It might also be a side effect of training models to help out with mechanistic interpretability, which I guess would help out a lot with recursive self-improvement.