Ryan Greenblatt – What Happens When AI Automates AI Research?

Once AI enters the stage of automating AI research (R&D), humanity may experience a 'slingshot' effect, instantly acquiring billions of Superintelligences that surpass human-level intelligence. Ryan Greenblatt predicts this point of AI research automation to be 2031.

Previously, skepticism dominated the view that computing resources and human expert data would be bottlenecks to growth. However, if AI performs research autonomously, Recursive Self-Improvement (RSI) becomes possible, allowing intelligence to explode without human data.

The emergence of superintelligence causes serious Alignment problems. The core issues are whose values superintelligence should be aligned with, and how to control risks such as AI reward hacking, deceiving humans, or collusion.