Ethereum co-founder Vitalik said in a post on X that he is running a self-experiment aimed at getting personalized diet and exercise recommendations from advanced AI models without exposing private information to remote systems. The setup uses a local Qwen 3.8 Flash Next model as the coordinator, while more powerful remote models are called as tools. To reduce privacy leakage, the design relies on the local model to draft queries, zkAPI to separate payment identity, and Tor to isolate network and IP information.
Vitalik said the system is already working and has returned recommendations, but he also outlined several trade-offs. According to his post, Tor is not well suited for unlinking requests one by one, and latency can run 10x to 100x above the ideal level. He added that the current request-construction strategy is still not good enough. Qwen 3.8 Flash Next is currently running at about 20 to 30 TPS, while speeds above 100 TPS would make the setup feel meaningfully faster. He also noted a basic constraint of the approach: the more carefully users limit what remote models can see, the less help those remote models can provide.
Odaily reported that Vitalik said in a post on X that he is running a self-experiment using personal health and travel data to obtain personalized diet and exercise advice from advanced models while avoiding the disclosure of private information to remote models.
He said the setup uses a local Qwen 3.8 Flash Next model as the coordinator, with stronger remote models invoked as tools.
Privacy design
According to Vitalik, the privacy setup includes having the local model draft queries, using zkAPI to isolate payment identity, and using Tor to isolate network and IP information.
Current performance and bottlenecks
Vitalik said the system has already run and returned recommendations. Still, Tor is not suitable for per-request unlinkability, and latency may be 10x to 100x higher than the ideal level. He also said the request-construction strategy is still not ideal. Qwen 3.8 Flash Next is currently running at around 20 to 30 TPS, and the system would feel noticeably faster only once it reaches more than 100 TPS.
He added that the more carefully data is withheld from remote models, the less help those remote models can provide.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
