Odaily Planet Daily reports that Vitalik posted on X saying that he is conducting a self-experiment, planning to use personal health and travel data to obtain personalized diet and exercise recommendations from cutting-edge models, while also avoiding the leakage of private information to remote models. The solution uses the local model Qwen 3.8 Flash Next to coordinate, and calls a powerful remote model as a tool; the privacy approach includes having the local model draft the queries, isolating payment identity via zkAPI, and isolating the network and IP information through Tor.

Vitalik said the方案 has already been run and has returned recommendations, but Tor is not suitable for un-linking on a per-request basis, and the latency could be 10 to 100 times higher than ideal. He also said that the request construction strategy is still not ideal: Qwen 3.8 Flash Next is currently running at about 20 to 30 TPS, and only when it reaches above 100 TPS will it be noticeably faster. He added that the more cautiously the data is provided to the remote model, the less help the remote model can provide.