Vitalik Buterin Describes Privacy-Focused AI Experiment for Personalized Health Advice

Vitalik Buterin said on Farcaster that he is running a personal experiment to generate personalized diet and exercise recommendations from health and travel data while preserving privacy through a combination of local and remote frontier models. According to ChainCatcher, the system uses a three-layer privacy architecture: a local model, Qwen 3.8 Flash Next, rewrites user queries to avoid exposing writing style, zkAPI hides payment information, and Tor masks IP addresses.Buterin said the system is already working, but he identified three shortcomings. Tor is inefficient and adds high latency when trying to de-link requests individually, the local model runs at only 20-30 TPS and would need more than 100 TPS for a smooth experience, and stricter data protection limits how much help the remote model can provide. He added that the related code has been submitted to the Ethereum zkAPI repository.