Vitalik combines local AI model with remote tool calls to build a three-layer privacy architecture for personalized health advice

PANews reported on October 4 that Vitalik shared a personal experiment on X, using his own health and travel data combined with frontier models to generate personalized diet and exercise advice while avoiding leaking any private information to remote models. The system uses a local model (Qwen 3.8 Flash Next) for orchestration, calling powerful remote models as tools to obtain higher-level reasoning and knowledge that the local model lacks. It adopts a three-layer privacy protection architecture:

• Identity layer: the local model (Qwen 3.8B) builds query requests on behalf of the user to avoid exposing identity through writing style
• Payment layer: uses zkAPI to hide payment information
• Network layer: hides the IP address through Tor

The system has run successfully and produced advice. Vitalik said the main shortcomings include: Tor is not optimized enough for per-request de-correlation and has high latency; the local model's speed is only 20-30 TPS (ideally 100+ is needed); the stricter the data protection, the more limited the help the remote model can provide.