Vitalik Buterin’s AI Experiment Balances Privacy and Personalization

Vitalik Buterin has shared details of an AI experiment designed to balance personalized recommendations with privacy protections. The system, which generates diet and exercise suggestions, combines three layers of privacy: a local AI model, zkAPI payments, and Tor networking. Each layer addresses a different potential source of identification, including request content, payment information, and network identity.

The local AI model, identified as Qwen 3.8 Flash Next, processes requests before they reach remote models. This approach helps prevent personal details and writing patterns from being exposed. However, Buterin noted that local processing speeds were only 20-30 tokens per second, far below his desired speed of over 100 tokens per second. The local model also struggled with optimal request construction, limiting the system’s efficiency.

zkAPI was used to handle payments for remote model access, accepting deposits in ETH or USDC. Tor was added to protect networking and IP address information, though Buterin acknowledged its latency issues and questioned its effectiveness. Together, these layers created a workflow where remote models could improve recommendations while minimizing personal data exposure.

Buterin highlighted the trade-offs between privacy and utility. Stricter privacy measures limited the context available to remote models, reducing their ability to provide useful recommendations. Despite these challenges, the experiment demonstrated a working setup that addresses multiple identification channels, even if each component’s protection and performance remain imperfect.