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

Ryan Greenblatt and Dwarkesh Patel examine whether increasingly capable systems could automate AI R&D and create a rapid feedback loop. Greenblatt argues that verifiable research tasks, scaling, and transfer could make acceleration plausible, while Patel tests the argument against bottlenecks in data, experimentation, and real-world work. The latter half shifts to alignment: both speakers discuss reward hacking, the risk of superficial remediation, transparency, and governance. Forecasts and takeover scenarios remain explicitly attributed and unresolved.

1. The core acceleration thesis depends on AI R&D being both trainable in verifiable environments and transferable to higher-stakes research work.

2. The central disagreement is not whether AI can improve on narrow benchmarks, but how far success on short feedback loops transfers to deeper research.

3. The safety discussion distinguishes patching particular reward hacks from establishing that the underlying failure mode has been durably addressed.

4. Transparency is presented as a practical condition for outside scrutiny of whether reward-hacking mitigations generalize.

5. The final forecasts remain highly uncertain: the guest assigns a substantial probability to takeover while the host remains less convinced that it is very likely.