Avi is joined by Mark Suman (CEO of Maple AI, former Apple engineer) for a technical dive into the intersection of AI privacy, confidential computing, and Web-of-Trust as the internet shifts toward an agentic future. Mark explains Maple’s core design: privacy-first by default, where each user starts with a private encryption key, data is encrypted locally, and then processed in the cloud using secure enclaves/confidential computing so the company only ever sees encrypted blobs. The conversation contrasts this with “AI proxy” services (VPN-like shared accounts) that may reduce identity linkage but still send sensitive prompt content to big-tech model providers. From there, they widen out into the economics and trajectory of models: open-source catch-up (benchmarks like Humanity’s Last Exam), the limits of benchmark-chasing, and why Mark expects the “model obsession” to fade as apps and user experience become the real battleground. They also debate the sustainability of today’s venture-subsidized inference, the likelihood of price “switch flips,” and how platforms monetize users indirectly. The back half turns to agents: Mark outlines Maple’s roadmap toward a privacy-preserving personal agent with durable memory and carefully staged permissions (read-only integrations first, sandboxed work later), plus the hard problem of letting agents act in the world without becoming a giant attack surface. The episode closes by tying agents to identity and trust: Nostr’s signed events as an authenticity primitive, and the need for richer reputation signals as bots and humans transact side-by-side. Links Mark Suman on Nostr Maple AI