Today's frontier models outperform humans at some AI research engineering tasks

Over the course of a year, in a task testing a model's ability to speed up training for small AI model, Anthropic's models went from slightly underperforming a skilled human researcher to significantly outperforming them.