My self-sovereign / local / private / secure LLM setup, April 2026

Vitalik Buterin describes an approach to AI that treats privacy, security and self-sovereignty as non-negotiable, built on local LLM inference, local files and sandboxed agents. He argues that the wider AI ecosystem, including local open-source projects, is cavalier about privacy and security, citing criticism of OpenClaw and the claim that roughly 15% of the skills seen contained malicious instructions. His setup runs Qwen3.5:35B through llama-server on a laptop with an NVIDIA 5090 GPU, reaching about 90 tokens per second, and uses the pi agent framework inside a bubblewrap sandbox. He also notes that he found the DGX Spark slower than a good laptop GPU and that the post is meant as a starting point rather than a finished, secure product.