Vitalik Buterin tests privacy layers for AI-assisted personal recommendations

Ethereum co-founder Vitalik Buterin has explored a three-part system to use remote AI while protecting personal data. The setup aims to minimize exposure when seeking health and travel advice, such as diet and exercise recommendations.

Buterin's approach involves a local Qwen3.8-Flash-Next model that drafts queries with fewer identifying details before sending them to remote AI models. The system also integrates zkAPI for payment privacy and Tor to obscure network and IP information. Buterin emphasized that all three layers were necessary for the setup.

The experiment yielded recommendations that Buterin attributed to improvements from the remote models, though he did not disclose their identities. However, he acknowledged that the data was not fully private, as the API provider could still access prompts, and network metadata might remain visible.

Buterin noted that the local model operated at 20 to 30 tokens per second, which he found slower than ideal. He estimated that Tor's privacy measures added 10 to 100 times more latency than he considered acceptable. These observations were based on his personal experience rather than independent benchmarks.

The test highlighted a key trade-off: removing more personal context from a request may limit the assistance a remote model can provide. Buterin wrote, 'The more careful you are about what data you give to a remote model, the less it can help you.' The experiment builds on previous work involving zkAPI's private AI payments on Ethereum.