"The last invention that man need ever make": The risk of self-improving AI

Researchers at Anthropic and OpenAI report AI model training is becoming increasingly automated, nearing recursive self-improvement capabilities. This automation could lead to increasingly capable AI systems that are harder for humans to control.