Dr Miles Cranmer from the Department of Applied Mathematics and Theoretical Physics has been announced as a Schmidt Sciences AI2050 Early Career Fellow. He is one of 28 researchers worldwide who are collectively being awarded more than $18 million to advance AI’s potential for the benefit of humanity.
The AI2050 program funds a global community of researchers working on hard problems in artificial intelligence that could offer game-changing opportunities for addressing social challenges or advancing discovery.
Over the next 3 years, the twenty-one early career fellows and seven senior fellows announced in this cohort will receive funding to tackle challenging problems in AI by building AI scientists, designing safer and more trustworthy AI models and improving the ability of AI to pursue biological and medical research.
Addressing an AI puzzle
Miles Cranmer is Assistant Professor of Data Intensive Science in the Department of Applied Mathematics and Theoretical Physics (DAMTP) and the Institute of Astronomy (IoA).
His research attempts to solve a puzzle: AI models trained on physics data sometimes make better predictions than our best theories, but scientists can't see how the AI arrives at its answers.
"AI models trained on physics are finding patterns our theories miss," says Cranmer. "We can see that these models work, but it's not always clear what they’re learning. What's more, the larger AI models get, the better they seem to do this! This funding from Schmidt Sciences will help us figure out what they've learned that science is missing."
Cranmer and his group develop techniques to distil this knowledge into interpretable theories. As part of his early career AI2050 fellowship, Cranmer will push these methods to large-scale, general models – ones similar in spirit to large language models such as OpenAI's ChatGPT or Google's Gemini – but trained on scientific data rather than text.
His goal is to discover new mathematical concepts and physical laws that explain how these AI models achieve their superior predictions.
Advancing AI through a community of support
Alongside the funding, the cohort of Schmidt Sciences AI2050 fellows also benefit from an annual meeting to share findings, exchange ideas and build networks. AI2050 also offers funding to support exceptional computational needs, enabling fellows to accelerate their research.
"AI is underhyped, especially when it comes to its potential to benefit humanity," said Eric Schmidt, co-founder of Schmidt Sciences with his wife Wendy. "The AI2050 fellowship was established to turn that potential into reality – by supporting the people and ideas shaping a healthier, more resilient, and more secure world."
"In four years, the AI2050 fellows have created a deep sense of community that we are grateful to be able to grow each year," said Mark Greaves, executive director of AI2050."We trust that the network they’ve created will remain a source of inspiration and support throughout their careers as they advance AI for the benefit of all."
This article is adapted from a University of Cambridge news story and Schmidt Sciences news announcement. Photo credit: Philipp Ammon for Quanta Magazine.
Discover more about Miles Cranmer's work in our article Teaching AI the language of science.