Lilian Weng argues near-term AI self-improvement runs through the harness, not the weights
Lilian Weng defines the harness as the system around a base model that orchestrates planning, tools, context, and evaluation, and argues near-term recursive self-improvement will come from evolving that layer rather than from models rewriting their own weights. No code or benchmark ships with it, but the three patterns are concrete enough to build against — plan-execute-test loops, the file system as durable memory instead of context, and sub-agent parallelism stored as files rather than transient chat. Weng names weak evaluators, context lifecycle management, and reward hacking as the open problems, and puts humans further up the stack rather than out of the loop.
Source: lilianweng.github.io ↗
the layer between the raw model and the real-world context seems to be as important as the model's raw intelligence
Why this matters
- → Near-term AI self-improvement comes through orchestration layers, not model weights
- → Harness engineering—planning, tools, context, evaluation—becomes the optimization target
- → Practical path avoids models rewriting themselves; keeps humans further up the stack