Vitalik: Obtain Personalized Health Advice via Local Models, zkAPI, and Tor

Odaily reports that Vitalik posted on X platform that he is conducting a self-experiment, planning to use personal health and travel data to obtain personalized diet and exercise advice through frontier models, while avoiding leaking private information to remote models. A local Qwen 3.8 Flash Next model is used for coordination, with powerful remote models called as tools; the privacy scheme includes having the local model draft queries, isolating payment identity through zkAPI, and isolating network and IP information through Tor.

Vitalik stated that the scheme is already running and returning suggestions, but Tor is not suitable for per-request unlinkability, with latency potentially 10 to 100 times higher than ideal; the request construction strategy is still suboptimal, and Qwen 3.8 Flash Next currently runs at approximately 20 to 30 TPS, and would only be noticeably faster above 100 TPS. He also stated that the more cautious the data provided to remote models, the less help remote models can offer.