Thinking Machines co-founder and chief scientist Lilian Weng argues that in the short term, AI self-improvement probably won't start with models directly rewriting their own weights. The more realistic path, she writes, is optimizing the peripheral system she calls the Harness.
The Harness is essentially the model's operating system. It manages prompts, tool calls, control flow, and persistent memory. For complex long-horizon tasks, traditional static prompts tend to fall apart.
The emerging direction: let the model act as a meta-optimizer, autonomously modifying and rebuilding the Harness's control-flow code — enabling the system to evolve itself.