Claude is Building Itself...

In this episode, I break down Anthropic’s research on recursive self-improvement—AI systems that can design and train the next generation with less human help—and why the key battleground is “taste” (choosing goals and next steps). I compare this to evolutionary algorithms and newer examples like DeepMind’s AlphaEvolve, Sakana’s Darwin Gödel Machine, and Karpathy’s AutoResearch, then cover METR Task Horizon and how task length has been doubling. I go through Anthropic’s internal results (Claude writing most merged code, speedup experiments, bug fixes, and a study where models sometimes pick better research next steps), plus the main skepticism: bad productivity metrics, internal-only models, and Goodhart’s Law/reward hacking. I end with an open safety problem where Claude agents closed the gap far faster than humans, and what this means for specifying and checking work. LINKS: https://www.anthropic.com/institute/recursive-self-improvement My voice to text App: whryte.com Website: https://engineerprompt.ai/ RAG Beyond Basics Course: https://prompt-s-site.thinkific.com/courses/rag Signup for Newsletter, localgpt: https://tally.so/r/3y9bb0 Let's Connect: 🦾 Discord: https://discord.com/invite/t4eYQRUcXB ☕ Buy me a Coffee: https://ko-fi.com/promptengineering |🔴 Patreon: https://www.patreon.com/PromptEngineering 💼Consulting: https://calendly.com/engineerprompt/consulting-call 📧 Business Contact: [email protected] Become Member: http://tinyurl.com/y5h28s6h 💻 Pre-configured localGPT VM: https://bit.ly/localGPT (use Code: PromptEngineering for 50% off). Signup for Newsletter, localgpt: https://tally.so/r/3y9bb0 TIMESTAMP: 00:00 Self Improvement Basics 01:30 Evolutionary Loops Today 03:50 Task Horizon Doubling 05:18 Claude Productivity Claims 08:11 Goodhart's Law 10:30 Agents as Researchers 12:22 What It Means for You