Intelligence Snacks 67 - Local Models, Private Agents, and Wingman

# Intelligence Snacks, Episode 67: Local Models, Private Agents, and Wingman*Hosts: Pete and Andy · Guest: Anthony*What does useful, sovereign AI look like when the frontier providers can change the rules overnight? Pete, Andy, and Anthony compare local models, private-agent appliances, and the hardware needed to keep capable AI close to home. The conversation then moves from personal automation to small-business adoption, why familiar interfaces matter, and how Wingman, Flight Deck, Autopilot, and agent delegation are evolving into a more human way to work with AI.**Key Moments:**- [00:00] Episode 67 opens live from the beach, with Anthony joining Intelligence Snacks for the first time.- [02:43] Anthony frames the episode: model sovereignty, changing OpenAI and Anthropic guardrails, and the risk of relying on any remote API.- [05:04] A practical local-model benchmark: roughly 27–30B parameter models can already handle private admin, documents, PDFs, and vision tasks.- [10:02] The private-AI appliance opportunity: a home system does not need frontier intelligence if it can reliably deliver the useful 90 percent.- [15:30] The group imagines a private household agent connecting calendars and financial feeds, spotting subscriptions and finding savings.- [20:18] Pete explains how Flight Deck separates passive feeds from agent work that genuinely requires attention.- [25:13] Buzz may educate the market about working with agents, leaving Wingman to offer a more sovereign and flexible implementation.- [30:00] Businesses are a natural container for AI because their people, tools, processes, and domains already provide structure.- [35:12] Agent adoption works better through familiar, multiplayer interfaces than through a completely new way of working.- [40:28] Wingman Apps hide the plumbing: companies want a well-designed interface while agents do the work underneath.- [45:00] Pete describes using spare Codex capacity for long-running research, while the group weighs subscriptions, API costs, and efficient models.- [50:00] Local compute economics bite: DGX Spark-class hardware is attractive, but still expensive and hard to source.- [60:01] Pete describes Rick as the manager agent that delegates work, tracks tasks, keeps him informed, and searches the graph.- [65:20] A Nostr-native Git interface could give agents shared access through npub identities and NIP-98 authorization.- [70:00] Flight Deck gives interrupted work a durable home so sessions can stop and resume without Pete keeping browser tabs open.- [77:04] Maple Agent inspires the next private Wingman: a more locked-down local agent working inside the wider system context.**Friends of the Pod:** Jared Griggs, Tim Corrick, Paul Itoi (Stakwork), Rod, Mark (Maple)**Projects Mentioned:** Wingman, Flight Deck, Autopilot, Wingman Apps, Buzz, Maple Agent, Stakwork**Concepts Introduced:** model sovereignty, private household agents, local inference appliances, manager-agent delegation, graph-backed code intelligence