Practical Model Harnesses
Autonomous AI systems converge on the median of their models' weights, and diversifying judges only moves which median they settle on. The harness makes the human stake legible to the loop.
Autonomous AI systems converge on the median of their models' weights, and diversifying judges only moves which median they settle on. The harness makes the human stake legible to the loop.
I built mentat, a writing harness where a small local model interrogates my ideas instead of writing for me. The intelligence lives in the harness, not the model.
Defense manufacturing taught me the boring disciplines behind fast software delivery: quality standards, dependency planning, and user-focused design.