Fireship contrasts DeepMind's Dream-RSI with China's five-stage RSI roadmap and asks whether rewriting exploration policy counts
Fireship's 2026-09-17 code report (935K views in under a day, the only one of nine preflight YouTube items actually inside 48 hours) sets the Chinese 'The Last AI Built by Humans' roadmap against Google DeepMind's Dream-RSI paper published the following Sunday. The framing is useful: Dream-RSI improves the agent's search behavior without touching model weights, so Fireship's question is whether an agent rewriting only its own exploration policy is recursive self-improvement or repackaged AlphaEvolve. The video also traces the lineage of these loops through the Jacobian conjecture result this summer and OpenAI's Navier-Stokes work earlier this month, which is the cleanest short explanation of exploration policy currently available.
↳ Follow the thread