AI systems like Fable are getting better at designing GPU kernels, a fundamental task in AI research and development. Fable achieved an 18.71X speedup by writing Cuda code, outperforming other attempts. This advancement hints at broader AI R&D automation capabilities. Additionally, AI systems are improving at automating online freelance projects, with success rates rising from 2.5% to 16.1% in less than a year. Why it matters: these developments signal significant progress in AI’s ability to build and improve itself, which could lead to recursive self-improvement and impact the economy.