Dwarkesh Patel revisits his earlier argument that AI systems need to learn from months of work experience before they can automate advanced AI research. He says he has become less certain that this limitation will prevent such automation by 2028.
He suggests that longer effective context could let a model learn within extended sessions, and that future models might solve continual learning as a research problemContinual learning lets an AI system acquire new knowledge or skills over time while retaining what it learned earlier. without already possessing that capability. Both are possibilities and predictions, not established results.
Watch on YouTube
