Upcoming Event

Speaker: Jacqueline Harding, Stanford University Abstract:  What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models’ capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and […]

Speaker: Prof Herman Cappelen, The University of Hong Kong Abstract:  This paper makes three main claims: 1) There is no universal set of core values for AI systems to align with – moral disagreements run too deep. 2) The increased causal potency that creates the problem of AI risk also implies that imperfect safety mechanisms

Speaker: Prof Darrell Rowbottom, Lingnan University Abstract:  According to the standard no miracles argument, science’s predictive success is best explained by the approximate truth of its theories. In contemporary science, however, machine learning systems, such as AlphaFold2, are also remarkably predictively successful. Thus, we might ask what best explains such successes. Might these AIs accurately

Speaker: Dr Abelard Podgorski, The National University of Singapore Abstract:  There is a choice point in ethics between theories which care about whether our actions make the world impersonally better or worse, and theories which care about whether our actions are better or worse for particular individuals. The distinction has important implications for issues in

Speaker: Dr Dan Hendrycks, Centre for AI Safety Abstract Dr Hendrycks will discuss principles for measuring capabilities of AI systems, and walk through popular general capabilities benchmarks. He will then discuss how general capabilities can be separated from the measurement of their safety, then overview new ways to measure safety.

Speaker: Duan Yawen, Concordia AI Background readings: Model evaluation for extreme risks Sociotechnical Safety Evaluation of Generative AI Systems Emerging Processes for Frontier AI Safety OpenAI Preparedness Framework

Speaker: Dr Peter Salib, The University of Houston Abstract:  Assume that we succeed in crafting effective safety benchmarks for frontier AI systems. By “effective,” I mean benchmarks that are both aimed at measuring the riskiest capabilities and able to reliably measure them. It would then seem sensible to integrate those benchmarks into safety laws governing

Speaker: Rob Long, Center for AI Safety Abstract:  AI systems could plausibly deserve moral consideration in the not-too-distant future. Although precise evaluations are difficult to devise in light of moral and empirical uncertainty about AI moral patienthood, they will be an important tool for handling this issue responsibly and for avoiding both under- and over-attributing