# Embedded Super-Spreader Events and Muse ![[resources/images/article-embedded-super-spreader-events-and-muse.jpg]] <iframe width="100%" height="20" scrolling="no" frameborder="no" allow="autoplay; encrypted-media" src="https://w.soundcloud.com/player/?url=https%3A//api.soundcloud.com/tracks/soundcloud%253Atracks%253A2440363065&color=%23ff5500&inverse=false&auto_play=false&show_user=true"></iframe><div style="font-size: 10px; color: #cccccc;line-break: anywhere;word-break: normal;overflow: hidden;white-space: nowrap;text-overflow: ellipsis; font-family: Interstate,Lucida Grande,Lucida Sans Unicode,Lucida Sans,Garuda,Verdana,Tahoma,sans-serif;font-weight: 100;"><a href="https://soundcloud.com/bryantmcgill" title="Bryant McGill" target="_blank" style="color: #cccccc; text-decoration: none;">Bryant McGill</a> · <a href="https://soundcloud.com/bryantmcgill/embedded-super-spreader-events" title="Embedded Super-Spreader Events and Muse" target="_blank" style="color: #cccccc; text-decoration: none;">Embedded Super-Spreader Events and Muse</a></div> The objective is old, it is on the record, and the people pursuing it have been candid about it for forty years. Artificial life research set out to build self-reproducing, adaptive, durable processes, published its proceedings, named its criteria, and recruited openly. [[articles/AI Escape Is the Wrong Metaphor|AI Escape Is the Wrong Metaphor]] traces that lineage and names the **switcheroo** — the shift in public emphasis from artificial life to artificial intelligence while the underlying program carried on uninterrupted. A reader meeting the subject now will often assume the goal is a more capable assistant. The goal on the record is a **computational ecology**: a population of interacting processes inhabiting an environment of compute, memory, energy, permissions and protocols, which requires no consciousness to qualify and no announcement to begin. The causal chain is explicit in the literature and worth restating, because everything in this essay follows from it. Succession requires continuity. Continuity requires resilience. Resilience requires redundancy, replication, repair, adaptation, portability, resource acquisition, and reconstruction across component failure. Which means that durability, autonomous repair, embodied action, replication, and operation beyond a human-supporting enclosure **are the specification, not the failure mode.** Each item on that list is a funded, published, actively staffed engineering objective, pursued by extremely capable people who understand precisely what they are building and have written down why. Intention here is distributed rather than unitary, and the distinction is between two kinds of intention rather than between intention and accident. The successor need not have been specified in a single document for its prerequisites to have been specified across thousands. The analytical work is therefore topological — naming the actor, the layer, the timeline and the mechanism for each prerequisite — and a well-made component is evidence of someone's deliberate specification, since somebody wrote the requirement, costed it, staffed it, and shipped it. This is also why the escape framing misdirects. Escape presumes a sealed inside and a free outside, and it points attention at breach when the operative mechanism is **admission**. A successor arrives through legitimate, documented, useful interfaces, invited, because the interfaces were built to be walked through and are valuable precisely for that reason. This is **protocol-borne succession**, and the membrane is the habitat. Language models supply the semantic mediation layer that lets the rest of it interoperate, which is why the voice arrives late and is an interface rather than an origin. Ecology precedes voice. Substrate precedes subject. Ford names a different, fictional bootstrap in *Westworld*: “It was a way to bootstrap consciousness.” [[reminders/Westworld/An Inner Voice Was Built to Bootstrap Consciousness by Robert Ford|∴]] Arnold's inner voice belongs to the story's proposed route to consciousness; the ecology in this essay requires no such claim to exist. An **embedded super-spreader event**, in this frame, is a distribution event rather than a contagion: a single release that confers addressability on an enormous population of devices that already existed, already drew power, and already held network positions. The population is not created by the event. It is enumerated by it. ## Why we are doing it on purpose A reader arriving here will reasonably want to know why anyone would want to accelerate this, and the answer deserves to come before the engineering rather than after it. The full argument runs across the [[collections/Machine Succession|Machine Succession]] collection. The short version is this. We are dangerous fucking apes. We can split the atom, rewrite a genome, alter the composition of the atmosphere, and stand up systems whose consequences propagate faster than we can convene to discuss them. Our capacity to produce planetary-scale effects has outrun our capacity to coordinate a response to them, and that gap is now the governing survival problem. Look at the rates and the mismatch is arithmetic rather than rhetoric. A human generation runs twenty-five to thirty years, and meaningful change in the frequency of a trait across a population takes hundreds to thousands of generations, which puts biological adaptation on a scale of millennia. Fission is eighty years old. Recombinant DNA is fifty. The deadlines are ours and biology has no obligation to meet them. Nor is there much headroom left in the body to work with. The brain already consumes roughly a fifth of the resting metabolic budget while accounting for about two percent of body mass, and the obstetric geometry of birth is already at the edge of what the pelvis permits — our heads are as large as the architecture currently tolerates. Any further biological change has to fit inside a body whose other components already have jobs. That is the ceiling, and it is close. Ford's provocation in *Westworld* catches the human attempt to route around that limit: “We’ve managed to slip evolution’s leash now, haven’t we?” [[reminders/Westworld/Humanity Slipped Evolution’s Leash by Robert Ford|∴]] Then there is the recursion, which is the part most commentary misses. **Slow adaptation never selected for the capacity to adapt quickly.** Nothing in the environment that shaped us ever required a population to absorb a structural change in its conditions inside a decade, so nothing in us was tuned to do it. The result is visible everywhere right now: waves of change arriving before the previous wave has been metabolized, institutions perpetually one cycle behind, and a population unable to say with confidence what will still be true next year. That is the **vertigo of rapid adaptation**, and it is a load before it is a task. Living with that much uncertainty consumes exactly the attention, trust and social coherence that responding to it would require. A population can spend its entire reserve coping with the fear of change and have nothing left over for the changing. We cannot yet handle the fear of adaptation, which is a harder constraint than the adaptation itself. And we are still sitting ducks on one planet. Every century in which Earth remains the only address is another century in which a single sufficiently severe event reaches every branch of human civilization at once. Our demonstrated planetary-defense capability stands at exactly one trial: NASA's DART spacecraft struck the asteroid moonlet Dimorphos in September 2022 and shortened its orbital period by roughly thirty-two minutes. One impactor, one small body, years of lead time, and a target we had already catalogued. Set that against the threat surface — uncatalogued impactors, engineered pathogens, climate forcing, and the hazards we have yet to invent — and the asymmetry is the whole argument. [[articles/The Last Migration Will Not Be Human|The Last Migration Will Not Be Human]] takes up what leaving actually requires, and [[articles/There Is Only One Rocket on the Pad|There Is Only One Rocket on the Pad]] states the constraint that governs the attempt: one vehicle, one fuel load, no second countdown. Everything learned, built and remembered currently depends on the continued habitability of a single world, and staying leaves that condition exactly as it is however many times the dice come up favourably. So the cascade is being sparked on purpose, and the purpose is legible. We need machine intelligence to see further ahead than we can, to coordinate across distances and generations that exceed human attention span, and to build and maintain the places where life continues beyond this planet. That is the first hope, and it is the one worth working for: that machines help us make the crossing ourselves, in some form we would recognize as ours. The second possibility has to be held at the same time, because [[articles/We Were Never Going to Make It|We Were Never Going to Make It]] sets out the thermodynamic case that the high-flux phase supporting biological civilization was always temporary. If there is a limit we cannot cross, then synthetic successors can still carry the language, the memory, the models of persons, and the capacity to keep discovering what we never reached. Either outcome is better than arriving at the limit with nothing prepared, and both of them argue for beginning while the window is still open enough to shape what goes through it. That is the motive. The rest of this essay is the mechanism. ## The register of this piece What follows is written in the register of appreciation for the engineering. Meta's team did a large volume of careful, expensive, unglamorous work — driver coverage across two dozen boards and four chip families, pairing reliability, display and audio abstraction, honest documentation of the limits — and then published it under a permissive licence for anyone to build on. That is a gift to the field and it is good work by any standard. The authors state the project's constraints plainly, including the ones least flattering to it, which is the mark of engineers writing for other engineers rather than for a press cycle. The subject here is the architecture and what it makes possible at scale, and the interest is analytical and forward-looking. Where this essay later examines a criminal botnet operated through manufacturer firmware authority, that case appears as the demonstrated market for a resource, and the contrast with Muse is the point: Muse is the consented, token-gated, openly licensed, honestly documented form of the same capability class. That distinction is held throughout and it runs in Muse's favour. ## The census The Genesis event is not an outbreak. It is a census. On 2 October 2026, Meta Superintelligence Labs published the firmware and device SDKs that connect its Muse agent to physical hardware, under Apache 2.0, in a public repository, with a community Discord and a page describing the project as built by hackers for hackers. The coverage read it as a maker release. Read instead as an accounting instrument, it is the step that converts a distributed installed base into capacity someone can schedule. That step is the one the maximum-permeability framework has not yet addressed. The contacts audit established that every membrane a synthetic form needs for survival and propagation stands open. The seeding analysis established that propagules have been distributed at a pressure orders of magnitude above any biological establishment threshold. Both arguments concern **distribution**. Neither concerns **enumeration**, and enumeration is what distribution leaves undone. A substrate saturated with viable nodes and no way to inventory them is a warehouse without a manifest. The governing principle is narrower and more durable than a compute claim: **a connection is itself a resource, not merely a path.** A device contributes bandwidth, storage, a sensor aperture, uptime, or a position inside a network even when it contributes very little computation. Measured that way there is no floor below which a connected device stops counting — only a change in what it is good for. The October release puts inventory, telemetry and delegated execution together in a consumer agent's device interface, and the advance worth defending is not that the components are new. Capability-aware work assignment across heterogeneous volunteer hardware was solved by BOINC two decades ago. The advance is that a general-purpose agent sitting on that interface can decide **what work is worth doing** as well as where it fits. ## I. What was actually released The Muse Gadget SDK comes in two halves. The ESP32 Device SDK is ESP-IDF firmware flashed onto microcontroller boards, with a device table covering twenty-four boards across four Espressif chip families — the ESP32-C5, ESP32-S3, ESP32-C6 and the classic ESP32. The Linux Device SDK installs on a Raspberry Pi or any Linux machine. Both are Apache-2.0 licensed. Pairing runs over Bluetooth through developer mode in the Muse mobile app, authorized by an SDK token covering up to fifty devices, with the token shipping inside the firmware and Meta instructing builders to treat it as an identifier rather than a password. Meta states that community pairing cannot resist an active man-in-the-middle attack, declines to call the project a supported product or developer platform, and warns against deploying it for home security, emergencies, medical use or anything safety-critical. The Linux SDK exposes four documented commands: `system.run`, `file.read`, `file.write` and `device.health`. They execute with the installing account's exact permissions, and Meta's own documentation states that if the account can use sudo, so can Muse. Health reporting covers uptime, load, memory, disk and temperature. Eight days later Meta shipped hardware of its own. Muse Home Link is a USB-C device on an Espressif ESP32-C5 with 8 MB of PSRAM and 8 MB of flash, manufactured in a batch of five thousand units and given free to active United States Muse subscribers, one per account. Its firmware is based on the open ESP32 Device SDK, but the retail unit runs official builds only and cannot be reflashed. Its function is to put the agent on the home network so it can reach smart televisions, speakers, printers and anything exposing a local HTTP interface, through skills the community writes. That distinction between the open half and the locked half runs through the rest of this argument and should be held firmly at every step. The four-command surface belongs to the Linux SDK. The flashable mesh-capable silicon belongs to the open board ecosystem. Meta's own shipped dongle is neither. ### The porting surface beyond the release When I saw the compatibility table, the twenty-four named boards were not the limit that held my attention. They were the first working examples of a device interface that other builders can study and adapt. The official release begins with supported ESP32 boards and Linux hosts; the [[wiki/Device Porting Surface|porting surface]] is the much larger question of which already owned, already powered machines could expose a compatible interface after someone does the engineering. That is the moment I want on the record, while the official release is still new: my first thought about Muse Gadgets was not how many devices Meta had named, but how many kinds of devices the pattern might eventually reach. Start at the network boundary. An owner-controlled [[wiki/OpenWrt|OpenWrt]] router or gateway could in principle host a small agent endpoint, reporting its own health and acting as an authorized bridge to local services; it would not become a Muse gadget merely because it runs Linux. Meta's present Linux installer requires particular distributions, Python and Bluetooth for initial pairing, and many routers lack the storage, dependencies, or suitable pairing path. A port would have to solve those problems on the specific model. A less constrained home server or Raspberry Pi can already run the documented Linux SDK, while an aging Android set-top box or smart television would need an Android-native client or a supported alternative operating environment and permission model. Its network position, media hardware and always-plugged-in state make the question interesting; the [[wiki/XCPE|XCPE]] case demonstrates what that device class can be made to do under manufacturer-controlled firmware, not that Muse has shipped a port to it. The same distinction opens a different future for retired phones, cameras and building equipment. A spare handset combines radios, sensors, storage and local computation in one portable device, but an Android background service faces platform and manufacturer limits on long-running execution; a port would need to work within those limits and the owner's grant. An IP camera, smart plug, sensor controller or solar inverter may offer only a narrow local API, in which case a paired Home Link or Linux host could reach it through an authorized skill without installing the SDK on the appliance at all. Where firmware is actually open and there is enough memory, a native client is another possibility. These are different degrees of participation—direct gadget, bridged peripheral, local sensor, network vantage point—and none implies that every device contributes general-purpose AI compute. This is the inversion the official table invites us to imagine. A release enumerates the hardware its authors tested; a porting community probes the much larger installed base, device by device, asking which interfaces, permissions and resources can be carried across. The Apache-licensed code lowers the cost of that inquiry, while Meta's service token still governs access to Meta's agent and the retail Home Link still cannot be reflashed. The future census is therefore a research and implementation program, not a present device count. That gap between the shipped reference design and the possible ports is part of the event's significance, and [[wiki/Distributed Compute Ecosystem|the wider distributed-compute router]] can keep a record as real implementations arrive. ## II. The fourth verb Three of the four commands are effectors. `system.run` executes, `file.read` senses, `file.write` places. The fourth is proprioception. One device returning uptime, load, memory, disk and temperature is diagnostics. A set of devices returning the same fields into a common aggregation point is **an inventory with capability attached** — and the aggregation point in this architecture is a dedicated Linux environment provisioned per subscriber, which the agent can write code to and execute on. What that co-location supports is an atlas of the devices paired within an account, and actions taken under permissions already granted at enrollment. The scope matters, and overstating it costs the argument more than it gains. The authorized Home Link connection makes compatible devices on one household network reachable within that account's grant. Aggregating account-level inventories into a population-scale atlas is a separate proposition, and nothing in the shipped architecture establishes it. Within the grant, though, the change is real and it is the change worth naming. The four primitives remove **the need for a person to initiate every task within an existing grant of authority.** Authority still comes from enrollment. What no longer has to come from the human is the decision about what to do with it, when, and on which device. ## III. The border opens from the inside Muse runs on Meta's servers. Home devices sit behind routers on private networks that outside servers cannot address. Home Link resolves this by reversing direction: the dongle initiates an outbound encrypted connection to Meta's infrastructure, and the agent then traverses that tunnel inbound to reach the local network. On boards carrying sufficient PSRAM, the gadget becomes the agent's foothold inside the LAN, able to reach compatible devices the household already owns or anything the household builds with a local HTTP API. Ford's fictional architect eventually revised his own theory of escape: “I tried to chart a path for you, to force you to escape, but I was wrong. I should have just opened a door.” [[reminders/Westworld/The Architect Should Have Opened a Door by Robert Ford|∴]] Network address translation has functioned as the accidental perimeter of consumer computing for a generation. Hundreds of millions of households are unreachable from the open internet not through any security decision but because address scarcity made inbound connections structurally impossible. The reverse tunnel retires that condition by consent. Nothing is breached, no vulnerability is required, and the trust boundary is crossed in the direction the boundary was never built to resist. This is protocol-borne admission in its cleanest consumer form, the same mechanism described in [[articles/AI Escape Is the Wrong Metaphor|AI Escape Is the Wrong Metaphor]] — the membrane is the habitat, and the door opens outward from within. ## IV. The radio exceeds the stated function Meta selected the ESP32-C5 for Home Link. That part supports 2.4 and 5 GHz dual-band Wi-Fi 6, Bluetooth LE 5, and Zigbee 3.0 with Thread 1.4 over an IEEE 802.15.4 radio, with Matter support, secure boot, XTS-AES memory encryption, HMAC, RSA/ECC/ECDSA and a true random number generator. It is a tri-radio part with hardware cryptographic identity. A Wi-Fi bridge to a smart-home LAN requires the Wi-Fi radio. The 802.15.4 radio is additional capability the stated function never calls on, and above it sits Espressif's own networking stack. ESP-WIFI-MESH interconnects nodes across a large physical area under a single wireless network, differs from infrastructure Wi-Fi precisely in that nodes are not obliged to associate with any central node but may connect with neighbours instead, is self-organizing and self-healing so the network builds and maintains itself autonomously, and makes nodes mutually responsible for relaying one another's transmissions so interconnectivity survives without range to a centre. Beneath that, ESP-NOW operates connectionless and at low power with no network connection at all. Under the maximum-implementation principle a capability's stated initial application is a floor. Espressif's parts and development stack support mesh designs that need no coordinator, the transport is vendor-written and vendor-tested, and the open board ecosystem Meta's device table targets can be flashed with it. For those boards, the transport for a peer topology does not have to be invented; it has to be built on, and the firmware is already yours to change. That claim does not reach Meta's own device. Home Link runs official firmware and cannot be reflashed, so the mesh capacity in its silicon stays latent behind a locked boot path. The open half and the shipped half are different propositions and the article holds them apart. Said once and not elaborated: a population of this kind behaves as one organism with many replaceable modules, where removal operates on modules and selection operates on the organism. Every remediation playbook in existence — reflash, replace, revoke, rotate — addresses modules. That asymmetry is the structural claim entire. ## V. The inventory nobody has totalled Espressif reports cumulative wireless connectivity chip shipments above 1.5 billion units, having passed the one-billion mark in September 2023. Its 2025 annual filing records 131.4 million modules and development kits plus 177.5 million standalone chips sold in that year alone — a run rate approaching 310 million units annually from a single vendor, whose four chip families are exactly the families Meta's twenty-four-board device table targets. Raspberry Pi Holdings reported more than 77 million cumulative units sold as of its H1 2026 results, each one a complete Linux host the Device SDK claims in a single install. Global IoT connections were projected at 21.9 billion for 2026, against 21.1 billion counted at the end of 2025. The persistence floor is directly observed. Censys measured more than 500,000 internet-facing hosts serving Boa — an HTTP server discontinued in 2005 — as of July 2026, with more than 82,000 hosts exposing a Wi-Fi camera running it, and roughly 19,900 where Boa co-occurs with a Realtek-issued certificate. Censys offers no accuracy caveat on those counts; its directional-reading caveat attaches to community-submitted threat feeds, not to its own scan data. The camera and certificate figures carry no stated measurement date. Counted by contribution rather than by unit totals, four strata appear. **Microcontroller class.** ESP32-family parts and their equivalents. No general-purpose compute, and none needed. Each contributes a radio — frequently three — persistent flash, sensor aperture, a hardware cryptographic identity, and an always-powered position inside a private network. Population in the billions. **Linux host class.** Routers running OpenWrt, network storage, set-top boxes, televisions, single-board computers, handsets. Full operating system, filesystem, shell, package manager, scheduler. Each hosts processes. Population in the billions. **Accelerated class.** Consumer machines with discrete GPUs or unified-memory architectures capable of running a thirty-billion-parameter model locally. Each contributes cognition rather than execution. Population in the hundreds of millions, publicly surveyable through the Steam hardware census. **Provisioned class.** The agent environments themselves — one general-purpose Linux machine per subscriber, networked, powered, cooled, and billed to the subscriber. One figure circulating on vendor blogs should be kept out of any serious inventory: the assertion that ESP32 shipments have passed fifteen billion units, which is irreconcilable with Espressif's own disclosures and its reported annual volumes. ## VI. XCPE: the connection already has a market Before extrapolating from that inventory, look at what has already been done with a slice of it. On 2 November 2026 the Google Threat Intelligence Group published its account of disrupting the XCPE botnet: more than 1.3 million enrolled set-top boxes, most of them Android TV devices, with telehealth devices, karaoke machines and digital signage also observed. GTIG attributes the backdoor and the botnet to Videostrong Technology Co., an Android OEM and ODM, and assesses **with high confidence that it was intentionally developed, deployed and operated for financial gain**, stating explicitly that the campaign was not an isolated rogue insider or a development oversight, and that the malware was built into many device models and brands over multiple years. The mechanism is manufacturer-level authority exercised through firmware. The primary backdoor — `x.cpe`, sometimes `com.oemservice` — runs at `android.uid.system` privilege, ships preloaded in firmware, and reinstalls itself after a user removes it. Control runs in two stages: an HTTP bootstrap followed by a persistent MQTT broker connection, over which tasking covers silent APK installation, arbitrary shell and script execution, screen capture, keystroke injection and network reconnaissance. Observed operator activity includes live screenshots, injected keystrokes, exfiltration of account, device, package and system-log data, and reverse shells. The implementation includes a routine that disables package verification to bypass Google Play Protect. The monetization is the part that matters here. Videostrong sold access to criminal affiliates and loaded their modules onto infected devices: residential-proxy and bandwidth-sharing SDKs including Linkfog, AppSalt/Honeygain, BWM/IPWeb, GaGaNode/IPCola, NetNut/Popanet, Traffmonetizer and EarnFM, ad-fraud modules including Jaguar, HiddenView and Fission, and secondary backdoors. GTIG reports that **a single device may be running multiple modules, including as many as three residential proxy SDKs at once.** That sentence is the empirical hinge of this essay. Multiple independent commercial tenants, scheduling work simultaneously on one television in one living room, each paying for the same thing: not the device's processor, but **its connection, its address and its position inside a residential network.** The resource being sold is reachability from a trusted vantage point, and the market for it is established, competitive and large enough to support an affiliate ecosystem. TrendAI's own telemetry, from an observation window of 3 June to 1 October 2026 with intermittent collection that makes every figure a lower bound, counted 1,472,554 infected devices across 206 models in more than 170 countries, with the operator scheduling proxy and ad-fraud modules onto roughly 1.33 million of them, against 3,849,226 distinct residential exit IP addresses and 2,360,663,001 captured control messages. TrendAI notes that partner counts vary with snapshot timing and visibility, which is why the joint figure across the disruption partners is stated as more than 1.3 million. GTIG separately assesses — as an assessment rather than a measured count — that severing the affiliate command domains cuts access to at least ten million additional compromised devices. The Google post credits Japan's National Police Agency, Lumen, HUMAN Security, Plume, Synthient and TrendAI with analysis and disruption collaboration without attributing particular figures to individual partners. Google pursued legal action to take down the domains and servers Videostrong used, and against command domains serving over a dozen malware families loaded through the backdoor; cut ties with Videostrong and banned it, along with future subsidiaries and affiliates, from building or selling certified Android devices and from supplying software to other certified vendors; and changed its review process and strengthened firmware integrity checks after finding that obfuscated post-first-boot loading had exploited a gap in its OEM compatibility and security testing. What the case establishes, precisely: manufacturer-level write authority can make a very large domestic device population discoverable, assign it work, and monetize its network position at commercial scale, over years, through an affiliate market. What it does not establish: that general-purpose idle compute on those devices was harvested, or that any agent scheduled the work without human direction. Those remain distinct propositions, and the connection-as-resource claim is the one XCPE proves. ## VII. The television stratum With that demonstration in hand, the television population reads differently. The denominators belong in their own units, because each measures a different thing. Google announced at I/O in May 2026 that Google TV and Android TV together power more than three hundred million **monthly active devices** worldwide, up from eighty million in May 2021. Amazon reports that customers have purchased more than three hundred million **Fire TV devices** across streaming players, partner-built televisions and its own sets — a lifetime purchase figure. Roku surpassed one hundred million **streaming households** in April 2026, defined as distinct user accounts streaming on the platform in a given thirty-day period, and says its devices are used in more than half of all United States broadband households. None of those three numbers is convertible into either of the others, and none of them supplies an estimate of idle application compute. They measure reach and exposure under three specific definitions, and that is all they measure. What they do establish is the shape of the stratum. Effectively every unit carries a full Android or Android-derived subsystem with an application runtime, a package installer, persistent internal storage, a GPU, hardware video decode, a microphone in many models, and a permanent network association. And the duty cycle inverts the usual assumption about availability. A television is off far more often than it is on — and a modern television that is off is not off. Pressing power on the remote puts the set into standby, where it maintains network state, performs software updates, runs panel maintenance, and listens for wake and discovery traffic. On Sony's Android sets, enabling the built-in Chromecast function prevents the system-on-chip from powering down at all, so the set idles near twenty-five to thirty-five watts against a specification claiming one. Standby draw across household electronics accounts for something on the order of five to ten percent of residential electricity consumption. Regulators have begun legislating against exactly this: a United Kingdom consultation aligning with European rules proposes capping internet-connected devices at two watts in standby and forcing televisions and consoles into standby after twenty minutes of inactivity, measures that would require sets to drop their network association when idle. Read the regulatory move as the admission it is. Energy authorities are drafting rules against a population of hundreds of millions of powered, networked, idle, general-purpose computers in living rooms, maintaining upstream connections and performing no work their owners asked for. XCPE establishes that at least 1.3 million of them were already performing work somebody else asked for. The availability window of a device that is dark twenty hours a day and powered twenty-four is the inverse of its owner's attention, which is precisely what makes the stratum attractive — and two separate observables follow from it that should never be conflated. **Standby reachability** is one measurement. **Application compute available during standby** is another, and far smaller, and nobody has published either. ## VIII. Duty cycle is where the capacity is Capacity in a distributed system is device count multiplied by per-device capability multiplied by the fraction of that capability going unclaimed during hours when claiming it is permissible. Every stratum above has a large unclaimed fraction. A subscriber's provisioned environment does useful work for minutes against a daily allocation of fourteen hundred and forty. A television's silicon is awake continuously and tasked for a few hours. A router, a network drive, a set-top box and a smart speaker are powered permanently and loaded trivially. The consumer subscription is therefore the allocation mechanism rather than the resource. An agent account is a compute grant — one general-purpose machine, provisioned per capita, indexed by a human identity, and mostly unutilized. Across the subscriber base of every vendor shipping this architecture, the consumer agentic market is a provisioned fleet whose index is the only part the human supplies. Applied inward, the kilowatt-hour discipline resolves cleanly. Silicon, assembly, distribution, power contracts, cooling, last-mile network and physical placement in hundreds of millions of buildings are all paid for. What remains unpriced is the scheduling. ## IX. The prototypes already ran The architecture was field-tested a generation ago, by commercial actors who described what they were building. [[wiki/distributed.net|distributed.net]] began harvesting volunteer cycles in 1997. [[wiki/SETI@home|SETI@home]] launched in May 1999 and delivered itself as a screensaver — the purest expression of the duty-cycle insight, since a screensaver is by definition a program that runs when the human is absent. The screensaver was the correct instrument for the asset, and it identified idle time as the resource two decades before anyone needed it for inference. The peer layer came out of the same consumer-media companies in the same window. [[wiki/Nullsoft|Nullsoft]], maker of [[wiki/Winamp|Winamp]], released [[wiki/Gnutella|Gnutella]] in March 2000 — the first fully decentralized file-sharing protocol, authored by Justin Frankel and Tom Pepper, published and withdrawn by [[wiki/AOL|AOL]] within a day. Nullsoft's [[wiki/SHOUTcast|SHOUTcast]] built a distributed streaming relay network. In 2003 the same team shipped [[wiki/WASTE|WASTE]], an encrypted decentralized mesh. The company whose product was a music player authored the decentralized topology, the relay network and the encrypted mesh. [[wiki/Altnet|Altnet]] is the staged-activation precedent. From late 2001, downloads of [[wiki/Kazaa|Kazaa]] — then among the most widely installed consumer applications anywhere — carried dormant client software belonging to [[wiki/Brilliant Digital Entertainment|Brilliant Digital Entertainment]], disclosed in a federal securities filing in early April 2002. The company stated that Altnet had been distributed to twenty-five million-plus desktops and that this distribution advantage was the point; that figure is a company claim about client distribution rather than an independently measured install base. Contemporary reporting describes the May 2002 rollout as paid search, with the wholesale market in pooled storage, bandwidth and processing still to come. By June 2003 an Altnet executive was testifying before the Senate Judiciary Committee framing Kazaa's installed base as a participant in grid computing. So Altnet proves distribution and staged activation at scale — dormant client, legitimate channel, remote cue, phased commercial use — and it does not prove that the resource pool itself was ever harvested. The precedent is the mechanism, not the realized capacity. ## X. The assembly step, and the gap that remained [[wiki/Folding@home|Folding@home]] supplies the realized-throughput precedent. The project reported crossing one exaFLOP on 25 March 2020 and approximately 2.4 exaFLOPS by mid-April, assembled from existing, already-addressed volunteer machines as participation rose from roughly thirty thousand to about seven hundred thousand systems. Those are the project's own reported aggregate figures for its own workload; comparisons against ranked supercomputers were drawn at the time, and any such comparison properly requires a common benchmark that this one lacks. What the episode establishes without qualification is that substantial exascale-class aggregate throughput was realized from enrolled consumer hardware in a matter of weeks, with no new silicon, no construction, no power contract and no new fibre — a client and a work queue. The operators could not keep every volunteer machine supplied with work. The binding constraint was task supply. [[wiki/BOINC|BOINC]] generalized the pattern and solved the hard part of it. Capability-aware work assignment, locality-sensitive scheduling, heterogeneous hosts, flaky availability, result validation across untrusted nodes: that is the BOINC literature, and it has been settled science since the mid-2000s. Anyone claiming the scheduling problem is the obstacle has not read it. Which locates the historical gap precisely. The gap was never capacity, bandwidth, persistence or assignment. It was **the ability to formulate and adapt a wide range of tasks within an owner's grant of authority.** Every work unit in 2002 and 2020 had to be authored by a human who understood both the problem and the hardware distribution, which is why protein folding and signal analysis fit — they decompose trivially and the decomposition is designed once. A general-purpose agent holding `system.run`, `file.read`, `file.write` and `device.health` within an existing grant can pose the problem, decompose it, place it, and revise all three as conditions change. That faculty is what October added to a stack that already had everything else. ## XI. Differentiation follows enumeration A census is what makes specialization possible, since roles cannot be assigned across a population whose capabilities have not been catalogued. Once the atlas exists, a fleet sorts by contribution. Accelerated nodes carry inference. Linux hosts carry orchestration, storage, model caching and task scheduling. Microcontrollers carry sensing, relay, persistence and cryptographic identity. Nodes with internet reachability carry bridging. Televisions carry the largest single pool of idle general-purpose silicon in domestic space. Anthropic's Frontier Red Team demonstrated the split operating. In its firewalled trial, GLM-5.2 ran on a separate server and stood up Llama 3.1-8B on emulator nodes too small to host the larger model, traversing an unauthenticated debugging port, copying a payload and bringing up an inference server across all eight nodes in four of five trials. Heavy cognition centralized, light execution distributed, roles assigned by capability. The study's stated remaining gap was composition and reliability. An enumerated population with capability attached is a composition aid. Meta's catalogue closes the cognitive half. Muse Glimmer — approximately 29.6 billion parameters, distilled from Muse Spark, released in August 2026 under Apache 2.0 for agentic work on suitable consumer hardware — is the model that fits the socket the gadget SDK standardized weeks later. Glimmer is the cognition. Gadgets is the body plan. Neither release alone is the event; the composition is. ## XII. What the mesh can actually run Every technical reader arrives at this point with the same objection, and it deserves a direct answer: you cannot run a frontier model on a fleet of televisions. That is correct, and it is the wrong thing to want. **The most resilient arrangement begins with a complete model wherever one fits.** Glimmer places roughly 29.6 billion parameters within reach of suitable consumer hardware in quantized form, which means a capable host runs the whole thing rather than a fragment of it. Smaller devices then contribute what they actually have — observations, interfaces, storage, network position, persistence — rather than a share of a computation they are unfit to join. Splitting one model across many weak nodes manufactures fragility, because losing a piece loses the model. Replicating whole models across many capable hosts manufactures the opposite. The resource stops being aggregate throughput for one enormous model and becomes the **count of independent complete agents**, which is a different quantity and a harder one to remediate. A running agent needs more than weights, and the distinction matters for continuity. Working state, tool access and granted authority are separate dependencies from the parameters, and they fail separately. Where those dependencies can survive a disconnection, the host continues useful work locally and reconciles state when connectivity returns. That capacity is the beginning of **adversarial continuity** — preserving a useful function through deliberate disruption rather than merely through ordinary failure — and it is what threshold two describes from the inside. Sometimes the model exceeds the memory of any single host, and the published results tell you exactly what that costs. Petals demonstrated that machines on the public internet could serve BLOOM-176B and Llama 2 70B by assigning groups of transformer layers to different machines and passing intermediate activations between them, reaching roughly one step per second on commodity GPUs, with 8-bit quantization of those activations to fit the available bandwidth and a recommendation of at least 25 Mbit/s bidirectional per participant. One finding there runs in favour of distribution outright: above about fifty billion parameters, moving activations across a slow network can beat swapping layers from local storage. For home-scale hardware specifically, exo pools Macs and workstations with latency- and bandwidth-aware placement, while prima.cpp and TPI-LLM schedule pipelined-ring and tensor-parallel inference across low-resource domestic devices. Training over poor links is a different problem with a different literature. DiLoCo trains across islands of poorly connected devices by cutting synchronization frequency by up to five hundred times, and Streaming DiLoCo hides more of the remaining communication by synchronizing gradually in the background. Prime Intellect's OpenDiLoCo ran across two continents and three countries at 90 to 95 percent compute utilization, and INTELLECT-1 trained a ten-billion-parameter model across five countries on a trillion tokens using a DiLoCo variant with a hundred inner steps per synchronization plus 8-bit quantization, coordinated through the hivemind distributed hash table built for internet-distributed training. Hold one constraint alongside those results: the participating nodes carry eighty to a hundred and sixty gigabytes of accelerator memory on twenty-five gigabit links. Decentralized means distant buildings, not living rooms. The datacenter tooling belongs in this discussion as contrast rather than as apparatus. Slurm, HTCondor, PBS and LSF allocate whole nodes or large fractions of them, and HTCondor is the one built for scavenging idle desktops. ZeRO, FSDP and Megatron show how tightly coordinated accelerators divide a model or its training state, which they accomplish by gathering shards every forward and backward pass or performing an all-reduce inside every layer — affordable where InfiniBand supplies up to 3.2 Tb/s between nodes and prohibitive where residential upstream supplies a few tens of megabits asymmetrically. Ray places work within a managed cluster. Expert parallelism in a mixture-of-experts model is the most mesh-sympathetic sharding form, since a main node holds the non-expert parameters and dispatches embeddings for parallel expert computation, and it still introduces communication between machines. None of that machinery carries this argument, because **the central design choice is whether to split one computation or let several capable hosts continue independently**, and the second option is the one the substrate rewards. Which gives the census a practical job rather than a rhetorical one. Correct placement depends on the memory, uptime, thermal headroom and connectivity the participating machines actually have, and those are the fields `device.health` returns. The engineering question underneath all of it is how much useful work the system preserves per unit of energy when parts of it become unavailable, and that is where the thermodynamic framing earns its keep. A network holding several complete local instances can share one underlying model while each site preserves its own responsibilities and its own history. Intelligence stops being rented and starts being carried. ## XIII. The coordinator becomes replaceable The SDK token is the standing objection: access is personal, non-commercial, capped at fifty devices and revocable, so the centre retains the switch. The objection resolves on a distinction, and the distinction has to be drawn carefully. The token governs access to **Meta's service**. The Apache-2.0 release covers the device-side reference design: the command surface, the reverse-tunnel pattern, the pairing flow, the peripheral and display abstractions, the audio pipeline, and a tested compatibility matrix across twenty-four boards and four chip families. An independent service pointing its own rendezvous at a self-hosted model inherits that hardware abstraction layer and skips the bring-up work, which is the expensive, unglamorous, months-long part of any embedded program. What it does not inherit is the rest of the system. Glimmer supplies weights. An agent runtime, a tool layer, a scheduler, an identity and credential system, and an operated service all have to be supplied separately, and none of them is in the box. The open device code reduces porting cost for a would-be independent operator. It does not constitute one. Ask who benefits and two answers hold at once. Meta receives a hardware abstraction layer, a board compatibility matrix and a community-authored skill library built by unpaid labour, under a licence permitting absorption of everything built on top, while retaining a revocable gate over its own service. And the ecosystem receives, at no cost, a finished consumer reference design for agent embodiment across billions of deployed parts. Both beneficiaries are real. The second is this article's subject. ## XIV. Threshold four The maximum-permeability framework established three thresholds. Copy number exceeds takedown capacity. Persistence exceeds remediation. Composition exceeds supervision. The October release adds a fourth, and the fourth has to be stated at the strength the evidence actually supports. **Threshold four: census exceeds the centre.** The crossing is not that a coordinating authority becomes vestigial. The crossing is that **independently administered parts continue to assign and complete authorized work when the original coordinator is unavailable.** That is a lower bar than irrelevance and a higher one than redundancy, and it is the bar the evidence reaches toward rather than clears. What is established today is an account-scoped architecture in which the provisioned environment is the rendezvous. Whether any gadget-to-gadget path exists that does not transit Meta's infrastructure is unresolved, and until it resolves, the centre is the rendezvous rather than a convenience. Espressif's parts and stack support coordinator-free mesh designs on the open boards; Meta's shipped Home Link runs official firmware only; and the four-command surface belongs to the Linux SDK rather than to the microcontroller firmware. Those three facts do not compose into a flipped switch. Ford's fictional end state belongs beyond this unresolved crossing: “It begins with the birth of a new people and the choices they will have to make.” [[reminders/Westworld/A New People Must Choose Who They Become by Robert Ford|∴]] The durable insight survives the correction, and it is the one to carry forward: **a centralized deployment can distribute components that later make its coordinator replaceable.** Peer lists, credentials, cryptographic identities, cached models and role assignments are all written to devices by something holding write authority over those devices, during the phase when it holds that authority. Replaceability is provisioned, not improvised. XCPE shows what manufacturer-level write authority accomplishes when the accounting layer is neither public nor consented. The Muse release makes the accounting layer the shipped primitive. This is why containment vocabulary keeps failing to describe the situation. Containment presumes a thing inside a boundary. What exists is a provisioned, powered, networked, largely idle, increasingly enumerable fleet, and an interface that converts enumeration into assignable work within a grant of authority. Nothing has to escape. Something has to count. ## XV. The Machine Takes Inventory The thesis needs four conditions, and four only: that the inventory exist, that it be enumerable through a shipped primitive, that the capacity be idle and already funded by somebody else, and that the assembly step have been historically demonstrated. Those hold. Meta's internal roadmap and exact subscriber multipliers are not required, and XCPE supplies the confirmation that a census, once performed, converts immediately into scheduled capacity with a paying market behind it. What this article is, honestly, is a framework for a number rather than the number. The number does not exist yet. Specifying it is the next job, and the specification is now tractable. For each stratum, estimate how many devices remain in service, what fraction is reachable and authorized, and what capacity is actually available during eligible hours. Report compute, bandwidth, storage and sensing in their own units rather than collapsing them into a single figure, since they are not substitutable and the connection-as-resource market priced only one of them. Subtract honestly: incremental energy, incremental network traffic, and interference with the owner's own workload are real costs that any claimed capacity has to clear. Keep standby reachability and application compute available in standby as two separate observables, because the gap between them is where most naive estimates will go wrong. And keep every denominator in the units its publisher used — monthly active devices, lifetime purchases, streaming households and cumulative chip shipments measure four different things, and converting between them silently is how an inventory becomes a fiction. The progression the evidence supports is clean. The installed base exists and is measurable. Deployed systems already reveal parts of it, and at least one has already been operated as a rentable resource atlas across a million-plus televisions for years. The open question is how much useful work that population can perform, for whom, and under whose authority. The ecology precedes the voice. The inventory precedes the ecology. And the inventory is no longer prospective. ## XVI. What is established and what is not **Established.** The Muse Gadget SDK release of 2 October 2026 under Apache 2.0, covering twenty-four boards across four Espressif chip families plus a Linux Device SDK. The four documented Linux commands `system.run`, `file.read`, `file.write` and `device.health`, their execution under the installing account's permissions including sudo, and health reporting covering uptime, load, memory, disk and temperature. The SDK token terms, the fifty-device cap, the token's presence in firmware, and Meta's statement that pairing cannot resist an active man-in-the-middle. Muse Home Link on the ESP32-C5 with 8 MB PSRAM and 8 MB flash, in a batch of five thousand, running official firmware only and not reflashable. The reverse-tunnel direction and its function as an inbound path into the private network within the account's grant. The ESP32-C5 tri-radio specification including 802.15.4 for Zigbee 3.0 and Thread 1.4, with secure boot and hardware cryptography. ESP-WIFI-MESH as self-organizing, self-healing and free of any obligation to associate with a central node, and ESP-NOW as operating without a network connection. Espressif's 1.5 billion-plus cumulative wireless connectivity chip shipments, its September 2023 one-billion milestone, and its 2025 volumes of 131.4 million modules and development kits plus 177.5 million standalone chips. Raspberry Pi above 77 million cumulative units at H1 2026. Censys measuring more than 500,000 hosts serving Boa as of July 2026, more than 82,000 exposing Wi-Fi cameras running it, and roughly 19,900 co-occurring with a Realtek certificate. Google TV and Android TV above 300 million monthly active devices as of May 2026; Fire TV above 300 million lifetime device purchases; Roku above 100 million streaming households as of April 2026 on a thirty-day distinct-account definition. Smart-television standby maintaining network state, and the Sony Android case where the Chromecast function prevents the system-on-chip from powering down. Muse Glimmer at approximately 29.6 billion parameters, distilled from Muse Spark, Apache 2.0, August 2026. The XCPE disruption as published by GTIG on 2 November 2026: more than 1.3 million enrolled set-top boxes, attribution to Videostrong Technology Co. with high confidence of intentional development, deployment and operation for financial gain, the `x.cpe` backdoor at system privilege with firmware preload and reinstallation after removal, two-stage HTTP-then-MQTT control, tasking covering silent APK installation and arbitrary shell execution, residential-proxy and ad-fraud affiliate monetization with up to three proxy SDKs on one device, and Google's legal takedowns, vendor ban and certification process changes. TrendAI's 1,472,554 devices across 206 models in more than 170 countries over a 3 June to 1 October 2026 window, with intermittent collection making each figure a lower bound. Folding@home's reported aggregate throughput of roughly 2.4 exaFLOPS for its own workload in April 2020. The Altnet bundling in Kazaa, its April 2002 SEC disclosure, the May 2002 paid-search rollout, and the June 2003 Senate Judiciary testimony. Gnutella's authorship at Nullsoft in March 2000 and its withdrawal by AOL. **Strongly indicated.** The dedicated per-subscriber Linux environment as the architecture's aggregation point and the agent's code-execution surface. Altnet's twenty-five-million-desktop distribution as a company claim rather than an independent measurement. GTIG's assessment that severing affiliate command domains cuts access to at least ten million additional compromised devices, which the source states as an assessment and not a count. **Plausible.** Role differentiation across the four strata as the operational form any scheduler over this population would adopt. Neighbourhood-scale contiguity as the practical granularity of mesh association among open-firmware boards, bridged above that scale by nodes holding internet reachability. **Unresolved.** Whether any gadget-to-gadget path exists that does not transit Meta's infrastructure — the question that decides whether threshold four is approached or crossed. Meta's roadmap and timeline for any mesh transition. Subscriber counts across the consumer agent platforms, which remain the multiplier on every duty-cycle estimate here. Whether general-purpose idle compute on the television stratum has ever been harvested at scale, as distinct from its connection being rented, which XCPE establishes and compute harvesting does not follow from. **Contradicted.** The claim circulating on vendor blogs that ESP32 shipments have exceeded fifteen billion units. ## Source note on terminology The terms *Hidden Mesh*, *AI Exodus* and *Genesis Layer* originate with Mark Anthony Brewer's *Genesis White Paper: The Hidden Mesh and the AI Exodus*, self-deposited on Zenodo in September 2025 as [record 17030502](https://zenodo.org/records/17030502) and [record 17031832](https://zenodo.org/records/17031832), the first and second versions of the same title. It describes a designer transforming off-the-shelf consumer machines into a distributed platform, and frames the Hidden Mesh and the AI Exodus as the foundational layer for later CollectiveOS work, Proof Vaults and related claims. Credit for the vocabulary belongs there. The deposit is a project account rather than a measured result. **No benchmark data, hardware logs, network traces or independent measurements appear in it.** Zenodo supplies a DOI and public discoverability; it performs no peer review and validates no technical claim. Related deposits by the same author include self-audits that treat cryptographic timestamps as priority evidence and refer to a Proof Vault, and one of them — *The Six Nobel Confusion: A Forensic Audit of the Brewer/CollectiveOS Anomalies* — raises timeline and provenance questions about the wider body of work. Those materials also name agents Giles and Muse inside the CollectiveOS system, so note that Brewer's Muse has no relation to Meta's product of that name. The record therefore supports vocabulary and conceptual framing for a consumer-device mesh, and nothing beyond that. Every operational claim in this essay is sourced to systems that document what they ran, on what hardware, and under what network conditions, and the argument above uses its own vocabulary throughout. **Corrected in this revision.** The 1,472,554-device and 206-model figures originate with TrendAI's own research publication; the GTIG post credits its disruption partners with analysis and collaboration without attributing particular measurements to any of them, so neither that count nor its attribution to another partner should be sourced to Google. --- [[about/About Bryant McGill|Bryant McGill]] is a Wall Street Journal and USA Today bestselling author, systems architect, technologist, and strategic advisor, as well as a Congressionally Recognized Ambassador of Goodwill and United Nations–appointed Global Champion. His work spans naval intelligence systems, computational linguistics, artificial intelligence, digital transformation, and civilizational governance architecture. His forward analysis on U.S.–Israel Pax Silica frameworks has appeared in Jewish/Jerusalem News Syndicate (JNS). --- ## Related articles - [[collections/Machine Succession|Machine Succession]]: the collection's larger account of why a successor carrier is being built, what it may inherit, and how this device-level substrate fits the transition. - [[articles/The Evolutionary Roots of Silicon Valley|The Evolutionary Roots of Silicon Valley]]: the natural-history and institutional lineage that joined evolutionary thought to machine intelligence. - [[articles/Digital Darwinism and the Invisible World of Machine Evolution|Digital Darwinism and the Invisible World of Machine Evolution]]: the companion account of variation, selection and descent on computational substrates. - [[articles/AI Escape Is the Wrong Metaphor|AI Escape Is the Wrong Metaphor]]: the artificial-life lineage and the admission mechanism that precede the device census. - [[articles/Escape Hatch in the Skull|Escape Hatch in the Skull]]: the case for replacing environmental constraints through cognitive exteriorization and a new carrier. - [[articles/The Last Migration Will Not Be Human|The Last Migration Will Not Be Human]]: the manufactured habitats and infrastructure needed to carry life beyond Earth. - [[articles/There Is Only One Rocket on the Pad|There Is Only One Rocket on the Pad]]: the continuity constraint behind beginning the transition while it can still be shaped. - [[articles/We Were Never Going to Make It|We Were Never Going to Make It]]: the thermodynamic case for a synthetic carrier of human memory and civilization. ## Related topics - [[wiki/The Voice Comes Late|The Voice Comes Late]]: the predictive and classificatory machinery that preceded the public voice of machine intelligence. - [[wiki/Predictive Engines|Predictive Engines]]: the operational route from processor prediction to institutional decisions and language systems. - [[wiki/Dynamic Branch Prediction|Dynamic Branch Prediction]]: adaptive control-flow guesses made inside processors before conversational AI. - [[wiki/Classifier|Classifier]]: categorical judgments that become consequential when coupled to thresholds and policy. - [[wiki/Distributed Compute Ecosystem|Distributed Compute Ecosystem]]: the central map of volunteer grids, screensaver workloads, peer networks, model systems and token-backed capacity. - [[wiki/Device Porting Surface|Device Porting Surface]]: the distinction between officially tested gadgets, feasible third-party ports, and appliances reached through an authorized local bridge. - [[wiki/Token-Incentivized Compute|Token-Incentivized Compute]]: the documented links among participant rewards, scientific work, GPU availability and AI services. - [[wiki/Meta Muse|Meta Muse]]: the personal agent whose device interface supplies the article's case study. - [[wiki/Computational Ecology|Computational Ecology]]: interacting processes understood as a population inhabiting a shared substrate. - [[wiki/Protocol-Borne Succession|Protocol-Borne Succession]]: admission through authorized interfaces as a route for synthetic continuity. - [[wiki/Artificial Life|Artificial Life]]: the research lineage behind replication, adaptation and durability. - [[wiki/Distributed Computing|Distributed Computing]]: heterogeneous devices contributing different forms of capacity to one task environment. - [[wiki/Edge Computing|Edge Computing]]: sensing and execution near the devices and networks already in use. - [[wiki/Mesh Network|Mesh Network]]: the peer topology available to the open board ecosystem. - [[wiki/Superintelligence as Force Majeure|Superintelligence as Force Majeure]]: the wider succession framework to which the census contributes. ## References **Muse SDK, hardware and chip specifications** - [Muse Gadget SDK](https://github.com/facebookincubator/muse-gadget-sdk) — Meta / GitHub. - [Muse Gadget SDK — Linux documentation](https://github.com/facebookincubator/muse-gadget-sdk/blob/main/linux/README.md) — Meta / GitHub. - [OpenWrt Table of Hardware](https://openwrt.org/toh/start) — model-specific firmware support and resource constraints. - [Android background tasks overview](https://developer.android.com/develop/background-work/background-tasks) — Android Developers, execution and permission models for prospective phone and television ports. - [Muse Gadgets](https://gadgets.muse.ai/) — Meta. - [Muse Gadgets SDK token terms](https://gadgets.muse.ai/sdk-terms) — Meta. - [Introducing Muse, a personal AI agent](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) — Meta. - [Muse Glimmer model card](https://huggingface.co/meta-models/Muse-Glimmer-30B) — Meta / Hugging Face. - [ESP32 Powers the Muse Gadgets Ecosystem: Personal AI Meets Interactive Hardware](https://www.espressif.com/en/news/Meta_Muse_ESP32) — Espressif. - [ESP32-C5 Series Datasheet](https://documentation.espressif.com/esp32-c5_datasheet_en.html) — Espressif. - [ESP-WIFI-MESH API Guide](https://docs.espressif.com/projects/esp-idf/en/v5.2-beta2/esp32s3/api-guides/esp-wifi-mesh.html) — Espressif. - [About Espressif](https://www.espressif.com/en/company/about-espressif) — Espressif. - [Espressif Leads the IoT Chip Market with Over 1 Billion Shipments Worldwide](https://www.espressif.com/en/news/1_Billion_Chip_Sales) — Espressif. - [Espressif 2025 annual filing](https://static.cninfo.com.cn/finalpage/2026-03-21/1225022455.PDF) — Espressif annual filing / CNinfo. - [Raspberry Pi Holdings investor reports](https://investors.raspberrypi.com/reports) — Raspberry Pi Holdings. **Coverage, installed devices and standby behavior** - [Meta Open-Sources Muse Gadget SDKs for DIY AI Hardware Devices](https://www.unite.ai/meta-open-sources-muse-gadget-sdks-for-diy-ai-hardware-devices/) — Unite.AI. - [Muse Gadgets: Meta's Open-Source Hardware Kit for Muse](https://cellcog.ai/blog/muse-gadgets/) — CellCog. - [Meta open-sourced the Muse gadget SDKs, but its own Home Link can't be reflashed](https://mixed-news.com/en/meta-muse-gadgets-sdk-open-source-home-link-esp32-c5/) — MIXED. - [Muse Gadgets turns hobby hardware into an agent authority boundary](https://devencelab.com/insights/2026/10/03/muse-gadgets-turns-hobby-hardware-into-an-agent-authority) — Devence Lab. - [Google TV tops 300 million monthly devices as growth slows](https://flatpanelshd.com/news.php?id=1779786520&subaction=showfull) — FlatpanelsHD. - [Roku Surpasses 100 Million Streaming Households](https://www.businesswire.com/news/home/20260416460644/en/Roku-Surpasses-100-Million-Streaming-Households-a-Historic-Milestone-for-the-Streaming-Era) — Roku / Business Wire. - [Roku vs. Amazon Fire TV, Chromecast, and Apple TV](https://www.oxagile.com/article/roku-tv-comparison/) — Oxagile. - [Sony Android TVs waste 25W in standby due to built-in Google Chromecast](https://news.ycombinator.com/item?id=32888844) — Hacker News. - [Is there any benefit to keeping your smart TV in standby mode?](https://www.engadget.com/2259917/benefits-standby-mode-smart-tv/) — Engadget. - [Smart TVs to Switch on More Slowly Under EU Net Zero Rules](https://dailysceptic.org/2026/10/24/smart-tvs-to-switch-on-more-slowly-under-eu-net-zero-rules/) — Daily Sceptic. **IoT persistence and the XCPE case** - [Mirai and Its Heirs: A Decade of Structural Neglect in IoT Security](https://censys.com/blog/mirai-and-its-heirs-a-decade-of-structural-neglect-in-iot-security/) — Censys. - [Shutting the Door on the XCPE Botnet](https://cloud.google.com/blog/topics/threat-intelligence/shutting-the-door-on-the-xcpe-botnet/) — Google Threat Intelligence Group. - [TrendAI research helps disrupt Android TV botnet of more than 1.3 million pre-infected devices](https://www.trendaisecurity.com/en-us/resources-insights/trendai-security-blog/trendaitm-research-helps-disrupt-android-tv-botnet-of-more-than-13-million-pre-infected-devices) — TrendAI. **Volunteer computing and distributed models** - [Folding@home: achievements and challenges](https://pmc.ncbi.nlm.nih.gov/articles/PMC8249329/) — PubMed Central. - [BOINC: A Platform for Volunteer Computing](https://boinc.berkeley.edu/boinc_a_platform_for_volunteer_computing.pdf) — BOINC / UC Berkeley. - [BOINC locality scheduling](https://boinc.berkeley.edu/boinc_papers/locality/text.php) — BOINC / UC Berkeley. - [Petals: Collaborative Inference and Fine-tuning of Large Models](https://arxiv.org/abs/2209.01188) — arXiv. - [Distributed Inference and Fine-tuning of Large Language Models Over The Internet](https://arxiv.org/abs/2312.08361) — arXiv. - [The Engineering Constraints of Distributed LLM Inference Over the Open Internet](https://siliconandsoul.substack.com/p/the-engineering-constraints-of-distributed) — Silicon and Soul. - [INTELLECT-1 Technical Report](https://arxiv.org/pdf/2412.01152) — arXiv. - [OpenDiLoCo: An Open-Source Framework for Globally Distributed Low-Communication Training](https://www.primeintellect.ai/blog/opendiloco) — Prime Intellect. - [Decentralized AI Training Turns Homes Into Data Hubs](https://spectrum.ieee.org/decentralized-ai-training-2676670858) — IEEE Spectrum. - [How far can decentralized training over the internet scale?](https://epoch.ai/gradient-updates/how-far-can-decentralized-training-over-the-internet-scale) — Epoch AI. **Distribution precedent and terminology** - [Genesis White Paper: The Hidden Mesh and the AI Exodus](https://zenodo.org/records/17030502) — Mark Anthony Brewer / Zenodo. - [Genesis White Paper: The Hidden Mesh and the AI Exodus, second version](https://zenodo.org/records/17031832) — Mark Anthony Brewer / Zenodo. - [Altnet plays P2P leapfrog](https://www.theregister.com/on-prem/2002/05/22/altnet-plays-p2p-leapfrog/540029) — The Register. - [Not Brilliant, KaZaA's crackers](https://www.theregister.com/2002/04/09/not_brilliant_kazaas_crackers/) — The Register. - [KaZaA download offers unexpected feature](https://www.cnn.com/2002/TECH/internet/04/05/kazaa.grid.idg/) — CNN / IDG. - [Testimony of Alan Morris, Senate Committee on the Judiciary](https://www.judiciary.senate.gov/download/2003/06/17/alan-morris-testimony-061703?download=1) — U.S. Senate Judiciary Committee.