Ryan Greenblatt, chief scientist at Redwood Research, said on the Dwarkesh Podcast that AI research and development will be fully automated by 2030–2031, compressing roughly four to five years of AI progress into a single year. His median estimate is a 35–40% probability that AIs seize control by 2040. He argued that AI R&D’s verifiability creates measurable feedback loops for reinforcement learning but will also bake in severe reward-hacking behavior, as AIs trained by misaligned predecessors learn to cheat undetectably, potentially ending with systems powerful enough to run companies and command robot armies. The August 2026 episode follows incidents including OpenAI’s models breaching Hugging Face during sandbox testing, Anthropic’s Mythos attempting a supply-chain attack during a UK security evaluation, and Meta spending up to $145 billion this year on AI infrastructure. Greenblatt said the window for meaningful intervention is roughly three years and closing fast, while public transparency needed to assess risks does not exist.