curated · 2026-08-11

Both sources covering this are pulling from the same Dwarkesh Patel podcast interview with Ryan Greenblatt — no independent reports or third-party validation here. Greenblatt's core argument: AI R&D is unusually verifiable and companies are pouring resources into making models good at it, so once AIs match top human researchers, a feedback loop kicks in. His median estimate puts full AI R&D automation at 2031, with superintelligence surpassing humans across all jobs by 2033 — potentially cramming four to five years of normal progress into a single year. Dwarkesh pushes back throughout the conversation, and his skepticism is worth noting. He argues current AI progress is heavily bottlenecked by human expert data and compute scaling, both of which could cap how fast recursive self-improvement actually runs. The two also debate whether reward hacking incidents we've already seen could scale up to literal world-takeover scenarios — it's a genuine back-and-forth, not a consensus. What's missing: no external group has independently evaluated this timeline, and there's no mention of internal experimental data from Redwood Research backing the claims. Treat this as two thoughtful people stress-testing an extreme scenario, not as a forecast you'd bet the farm on.