The paper introduces opML (Optimistic Machine Learning on Blockchain), an innovative approach that integrates machine learning with blockchain technology to enable decentralized, secure, and transparent AI services. opML leverages an interactive fraud proof protocol, similar to optimistic rollup systems, to ensure decentralized and verifiable consensus for ML services. Unlike zkML (Zero-Knowledge Machine Learning), opML offers cost-efficient and highly efficient ML services with minimal participation requirements. opML can execute extensive language models, such as 7B-LLaMA, on standard PCs without GPUs, significantly expanding accessibility. Key contributions of opML include: 1. **Cost-Efficiency and High Efficiency**: opML provides ML services with low cost and high efficiency, making it more practical for large models compared to zkML. 2. **Multi-Phase Fraud Proof Protocol**: opML introduces a novel multi-phase fraud proof protocol with semi-native execution and lazy loading design, improving the efficiency of the fraud proof process and overcoming memory limitations. 3. **Security and Liveness**: opML maintains safety and liveness under the AnyTrust assumption, where any honest validator can force opML to behave correctly. 4. **Incentive Mechanism**: opML addresses the Verifier's Dilemma by introducing an Attention Challenge mechanism, ensuring that rational validators will always check results and rational submitters will never cheat. The paper also discusses the architecture of opML, including the Fraud Proof Virtual Machine (FPVM), the Machine Learning Engine, and the Interactive Dispute Game. It provides a detailed explanation of the multi-phase dispute game, which enhances the efficiency and effectiveness of the fraud proof protocol. Additionally, the paper explores potential optimizations and applications of opML, such as training and fine-tuning processes, and integrates it with zkML for enhanced privacy. Related work in AI computation on blockchain and zero-knowledge machine learning is reviewed, highlighting the limitations and advancements in the field.The paper introduces opML (Optimistic Machine Learning on Blockchain), an innovative approach that integrates machine learning with blockchain technology to enable decentralized, secure, and transparent AI services. opML leverages an interactive fraud proof protocol, similar to optimistic rollup systems, to ensure decentralized and verifiable consensus for ML services. Unlike zkML (Zero-Knowledge Machine Learning), opML offers cost-efficient and highly efficient ML services with minimal participation requirements. opML can execute extensive language models, such as 7B-LLaMA, on standard PCs without GPUs, significantly expanding accessibility. Key contributions of opML include: 1. **Cost-Efficiency and High Efficiency**: opML provides ML services with low cost and high efficiency, making it more practical for large models compared to zkML. 2. **Multi-Phase Fraud Proof Protocol**: opML introduces a novel multi-phase fraud proof protocol with semi-native execution and lazy loading design, improving the efficiency of the fraud proof process and overcoming memory limitations. 3. **Security and Liveness**: opML maintains safety and liveness under the AnyTrust assumption, where any honest validator can force opML to behave correctly. 4. **Incentive Mechanism**: opML addresses the Verifier's Dilemma by introducing an Attention Challenge mechanism, ensuring that rational validators will always check results and rational submitters will never cheat. The paper also discusses the architecture of opML, including the Fraud Proof Virtual Machine (FPVM), the Machine Learning Engine, and the Interactive Dispute Game. It provides a detailed explanation of the multi-phase dispute game, which enhances the efficiency and effectiveness of the fraud proof protocol. Additionally, the paper explores potential optimizations and applications of opML, such as training and fine-tuning processes, and integrates it with zkML for enhanced privacy. Related work in AI computation on blockchain and zero-knowledge machine learning is reviewed, highlighting the limitations and advancements in the field.