An Anthropic researcher just gave us a peek at self-improving AI | TechCrunch

The team led by Anthropic researcher Chen Yuehan published the paper “Automated Researchers Can Reliably Mitigate Alignment Failures,” which demonstrates early practices of self-improvement using other AI models. In 10 specific behavior alignment benchmarks, the system improved all individual performance without reducing overall performance. The Automated Alignment Researcher (AAR) led by Chen Yuehan effectively identified efficient solutions through literature search, method development, and model training. The paper indicates that the best AAR methods outperformed directions proposed by human experts in an average of six hours, with an API inference cost of about 4 dollars per hour, far lower than 150 dollars per hour for human researchers. Although the benchmarks have limitations in reflecting actual alignment goals and literature maintenance, the results prove that automated alignment makes training近期 feasible, promoting a recursive self-improvement process and potentially leading to the gradual replacement of human AI researchers.