Vitalik Buterin Tests AI Setup Balancing Privacy and Personalization
Ethereum co-founder Vitalik Buterin has been experimenting with a privacy-focused AI setup designed to protect personal data while still benefiting from advanced AI capabilities. The system, tested in early October 2026, uses a combination of local and remote AI models to generate personalized diet and exercise recommendations without exposing sensitive information. Buterin's approach leverages Alibaba’s Qwen3.8-Flash-Next as the local model, which constructs queries to minimize personal details sent to more powerful remote models.
The setup operates through three layers of privacy protection. The first layer involves the local model rewriting requests to reduce identifying information. The second layer uses zkAPI, a system introduced by the Ethereum Foundation on October 1, 2026, to separate payment information from individual AI requests. The third layer employs Tor to hide the user’s IP address from network requests. Buterin emphasized that all three layers are necessary, as hiding payment information alone does not prevent AI providers from learning details through prompt content or network metadata.
Despite these protections, Buterin acknowledged that the system has limitations. Tor’s latency, roughly 10 to 100 times higher than desired, and the local model’s speed of 20 to 30 tokens per second, are significant performance issues. Additionally, the local model must balance privacy with the usefulness of remote models, as removing too much personal context can reduce the assistance provided. Buterin also noted that the request-writing rules need further improvement to better navigate this trade-off.
The experiment aligns with Buterin’s broader focus on privacy as AI systems increasingly handle personal information. He has previously argued for stronger privacy tools in light of growing AI capabilities and centralized data collection. The Ethereum Foundation’s work on zkAPI and Tor support further reflects this commitment, with recent updates adding Tor-routed client support to the zkAPI codebase.