I think the idea is fundamentally improved architectures. For example, transformer-based models were an incredible stepwise improvement. Self improvement would be a model discovering a stepwise improvement similar to the transformer. And presumably the improved models from that would be more likely to make further advances still.

Learning from training data is technically self-improvement but not the sort that is typically meant in this context.