Ryan Greenblatt suggests that a feedback loop could begin once AIs match top human experts in R&D, where AIs doing AI research produce smarter AIs.