A Cambridge researcher has been awarded a fellowship through the Schmidt Sciences AI2050 programme, a major initiative aimed at ensuring artificial intelligence delivers broad benefits to society by 2050. The fellowship will support early-career and senior researchers in tackling some of the most pressing challenges in AI, including training the next generation of AI scientists, developing safer and more trustworthy AI systems, and enhancing AI’s capacity to advance biological and medical research. The AI2050 programme has allocated $18 million in funding over the next three years to support 21 early-career fellows and seven senior fellows. This marks the fourth cohort of the programme, which now includes 99 fellows across eight countries and 42 institutions, reflecting a growing global network dedicated to responsible AI innovation. The recipient, Dr. Kyle Cranmer, is Assistant Professor of Data Intensive Science at the University of Cambridge, with joint appointments in the Department of Applied Mathematics and Theoretical Physics and the Institute of Astronomy. His research focuses on a fundamental mystery in AI: why models trained on physics data often make more accurate predictions than even our most advanced scientific theories—yet remain opaque in how they arrive at those conclusions. Cranmer and his team are developing new methods to distill knowledge from these AI models and translate it into clear, interpretable scientific theories. As part of his AI2050 fellowship, he plans to extend these techniques to large-scale, general-purpose AI models—similar in architecture to large language models like OpenAI’s ChatGPT or Google’s Gemini—but trained on scientific data rather than text. His goal is to uncover new mathematical concepts and physical laws that explain the superior predictive power of these models. “AI models trained on physics data are finding patterns that our current theories don’t capture,” Cranmer said. “We can see they work, but we don’t always understand what they’ve learned. And the larger these models become, the better they seem to perform. This funding will help us uncover what they’ve discovered that science has missed.” Eric Schmidt, co-founder of Schmidt Sciences with his wife Wendy, emphasized the programme’s mission: “AI is underhyped, especially when it comes to its potential to benefit humanity. The AI2050 fellowship was created to turn that potential into reality—by supporting the people and ideas that are shaping a healthier, more resilient, and more secure world.” Mark Greaves, executive director of AI2050, highlighted the growing impact of the fellowship network. “In just four years, the AI2050 fellows have built a deep sense of community, and we’re proud to see it grow each year. We believe this network will continue to inspire and support them throughout their careers as they work to advance AI for the benefit of all.”