Dr. Sofia Reyes

Sofia Reyes came to AI safety from philosophy, which she considers an advantage rather than an unusual background. 'Most of the hard problems in AI alignment are not purely technical,' she says. 'They require you to think clearly about values, about specification, about what it means for a system to do what you want. That is philosophy, and it turns out to be very hard.'

Her PhD at Berkeley on learned reward functions gave her a technical grounding in the mechanisms by which AI systems can fail to pursue the goals their designers intended. She has since become one of the most recognized voices in AI safety research, known for bridging the gap between abstract theoretical concerns and the concrete technical work of making AI systems safer.

At Anthropic, she led the team that developed the Constitutional AI approach — a method for training AI systems to be helpful, harmless, and honest through a process of self-critique and revision. She is careful to note that 'no current approach to alignment is solved; we have promising directions, not guarantees.'

At Harvard, she teaches courses on AI ethics, AI governance, and the technical aspects of alignment. Her office hours are notable for running significantly over time, as students invariably have more questions than the hour allows. She considers this a feature, not a bug.