Human AI collaboration for science and engineering

Abstract
Powerful AI, such as LLMs, holds immense potential to enhance human capabilities, particularly in science and engineering. But how can we truly collaborate with AI? Unlike traditional tools such as simulations, AI is more adaptive and can interpret our intentions, much like human colleagues. This makes alignment a unique and critical challenge for effective collaboration. However, achieving alignment faces deep mathematical and philosophical challenges. AI decisions can be unreliable in high-stakes applications, requiring human oversight, yet human feedback itself is often imperfect. In this talk, we explore approaches to understanding and addressing these challenges in both single- and multi-agent systems, with a focus on applications in Bayesian optimization.

Biography
Masaki Adachi is a final-year PhD student in the Machine Learning Research Group at the University of Oxford, under the supervision of Michael A. Osborne. He also serves as an assistant manager in the data science team at Toyota Motor Corporation. His research interests include Bayesian optimization, Bayesian quadrature, preference learning, social choice theory, and materials discovery.