ETH creator reveals privacy-focused AI experiment

TLDR

Ethereum creator Vitalik Buterin is testing a privacy-focused AI setup that uses his personal data while trying to keep it hidden from powerful remote models.

1. Vitaliks experiment combines a local AI model, zkAPI payments, and Tor to get personalized health advice without exposing his identity or raw data to frontier models.
2. The system works but is still experimental, with high latency, limited speed, and the reality that prompts and metadata remain visible to API providers.
3. For crypto users, this showcases how Ethereum-based tools like zkAPI could become building blocks for private AI agents that pay for services on chain.

Deep Dive

1. How The Experiment Works

Vitalik Buterin described a personal experiment that feeds his health and travel data into a local model, which then selectively queries more powerful remote AI models for diet and exercise recommendations.

The local model is Alibabas Qwen 3.8 Flash Next, running on his machine and rewriting prompts to strip out personally identifiable details and his writing style before sending them on to remote frontier models. As reported by U.Today, those remote models improved the recommendations while the sensitive source data stayed local to Vitaliks setup.

On the payments side, he uses the Ethereum-based zkAPI system, wrapped with Tor, so that API providers see metered usage but not a straightforward link between his wallet and each individual request.

2. Privacy Tradeoffs And Technical Limits

The architecture uses three layers: the local model to mask content and style, zkAPI to hide payment identity, and Tor to add network and IP privacy, which Vitalik summarized as you need all three in his post and Farcaster comments.

However, this is not perfect privacy. TokenPost notes that API providers can still see prompts, and network metadata and timing information can remain observable, meaning sophisticated correlation is still possible.

Vitalik also highlighted performance issues: Qwen runs at about 2030 tokens per second locally, while more than 100 tokens per second would feel smooth, and Tor adds roughly 10100 times more latency than he considers acceptable for request-by-request unlinking. Tight privacy rules also reduce how helpful the remote models can be.

3. Why It Matters For Ethereum And AI

The experiment builds directly on zkAPI, an Ethereum mainnet, private metered-API payment system coauthored by Vitalik and Davide Crapis and implemented with Open Anonymity, which Cryptoslate reports went live on October 1.

This points to a future where AI agents can pay for model access using Ethereum while minimizing the link between wallet identity and each API call, potentially important for health, finance, and other sensitive use cases. It also shows Ethereums role as infrastructure for privacy-preserving computation, not just simple token transfers.

Conclusion

Vitaliks privacy-focused AI experiment is a live test of how Ethereum-native tools can protect identity and payments while still tapping powerful remote models. It is far from plug-and-play, with clear latency and privacy tradeoffs, but it offers an early blueprint for how crypto rails and AI systems might combine to deliver more private, agent-driven services in the future.