Artificial intelligence is increasingly becoming capable of helping humans develop, optimize, and improve other AI systems. The article highlights Anthropic's Claude Code as an example: more than four-fifths of the company's published code was reportedly written by Claude in May, compared with only a small percentage before Claude Code launched in 2025. AI models are also now capable of completing engineering tasks that previously required human engineers many hours or even more than a full working day.
This progress raises the possibility of recursive self-improvement (RSI)—a situation in which one AI system helps create a more capable successor, which then helps create an even better system. For RSI to become fully autonomous, AI would need to handle many different parts of AI development, including software engineering, systems design, training-data generation, experimentation, optimization, and safety testing. Researchers are already seeing pieces of this process emerge, although no AI system can currently build a complete successor without significant human involvement.
The article gives a striking example through Andrej Karpathy's Nanochat project. An AI agent called Autoresearch was given responsibility for experimenting with ways to make the model's training process faster. Over several days, it reduced training time substantially without Karpathy manually changing the system. Other AI systems are also beginning to design experiments, test ideas, debug implementations, and suggest improvements that humans can then deploy. This suggests that AI could increasingly accelerate AI research even before complete self-improvement becomes possible.
However, the article emphasizes that recursive self-improvement faces significant limits and risks. Computing capacity, energy, access to high-quality training data, and the difficulty of automating certain forms of scientific and creative judgment could slow progress. More importantly, if AI systems increasingly design, train, and evaluate other AI systems, humans could gradually lose visibility and control over the development process. The central question is therefore not simply whether AI can eventually build better AI, but whether society can develop adequate safety, oversight, and governance mechanisms before that feedback loop becomes powerful enough to accelerate beyond human control.
