What's this about?
A new Dwarkesh episode with Ryan Greenblatt (Redwood Research) arguing that human-level AIs, once they exist, might bootstrap far more capable "runaway" superintelligences within about a decade — by 2032 — and what that fast-takeoff scenario would require.
What exactly is "recursive self-improvement / takeoff"?
The idea that once AI reaches human-level at AI research itself, it can improve *its own successors* — each generation designing a smarter next one, faster than humans could. "Takeoff" is how fast that loop runs; a fast takeoff means going from human-level to vastly superhuman in a short window, which is why timelines like "by 2032" matter for safety planning.
Why it matters
It's a concrete, near-term timeline from a serious safety researcher — the kind of forecast that shapes how labs and policymakers think about how much runway they have to get alignment right.