25: White Noise, MLS, and marmot w/ Jeff G

"What we're actually building is an unstoppable messenger." Jeff G & Gigi take a walk in Oslo. Listen on sovereignengineering.io In this dialogue: Jeff G explains how White Noise started as an attempt to build the most private messenger possible, then slowly mutated into something else: an unstoppable messenger that can survive ugly network conditions and hostile environments Why MLS on Nostr is mostly an ordering problem: identity and delivery are the easy parts, group-state evolution is where everything breaks Forked histories, missing commits, and the nightmare scenario where a group silently splits into incompatible realities without anybody noticing The temptation of coordinators, sequencers, and Pablo's hilariously named Serial Killer Relay, plus the deeper question of when a relay stops being a relay and quietly becomes a server Calle's Cashu headaches, NIP-60 and NIP-61 edge cases, and the "pocket with a hole" problem where balances seem to disappear, reappear, and generally behave like loose change under a couch Jeff G's current bet for Marmot: a deterministic state machine, tunable convergence rules, and automatic re-init flows that make occasional desync tolerable Why large-group privacy degrades fast, and why "perfect privacy" stops meaning much once a chat has hundreds or thousands of people in it Why Pika felt more reliable than NIP-17 for one-to-one chats, and where simple tools still beat more ambitious systems The White Noise v2 direction: replaceable key packages, smaller app components, and decoupling Nostr identity from the transport layer so relays become just one option among many How FIPS pushed the design forward by making transport agility feel real: relays, Wi-Fi, Bluetooth Direct, and whatever else can move packets when the normal internet is gone BitChat as a useful proof that narrow, situational tools matter, especially when the internet is overloaded, absent, or simply the wrong abstraction for the problem Other experiments blooming around the edges: Tubestr as a weirdly great MLS-backed permission system, plus gaming demos, new clients, and other signs that the building blocks are escaping the original chat use case OpenClaw, Fabian's OpenClaw NIP-17 plugin, and why cryptographic identities make one-agent-per-project workflows feel native Jeff G's current AI workflow: multiple agents, multiple boxes, very little hand-written code, and a strong belief that the implementation middle is collapsing while architecture, taste, and judgment matter more than ever Why "build it right" is getting harder to teach and more important to teach, because now you can vibe a terrible idea into existence in an afternoon just as easily as a good one Apprenticeship, elders, and why people like Johnathan Corgan matter: some engineering instincts only show up after you have lived through brittle systems, bad assumptions, and real adversaries Notification servers that should know almost nothing, the coming KYC of the internet, and the broader goal of building systems with less data to seize and fewer chokepoints to attack The argument over easy abstractions at the edge of Nostr, from "Coinbase for Nostr" worries to the closing detour on onchain zaps and second-order effects. For more on that last part, see Gigi's rebuttal, Careful, Icarus People mentioned: Justin Moon (Pika, small-group messaging trade-offs, and prior White Noise conversations) Pablo (Serial Killer Relay, naming crimes, and forcing the ha