Claim◆Audio · 0:00 — 11:00

AI R&D is a uniquely verifiable and containerizable domain, so once AIs match top human experts it can kick off a self-improving feedback loop yielding roughly four to five years of AI progress per single year.

Ryan argues AI R&D is especially well-suited to AI training because it is verifiable and containerizable, and that once AIs match top humans this could produce a feedback loop with a median expectation of four to five years of progress in one year. ✦ AI generated

Ryan Greenblatt · Dwarkesh Podcast · 2026-08-11 · original ↗

Share card ↓Export .md Export .txt

plays this moment only · 0:00 — 11:00

Elicited by

“I want to talk to you about recursive self-improvement... And historically, I've been quite skeptical that this kind of thing happens, but, you seem to think that it might be plausible, and so I wanted to hear the case for it.”

I think once you have AIs which are roughly matching the top human experts in AI R&D, that could kick off a feedback loop where the AIs are doing AI research. That produces smarter AIs. That feeds back in. That feedback loop could be strong enough that you end up with a lot of progress in a short period of time. Maybe my median expectation is something like four or five years of AI progress in a single year. This requires really overcoming a huge amount of diminishing returns in research and basically doing the equivalent of the progress we would have gotten after a really large compute scale-out.

verbatim transcript · starts at 0:00

Transcript · around this moment

0:00– Is AI R&D verifiable enough to unlock recursive self-improvement?

▦ Redwood Research ◉ Ryan Greenblatt

Around this claim

This moment responds to

explains mechanism → Once AIs roughly match top human experts in AI R&D, could kick off a feedback loop where AIs doing AI research produce smarter AIs, yielding four or five years of AI progress in a single year.Ryan Greenblatt · Dwarkesh Podcast supports → AI progress is not bottlenecked by expert human data; what matters is compute, the science of constructing RL environments, and algorithmic improvements, so automating AI R&D does not require replicating the human data industry.Ryan Greenblatt · Dwarkesh Podcast supports → Even if skills crucial to real-world domains are hard to train for, radical world transformation is achievable through verifiable R&D alone—spanning chips, fabs, robots, and further AI development—without needing AIs to master politics or negotiating.Ryan Greenblatt · Dwarkesh Podcast supports → Recursive self-improvement and continual learning, combined with the work of singular talents like Andrej Karpathy, could pull the future of AI forward in a very real way, potentially making annual order-of-magnitude improvements seem conservative.Gavin Baker · All-In Podcast supports → Recursive self-improvement, where an AI model has input into its own training, could put model improvement on a parabolic trajectory resembling a new form of Moore's law.Chamath · All-In Podcast