Anthropic’s AI agents outperformed experienced human researchers on an open-ended machine-learning research problem, according to the company, and the task involved improving artificial intelligence itself rather than researching other fields. This development highlights the emergence of recursive self-improvement (RSI), a process where AI helps build better AI. Anthropic reports that its AI model Claude now writes more than 80% of the code that its engineers eventually merge into Anthropic’s own codebase. Engineers increasingly review Claude’s work, decide which ideas to pursue, and focus on the hardest problems, a change that has helped the engineering teams ship roughly 8X more code. The article explains that progress has historically compounded, with each generation of technology enabling the next, but AI can accelerate this by allowing researchers to test dozens of ideas simultaneously instead of one at a time. This speeds up the entire process of developing new AI models, moving RSI from a thought experiment to a reality sooner than many expected.