We pre-registered, ran, and verified the macro ablation: information per joule, measured
A new study on Macrokit, an open-source tool for LLM development, demonstrates that encoding workflows at design time significantly improves efficiency. The research found that this "macro-on" approach, compared to real-time workflow composition ("macro-off"), increased information density by 1.24–1.62x and compute efficiency by 2.0–5.1x for several local models. These findings validate predictions from a related mathematical theory on value and computation. AI
IMPACT Demonstrates a method to significantly improve LLM efficiency, potentially impacting how developers build and deploy AI applications.