What happens after we solve continual learning?

Oct 13, 2025 // as a tweet summary

I recently gave this keynote at CoLLAS 2025, and also in Feb 2026 as an invited talk at Harvard (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 evaluation 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 are below, and are the most complete version of the talk. (The audio on the Harvard talk recording is a bit broken, and the CoLLAS talk recording doesn’t have the intro primer on continual learning.)

Cite as:

Chan, Stephanie C. Y. (2025). What happens after we solve continual learning? Keynote at the Conference on Lifelong Learning Agents (CoLLAs). https://scychan.github.io/2025/10/13/collas-keynote.html

@misc{chan2025continual,
  author = {Chan, Stephanie C. Y.},
  title  = {What happens after we solve continual learning?},
  year   = {2025},
  note   = {Keynote at the Conference on Lifelong Learning Agents (CoLLAs)},
  url    = {https://scychan.github.io/2025/10/13/collas-keynote.html}
}