AI-assisted discovery

New technologies have always reshaped the way we do research and expanded the scope of what we can discover. Artificial intelligence is the latest example, although the transformation it has triggered has been unusually swift. At the London Institute, we do not merely adapt, we help set its direction. We treat AI as an epistemic instrument that amplifies human intention and enlarges the space of ideas we can explore.

Mathematics and theoretical physics are especially fertile ground for human–machine collaboration. Their objects—numbers, series, graphs and geometric shapes—form an immense, exact and highly organised body of Platonic data. Using machine-learning methods, we can search this landscape for regularities and exceptional structures. Among the most evocative examples are the “murmurations” of elliptic curves—intricate, previously unseen patterns in arithmetic data that point to a deep underlying mathematical order. In another line of work, we have developed a framework for the systematic discovery of new quantum integrable models.

Large language models offer a different mode of collaboration. They can survey literature, accelerate calculations, write and debug code, and sustain exploratory dialogue across disciplines. They can help to turn half-formed intuitions into testable results, but mathematical taste—choosing fertile questions, recognising illuminating arguments and inventing the right concepts—remains decisive. The art of discovery increasingly unites machine power with human judgement.

We also turn mathematics back upon AI. Using mathematically controlled toy models, we investigate why learning systems favour simple solutions, how generalisation emerges, and what intelligence owes to data, optimisation and architecture. A science increasingly mediated by machines must also develop a science of those machines and understand their limits.

Our aim is not automation for its own sake, but a broader, faster and more deeply connected science—one in which human purposes continue to guide increasingly powerful forms of machine intelligence.