Meet the Physicist Revolutionizing AI with Science Literacy | Quanta Magazine

From a young age, Miles Cranmer was fascinated by physics. His grandfather, a professor at the University of Toronto, ignited his interest by gifting him books. His parents also fueled his passion by taking him to university open houses in southern Ontario. One of these places was the Perimeter Institute for Theoretical Physics. Cranmer recalls how thrilling it was to hear someone talk about infinity when he was just a kid. In high school, he interned at the University of Waterloo’s Institute for Quantum Computing, calling it the best summer of his life. This experience led him to pursue physics at McGill University.

During his second year, he came across an interview with Lee Smolin in Scientific American. Smolin suggested it could take generations to connect quantum theory with relativity. This notion was a wake-up call for Cranmer. He thought, “That can’t be right; we need to move faster.” He quickly realized that artificial intelligence could be the key to accelerating scientific discoveries. That moment sparked his journey into combining AI with his research in astrophysics at Princeton University.

Nearly ten years later, Cranmer is now at the University of Cambridge. He has witnessed AI starting to change the landscape of science, but he believes there’s much more to achieve. While tools like AlphaFold can make precise scientific predictions, researchers still lack general-purpose models designed for broader scientific discovery. Cranmer imagines creating something akin to a scientific ChatGPT that can generate simulations and predictions across various disciplines. In 2023, he and over two dozen other scientists kickstarted the Polymathic AI initiative to work on these foundational models.

Cranmer emphasizes that the first challenge is to equip AI with the specialized skills needed for scientific tasks that current systems struggle with. He recalls how some wanted to develop a language model specifically for astrophysics, but he was doubtful. He noted that simulating complex systems requires strong numerical processing skills, something that many AI models lack. Additionally, AI struggles to condense its predictions into neat formulas, like E = mc². The scientific data necessary for training these models isn’t as readily available online compared to other types of information, making this an even bigger hurdle.

Source link