Ryan Greenblatt – What happens once AI can automate AI research?

Ryan Greenblatt, chief scientist at Redwood Research, a technical AI safety group, argues that automating AI research could compress years of progress into one. On Dwarkesh Patel's podcast, Greenblatt says that once AI matches top human experts at AI research, a feedback loop could start in which AI builds better AI.

The reasoning rests on verifiability. Greenblatt says much AI research can be turned into containerised tasks with a measurable score, such as getting a small model to a target training loss faster, and models can be trained with reinforcement learning on many of them. Greenblatt compares this with mathematics, where verifiable problems have yielded real progress, and argues machine learning gives clearer intermediate signals and gains that tend to stack.

The stated median expectation is full automation of AI research around 2030 to 2031, then four or five years of AI progress in a single year, and AI that beats humans at essentially any job around 2033. These are forecasts, and Greenblatt concedes that transfer from small tasks to the hardest research is uncertain.