex material properties millions of times faster than traditional solvers.

As highlighted in Scientific Discovery in the Age of Artificial Intelligence, the fusion of AI with traditional scientific inquiry forms a methodology that pushes the very boundaries of what is discoverable.