Dwarkesh Patel argues that AI is already contributing to research and says recent progress has shortened his forecast for AI systems matching leading human AI researchers. He presents 2028 as his median estimate while allowing a substantial chance of a much later result.
He compares mathematical problem solving with AI research and argues that measurable training and evaluation outcomes could enable automated researchers to test ideas in parallel. His account contrasts a rapid software-only feedback loopRecursive self-improvement is the proposed process in which an AI system helps improve its own capabilities, then uses those improvements to support further advances. with a slower burst constrained by harder ideas and compute.
Patel discusses objections involving continual learningContinual learning lets an AI system acquire new knowledge or skills over time while retaining what it learned earlier., research judgment, data and experimental compute. These are forecasts and contested arguments about future capabilities, not established outcomes.
Watch on YouTube
