AI in Science: From Deep Thought to CERN Diffusion Models and the Nobel Turing Vision

Summary

The Royal Institution talk explores how AI and machine learning are transforming science, from basic definitions to concrete applications in physics and biology. It contrasts scientific philosophies, explains supervised and unsupervised learning, and demystifies deep learning and transformers. It highlights real-world cases like AlphaFold and CERN data workflows, and discusses whether AI can truly participate in science or merely accelerate it. It ends with a provocative Nobel Prize–level vision for AI as a collaborative scientist under the Nobel Turing Challenge.

• AI in science is a powerful predictor, not a thinker with true understanding.
• Supervised vs unsupervised learning and the rise of deep learning and transformers.
• AlphaFold and diffusion-based simulations show rapid progress and new capabilities.
• The Nobel Turing Challenge envisions AI as a synthetic scientist with oversight and collaboration.

AI and Science in Context

The talk situates artificial intelligence within the scientific method, outlining how AI systems are used to accelerate discovery while raising questions about what counts as genuine understanding. It presents a framework for defining science through testability, skepticism, and shared evidence, then maps how AI tools align with or challenge that framework. The discussion moves from foundational ideas in philosophy to practical machine learning concepts, including supervised and unsupervised learning, deep learning, and the transformer architecture. It also delves into concrete science cases where AI is making an impact, such as AlphaFold's protein structure predictions and CERN’s data-intensive research at the LHC, illustrating both the opportunities and limitations of current AI capabilities in science.

Key Insights

• Science relies on testable explanations and reproducible evidence; AI adds speed and scale but not necessarily understanding.
• Two major learning paradigms shape AI in science: supervised learning with labels and unsupervised discovery of patterns.
• Transformers and self-attention enable language-like modeling and scalable pattern recognition across data types.
• AlphaFold and diffusion-based simulations at CERN demonstrate how AI changes the way we model biology and detector physics.
• The Nobel Turing Challenge envisions AI systems that can autonomously contribute to scientific discovery while remaining under human oversight.