ONG: The real path of the big model to self-improvement is to optimize Harness

The contribution by the chief scientist of Honolulu, Thinking Machines, states that it is unlikely that the starting point for self-improvement in a large model in the short term will be a direct rewriting of weights, and that a more realistic path is to optimize the peripheral system known as Harness. Harness has an operating system similar to that of a large model, which manages prompts, tool calls, traffic control and lasting memory. In the face of complex long-range missions, traditional static phrases are extremely prone to collapse. The current direction of evolution is to allow large models to act as meta-optimizers, to modify and re-engineer Harness's control stream code and to self-evolve the system. The real RSI closed the loop and had to solve at least three things: run long, change and verify。