A new line of work suggests large language models can read and write the numerical weights of neural networks directly, a technique described as "Weight Space Programming." The approach is presented as a way to remove the autoresearch step from recursive self-improvement (RSI) loops, where AI systems refine AI systems. Instead of running experiments to find better model designs, the model edits the weights themselves. The same episode covers a demonstration of editing the weights of Alibaba's Qwen model to reduce repetitive reasoning loops, in which a model gets stuck cycling through the same steps. It also reports on a prototype "super-agent" from DeepSeek and the GDPx benchmark. The show notes do not include benchmark scores or full methodology, so the scope of these results should be checked against the original sources before they are relied on. For Australian businesses and professionals, this is research-stage work and changes nothing in day-to-day operations. There is no new cost, compliance obligation, or product to adopt. Teams that fine-tune or host open-weight models should keep an eye on the area, and any use of Chinese-developed models such as DeepSeek or Qwen should go through the usual vendor, security and data-handling checks, including obligations under the Privacy Act where personal information is involved.