Vitalik Buterin said he is running a personal experiment that uses health and travel data to generate diet and exercise recommendations while trying not to expose private information to frontier AI models. In the setup he described, a local model, Qwen 3.8 Flash Next, acts as the orchestrator, while remote large language models are used as tools to make up for weaker reasoning and knowledge on the local side. The privacy design has three layers: the local model rewrites queries to reduce identity and writing-style leakage, zkAPI is used so payment rails do not reveal identity, and Tor is used to hide network metadata and IP addresses. Buterin said he has already received recommendations and that remote models improved the output. He also pointed to several shortcomings in the current design, including Tor being a poor fit for linking requests, privacy that may still fall short, latency running 10x to 100x above what it should be, suboptimal skill-file strategy, local-model throughput of roughly 20 to 30 TPS, and a trade-off where giving remote models less data also reduces how much they can help.
Vitalik Buterin said in a post on X that he is running a personal experiment to generate diet and exercise recommendations from health and travel data while trying to avoid leaking private information to frontier AI models.
In the setup he described, a local model, Qwen 3.8 Flash Next, handles orchestration, while remote large models are called as tools to compensate for the local model's weaker reasoning and knowledge.
A three-layer privacy design
Buterin said the privacy approach is split into three layers.
- The local model rewrites queries so identity details and writing style are less exposed.
- zkAPI is used so payment channels do not reveal identity.
- Tor is used to hide network information and IP addresses.
He added that a skill file guides the local model to build requests that leak as little data as possible, and uses a Tor-wrapped zkAPI as a command-line tool.
Early results and limits
According to Buterin, he has already received recommendations, and the remote model helped improve the results. He also listed several constraints in the current version.
Tor, he said, is not well suited to linking requests one by one, which means privacy may still be insufficient. It also introduces latency that is 10x to 100x higher than it should be. He added that the skill-file strategy is not yet optimal, the local model runs at about 20 to 30 TPS and is still slow, and the less data the remote model receives, the less useful its output becomes.
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.
