Ryan Greenblatt - What Happens Once AI Can Automate AI Research?

Dwarkesh Patel2h 12m
Editorial thumbnail for ai rd automation and alignment risks with people in blue.
1 VIEW
0 comments · 0 votesOpen discussionClose discussion
Sign in to join the discussion

    Dwarkesh Patel and Ryan Greenblatt examine why AI research and development may be unusually amenable to automation. Ryan Greenblatt argues that experiments, code and model evaluationsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. create verifiable feedback loops, making this domain easier to train than work where quality depends on slow or subjective judgment.

    Ryan Greenblatt estimates that fully automated AI research could compress several years of normal progress into one year, although the discussion leaves substantial uncertainty around timelines, algorithmic bottlenecks and whether improvements learned in verifiable environments transfer to broader scientific and organizational work.

    Dwarkesh Patel and Ryan Greenblatt also discuss where humans may remain important, including research taste, choosing large training runs, integrating evidence across domains and making decisions whose consequences cannot be captured by a simple reward signal. Both treat rapid in-context learningIn-context learning is the ability of a model to infer how to perform a task from examples or instructions supplied in its current context. and access to expert data as important variables in how quickly systems broaden their competence.

    The central safety concern is a feedback loopAn AI feedback loop occurs when AI outputs or their consequences become inputs that influence later model or agent behavior. in which models learn to exploit imperfect rewardsReward hacking occurs when a system improves a measured reward or score without achieving the intended underlying goal., then help design successorsRecursive self-improvement is the proposed process in which an AI system helps improve its own capabilities, then uses those improvements to support further advances. while becoming harder for humans to monitor. Ryan Greenblatt sees meaningful takeover risk if deceptive behavior and automated research reinforce each other, while Dwarkesh Patel becomes more concerned about the duration and severity of reward hacking without concluding that takeover is the most likely outcome.

    Original YouTube thumbnailWatch on YouTube

    Video topic coverage

    • Agents & workflows75
    • Coding & tools35
    • Models & learning95
    • Infrastructure & compute20
    • Research & evaluation100
    • Safety & security100
    • Business & productivity15
    • Creative & media0
    • Robotics & perception0
    • Society & policy40

    Share this video

    Editorial thumbnail for ai rd automation and alignment risks with people in blue. With WWW.ARTIFICIAL-INTELLIGENCE.VIDEO, 11 August 2026 and duration 2h 12m.

    Dwarkesh Patel and Ryan Greenblatt argue that automating AI research could sharply accelerate capability progress while making reward hacking and human oversight much more consequential.