An international cohort of leading scientists has embarked on an ambitious new research initiative called Polymathic AI. Their aim is to develop an artificial intelligence system capable of aiding discovery across diverse scientific domains.
The researchers believe that “pre-training” an AI on massive multidisciplinary datasets will produce a more capable scientific assistant. Just as mastering multiple languages can make it easier to learn another tongue, exposure to data from physics, astrophysics, climate science and more may allow an AI to find connections that human experts tend to miss in today’s hyperspecialized world.
“This will completely change how people use AI and machine learning in science,” said Shirley Ho, principal investigator of Polymathic AI and group leader at the Flatiron Institute’s Center for Computational Astrophysics.
Leveraging Scale for Broadly Capable Models
Polymathic AI aims to create foundation models for science – a category of large neural networks that learn general skills and knowledge from vast datasets. The best-known example is ChatGPT, which was trained on internet text data.
But while ChatGPT handles words, Polymathic AI will ingest numerical data and physics simulations. This physics-grounded approach could avoid inaccuracies that plague chatbots relying solely on text associations. Additionally, you can also read about- Top 10 Countries Investment in Artificial Intelligency
The team has access to considerable computing power donated by the Simons Foundation, a nonprofit advancing scientific research. This will allow them to experiment with training foundation models at sizes previously infeasible in academic settings. The resulting system could then be served to the wider scientific community.
“It’s been difficult to carry out academic research on full-scale foundation models due to the scale of computing power required,” said Cambridge co-investigator Miles Cranmer. “Our collaboration with Simons Foundation has provided us with unique resources to start prototyping these models for basic science.”
Interdisciplinary Teams Combining Diverse Expertise
The Polymathic AI collaboration includes researchers from institutions like Cambridge University, the Flatiron Institute, Princeton University, and Lawrence Berkeley National Laboratory. The group encompasses experts in physics, astrophysics, mathematics, artificial intelligence, and neuroscience.
Francois Lanusse, a cosmologist at France’s Centre national de la recherche scientifique, explained: “Despite rapid progress of machine learning in recent years in various scientific fields, solutions are developed for specific use cases and trained on very narrow data…This creates boundaries both within and between disciplines.”
By combining strengths across domains, Lanusse and his colleagues hope Polymathic AI will uncover connections and insights that might otherwise be missed in today’s hyperspecialized research environment. Just as historical polymaths innovated by linking concepts across fields, the tool could rediscover this cross-disciplinary approach.
Democratizing Access to Benefit Wide Scientific Community
The researchers aim to make their methods and models publicly accessible to democratize AI for a broad community of scientists. As Ho explained, “We want to make everything public. We want to democratize AI for science in such a way that, in a few years, we’ll be able to serve a pre-trained model to the community that can help improve scientific analyses across a wide variety of problems and domains.”
This transparency and openness stand in contrast to tech giants that often hoard their AI advances. The team believes openly sharing their foundation models will accelerate discovery across disciplines. If you want you can also read- Character.ai Sees Rapid Growth in US, Catching Up to ChatGPT
But What Are the Risks?
Polymathic AI offers intriguing potential, but some experts urge caution around broadly capable AI systems. There are concerns about misuse, bias perpetuation, and other pitfalls if development outpaces safety research.
Yoshua Bengio, a pioneer in deep learning, recently argued we are not ready for “artificial general intelligence” that can match humans across capabilities. Others emphasize the need for teamwork between AI and human experts who provide wisdom the machine lacks.
The Polymathic AI team will need to take care to address these risks in their design choices. But if done thoughtfully, openly sharing this technology could profoundly enhance human collaboration and ingenuity.
You May Find Interest:Â 20 Best Scientific Discoveries in 2023
The Future of AI for Science
Polymathic AI represents a fascinating new frontier – but only one effort in increasingly AI-powered science. From tailored drug discovery AIs to autonomous telescopes scanning the cosmos, researchers across fields now routinely leverage artificial intelligence in their work.
As Ho remarked, “This will completely change how people use AI and machine learning in science.” But it remains to be seen whether an AI polymath can truly match the creativity and intuition of humans who have spent lifetimes mastering their disciplines.
Regardless, one thing is clear: the interplay between human and artificial intelligence in scientific discovery is only just getting started. Collaborations like Polymathic AI promise to push boundaries in unexpected ways – if pursued thoughtfully. The quest for knowledge marches on, increasingly augmented but not yet replaced, by thinking machines.
