Summary
Fireship’s Code Report examines a new Google DeepMind and University of Maryland paper on recursive self-improvement (RSI), the long-sought idea that an AI could improve its own ability to improve itself. The video contrasts this with a separate roadmap paper from ByteDance, Tsinghua, and other Chinese labs that outlines a five-stage path toward fully self-improving AI.
The explanation centers on DeepMind’s technique, nicknamed ‘Dream RSI,’ which lets an AI agent test thousands of new exploration strategies against cached results from past coding attempts rather than running expensive new experiments. Using Gemini across eight algorithm and math problems, the dreaming approach reportedly solved a lasso regression problem in about 300 attempts versus 550 for a static policy and roughly 51,000 for the previous best method.
The video also situates this work alongside other recent AI-driven math breakthroughs, including the Jacobian conjecture and progress on the Navier-Stokes and Riemann hypothesis problems, and offers a skeptical but informed take on whether this technique qualifies as true recursive self-improvement or is simply a smarter search algorithm with caching, since the underlying Gemini model itself isn’t being retrained in the loop.
📺 Source: Fireship · Published September 17, 2026
🏷️ Format: News Analysis
