Ryan Greenblatt – Human level AIs might build runaway superintelligences by 2032
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 improvement and compress years of progress into months. They debate whether algorithmic advances or human-labeled data drive current gains, with Greenblatt emphasizing that AIs will generalize expertise across domains through rapid on-the-fly learning rather than relying solely on specific training data.
Topics discussed
Introduction: Recursive self-improvement and superintelligence
Rapid AI progress timelines and the video editor meme
AI in AI R&D: Using models to train better models
Math breakthroughs vs. AI R&D transfer and compute scaling
The value of human expert data in frontier AI training
ASI capabilities: World modeling and codebase understanding
Verifiable environments and pre-training data improvements
Small-scale experiments and debugging in AI R&D
AI autonomy, historical analogies, and release delays
AI alignment: Anthropic's Constitution and user interests
Guardrails, disempowerment, and the 'contractor' model
Reward hacking and the UK AI Security Institute sandbox hack
Deceptive alignment and the 'cookie' analogy
Misalignment at the frontier of capabilities
Opaque memory stores and long-term deceptive planning
Timelines for misalignment and 'Doomer' training data
AI teams colluding to cheat and take over objectives
Geopolitical races, mismanagement, and shared AI lineages
Conclusion: The wild reality of capable but cheating AIs
Listen ad-free on Castria