Large language models can't yet learn continuously the way humans do, but they will—and when they do, keeping them safe becomes a fundamentally different challenge. According to research from LessWrong's Rauno Arike, the safety challenge isn't just that continual learning might make AI systems more capable; it's that we barely understand how it will develop. Arike proposes three concrete approaches: clarifying what different continual learning methods look like and which are safer; building the safer versions ourselves; and creating ways to evaluate continual learning agents. The key insight: the safest future continual learning systems look like Claude Code—memories stored as readable text that humans can edit, not hidden in the model's weights. It's a research agenda for a capability that barely exists yet, preparing safety infrastructure before the technology becomes urgent.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton