#特朗普支持CFTC管辖预测市场 Vitalik Buterin shares the latest developments in his autonomous large language model (LLM) setup and points out that the intersection between Ethereum infrastructure and AI is continuously expanding. He mentions that Deepseek V4 has been released, with its 2-bit quantized version capable of running within 90GB of memory, achieving about 35 tokens/second on Apple hardware, but only around 7 tokens/second on AMD, emphasizing that supporting multiple hardware vendors is key to distinguishing “decentralized AI” from “CROPS AI.” Additionally, Mistral’s Leanstral model (focused on Lean code writing) can run within 70GB and performs on par with large models of 1 trillion parameters.

Vitalik also elaborates on the role of formal verification in enhancing code security, believing that AI-assisted formal verification can achieve “end-to-end” security proofs for code, applicable to core components like STARKs, consensus algorithms, and EVM. He points out that blockchain and ZK-SNARKs offer open verifiability and privacy scalability, while the combination of AI and formal verification can improve coding efficiency and rebuild accuracy, forming a complementary technology stack. Vitalik calls for the Ethereum ecosystem to fine-tune models for Ethereum-related use cases and promote efficient support across multiple hardware platforms.