I am not smart enough to know the answer to this, so I am wondering if someone can help me with this question: if Astra is able to decipher structures and patterns within model weights, would an LLM be able to directly modify its own weights on-the-fly to achieve recursive self-improvement using a reasoning loop?