According to CoinMeta, Weng Li, the chief scientist of Thinking Machines, wrote that in the short term, it is unlikely that large models will improve themselves by directly rewriting their weights. A more realistic approach is to optimize what is known as the peripheral system, Harness. Harness is similar to an operating system for large models, responsible for managing prompts, tool calls, control flow, and persistent memory. In the face of complex, long-term tasks, traditional static prompts are highly prone to failure. The current evolutionary direction is to make large models act as meta-optimizers, autonomously modifying and reconstructing the control flow code of Harness to achieve self-evolution of the system. Weng Li emphasizes that a true closed-loop for RSI must address at least three issues: being able to run for a long time, being able to make changes, and being able to verify those changes accurately.