Pete and Andy are joined by Mark from Maple to talk about privacy in AI, why everyday users may finally be starting to care, and what it takes to build private AI products that can still compete on user experience. The conversation ranges from end-to-end encrypted AI and open source model stacks through to agent security, business control, workflow automation, and why most of today's agent hype still feels more like fake sizzle than finished sausage.## Chapters and Themes- `00:00-03:26` Mark introduces Maple as a private AI product, and the conversation opens on why privacy matters online and whether normal users actually care.- `03:26-07:51` Competing with ChatGPT and Claude means winning on UX first, with privacy as the extra unlock for more personal or sensitive use cases.- `07:51-14:46` Kids, AI companions, model bias, and the quieter long-term risk of AI shaping how people think rather than just what they know.- `14:46-22:34` Open source models, open source harnesses, and why visibility into prompts, middleware, and agent behavior matters.- `22:34-29:50` Maple's roadmap, Wingman's architecture, and the difference between consumer AI products and SME-focused agent orchestration.- `29:50-38:14` Why privacy is often mispriced by businesses, and why control may be the stronger commercial framing than privacy alone.- `38:14-49:38` Are AIs actually replacing jobs, or just making small teams more capable and more capital efficient?- `49:38-58:07` OpenClaw, determinism, pipelines, memory, and the "lethal trifecta" of private data, inbound internet, and outbound internet access.- `58:07-01:11:16` Segregated scopes, agent permissions, enterprise information boundaries, and whether central AI intelligence is the right architecture at all.- `01:11:16-01:44:23` Jack Dorsey's intelligence layer, agent gossip, software-defined businesses, and a closing detour into British accents, Siri, and bedtime podcast energy.## Key Takeaways- Privacy only matters commercially if the product experience is good enough to compete.- For many businesses, `control` may be a clearer selling point than privacy on its own.- Open source models are not enough; the harness and surrounding tooling matter just as much.- Most agent hype still breaks down when reliability, repeatability, and permissions matter.- The real opportunity may be software and pipelines generated by AI, not agents acting unchecked.- Businesses will need scoped agents, clear approvals, and tighter boundaries than today's demos suggest.## Notable Lines- "It's similar to ChatGPT... but we do it end-to-end encrypted."- "I think the agent is the sizzle, not the sausage."