"Helped build itself" is a bit of a stretch here, it makes it sound as if the model was doing lasting self-improvements. What the article describes is that the

"Helped build itself" is a bit of a stretch here, it makes it sound as if the model was doing lasting self-improvements. What the article describes is that the model was able to tweak to its own deployment harness (memory, skills, experimental loop etc) to improve performance on benchmarks. While impressive, it's not doing any modifications to its own weights by e.g. modifying the training code.