AI agents ace 79% of training-run checks but stumble on core software engineering
Yahoo
OpenAI measured how often its AI agents finish research tasks without help. From March to October, success on jobs needing two to four hours of human work rose from 23% to 46%, and on four-to-eight-hour jobs from 13% to 35%. Agents did best at monitoring training runs, at 79%, and struggled with coding for inference systems, at 25%.
Summary written with AI by HowTechMoves, based on reporting by Yahoo.