@Jsevillamol - A conversation with @ryangreenblatt : can AI automate AI

A conversation with @ryangreenblatt : can AI automate AI R&D before 2030, and will this lead to a fast takeoff? Highlights: 1. We discuss whether we will achieve AI that can automate AI R&D work by 2030, and whether this could lead to a million-fold improvement in training efficiency within 1-2 years. 2. Both of us consider both claims plausible, but we disagree on how likely they are. 3. Ryan argues that after automating AI R&D, we will be able to run the equivalent of millions of copies of AI researchers, each of them operating 15x faster than humans, and each matching the best AI engineers and researchers. 4. I argue that returns to parallelization in research are low, and a large gap between AI that can automate R&D as well as the 50th percentile software engineer and a superhuman engineer. 5. We both consider each other's view plausible, but disagree on the likelihood of each subclaim, so that the most likely scenario in our heads ends up being quite different. We recorded this podcast on July 4th. Since then, we have seen the release of GPT-5 and a handful of other important AI developments. For a more recent update from Ryan on his views see this recent post of his: https://www.lesswrong.com/posts/2ssPfDpdrjaM2rMbn/my-agi-…