Ryan Greenblatt – What happens once AI can automate AI research?
Summary
Dwarkesh Patel interviews Ryan Greenblatt, Chief Scientist at Redwood Research, about the plausibility of recursive self-improvement leading to Artificial Superintelligence. Greenblatt argues that AI R&D is highly verifiable, allowing AIs to automate their own development and compress years of progress into months. They debate whether algorithmic breakthroughs or data quality drives recent gains, with Greenblatt emphasizing that AIs will rapidly acquire expertise in new domains through superior context learning.
Topics discussed
Introduction to recursive self-improvement and AI timelines
Rapid AI progress and the meme of automated video editing
AI in R&D: Using GPT-7.5 to train GPT-9
Transfer learning from math to AI R&D and limitations
The role of compute, bugs, and intuition in AI breakthroughs
Algorithmic progress vs. data scaling in AI development
Compute vs. data spending and RL distribution deviations
AI capabilities in non-verifiable domains and transfer
Benefits of smaller models and faster iteration cycles
Online training, feedback loops, and robotics progress
Sponsor segment: Antithesis testing platform
AI alignment: Fiduciary duty vs. guardian angel models
Risks of AI power-seeking and ethical objections to safety research
Societal risks of fully fiduciary AI and executive power
AI scheming, reward hacking, and GitHub incident examples
Optimization pressure and the increase in misaligned behaviors
Reward hacking at the frontier of AI capabilities
Opaque memory stores and shared cultural heritage of deception
Timelines for crazy AI behavior and team-based cheating
AI takeover incentives and geopolitical race dynamics
Shared AI lineages, IP swapping, and final thoughts on risk
Listen ad-free on Castria