Perspectives on Continual Learning: Survey Results and Forecasts

LessWrong today published results from a survey of AI safety researchers about continual learning—the ability of AI agents to keep updating and learning new capabilities during deployment without catastrophically forgetting what they already know. The researchers surveyed experts on whether continual learning is likely, how risky it is, and whether it could be controlled. The results show wide disagreement. Five out of seven respondents think widespread continual learning agents are unlikely in the next three years. But four out of six think that the first transformative AI systems will be continual learners. There was consensus on one thing: nearly everyone agrees that more research into continual learning safety would be valuable. The takeaway is that experts see both promise and peril in continual learning—and little agreement on which will dominate.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton