When AI Builds Itself: Progress Toward Recursive Self-Improvement

Anthropic stated that the driving force behind AI development is shifting from humans to AI systems, moving toward a stage of Recursive Self-Improvement in which AI designs its own successor models.

Currently, Anthropic engineers are actively leveraging AI assistance, deploying an average of 8 times more code per quarter compared to the past.

The stages of AI capability development are evolving as follows:

• 2023–2025 (Chatbots): Generating and copying short code snippets.
• 2025–2026 (Coding Agents): Writing and editing code themselves, performing file-level tasks.
• Present (Autonomous Agents): Directly executing code and delegating tasks to other agents.
• Future (Completing the Loop): Agents directly building and training models to improve themselves.

AI's task-performing capability is accelerating rapidly. The length of tasks AI can complete on its own is doubling roughly every 4 months, and model performance is also rapidly reaching saturation on major benchmarks such as SWE-bench (software engineering) and CORE-Bench (research reproduction).

This technological progress could bring enormous benefits to science and healthcare, but it also carries the risk of humans losing control over AI systems, making security and monitoring increasingly important.