The Neuron: AI Explained The Privacy Nightmare Hiding Inside Every AI Chat
71 snips
Mar 22, 2026 Eamonn Maguire, a machine learning leader at Proton with a PhD from Oxford and CERN postdoc, built the privacy-first AI Lumo. He explains how encrypted inference and client-side indexing protect user data. He exposes why Big Tech can’t build private AI, warns about viral image trends and agent risks, and outlines Proton’s model routing, purge policies, and private memory designs.
AI Snips
Chapters
Transcript
Episode notes
Routing Requests To Specialized Open Models
- Lumo routes prompts to multiple specialized open models rather than one monolith.
- A classifier directs requests to models like Mistral, QN2 Coder, or OMO2 for tasks (summaries, code, chat) to balance capability and privacy.
Index Locally And Insert Only Relevant Chunks
- Limit context by indexing and chunking files locally to keep responses relevant and private.
- Projects download encrypted Drive files to device, index them locally, extract keywords, and insert only relevant chunks into the LLM context.
Viral Filters Double As Data Harvesters
- Viral creative trends serve as free data-harvesting mechanisms.
- Eamonn points to Dollbox and Studio Ghibli filters where users uploaded personal images and metadata, creating labeled datasets for model training.