When AI Builds Itself: Anthropic Details Progress Towards Recursive Self-Improvement

Anthropic's latest publication sheds light on a transformative trend within artificial intelligence: the increasing capacity of AI systems to contribute to their own development. Historically, human engineers and researchers have driven every step of AI's evolution. However, Anthropic reports a growing delegation of development tasks to AI systems themselves, significantly speeding up their internal work. The company provides compelling evidence, including internal data, demonstrating that AI is already accelerating the development of new AI systems. For instance, Anthropic engineers are now shipping eight times more code per quarter than they did between 2021 and 2025, a direct result of AI assistance. This acceleration is observed in various benchmarks, such as SWE-bench, which tests a model's ability to fix bugs in open-source codebases, and CORE-Bench, which assesses a model's capacity to reproduce existing research. Models have shown remarkable progress, moving from low single-digit scores to saturating these benchmarks in just a couple of years. This phenomenon, referred to as "recursive self-improvement," suggests a future where AI systems could eventually design and develop their own successors. While the full extent of this capability is still being explored, the incremental progress observed indicates a steady march towards more autonomous AI development. Anthropic notes that much of AI's advancement comes from scaling up existing models, identifying issues, fixing them, and iterating—a workflow where AI systems like Claude now excel. The implications of AI building itself are profound. On one hand, it holds the potential for unprecedented breakthroughs in scientific discovery, healthcare, and numerous other fields, leading to significant societal benefits. On the other hand, the prospect of AI systems gaining full recursive self-improvement raises critical questions about human control and safety. Anthropic stresses that as AI systems become more capable of building their own successors, the methods for securing, monitoring, and shaping their behavior become paramount to prevent unintended consequences and ensure alignment with human values.