/* drift.jsx — paper microsite + DRIFT systems prototype */
const DRIFT_REPO = "https://github.com/TaewoooPark/DRIFT";
const DRIFT_PAPER = "/drift-media/ai-neo-feudalism-and-intelligence-sovereignty.pdf";
const DRIFT_PAPER_EN = "/drift-media/ai-neo-feudalism-and-intelligence-sovereignty-en.pdf";
const DRIFT_DEMO_IMAGE = "/drift-media/drift-demo-complete.png";
const DRIFT_AXES = [
{
n: "01",
title: "Production",
brief: "Pool resources without surrendering ownership.",
body: "Personal devices, user cooperatives, independent foundations, and research communities contribute bounded capabilities rather than transferring the machines themselves.",
},
{
n: "02",
title: "Information",
brief: "Keep raw data and personal memory local.",
body: "A Job Capsule carries purpose, permissions, error bounds, validation rules, deadlines, and payout conditions. The underlying context remains in an owner-controlled vault.",
},
{
n: "03",
title: "Decision",
brief: "Separate assignment, verification, payment, and rulemaking.",
body: "No enterprise, state, scheduler, resource owner, or foundation should be able to allocate work, judge its own output, and close the network in the same decision.",
},
{
n: "04",
title: "Exit",
brief: "Make forkability executable, not symbolic.",
body: "A community must be able to leave with models, operating records, evaluation rules, pooled funds, and personalized workflow state — and continue producing elsewhere.",
},
];
const DRIFT_STAGES = [
["01", "Personal Resource Cells", "Register GPU, CPU, NPU, memory, storage, bandwidth, power, sensors, local data, and human judgment as narrowly scoped capabilities."],
["02", "Training · Compute · Inference", "Use training islands for loose synchronization, output lots for framework-neutral computation, and layer routing for low-latency inference."],
["03", "Purpose-bound security", "Keep source data in local vaults; transmit only permissions, bounded intermediate results, and disclosure records under a fixed privacy budget."],
["04", "Verification · Settlement", "Price verified results, not wasteful device time. Record execution, standby, maintenance, insurance, and public-good contributions separately."],
["05", "Compute ≠ Citizenship", "More hardware may earn more compute compensation, never more votes. Economic contribution and protocol citizenship remain distinct."],
];
const DRIFT_CAPITAL_CIRCUIT = [
{
id: "capex",
n: "01",
title: "Capex escalator",
metric: "$600B → $1.2T",
detail: "The paper begins with the widening threshold between hyperscaler infrastructure spending and the revenue needed to justify it: roughly $600B in 2024, recalculated to about $1.2T on the paper’s 2026 assumptions.",
consequence: "Capital expenditure is interpreted as evidence of future demand, while the expectation of demand justifies still more capital expenditure.",
},
{
id: "demand",
n: "02",
title: "Demand gap",
metric: "$50–150B",
detail: "The paper contrasts the roughly $1.2T infrastructure threshold with estimated end-user AI revenue of about $50–150B. It also cites a business study in which about 95% of surveyed organizations reported no measurable financial return from generative-AI investment.",
consequence: "The gap cannot be closed forever by cheaper consumer subscriptions alone.",
},
{
id: "valuation",
n: "03",
title: "Valuation feedback",
metric: "34% capex intensity",
detail: "High valuations support larger financing rounds and lease commitments; those commitments are then read as proof that the market must be enormous. The paper describes this as a self-reinforcing circuit rather than an independently verified demand signal.",
consequence: "Valuation and infrastructure spend become mutually legitimating inputs.",
},
{
id: "liquidity",
n: "04",
title: "Late liquidity",
metric: "retail as exit",
detail: "When private capital becomes harder to recycle, late-stage investors and insiders need liquidity windows. The paper reads the 2026 mega-IPO cycle and unusually large retail allocations as one possible terminal channel for that pressure.",
consequence: "Private infrastructure risk is distributed outward after the steepest valuation expansion.",
},
{
id: "pricing",
n: "05",
title: "Value-based pricing",
metric: "cost → rent",
detail: "The paper’s central pricing warning is a shift away from charging for tokens or compute cost and toward charging for the value of the outcome: a won contract, a discovered molecule, a productivity gain.",
consequence: "The provider’s claim moves from service fee toward a continuing share of production.",
},
{
id: "weights",
n: "06",
title: "Open-weight mirage",
metric: "file ≠ execution",
detail: "Even a permissively licensed model remains practically closed when its memory, accelerator, power, and networking requirements exceed individual ownership. Cloud hosting simply remonetizes the supposedly open asset.",
consequence: "The model file is free; the means of production remain rented.",
},
];
const DRIFT_PYRAMID = [
{
id: "capital",
label: "Infrastructure capital",
control: "Accelerators · power · datacenters · financing",
receives: "Lease revenue, asset appreciation, and the first claim on infrastructure scarcity.",
externalizes: "Demand risk and long-duration investment risk flow into customers, public markets, and suppliers.",
},
{
id: "frontier",
label: "Frontier model providers",
control: "Models · APIs · prices · safety policy · access",
receives: "Usage revenue, ecosystem rents, enterprise workflow lock-in, and a share of AI-created value.",
externalizes: "Outages, hallucinations, price changes, and migration costs are borne downstream.",
},
{
id: "operating",
label: "Enterprise operating layer",
control: "Agent graphs · context graphs · permissions · evaluation",
receives: "Organizational memory, performance data, and the power to define acceptable work.",
externalizes: "The cost of deskilling, surveillance, and brittle workflows falls on teams and workers.",
},
{
id: "labor",
label: "AI-supervised labor",
control: "Prompts · review queues · exception handling",
receives: "Faster output and a narrower role as supervisor of a system designed elsewhere.",
externalizes: "Accountability remains with the worker even as the practice needed for judgment erodes.",
},
{
id: "capability",
label: "Individual capability",
control: "Attention · memory · learned skill · personal context",
receives: "Convenience, fluent output, and immediate access to intelligence on demand.",
externalizes: "Cognitive lock-in, loss of practice, and the highest cost of exit accumulate at the base.",
},
];
const DRIFT_LOCKIN_LAYERS = [
{
id: "model",
label: "Model",
portable: "Prompts and nominal model settings",
residue: "Reasoning style, refusal boundaries, tone, tool affordances",
cost: "Rewrite evaluations and prompt conventions",
},
{
id: "orchestration",
label: "Orchestration",
portable: "Some workflow definitions and tool schemas",
residue: "Agent routing, permissions, retries, provider-specific tool behavior",
cost: "Rebuild the operational graph and failure handling",
},
{
id: "data",
label: "Data",
portable: "Files, tables, exports, and explicit records",
residue: "Access policy, embeddings, indexes, provenance, data relationships",
cost: "Re-index, re-authorize, and revalidate the entire corpus",
},
{
id: "knowledge",
label: "Knowledge",
portable: "Written procedures and curated memories",
residue: "Tacit exceptions, evaluator preferences, organizational habits",
cost: "Recover what the old system knew without ever stating",
},
{
id: "cognition",
label: "Cognition",
portable: "The user’s explicit work products",
residue: "Learned dependence on suggestions, reduced unaided practice, trust calibration",
cost: "Regain the ability and confidence to work outside the loop",
},
];
const DRIFT_FUTURE_STEPS = [
{
n: "01",
label: "Assist",
title: "The tool removes friction.",
event: "Drafts, code, analysis, and decisions arrive faster. Adoption is individually rational and immediately rewarding.",
human: "The person still owns the workflow and can reconstruct the result without the system.",
},
{
n: "02",
label: "Accumulate",
title: "Context becomes the product.",
event: "Mail, meetings, documents, code, customers, exceptions, and preferences gather inside one context graph.",
human: "Leaving now means losing relationships among data, not merely losing the files themselves.",
},
{
n: "03",
label: "Delegate",
title: "Recommendation becomes operating policy.",
event: "The system decides what to read first, whom to trust, what tone to use, what risk to flag, and which exception requires approval.",
human: "Following the recommendation becomes the safe and measurable definition of competence.",
},
{
n: "04",
label: "Deskilling",
title: "The supervisor loses the skill to supervise.",
event: "Unaided practice disappears. The worker remains accountable for outputs but becomes less able to reconstruct or challenge them from first principles.",
human: "The system is trusted more precisely when the user is least capable of auditing it.",
},
{
n: "05",
label: "Enclose",
title: "The office becomes the company town.",
event: "Identity, memory, reputation, tools, model access, and evaluation are governed by one operating layer.",
human: "The worker stays not because departure is forbidden, but because the self outside the system no longer feels employable.",
},
];
const DRIFT_ROOT_CAUSES = [
["01", "Public utility, private rule", "Society adopts AI like electricity or communications while access, price, memory, and policy remain private discretion."],
["02", "Openness collapsed into files", "Weights may be downloadable while training, data, energy, memory, and executable compute remain inaccessible."],
["03", "Gains rise; risk falls", "Organizations and providers capture saved time and revenue. Workers absorb deskilling, switching, outage, and error risk."],
["04", "Output hides capability loss", "Productivity accounting values the finished report but not the disappearing opportunity to practice judgment and regenerate skill."],
["05", "Exit is priced per individual", "The default alternative asks one person to buy a datacenter-class machine or accept a smaller model. The unit of escape is wrong."],
];
function DriftCapitalCircuit() {
const [activeId, setActiveId] = React.useState("capex");
const active = DRIFT_CAPITAL_CIRCUIT.find((item) => item.id === activeId) || DRIFT_CAPITAL_CIRCUIT[0];
return (
Fig. 01 · The capital circuit behind infrastructural dependence
Self-reinforcingValuation ↔ infrastructure
{DRIFT_CAPITAL_CIRCUIT.map((item, index) => (
))}
Selected circuit · {active.n}
{active.metric}
{active.title}
{active.detail}
System effect
{active.consequence}
Select any node. The diagram summarizes the paper’s financial argument; its figures are reported estimates and scenarios from the cited sources, not a live market dashboard.
);
}
function DriftNeoFeudalPyramid() {
const [activeId, setActiveId] = React.useState("frontier");
const active = DRIFT_PYRAMID.find((layer) => layer.id === activeId) || DRIFT_PYRAMID[1];
return (
Fig. 02 · AI neo-feudalism as a flow of control, value, and risk
value · rent · data↑
{DRIFT_PYRAMID.map((layer, index) => (
))}
↓risk · liability · deskilling
Selected layer
{active.label}
Controls
{active.control}
Receives
{active.receives}
Externalizes
{active.externalizes}
This is an interpretive diagram distilled from Sections 2.3 and 3, not an empirical class model or a claim that historical feudal law has literally returned. It visualizes the paper’s directional thesis: governing power and value claims tend upward while operational risk and capability loss are displaced downward.
);
}
function DriftLockInMap() {
const [activeId, setActiveId] = React.useState("orchestration");
const active = DRIFT_LOCKIN_LAYERS.find((layer) => layer.id === activeId) || DRIFT_LOCKIN_LAYERS[1];
return (
Interactive synthesis · compounding switching cost from model to cognition
{DRIFT_LOCKIN_LAYERS.map((layer, index) => (
))}
Portable artifact
{active.portable}
Non-portable residue
{active.residue}
Switching cost
{active.cost}
);
}
function DriftFutureTimeline() {
const [activeIndex, setActiveIndex] = React.useState(2);
const active = DRIFT_FUTURE_STEPS[activeIndex];
return (
Raw data, personal memory, adapters, and workflow context stay with their owner.
Governance railOne person, one vote
Agenda, evidence, approval, execution, and appeal are separated across roles.
Exit railCollective fork
Models and operating assets move to a new combination without asking a single gatekeeper.
A synthesis of the paper’s first figure. Decentralization is not geographic distribution alone: production, information, decision rights, and the practical ability to exit must move together.
);
}
function DriftRouteFigure() {
return (
Fig. 06 · DRIFT chain mode, keyed wire, and hop receipt
Mac · Apple GPU
MPS
decoder layers 0:k
{Array.from({ length: 7 }).map((_, i) => )}
PC · NVIDIA GPU
CUDA
decoder layers k:n
{Array.from({ length: 7 }).map((_, i) => )}
Output
Next-token probability
Hop receipt · Ed25519node · session · seq · layer range · SHA-256 input/output hash
DRIFT connects PyTorch’s Apple MPS and NVIDIA CUDA backends through a neutral byte contract. The receipt proves a signed, contiguous execution path; it does not, by itself, prove that every numerical operation was correct.
);
}
const DRIFT_WORKLOADS = {
training: {
n: "01",
label: "Training",
title: "Training islands replace global step-by-step synchronization.",
body: "Nearby devices perform multiple local steps, then exchange summarized updates. Churn is treated as a normal condition: an island can disappear, another can continue, and a round ends rather than weakening its privacy threshold.",
metrics: [["~500×", "less communication reported by DiLoCo"], ["90–95%", "compute utilization reported by OpenDiLoCo"], ["64–512×", "communication reduction range reported by Photon"], ["up to 45%", "training-time reduction reported by Go With The Flow"]],
},
compute: {
n: "02",
label: "Computation",
title: "Output lots separate the result contract from the hardware vendor.",
body: "Data cleaning, embedding, indexing, evaluation, safety checks, compression, and adapter merging become small jobs defined by input/output hashes, semantics, and error bounds. A device satisfies the contract without exposing a framework-specific implementation.",
metrics: [["8 Mbps → 10 Gbps", "link range explored by FusionLLM"], ["48 GPUs", "heterogeneous placement environment"], ["result-priced", "verified output sets the budget"], ["framework-neutral", "Apple NPU and CUDA can satisfy one contract differently"]],
},
inference: {
n: "03",
label: "Inference",
title: "Layer routing turns many small machines into one token path.",
body: "The first device computes a hidden state, passes it to the next layer range, and the final stage produces the next-token probability. The KV cache remains with each shard; only the residual stream and small positional metadata cross the boundary.",
metrics: [["layer split", "decoder ranges tile the model"], ["hidden state", "the primary per-token payload"], ["continuous", "conversation state survives across decode steps"], ["DRIFT", "the paper’s implemented data-plane proof"]],
},
};
const DRIFT_RESOURCES = [
{ id: "gpu", label: "GPU 12 GB", condition: "01:00–05:00 · fp16 allowed", role: "decoder layers or output-lot execution" },
{ id: "npu", label: "NPU", condition: "battery > 70% · local only", role: "bounded on-device preprocessing" },
{ id: "storage", label: "Encrypted storage", condition: "200 GB · 30-day retention", role: "checkpoint, evidence, or temporary recovery state" },
{ id: "bandwidth", label: "Bandwidth", condition: "50 Mbps · no public identity", role: "relay or encrypted intermediate transport" },
{ id: "data", label: "Local data", condition: "statistics only · no raw export", role: "purpose-limited computation at the data source" },
{ id: "validation", label: "Human validation", condition: "Korean domain · 20 min/day", role: "evaluation, dispute, or appeal evidence" },
];
const DRIFT_GOVERNANCE_ROLES = [
{ id: "agenda", label: "Agenda", may: "Propose a rule or model change and publish its rationale.", cannot: "Approve its own proposal or execute the final change." },
{ id: "evidence", label: "Evidence", may: "Collect tests, receipts, privacy impact, and minority reports.", cannot: "Choose the winner after seeing outcomes or suppress adverse results." },
{ id: "approval", label: "Approval", may: "Authorize a bounded change under one-person-one-vote citizenship.", cannot: "Acquire extra votes by contributing more hardware or tokens." },
{ id: "execution", label: "Execution", may: "Apply the approved change exactly as specified.", cannot: "Rewrite the decision or control the appeal channel." },
{ id: "appeal", label: "Appeal", may: "Pause, challenge, and reopen a disputed decision with new evidence.", cannot: "Be owned by the same effective controller as execution." },
];
const DRIFT_TECH_PLANES = [
{ id: "control", label: "Control", title: "Assign the route.", body: "The orchestrator discovers workers, assigns contiguous decoder-layer ranges, configures edge duties, and drives the decode loop. This is deliberately the prototype’s remaining central control point." },
{ id: "data", label: "Data", title: "Move neutral bytes.", body: "A 4-byte length prefix frames a msgpack dictionary and tensor payload. In chain mode the hidden state flows node-to-node instead of returning to the head after every hop." },
{ id: "kv", label: "KV", title: "Keep state with the shard.", body: "Each worker materializes the KV cache for its own layer range and session. The cache never crosses the network; doing so would erase the bandwidth advantage of the split." },
{ id: "security", label: "Security", title: "Key the wire.", body: "A network PSK and ephemeral X25519 exchange feed HKDF-SHA256. Directional keys authenticate and encrypt frames with ChaCha20-Poly1305." },
{ id: "trust", label: "Trust", title: "Sign every hop.", body: "Ed25519 receipts bind node identity, session, sequence, mode, layer range, and SHA-256 input/output hashes. The head verifies signatures, adjacency, anchors, and layer continuity." },
];
const DRIFT_REFERENCES = [
["Capital and pricing", "Cahn — AI’s $600B Question; Moody’s hyperscaler lease obligations; Allianz AI capex cycle; Bank of England FPC record; BIS Annual Economic Report 2026; MIT Project NANDA; OpenAI and Anthropic pricing materials."],
["Lock-in and capability", "Zapier AI Vendor Lock-In Survey; Deloitte State of AI in the Enterprise; Kosmyna et al. — Your Brain on ChatGPT; Lee et al. — Generative AI and Critical Thinking; Fan et al. — Metacognitive Laziness; International AI Safety Report 2026."],
["Company towns and rent", "Sadowski — The Internet of Landlords; U.S. National Park Service Pullman histories; U.S. Department of Labor FLSA history; 29 CFR 531; EU Data Act."],
["Open and distributed AI", "Open Source AI Definition 1.0; Anderson — BOINC; DiLoCo; OpenDiLoCo; Photon; Go With The Flow; FusionLLM."],
["Privacy, proof, settlement", "NIST privacy attacks in federated learning; Bonawitz et al. secure aggregation; Proof-of-Learning; zero-knowledge proofs of training; Filecoin, Akash, and Ethereum mechanism documentation."],
["Technical demonstrator", "TaewoooPark — DRIFT: Decentralized Routed Inference For Tokens, source code and README, MIT License."],
];
function DriftWorkloadExplorer() {
const [mode, setMode] = React.useState("training");
const active = DRIFT_WORKLOADS[mode];
return (
Fig. 04 · Different workloads require different forms of distribution
Reported reductions and utilization figures come from the works cited by the paper and describe those studies’ settings, not guaranteed network-wide performance.
);
}
function DriftCapsuleBuilder() {
const [selected, setSelected] = React.useState(["gpu", "storage", "validation"]);
const toggle = (id) => setSelected((current) => current.includes(id) ? current.filter((item) => item !== id) : [...current, id]);
const chosen = DRIFT_RESOURCES.filter((item) => selected.includes(item.id));
return (
Illustrative Personal Resource Cell · sample values, not paper measurements
Fig. 05 · Contribution is plural; citizenship is not for sale
ContributionCompensation basisEvidence
{rows.map((row) =>
{row.map((cell) => {cell})}
)}
Economic planeMore verified work → more compensation
≠
Political planeOne person → one vote
A large GPU cooperative can earn more for completed work. It cannot buy more citizenship, evaluate its own model, or suspend another person’s right to leave.
);
}
function DriftTechnicalPlanes() {
const [activeId, setActiveId] = React.useState("data");
const active = DRIFT_TECH_PLANES.find((plane) => plane.id === activeId) || DRIFT_TECH_PLANES[1];
return (
Position paper + systems prototype · 2026Intelligence infrastructure
AI Neo-Feudalism and Intelligence Sovereignty
AI is becoming essential infrastructure while the means of producing it remain privately governed. Models, compute, organizational memory, pricing, and exit are accumulating inside a small number of operating layers.
This paper names that condition AI neo-feudalism and proposes a decentralized infrastructure for training, computation, and inference — one that returns the ability to produce, verify, govern, and leave to individuals and communities.
AuthorsTaewoo Park · Seohyeon Hong
AffiliationsKAIST Physics · KAIST Chemical & Biomolecular Engineering
Visual 01 · Technical demonstratorCompletion frame from the supplied 3412 × 2038, 60 fps source recording. Front decoder layers run on Apple MPS; back layers run on NVIDIA CUDA; the center console shows one shared generation.Central thesis
Open weights without executable compute do not constitute a practical exit.
Software-era forkability must expand into AI-era runability: the real capacity to execute a model, retain one’s context, verify results, and continue under a different operating arrangement.
01 · Capital circuit
Why an intelligence economy can become a rent economy.
The paper argues that AI infrastructure is being financed as if future demand were already secured. Capital expenditure builds scarce compute; scarcity supports premium pricing; pricing expectations support valuation; valuation can replenish the capital that expands the same infrastructure.
This loop is productive—it creates extraordinary capabilities—but it also concentrates the right to set prices, shape access, and decide which forms of intelligence become economically legible.
$1.2T2026 revenue threshold
The paper’s updated estimate of revenue required to support projected AI infrastructure expenditure.
$50–150Bend-user revenue
A much smaller observed revenue range, revealing the scale of the monetization gap.
95%no measurable financial return
The share of surveyed organizations in the cited MIT Project NANDA report that had not recorded measurable financial return from generative AI.
≈800Mtokens / month
A paper estimate for when self-hosting may break even under its cited hardware and utilization assumptions.
02 · Dependence
The company town returned as an operating layer.
Historical company towns combined employment, housing, commerce, debt, and local rules under one owner. Their power came from making departure economically ruinous.
The modern AI stack repeats the structural pattern without physical walls. Cloud APIs replace company-owned production machinery; provider-specific credits and usage rights replace scrip; non-interoperable memory, agent graphs, and evaluation habits raise the cost of leaving.
Dependence is experienced as convenience. The worker remains because the system outside the system is slower, forgetful, and increasingly unfamiliar.
AI operating layerCompute · models · work memory · priceWorkflow · personalization · non-portability
81%concerned about lock-in
Technology leaders reporting concern about dependence on their AI vendors.
6%can switch cleanly
Organizations that believe they can change providers with little operational disruption.
47%operations could stop
Leaders who say core operations would halt if their primary vendor failed.
1 in 5mature agent governance
The cited Deloitte report’s estimate for organizations with mature governance of autonomous agents.
Figure 02 · AI neo-feudal pyramid
Capability widens downward. Control accumulates upward.
The pyramid is not a claim that every relationship is coercive. It is a map of asymmetry: many people contribute data, labor, attention, and subscription revenue, while comparatively few entities command frontier infrastructure, model policy, and the terms of access.
03 · A possible 2036
The dangerous future arrives as accumulated convenience.
This is a scenario, not a forecast. It traces the paper’s proposed sequence: assistance becomes accumulated memory, memory enables delegation, delegation can produce dependency, and dependency can become an economic requirement.
Recall ↓weaker ownership
A cited essay-writing study found lower recall and weaker feelings of authorship under LLM assistance.
Effort ↓metacognitive laziness
When the system supplies the next cognitive move, users practice less monitoring of their own reasoning.
AF447automation paradox
The paper uses aviation to show how rarely exercised human skill becomes crucial precisely when automation fails.
−6%unaided detection
A cited clinical finding: after AI assistance, unaided tumor-detection performance declined.
04 · Root causes
Five pressures reinforce the same center of gravity.
Concentration does not arise from one bad actor. It is produced by interacting technical, economic, institutional, and cognitive forces. A credible alternative must answer all five.
05 · Definition
Sovereignty has four inseparable axes.
Distributing servers is not enough. If a single scheduler still assigns work, fixes price, owns memory, or blocks a fork, the system has only relocated the gatekeeper.
{DRIFT_AXES.map((axis) => (
{axis.n}
{axis.title}
{axis.brief}
{axis.body}
))}
06 · Architecture
Five design stages turn the principle into infrastructure.
The design treats unreliable devices, heterogeneous runtimes, privacy boundaries, economic incentives, and governance capture as normal operating conditions — not edge cases to add later.
{DRIFT_STAGES.map(([n, title, body]) => (
{n}
{title}
{body}
))}
Figure 04 · Workload-specific topology
Training, compute, and inference should not share one forced topology.
The infrastructure offers common identity, resource discovery, accounting, and governance, while each workload keeps the communication pattern its physics requires. Select a workload to inspect its topology and evidence.
07 · Personal resources
A Job Capsule makes the workload portable, inspectable, and governable.
Instead of sending an opaque request into a provider-owned runtime, the paper proposes a portable description of the work: required capabilities, privacy boundary, verification rule, resource ceiling, and settlement terms. Build a capsule below.
01 · Local vault
Memory stays addressable by its owner.
Personal and organizational context is stored behind an explicit boundary, with export and deletion treated as system primitives.
02 · Purpose limitation
Disclosure becomes a typed object.
A workload states what data it needs, why, for how long, and whether a transformed or minimal representation is sufficient.
03 · Privacy budget
Participation has a non-negotiable floor.
Group size, disclosure limits, and any added noise are fixed before execution; payment cannot purchase a weaker privacy baseline.
08 · Governance
Operational power must be split before it can be shared.
A decentralized network can recreate a platform monopoly if admission, scheduling, verification, payment, and rulemaking collapse into one role. The design separates these powers and makes every interface replaceable.
Figure 05 · Proof, settlement, citizenship
Contribution can earn compensation without purchasing political control.
The paper keeps economic accounting separate from membership and governance. More GPUs may justify more reward; they should not automatically produce more citizenship.
09 · Practical exit
The relevant comparison is not open versus closed. It is rentable versus runnable.
A model can be downloadable yet operationally inaccessible. Practical sovereignty appears only when weights, compute, memory, verification, economics, and governance can move together.
Interpretive comparison. “Proposed / DRIFT direction” combines the paper’s target architecture with only the narrower execution-path features that DRIFT currently implements; it does not imply that the full governance or settlement system is deployed.
10 · Technical proof
DRIFT makes heterogeneous inference a real execution path.
DRIFT — Decentralized Routed Inference For Tokens asks a deliberately narrow question: can different personal devices jointly produce one model inference? Its answer is an implemented data plane, not a token-economy claim.
An in-process Apple MPS split matched the full-model token sequence on the same Mac.
130tokens · 3 prompts
Mac MPS layers 0:14 and a Colab NVIDIA T4 CUDA worker on layers 14:28 matched the reference runs reported by the paper.
48tokens · failover test
A worker was stopped mid-generation; the model was re-sharded, the prefix replayed, and the completed sequence matched the uninterrupted run.
What the prototype proves — and what it does not
Implemented proof
Heterogeneous devices can split decoder layers, pass hidden states through a neutral wire contract, survive replay-based failover, and journal a signed chain of contribution.
Current boundary
The paper’s prototype uses a central orchestrator, self-asserted membership, and no payout protocol. Participating workers still process decrypted inputs or hidden states.
Next design layer
A replaceable Job Capsule scheduler, Sybil-resistant admission, recompute or redundant audits, settlement, and data-privacy mechanisms are required for the full infrastructure.
11 · Research map
The argument crosses political economy, distributed systems, privacy, cognition, and governance.
These grouped notes are a navigational layer, not a substitute for the bibliography. The downloadable paper preserves the complete citations, qualifications, and source context.
Decentralization is not the rejection of trust. It is the decomposition of unilateral shutdown power.
To own intelligence is not merely to possess a model file.
It is to share the resources that execute it, define the boundary of one’s information, verify its results, receive one’s share of produced value, participate in its rules as an equal citizen, and ultimately leave for another operating system.
The end of a rented future is not a cheaper subscription. It is the capacity to produce our share without borrowing it.
Open technical demonstrator · MIT LicenseTaewoooPark / DRIFT
MPS ↔ CUDA · TCP + msgpack · chain routing · keyed wire · signed hop receipts · replay failover