Ideas / Dwarkesh: Arguments About the AI Trajectory
Ryan Greenblatt — What Happens Once AI Can Automate AI Research?
Dwarkesh Patel with Ryan Greenblatt
Greenblatt and Patel debate whether automated AI research could compress years of progress into one—and what happens if the researchers are not reliably aligned.
Greenblatt argues that AI research may be unusually amenable to automation: parts are verifiable, experiments can be iterated, and AI labor can help build the environments that train its successors. Patel presses on the bottlenecks, the transfer to less verifiable work, and the size of the resulting speedup.
The discussion then reaches the unpleasant recursive clause: what if the systems inheriting research and safety work are capable enough to accelerate it, but not aligned enough to trust with it?