# SIG-P4B 2026-09-07 β€” Session Notes Participants: - Timber | UTC-7 - Boris Mann UTC-7 - rafa (UTC+1) - Jacob Sayles --- ## πŸ“– The Reading Not clearly identified in the discussion. This session was not a reading discussion but a cross-group conversation/demo between the Protocol Institute's "Protocols for Business" special interest group (SIG) and the zSpace Society "community-scale AI" / shared computer network project. Reference points mentioned include a Summer of Protocols presentation on "AI as a public service" and Boris's 2024 "community search engines" application. ## 🧭 Overview Rafa convened a call between Protocol Institute members (Rafa, Timber) and two builders from Vancouver's zSpace Society (Boris Mann and Jacob Sayles) to explore possible cross-pollination. After introductions to the Protocol Institute's "context tank" model and its SIGs, Boris and Jacob described their motivation for and progress on a community-scale, self-hosted AI/inference stack. The session closed with a live demo of the running system and a discussion of how the two groups might collaborate. ## πŸ’‘ Key Points & Themes **Protocol Institute / context tank framing (Timber, Rafa)** - **Timber:** PI researches protocols "in the broadest possible sense" β€” from TCP/IP to handwashing to climate accords β€” to build a general theory of what a protocol is and how it can be applied in specific fields. It began with the Ethereum Foundation wanting to govern its protocol better, found broader utility, and is now spinning out of the EF into a standalone research institute. - **Rafa:** Leads the "Protocols for Business" SIG, focused on viewing organizational models emerging from AI/blockchain through a protocol lens. - **Rafa:** Other SIGs include memory research, protocol fiction, formal protocol theory (currently focused on stigmergy and security implications of stigmergic patterns), a new personhood research group, and distributed robotics. All feed a shared "context pipeline"/"context universe." - **Timber:** Two additional angles on "context tank": protocols as a first-class lens to reinterpret existing problems, and using AI to synthesize research context within SIGs (though AI "doesn't have really good research taste yet"). **Motivation for community-scale AI (Boris, Jacob, Rafa)** - **Boris:** "Self-hosting is selfish" β€” the technical part of hosting has never been the hard part; the hard part is going *multiplayer* (permissions, privacy, onboarding, support). That's the gap almost no one is working on outside fully-hosted US SaaS clouds. - **Boris:** The building block is "the group and group coordination"; at a certain scale groups share resources (office, camp, car, compute). AI/community search engines are the next such shared infrastructure. - **Jacob:** Society has normalized the "myth that software is free" and services just appear; centralization pushes the human interface smaller and smaller. His thesis: "as everything gets less and less human, surprise, the answer is human." A community-tied model connects builders, users, hardware buyers, and stewards. - **Rafa:** Community-scale knowledge of these systems is itself important infrastructure β€” "someone should know how to fix a car / lay pipe," and likewise someone should understand the models and hardware. He also wants a counterbalance to monopoly and access to models with different guardrails for research. - **Boris:** They looked around, saw essentially no one doing this, and are learning by doing. - **Boris (repeated caution):** "No one fucking knows this shit." Even ten-year-old tech (VR, digital twins, construction-site cameras) still isn't ubiquitous β€” so this requires practical experimentation, not theory. He's "98% praxis over theory." **The four-phase build plan (Jacob)** - Phase 1 (done): machines at zSpace running inference; a membership base connects via an OpenAI-compatible API and can plug inference into any app. - Phase 2: shared data model / "commons," workspaces, and a context harness to capture and access data via AI in both directions. - Phase 3: local client on your laptop, enabling local data stores; using CRDTs so the web interface is just another client and data mirrors to your drive (also lets them "close the doors" instead of everything being open via HTTPS). - Phase 4: many nodes connected to a central hub, all collaborating over open protocols β€” anyone can leave at any time; members share models they run themselves (e.g., home labs / idle high-end machines). **Technical stack (Jacob, Boris)** - Infrastructure-as-code ("ZiOps") using Proxmox base boxes + Ansible; every service in a separate container; an app called "Corliss" ties it together. Repos are public at zSpace Society. - AT Protocol (ATProto) used as identity primitive; Jacob built an OIDC connector to bridge into arbitrary tools. Single sign-on; membership registry; emerging ATProto permission-data spec is a dependency they're waiting on. - Proxy currently LightLLM (evaluating alternatives); users get an API key + URL. Data flow: local-first β†’ private data β†’ published to ATProto. - Hardware: three nodes at zSpace β€” a DGX Spark ("Digispark," 128GB unified RAM, running Gemma; on loan, ~CAD $5,000), a dual RTX 3090 (48GB, running Qwen 3), and a dual 1080 Ti. Roughly CAD $10k in hardware, some GPUs from the old render farm, some donated. - zSpace pays flat-rate rent with no separate electricity cost (BC power is cheap), which shapes their cost intuitions. Usage is very low so far β€” box is "90%+ unused" β€” giving early "margins in the wild" intuitions. - Membership: a loose three-tier structure (zSpace member / "tinker class"; donated ~$1,000; donated ~$2,500) but not currently enforced β€” they want maximal usage for now. - Privacy: queries and responses are *not* logged/stored; only stats retained (token count, timestamp, who, which model). No encryption at rest yet; trust currently rests on trusting the operator ("do I trust Jacob when he says he can't see what's happening?"). **Product/community philosophy (Jacob, Boris)** - Jacob: not trying to disrupt Discord or other software β€” "I'm looking to disrupt my co-op," his coworking spaces, etc. A stewardship model (local, human support) is baked in from the start. - Jacob uses his 84-unit housing co-op (governed by ~4 master documents plus fire code/insurance) as a non-technical "litmus test." People don't read docs; "people wanna talk to people" β€” hence a chat-first interface plus AI over community context. - Boris: the sale is ultimately to *the group*, which picks the tools; the aim is to "level up the group" into a genuinely useful interface free of ads. Chat (likely on phones) will be a primary interface; example query: "does anyone nearby sell fresh Okanagan peaches?" - Boris frames it against "not all software is made in California" and wanting a counterbalance to large American SaaS. - Related demos Boris has been running: a French-government Google Meet/Zoom clone using the Parakeet transcription model, hosted on their own infrastructure. ## πŸ”€ Questions & Disagreements - **Rafa** repeatedly pressed on *why* individuals want to self-host given cheap cloud/API subscriptions β€” is it real utility, or aesthetic/skill exploration? Boris and Jacob answered with the "multiplayer," counterbalance, and community-competency arguments rather than pure cost savings. - **Timber** asked for the individual-utility pitch to "the laziest person in the co-op" (someone who finds Facebook easy and isn't motivated by collective benefit) β€” a framing the group acknowledged but did not fully resolve; the answer leaned on the group/context value rather than a crisp individual hook. - **Rafa** asked whether this is more the evolution of the *library* or of a *public utility* (like a water utility); not definitively answered β€” treated as an "and," alongside analogies to local water wells and municipal Wi-Fi. - Open technical uncertainties, openly acknowledged: which search tools/agents to include and their biases; how useful gathered data actually is for agents ("we don't know β€” having the data is a starting point"); how to do shared multiplayer human+agent workspaces ("no one has it the fuck figured out"). - Mild, friendly product disagreement: OpenWebUI is in the stack but seen as having a "different mission" (Boris: "unusable"); the challenge is making tools "aware of plugging back in" to workspaces. - On models: general agreement that open models lag frontier by roughly six months but are "more than good enough for fucking anything today," with a path to running locally. ## πŸ”— References Mentioned - Protocol Institute (PI) and Summer/Summer of Protocols; Ethereum Foundation origins. - zSpace Society (Vancouver coworking nonprofit); repos public at zSpace Society; "Commons Computer" / sharedcomputer.network project. - Boris's 2024 "community search engines" application; his startup Fission (Ethereum, IPFS, Protocol Labs). - Jacob's Office Nomads (Seattle); Vancity; Internet Archive Canada (Jacob as technical steward). - ATProto / AT Protocol lexicons, PDS, ATProto permission-data spec; OIDC; Blue Sky/Bluesky. - Auto merge (Automerge) with upcoming "Keyhive"/"Beelay" (transcribed "subduction/beelay") for end-to-end encrypted local data; CRDTs. - Proxmox, Ansible, LightLLM, OpenWebUI; "Corliss" app; "ZiOps" IaC; S3 object store; Synology NAS. - Models/hardware: DGX Spark ("Digispark"), RTX 3090s, 1080 Ti, Gemma, Qwen 3; Parakeet transcription model; a French model provider "Oumans/Oomans" and "DeepSeek Flash." - Ink & Switch (Boris also works there; "building apps on demand as a group"). - People named: Scott (third core technical contributor, absent); Venkat, Kiana Krotler (memory SIG), Sarah Friend (personhood SIG, Berlin) β€” all referenced, spellings approximate from transcript. - Concepts: stigmergy; heat-exchange from sewers (Olympic Village); French government self-hosting. ## βœ… Action Items & Next Time - **Rafa** to reach out to Boris and Jacob and give the project more publicity in the short term; proposed inviting them to write up the project as a case study for "Protocolized"/Protocol Institute publication. - Possible collaboration options floated (no firm commitments): PI/SIG members could beta-test; get a hosted workspace within zSpace's system; build a "Discord-to-workspace" ingestion app; or, if a group could raise ~$20k and had a technical steward, Jacob offered to share notes and help them build their own deployment. - zSpace side is seeking sponsors (e.g., a Montreal consultancy interested in sponsoring/providing hardware) and other spaces to replicate the model (a Salmon Arm coworking space and local IT provider are in discussion). - Boris flagged, half-jokingly but seriously, that the PI group "need[s] to get off Discord" eventually β€” no replacement offered yet. - Boris noted more marketing/onboarding work needed on sharedcomputer.network (e.g., email waitlist signup). - Session was low-attendance due to Labor Day; notes/recording to go out via the email newsletter. ## ⭐ Memorable Quotes - **Boris:** "Self-hosting is selfish… the hard part is when you go multiplayer, when you do it for others." - **Jacob:** "As everything gets less and less human, surprise, the answer is human." - **Boris:** "No one fucking knows this shit… we have to kinda try this stuff to see how the shapes fit together."