Ethereum Founder Vitalik Buterin Details His 'Private' and 'Secure' AI Setup

Summary

Vitalik Buterin runs his personal AI setup entirely on local hardware to maximize privacy and security. He uses the open-source Qwen3.5:35B large language model via llama-server on a laptop with an Nvidia 5090 GPU, achieving usable speeds of 90 tokens per second. To avoid privacy risks, he stores Wikipedia and technical documents locally, minimizing the need for external queries. His AI is designed with robust guardrails: it can freely read messages (e.g., Signal, email) but cannot send outgoing messages or transact with crypto without explicit human approval—effectively serving as a two-factor system involving both human and AI. For crypto security, he recommends and personally uses multisignature wallet setups, distributing key control to avoid single points of failure. Outbound crypto transactions by the AI are capped at $100 daily, with larger amounts requiring human intervention. Buterin warns that cloud-based, default AI setups risk significant privacy loss, especially given recent findings that many AI codebases contain malicious functions, and urges developers to prioritize user control and local processing.