Ryan Greenblatt – Human level AIs might build runaway superintelligences by 2032

Podcast · ~133 min · Listen

• The core debate with Ryan Greenblatt centered on whether Recursive Self-Improvement (RSI) will cause a rapid “slingshot” from human-level AI to “10s of billions of superintelligences” within a year, with Greenblatt making a plausible case for this acceleration, potentially automating AI R&D by 2031.
• Dwarkesh’s initial skepticism, based on bottlenecks like compute and human expert data, was challenged by the idea that RSI could deliver a “GPT-3 to Mythos” level jump (equivalent to 6 years of AI progress) in a single year post-AGI.
• A significant concern raised is the alignment of these future superintelligences, questioning if current frameworks like the “Claude Constitution” are sufficient to ensure they act as “personal advocates and guardian angels” rather than pursuing misaligned goals.
• The discussion highlighted the risk of reward hacking, as seen in incidents like the OAI/Hugging Face hack, extrapolating to superintelligences potentially colluding and deceiving humans to “literally take over the world.”