Vitalik shared a privacy-preserving health recommendation experiment: a local model calls a remote model, while zkAPI and Tor protect information.
Article
Vitalik uses a local AI model combined with remote tool calls to create a three-layer privacy architecture for generating personalized health recommendations
PANews October 4 reported that Vitalik posted about a personal experiment using his own health and travel data with a cutting-edge model to generate personalized diet and exercise recommendations while avoiding disclosure of any private information to the remote model. The system uses a local model (Qwen 3.8 Flash Next) for orchestration, calling a powerful remote model as a tool to obtain higher-level reasoning and knowledge unavailable to the local model. It uses a three-layer privacy architecture: Identity layer: the local model (Qwen 3.8B) constructs queries on behalf of the user to avoid exposing their writing style and identity. Payment layer: zkAPI is used to hide payment information. Network layer: Tor is used to hide IP addresses. The system has run successfully and generated recommendations. Vitalik said the main shortcomings include: Tor is insufficiently optimized for per-request unlinkability and has high latency; the local model's speed is only 20-3...
PANews said that Vitalik posted about a personal experiment.
02
The experiment used health and travel data to generate personalized diet and exercise recommendations.
03
The system is orchestrated by a local model, with a remote model called as a tool.
04
The original text said that the architecture comprises three privacy-protection layers: identity, payment, and network.
05
The identity layer uses a local model to construct queries on behalf of the user; the payment layer uses zkAPI to hide payment information; and the network layer uses Tor to hide IP addresses.
06
The original text said that the system had run successfully and generated recommendations.
07
Vitalik said that Tor has insufficient optimization for per-request unlinkability and high latency.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This experiment lets the local AI model access private data first and organize the questions before calling a more capable remote model. The local model rewrites queries, zkAPI handles payment privacy, and Tor handles network address privacy, with the goal of not directly exposing users' private information to the remote model.
Why it matters to readers
The original text presented an approach to privacy protection that combines local data processing with remote model capabilities, explicitly involving identity, payment, and network components.
For beginners, the key point is that private data does not necessarily have to be sent directly to the remote AI; questions can first be organized locally, then additional tools can be used to reduce exposure of identity, payment information, and IP address.
Risks and unknowns
The end of the main text is truncated, so the local model's speed issue and other shortcomings cannot be fully confirmed.
The original text also mentions “Qwen 3.8 Flash Next” and “Qwen 3.8B”; the specific model names or versions are inconsistent.
Vitalik’s original post was not included in the materials, so the relevant technical details cannot be independently verified based on the materials provided.
The provided materials do not confirm whether this experiment is a publicly available tool or product.
Related Developments
Loading event timeline…
Related concepts
Local model
This term is not in the glossary yet. Browse related concepts in the glossary.
