I call BS on this. For a LLM to recursively improve itself it would need to (small step) improve the training data and/or (big step) come up with fundamentally new architectures superior to transformers. The small step improvements might be doable. But nobody is making any claims about the big step improvements.