Ethereum co-founder Vitalik Buterin announced a self-experiment: generating diet and exercise recommendations using personal health and travel data, utilizing cutting-edge models without directly sharing private information with them. The orchestrator is the local model Qwen 3.8 Flash Next, while remote strong models are only used as tools to supplement the high-level reasoning and knowledge lacking in the local model.

Three layers separate identity. The first layer blocks identifiable information and writing style: queries are written by the local model, not by him personally. The second layer blocks payment channels: using zkAPI. The third layer blocks network and IP: using Tor. He specified that all three layers are necessary, and at least to some extent, all three are currently in place. A skill file teaches the local model when and how to construct requests that expose minimal data, with a command-line tool that includes Tor and zkAPI.

The experiment ran successfully. Recommendations were returned, and the information from the cutting-edge model improved the results. He did not disclose the name of the remote model, nor did he provide an independent security audit. There are four flaws: Tor has not optimized for unlinkability of successive requests, which he believes is the only reasonable network-layer privacy today; long-lived identifiers are too fragile; this method may not be private enough, with delays 10 to 100 times higher than expected; and the request construction in the skill file is far from optimal. The local model runs at 20 to 30 tokens per second, and it only feels fast at over 100 tokens per second. The less data given to the remote model, the less help the remote model can provide.

zkAPI is a payment layer introduced by the Ethereum Foundation on October 1 by Vittorio Rivabella of the dAI team, designed from a proposal by Davide Crapis and Vitalik Buterin on the Ethereum research forum, and implemented by the Open Anonymity team as a client, server, and contract. The usage involves depositing a quota into the Ethereum vault once, and then using zero-knowledge proofs to authorize limited usage without revealing identity. The server only knows that a valid payment exists and how much each session costs, without knowing who you are, what you asked, or which deposit is being used for payment. The model provider can see prompts and responses because the model runs on their side, but cannot see who is making the payment. What is publicly available on-chain are deposits, closures, and withdrawals, not which balance purchased what.

The client exposes an interface compatible with OpenAI and Ollama on the local machine. The vault is an Ethereum contract, not a corporate account; even if the server disappears, the balance can still be closed and withdrawn on-chain. Measurement units can include model calls, blockchain node queries, image and video tasks, bandwidth periods, or one-time machine-to-machine expenses. The post has visible interactions of approximately 1,724 likes, 330 replies, and 329,000 views.

This is a payment and network isolation for personal data calls, not a new model release. The buyer is a user who does not want to share health and travel records along with their identity with a remote lab, while the seller remains the provider running cutting-edge models. Money is first deposited into the Ethereum vault, and then calls are made using proofs instead of account numbers. The beneficiaries are those who want remote knowledge but do not want their bills and IPs linked to their identities; the burden is on interfaces that tie usage, prompts, and identity together through accounts, credit cards, and IPs. The remote side still receives rewritten prompts, and the successive unlinking of Tor is also noted as potentially insufficient.

The local model is responsible for deciding what the remote needs and then rewriting the requests. He did not disclose the fields of health data or the specific content of the recommendations.

Source: Public Information