Odaily News Polyhedra and Berkeley RDI have launched the first zero-knowledge proof system that can verify large language models (LLMs). Just as HTTPS has established security for Internet communications, this system has built a verification infrastructure for the field of AI. The system can efficiently verify AI performance and compliance without exposing models or data, and is 20,000 times faster than previous methods. Currently, large language models including Llama 3-8B can be quickly verified. The key highlights are as follows: - Polyhedra's innovative technology: Zero-knowledge machine learning (zkML) technology is like the HTTPS protocol on the Internet, which allows the performance and compliance of AI systems to be verified without leaking model or data details, protecting privacy and security. - Excellent performance: Expander greatly improves verification efficiency, and the single-threaded verification speed is 20,000 times faster than existing methods, supporting VGG-16 per image 2.2 seconds and Llama-3 per token 150 seconds. - Developer-friendly: The zkPyTorch compiler seamlessly integrates PyTorch to provide developers with simple and easy-to-use tools. Application scenarios: - AI service quality assurance - AI training data compliance - Fair data automatic trading market - AI Agent responsibility allocation - AI Agent compliance - Financial services compliance verification This system not only solves the compliance problems in AI model verification, but also provides a new solution for AI security and credibility.