In episode 3 of Say Wot?, guest host David Strayhorn sits down with Matthias DeBernardini, a software developer and “agentic engineering” tinkerer who’s just joined NosFabrica to help build open-source Web of Trust tooling on Nostr.
They trace Matthias’ path from materials engineering into Bitcoin (including an early “$90 BTC is too expensive” family moment), then into graph theory, Lightning experiments, and Rust-heavy open-source work (Fedimint, AnchorWatch). From there, the convo zooms into the core question: can we build decentralized recommendation systems without recreating the extractive, centralized incentives of Big Tech?
They unpack why today’s large-scale AI training tends to favor hyperscalers (hardware, bandwidth, overhead), why “decentralized labeling” often still collapses into centralized control, and where a better hybrid might live: community-curated, topic-structured data (a “grapevine” Web of Trust) paired with local models for narrow tasks, fine-tuning, and personal assistants.
The episode ends with a clear thesis: the internet’s signal-to-noise problem is incentive-driven, and the way out is opt-out plus better tools, built around user control and delegated trust.
Links
