Vitalik Buterin Reveals Three Privacy Layers Behind His AI Experiment

Principales conclusiones

• Vitalik Buterin combined local AI, zkAPI and Tor to limit personal data exposure.
• Local AI processed 20–30 tokens per second, below his desired speed of over 100.
• Stricter privacy limited the context remote AI models could use to improve advice.

Vitalik Buterin has outlined three privacy layers for an AI experiment that uses personal health and travel data. His goal was to generate personalized diet and exercise recommendations while limiting the private information shared with powerful remote models.

The system used a local model, zkAPI payments, and Tor networking to address different ways services could identify him.

Each aspect handled a separate source of exposure within the same workflow. The local model created requests, zkAPI addressed payment identification, and Tor targeted network information.

Vitalik Buterin said the system returned recommendations, with remote models contributing information that improved the results.

Buterin also described drawbacks related to speed, request construction, and privacy. Those drawbacks included a trade-off between disclosing personal information and giving remote models with enough data to help.

Vitalik Buterin Uses Local AI to Prepare Requests

The first layer was built on a local model that Vitalik Buterin identified as Qwen 3.8 Flash Next. Instead of sending his own phrases directly, he used that model to write inquiries for remote systems.

This approach was intended to prevent the disclosure of personal details and identifying patterns in his writing.

A skill file guided the local model on when to consult remote models and how to minimize the information disclosed.

That arrangement allowed the experiment to draw on capabilities beyond the local model’s knowledge and reasoning. Meanwhile, the local system coordinated those requests as part of the broader recommendation process.

Vitalik Buterin said the remote models helped improve the recommendations, but he found the request strategies far from optimal.

He also reported local processing speeds of approximately 20–30 tokens per second. For the experience to feel fast, he wanted speeds above 100 tokens per second.

zkAPI and Tor Address Payment and Network Identity

The second layer used zkAPI to verify address through payments for remote model access. The included documentation explains that zkAPI accepts deposits in ETH or USDC to cover usage costs.

This payment option routes the test to the Ethereum crypto network, while AI models handle the prescribed procedure.

Payment privacy alone, however, did not address every identification channel described in the experiment.

Buterin therefore added Tor as the third layer to protect information associated with networking and IP addresses. He accessed the combined zkAPI and Tor setup through a command-line tool.

Together, these layers addressed request content, payment information, and network identity. Vitalik Buterin emphasized that the experiment required all three protections.

His description nevertheless acknowledged that their availability did not mean each component delivered sufficient protection or performance.

Privacy Limits Reduce Speed and Available Context

Tor presented a particular difficulty because Buterin said it poorly supported the separation of individual requests.

He questioned whether its privacy protections were sufficient for that purpose. He also described latency as 10–100 times higher than it could be.

That comparison concerned possible performance, rather than a stated measurement against direct connections.

Alongside the network delays, the local model’s processing speed made the workflow slower than he wanted. Improvements to request construction also remained necessary, according to his account.

Beyond those technical issues, the experiment exposed a limit on how much information remote models could use.

More cautious sharing left those systems with less context for generating useful recommendations. Buterin reported a working setup while noting that the privacy-and-utility trade-off remains an unresolved constraint.

This article is for informational purposes only and does not constitute financial or investment advice. Privacy technologies and experimental AI workflows may contain technical limitations and do not guarantee anonymity.

El artículo Vitalik Buterin Reveals Three Privacy Layers Behind His AI Experiment apareció primero en The Coin Republic.