Mesh News Network is a prototype for a different kind of news infrastructure: instead of loading heavy websites or platform feeds, people ask questions and receive concise answers generated from locally cached, signed bulletins. In this session, we’ll explore how a stack of "news cells", Raspberry Pi gateways, Meshtastic radios, and small language models (SLMs) can keep trusted information flowing when the usual web stack is slow, blocked, or broken.

Modern news consumption leans on massive, responsive sites and apps: interoperable apps, health monitoring, and platform‑specific clients. Every layer is a potential point of failure or censorship—domains can be blocked, app updates can be withheld, and large media payloads become unusable on congested, metered, or disaster‑strained networks. Yet most people don’t actually need that much surface area to stay informed. In practice, many just want the top handful of stories that affect them, or answers to very specific questions about safety, access, and current events. Mesh News Network starts from that reality and asks: what if the default unit of news was not a website, but a small, signed text object that can be moved cheaply, cached locally, and queried with a tiny model, i.e. verified RSS?

The conceptual core of the session is “bulletins, not websites.” We imagine newsrooms (or community media cells, "news cells") producing compact bulletin objects instead of designing a complete web experience around each story. A bulletin is a structured text envelope: an ID, publisher key, timestamp, region and topic tags, urgency, a short summary, an optional longer body, references, and a signature. These bulletins are published into a distributed transport layer, Nostr relays, RSS endpoints, or other decentralized feeds, and ingested by local hubs: Raspberry Pis that live in communities, camps, newsrooms, or mutual aid spaces. The Pi verifies signatures, stores bulletins in a small database, and indexes them by region, topic, and urgency.

From there, two different network paths emerge. When “real internet” is available, the Pi syncs with upstream relays and fills its local cache. When connectivity degrades and only a low‑bandwidth mesh like Meshtastic remains, the system stops trying to be “the web” and instead uses LoRa for what it’s good at: propagating state, alerts, and small deltas. Over the mesh, nodes can trade compact announcement packets: which bulletins exist, their hashes, topics, priorities, and how fresh they are. Full bulletin bodies don’t need to cross the mesh continuously; they can be pre‑seeded via sneakernet or periodic higher‑bandwidth syncs. The mesh layer effectively behaves as a low‑bandwidth metadata and alert channel layered under the heavier protocols.

At the edge of this system sits the user, who never has to think about any of that. They don’t open a news site; they open a simple chat‑like interface on their phone. That interface might be BitChat, a minimal browser UI over local Wi‑Fi, or a native shell, but the pattern is the same: a single text box and a few big buttons. The user can ask free‑form questions, "Anything new about road closures?”, “Is there a fire risk near here?”, “What changed since yesterday?”, or tap preset prompts like Latest, Safety, Transport, Shelter, and Local. Their query is handled locally by a small language model that lives on the Pi (or on‑device), which doesn’t go out to the cloud. Instead, it pulls from the cached bulletins, filters them according to topic, region, time, and urgency, and synthesizes a concise response with citations or references. The user receives an answer and, if they want, can drill down to see the raw bulletins that answer was built from.

This architecture aims at three things at once: bandwidth efficiency, censorship resistance, and cognitive ergonomics. Bandwidth efficiency comes from the fact that the “heavy lifting” (publishing, syncing) is done in text, and user interactions are always just small queries plus short answers. Censorship resistance comes from not depending on any single domain, app store, or platform skin: the crucial units, the bulletins and signatures, can move over many transports and be stored by anyone. Cognitive ergonomics comes from aligning the interface with what people usually want: answers and briefings, not an infinity scroll of under‑prioritized content while you're on the toilet.

The session itself will be structured in three main parts.

First, we’ll map the current “normie path” for consuming news: open browser or app, resolve domain, hit a complex CDN, negotiate JS, ads, trackers, fonts, and layouts, then finally read a few top stories. We’ll identify all the failure and control points along that path, from DNS to TLS termination to third‑party CDNs and platform moderation, and relate them to real‑world scenarios where this path fails: network shutdowns, bandwidth throttling, disaster zones, and targeted blocking. This sets the problem: news as we know it assumes abundance of bandwidth and political goodwill, neither of which can be taken for granted.

Second, we’ll walk through the Mesh News Network architecture from the perspective of each actor involved:

Newsroom cells (or community media groups) that author bulletins, sign them, and publish them into decentralized feeds instead of relying solely on web CMSs.

Pi operators who run local hubs, handle initial configuration, choose upstream relays or feeds, and physically locate devices in community spaces.

Mesh enablers who maintain Meshtastic or other low‑bandwidth meshes, ensuring that at least small bulletin announcements and alerts can hop across a neighborhood, forest, camp, or city.

Ordinary phone users who just open a chat, ask for what they need, and receive a readable answer.

We will diagram how bulletins move from a publishing laptop in a newsroom to a Pi under a desk in a neighborhood center, how they are summarized or curated, how priority alerts propagate over mesh, and how queries find the right pieces of text. I'll talk about pivoting news media tech vs local bulletins, about the constraints of LoRa vs. Wi‑Fi and occasional broadband, and how different grades of connectivity change the system’s behavior without breaking it.

Third, we’ll dive into the role of small language models as local “news clerks.” Instead of framing the model as a general‑purpose AI assistant, we treat it as a specialized retrieval and compression agent with four core jobs: daily briefing, “what changed” since X, topic‑specific digests, and rumor checks. We’ll explore how to constrain the model’s behavior so that it mostly stitches together and compresses known bulletins instead of hallucinating, how to expose provenance in the UI (e.g., “this answer is based on 3 bulletins from sources A, B, and C, last updated N hours ago”), and how to maintain trust boundaries when some bulletins are official and others are community‑generated. This section opens up questions of epistemology and power: who gets to publish, which keys are trusted by default, and how do we avoid accidentally re‑centralizing editorial control in the model?

Throughout the session, we will be explicit about trade‑offs and open problems rather than pretending the model is solved. For example: how do you design a bulletin schema that is expressive enough for many news types but compact enough to move over constrained links? How do you handle multilingual communities? How do you do personalization without building a surveillance profile, keeping preferences and filters local, while still giving people the sense that “this feed is for me”? What happens when two PIs disagree about trust or filtering policy? How do you design UX for people under stress (evacuations, outages) who may have one or two minutes of battery left?

In terms of format, the session is planned as a short presentation with an interactive edge. I’ll present the conceptual model and a reference architecture using diagrams rather than code, to keep it accessible to both technical and non‑technical participants. I’ll invite a brief discussion focused on use cases participants care about, independent media, disaster response, protest communication, rural connectivity, or cultural preservation. If connectivity and logistics allow, I’ll also show a basic prototype flow: a mock bulletin store, a simple query, and the kind of answer a local SLM could generate. The emphasis, though, is not on one specific implementation, but on the design pattern: local, query‑driven news stack over small text objects and cheap radios.

By the end of the session, participants should come away with a mental model of an alternative news pipeline that does not depend on big web stacks or centralized platforms: they will understand the roles required (publisher, hub operator, mesh maintainer, end user), the technical primitives (signed bulletins, local caches, low‑bandwidth relays, SLM‑based summarization), and the social questions that still need collective answers. Ideally, this becomes a seed for further collaboration at DWeb Camp: people who run meshes, who work in independent media, who build local AI systems, and who care about network sovereignty can find each other and iterate on the idea, adapting Mesh News Network to their own contexts and communities.