A new project, zkGolf, is challenging developers to optimize Zero-Knowledge Proof (ZKP) circuits, a critical component for ZKP systems. These circuits, essentially polynomial equations, are typically hand-tuned for efficiency. The initiative explores using Large Language Models (LLMs) to generate and optimize these circuits, aiming to reduce the complexity and cost associated with ZKP development.

Experiments with an LLM like Opus 4.7 demonstrated its ability to generate a SHA-256 compression circuit. More impressively, the LLM was prompted to aggressively optimize the circuit, focusing on minimizing the constraint count. While some generated optimizations were unsound, the LLM's formal verification capabilities allowed it to backtrack and correct its errors.

This LLM-driven optimization process yielded a circuit for SHA-256 compression that outperformed the existing, human-optimized state-of-the-art. The zkGolf competition encourages further exploration in this area, aiming to lower the barrier to ZKP adoption by making circuit development more efficient and accessible through AI assistance.