Vitalik Buterin has tested a system designed to obtain AI-generated recommendations while minimizing personal data exposure. The experiment used health and travel information to generate personalized diet and exercise suggestions, leveraging both local and remote AI models with privacy protections at each stage.
Three-Layer Privacy Approach
The system combined three complementary privacy mechanisms targeting different potential sources of identification.
The first layer addressed prompt content and writing style. A local model, identified as Qwen 3.8 Flash Next, composed questions and reduced the risk that remote services could identify the user through phrasing or personal details.
The second layer covered payments. The experiment used zkAPI, a system designed to separate payment authorization from user identity. According to an Ethereum Foundation announcement from October 1, zkAPI enables private usage credits where users fund a vault and authorize spending through zero-knowledge proofs rather than revealing which deposit paid for requests. Temporary API keys cap spending, while signed usage receipts determine actual charges.
The third layer targeted network information such as IP addresses. Buterin accessed zkAPI through a Tor-wrapped command-line tool, combining payment privacy with network routing.
Performance Gains and Identified Limitations
Buterin reported that recommendations benefited from knowledge provided by remote models. However, he identified four significant weaknesses in the current implementation.
Tor provided insufficient privacy for individual request separation, according to his assessment. Network latency was estimated at 10 to 100 times higher than theoretically possible. The local model generated roughly 20 to 30 tokens per second, below the 100-tokens-per-second threshold he considered necessary for comfortable usability.
The strategy for determining what information remote models should receive required further refinement. Additionally, a fundamental tradeoff emerged: withholding more context to enhance privacy reduced the quality of personalized recommendations from remote systems.
Broader Context
The experiment follows a September 27 essay in which Buterin described Ethereum's evolution toward a cryptographic world computer combining blockchain security with cryptographic privacy and verification. In that essay, he identified Hegotá, planned for the following year, as the likely last normal fork before broader verification advances and quantum-safe technology. He also cited PeerDAS as an early step toward this architecture.
