What happens after we solve continual learning?

I recently gave this keynote at CoLLAS 2025, and also in Feb 2026 at Harvard and Stanford (updated slides below).

Abstract: Researchers often point to continual learning as a major missing component for modern AI models. With increased focus on this research area, we may soon find ourselves in a world with widely deployed continual learning agents. The benefits are endless, but continual learning also poses major challenges for AI safety and alignment — many existing techniques assume a single static base model (e.g. RLXF-based post-training), and are not suited for dynamically changing models. In this talk, I will lay out some challenges and examples. I will also describe potential starting points for technical solutions, drawing connections to catastrophic forgetting and to Quine’s “web of ideas”.

—-

Slides below. Tweet summary here.

The Harvard talk recording is here, but there were some issues with the audio. The CoLLAS talk recording is here, but it doesn’t have the introductory primer to continual learning. The slides are the most complete.