15-889 Science of Frontier AI: Capabilities and Safety

12 units, Carnegie Mellon University

This advanced PhD-level seminar takes a scientific approach to frontier AI systems: what we know, what we assume, and how to find out. We follow these systems from training data to learning objectives to scaling laws, and touch on emerging paradigms such as synthetic data and AI agents. Capabilities have outpaced principled understanding of how training shapes behavior, leaving open questions about the limits of what these systems can do: meaningfully accelerating scientific progress, and learning continually and improving themselves. Increasingly capable AI also raises pressing questions of safety, which we distill into technical ML questions: how to evaluate model behavior, scale oversight, and detect dangerous capabilities before deployment. Lectures are combined with role-playing paper seminars and a final research project.

Prerequisites: none listed

Courseweave needs JavaScript to show sections, workloads and planning tools.