Neel Nanda

Neel Nanda did not set out to lead a team at one of the world’s most influential AI labs. After studying pure mathematics at Cambridge, he followed a well-trodden path into quantitative finance, with internships at Jane Street and Jump Trading. The jobs paid well, looked impressive, but left him cold. So he walked away and spent a year focused on a question that is now becoming impossible to ignore: as AI systems grow more powerful, which ones should we trust, and why?

That question sits at the centre of his work today. Nanda leads the mechanistic interpretability team at Google DeepMind, where his research focuses on understanding what happens inside advanced AI systems as they make decisions. Rather than treating AI as a black box, his team works to reverse-engineer the internal structures and patterns these models rely on — knowledge that is increasingly central to judging reliability, risk, and appropriate use.

Neel Nanda is widely recognised as one of the early figures in mechanistic interpretability, helping to grow the field from a handful of researchers into a serious area of frontier AI research. He has built some of its most widely used tools, including TransformerLens, and is known for three-hour YouTube walkthroughs where he patiently unpacks dense research papers line by line. Nanda often emphasises the need for layered safeguards. Achieved by combining interpretability with other approaches to decide when, where, and whether powerful AI systems should be trusted. His talks are valued not for bold predictions, but for its clear thinking about uncertainty, incentives inside large organisations, and how responsible progress actually happens.

Through his MATS programme, Neel Nanda has supported dozens of early-career researchers in finding their footing in AI safety. A sought-after adviser, he works with leaders, technologists, investors and policymakers to make sense of where AI is heading, how decisions are really made, and what it means to act responsibly when the future is unclear but the stakes are very real.