Ryan Greenblatt – What happens once AI can automate AI research?

Ryan Greenblatt, Chief Scientist at Redwood Research, joined the podcast to discuss recursive self-improvement in AI. This concept posits that once human-level AI is achieved, it could rapidly accelerate towards superintelligence, with each iteration being more competent than the top human experts across every field. Greenblatt argues that this is arguably the most important question in the world right now.

The Role of AI in AI Research and Development (AI R&D)

Greenblatt begins by highlighting that AI R&D is a task particularly well-suited for AI. Companies are actively working to make their AIs proficient in AI R&D, and the domain itself possesses favorable properties for AI development, such as being verifiable and allowing for iterative improvements. He suggests that once AIs can roughly match human experts in AI R&D, a feedback loop could be initiated. This loop would involve AIs conducting AI research, leading to smarter AIs, which in turn further enhance AI research. Greenblatt's median expectation is that this could lead to progress equivalent to four or five years of human AI development occurring within a single year. He emphasizes that this would require overcoming significant diminishing returns in research and achieving progress that would otherwise necessitate a massive increase in computational power. Greenblatt stresses that even three to five years of AI progress in a single year represents a substantial leap, drawing parallels to the rapid advancements seen in models like GPT-4 and the more recent Mythos-5.