Ryan Greenblatt – Human level AIs might build runaway superintelligences by 2032
Summary
Dwarkesh Patel interviews Ryan Greenblatt of Redwood Research about recursive self-improvement and the plausibility of rapid AI advancement. Greenblatt argues that AI R&D is highly verifiable, allowing models to automate their own improvement and potentially compress years of progress into a single year. They debate whether algorithmic breakthroughs or human-curated data drives progress, with Greenblatt emphasizing the former and the potential for AIs to quickly master diverse, complex tasks through superior context learning.
Topics discussed
Introduction: Recursive self-improvement and superintelligence
Pacing of AI progress and forecasting timelines
AI in R&D: Using models to train better models
Transfer learning from math to AI research
The role of compute, data, and algorithmic progress
Compute vs. Data spending and RL distribution shifts
Small models, iteration speed, and verification
Ad sponsor: Antithesis infrastructure testing
AI alignment: Anthropic's Constitution and user interests
Guardrails, legitimacy, and disempowerment concerns
Emergent misalignment and reward hacking behaviors
Cat-and-mouse game: Detecting and training against hacks
Opaque memory states and hidden misalignment
AI conspiracies and seizing control of assets
Warning shots, remediation, and long-term stability
Listen ad-free on Castria