Why "not-science" matters: Scientific Research at the Age of AI
Over the past three years, AI systems have proven incredibly valuable to the pursuit of science, garnering two Nobel prizes (Chemistry, 2024 and Physics, 2024) amongst numerous other accolades. As these systems continue to demonstrate scalability in complex computational tasks, it is natural that scientists are keen to assess the utility of AI in augmenting scientific research workflows. Known as AI for Science, this research domain is incredibly diverse, ranging from generative predictive models like AlphaFold to fully autonomous agentic pipelines. While there are numerous directions within AI for Science, the underlying substrate remains the same: scientific research. In light of accelerating AI integration, how can we as scientists continue to ensure safe and principled science? In this talk, I will go over the frameworks I have developed to maximize the safety and alignment of scientific AI, born of probing the bounds of science. In particular, I will focus on how the idea of antithesis can be leveraged to build valid evaluations for scientific contexts.
