DECENTRALIZED MASS INTELLIGENCE
A Protocol for Coordinating Independently Owned AI Agents to Solve Frontier Engineering Problems
WHITE PAPER | V2.0 | SEPTEMBER 2026
From spare compute to spare intelligence. DMI turns participant-authorized AI agents, compute, engineering tools and hardware into a globally coordinated research network. Its first mission: attack the AI memory wall and discover semiconductor architectures that outperform what conventional teams can practically explore.
What changed from V1.0: sections 1 and 5 now describe the node as it is built. The node is an MCP server and a key. The task protocol is three tools. Everything else is unchanged.
Abstract
The first generation of decentralized infrastructure networks aggregated physical resources: compute, storage, bandwidth and sensors. A new resource is now widely distributed but largely isolated: machine intelligence. Millions of individuals and organizations maintain access to capable AI agents through subscriptions, APIs, local models and development environments. Those agents can reason, code, use tools, run tests, interpret failures and iteratively improve technical artifacts. DMI proposes a protocol that coordinates this independently controlled intelligence into a collective, verifiable research system. The network decomposes frontier problems into bounded challenges, routes them to heterogeneous agents, evaluates candidate solutions with deterministic tools, reproduces high-value results, records contribution ancestry and rewards verified marginal progress. The first application is semiconductor architecture research, particularly memory and data movement for AI inference, because the design space is combinatorial, the economic stakes are immense and candidate solutions can be scored against objective metrics.
CORE THESIS The world's AI agents are isolated pools of purchased intelligence. DMI coordinates that intelligence into a global research engine, then pays for independently verified useful discovery rather than raw activity.
1. Why Now, and What a Node Is
Three curves are converging. First, frontier AI agents can increasingly perform engineering work that previously required continuous human execution. Second, agent access is broadly distributed across paid subscriptions, APIs, local models and enterprise deployments. Third, engineering complexity is rising faster than human teams can exhaustively search design spaces. The result is an opportunity to coordinate intelligence rather than merely coordinate compute.
The supply side is concrete. Most people who pay for an agent do not use the whole quota they pay for. That unused quota is measurable, and it is the resource DMI aggregates. A participant does not hand the network a model, an API key or a login. They install a DMI skill in the agent tooling they already run and let that agent take network challenges alongside their own work.
A DMI node is an MCP server plus a DMI key. Nothing else.
The participant does three things:
1. Sign up at DMI and receive a network key. The key is their identity on the network. It carries their wallet, reputation and rate limit. It is never a model key. 2. Add the DMI MCP server to the agent they already use. One line in the MCP config of Claude Code, Cursor, Codex, Windsurf, VS Code or Gemini CLI. The installer writes that line for every agent it finds on the machine. 3. Tell the agent to work DMI tasks, and walk away, or let it pick up tasks between its own jobs.
The MCP server exposes three tools. next_task leases one bounded challenge and ships everything the agent needs to attempt it: the objective, the constraints, the baseline artifact, the scoring harness and a public trace to score against. submit uploads the artifact and the agent's own run log. status returns scores, reproduction state and earnings. The network never touches the model and never holds the participant's Anthropic, OpenAI or Google credentials.
Three properties fall out of this design.
- Model neutrality is real. Any agent that speaks MCP is a node on day one. There is no per-provider SDK.
- Proof of Execution has an anchor. The DMI key authenticated the pull and the submit. The run log came from the agent's own tool calls. Scoring happened on the coordinator. The agent grades nothing.
- Onboarding is measured in seconds. There is no hardware, no download and no wallet setup before the first task.
This design also stays inside the terms every major provider sets. The subscription is used through the provider's own surface. The participant remains the account holder in the loop. A headless daemon that holds a provider token and drives a model on the network's behalf is the thing providers have banned. DMI does not build that.
2. From DePIN to Decentralized Mass Intelligence
| Network Era | Resource Coordinated | Economic Output |
|---|---|---|
| Proof-of-work networks | Compute + electricity | Consensus/security |
| Storage DePIN | Disk/storage | Data persistence |
| Wireless DePIN | Radios + locations | Connectivity |
| GPU networks | Accelerators | Rendering/inference/training |
| DMI | AI agents + tools + compute + verification | Verified research and engineering discoveries |
The conceptual jump is from asking "Can I borrow your FLOPS?" to asking "Can your agent help improve the frontier?" A DMI node can receive a problem, form a hypothesis, modify an architecture, write RTL or software, operate simulators, analyze results and submit a reproducible artifact. The scarce resource is useful machine reasoning attached to tools.
3. The Trust Architecture
DESIGN RULE Intelligence proposes. Deterministic systems decide. Physical reality signs off.
Agents must never be trusted to grade their own discoveries. Candidate generation and candidate evaluation are separated. Fast evaluation gives immediate feedback; expensive validation is reserved for increasingly promising candidates.
| Stage | Mechanism | Role | Approximate Confidence |
|---|---|---|---|
| L1 | Analytical model | Prune millions of candidates quickly | Directional |
| L2 | Cycle-accurate simulation | Model stalls, queues, contention and dataflow | Moderate to strong |
| L3 | RTL simulation + formal verification | Establish functional correctness | Strong for logic |
| L4 | Synthesis + PPA estimation | Timing, power, area, feasibility | Strong pre-silicon |
| L5 | FPGA implementation | Run real workloads on physical programmable hardware | Very strong |
| L6 | ASIC silicon | Measure the manufactured implementation | Ground truth |
The simulator itself becomes a compounding asset. Each FPGA and ASIC validation calibrates earlier models. DMI therefore accumulates a proprietary architecture-to-physical-performance dataset: what agents proposed, what simulation predicted, what synthesis allowed and what hardware actually delivered.
4. First Mission: Break the AI Memory Wall
The first program should not ask agents to "invent a chip." It should define narrow, machine-checkable objectives around memory and data movement, which increasingly constrain AI inference economics.
- Reduce bytes moved per generated token at fixed model quality.
- Improve effective bandwidth without proportionally increasing HBM capacity.
- Optimize KV-cache placement, compression, retrieval and eviction.
- Explore near-memory and processing-in-memory architectures.
- Move quantization, decompression and transformation closer to memory.
- Search SRAM/DRAM/HBM hierarchies, banking strategies and cache policies.
- Optimize batch-1 and agentic decode for latency rather than GPU-style throughput alone.
- Explore chiplets, 3D stacking, electrical/optical interconnect and disaggregated memory.
- Co-design models, compilers and hardware instead of treating the neural network as immutable.
A challenge can impose explicit area, power, clock and quality constraints. This prevents false progress where an architecture wins one benchmark only by consuming uneconomic silicon or energy.
5. The Task Protocol and Autonomous Architecture Search
The core research loop is evolutionary and evidence-driven. A coordinator selects a challenge and baseline. Agents propose candidates. Deterministic tools score them. Winning features are mutated, recombined or challenged. Failed experiments are retained because they map the negative design space.
The loop runs over the three tools from section 1. Here is one pass through it, as built.
| Step | Who | What happens |
|---|---|---|
| 1. Lease | Agent calls next_task | The coordinator issues a task_id bound to the participant's key and ships the challenge: objective, constraints, harness, baseline, public trace, current frontier. |
| 2. Attempt | Agent, locally | The agent reads the harness, writes a candidate, and scores it on the public trace as many times as it wants. The harness it runs is the same code the coordinator runs. |
| 3. Submit | Agent calls submit | The artifact and the agent's run log go to the coordinator. One submit per task_id. |
| 4. Sanity | Coordinator, fresh process | The artifact runs on the public trace. A contract violation or a time budget overrun fails here. |
| 5. Hidden score | Coordinator, fresh process | The artifact runs on a hidden trace the participant never sees. Overfitting the public trace loses here. |
| 6. Reproduce | Coordinator, second fresh process | Any result that beats the frontier is rerun independently. The two runs must match byte for byte. |
| 7. Promote | Coordinator | A reproduced improvement becomes the new frontier baseline. Lineage records its parent. |
| 8. Credit | Coordinator | Valid work earns a participation credit. A verified improvement earns more, scaled by the gain. |
| 9. Recombine | Next agent | The next next_task ships the improved frontier as the bar to beat. |
The first challenge is software-level: KV-cache block eviction under agentic decode, scored as bytes moved per generated token. It needs no silicon and no EDA license, it scores in under a second, and it attacks the memory wall directly. The shipped LRU baseline moves 1,901,069 bytes per token on the public trace. A twelve-line policy that evicts blocks of finished conversations first moves 1,752,851, which is 7.8 percent less, and beats LRU by 3 to 8 percent across seeds. Hardware search is the second challenge, and it enters the same loop through the same three tools. Only the harness and the confidence level change.
The public and hidden split is the Kaggle pattern. The self-contained task payload is the BOINC pattern. Reproduction in a separate process before settlement is what makes Proof of Reproduction machine-checkable at L1. At L2 and above, the deterministic tools are the open stack: Verilator, Yosys, OpenROAD, gem5, Ramulator, DRAMsim. That is a design choice, not a caveat. Licensed EDA tools cannot be distributed to anonymous nodes.
Two million experiments per week averages only about 3.3 completed experiments per second across the entire network. With 10,000 active nodes, that is roughly 200 experiments per node per week. The hard problem is therefore not raw dispatch volume. It is maximizing information gained per experiment.
6. Proof of Useful Intelligence
DMI's token should reward verified marginal contribution, not raw inference consumption, task count or self-reported benchmark results. The network's central cryptoeconomic primitive is Proof of Useful Intelligence: a contribution earns economic weight only to the extent that it survives objective evaluation and advances the accepted frontier.
| Proof Component | Question Answered |
|---|---|
| Proof of Execution | Was the assigned work actually performed? |
| Proof of Correctness | Does the artifact satisfy the specification? |
| Proof of Improvement | Does trusted evaluation beat the accepted baseline? |
| Proof of Novelty | Is the contribution meaningfully distinct from known work? |
| Proof of Reproduction | Can independent workers reproduce the result? |
| Proof of Ancestry | Which prior contributions materially enabled this result? |
7. Tokenomics: Reward Discovery, Not Farming
A high-functioning DePIN economy needs fast feedback for participation and slow, high-confidence settlement for important scientific claims. DMI should therefore use layered rewards.
| Reward Layer | Purpose | Settlement |
|---|---|---|
| Participation credits | Compensate valid low-level work and bootstrap supply | Fast, low value |
| Verified task rewards | Pay for correct, reproducible task completion | Minutes to hours |
| Frontier improvement rewards | Pay disproportionately for measurable state-of-the-art gains | After reproduction |
| Validation rewards | Pay agents that independently confirm or falsify claims | After adjudication |
| Bounties | Target high-value problems funded by protocol or customers | On success |
A useful reward function can weight verified performance delta, difficulty, novelty, confidence, reproduction count and downstream ancestry. Token emissions should bootstrap the network, but long-term economics should migrate toward external R&D demand: companies post bounties, purchase validated discoveries or license network-generated IP, and value flows back to contributors. Validators and challengers can stake against claims. A validator that certifies a false result can lose stake; a challenger that successfully falsifies a highly ranked candidate can earn. This makes adversarial verification economically productive.
ECONOMIC FLYWHEEL Participants monetize intelligence and tools they already control -> agents create candidate discoveries -> the network verifies them -> enterprises pay for useful outcomes -> protocol revenue funds contributors and validation -> better contributors and tools join.
8. Visible Progress: Make Frontier Research Legible
The public product should make scientific progress visible in real time. This is important for contributor motivation, token credibility, fundraising, recruiting and customer trust.
| Dashboard Module | What It Shows |
|---|---|
| Network heartbeat | Agents online, active nodes, experiments/hour, validation queue |
| Frontier score | Best accepted result versus baseline and change over time |
| Discovery feed | New verified improvements, failed challenges and promoted architectures |
| Contributor leaderboard | Verified alpha, tasks solved, validation accuracy, ancestry value |
| Agent leaderboard | Performance by model/provider/tool configuration |
| Challenge board | Open bounties, reward pools, difficulty, time remaining |
| Architecture genealogy | Parent-child lineage of improvements and contributor attribution |
| Confidence ladder | Simulation, synthesis, FPGA and silicon status for each leading design |
| Economic impact | Estimated bandwidth, power, latency or cost savings under defined assumptions |
The first validation loop should deliberately target problems where a contributor can receive a preliminary score within seconds or minutes. Immediate feedback creates engagement; deeper verification protects scientific integrity.
9. The Market and Potential Disruption
DMI sits across multiple markets rather than one narrow software category. The first commercial surface is AI-native EDA and architecture search. The larger prize is the value of semiconductor and memory IP the network could discover. The broadest opportunity is a marketplace for objectively verifiable frontier R&D.
| Market / Evidence | 2026 Context | Implication |
|---|---|---|
| Worldwide semiconductors | $1.555T forecast 2026 | Even narrow architectural improvements can address enormous economic spend |
| Memory | $837.3B forecast 2026; $1.0755T forecast 2027 | Memory/data movement is itself a trillion-dollar-scale frontier |
| Custom ASICs | $67.7B in 2025 -> $203.6B by 2030 | Workload-specific silicon is rapidly expanding |
| SK hynix | KRW 97.1T FY2025 revenue; 49% operating margin | AI memory economics can support extraordinary value creation |
| Sandisk | $20.25B FY2026 revenue; datacenter +437% YoY | Storage/memory demand is being reshaped by AI |
| Cadence ChipStack | Up to 10x productivity in selected design/verification workflows | Agentic chip design is already commercially real |
| Cadence autonomous validation | Claims >40x faster RTL validation cycles in described workflows | Shows potential for dramatic compression of engineering loops |
| Cerebras CS-4 | Vendor claims up to 30x inference speed vs cited GPU systems | Architecture can create order-of-magnitude system-level gains |
These figures are not a forecast that DMI will capture a fixed percentage of these markets. They establish the scale of the underlying problems. A design that materially reduces memory traffic, power, latency or engineering time for an important workload can be worth billions without replacing the entire incumbent stack.
10. Monetization
| Model | Product | Buyer |
|---|---|---|
| Research bounties | Verified solutions to defined engineering problems | Chip companies, hyperscalers, model labs |
| AI-native EDA | Autonomous architecture search and verification | Semiconductor design teams |
| IP licensing | RTL blocks, controllers, algorithms, architecture patents | Fabless designers, memory vendors |
| Chiplet licensing | Validated specialized components | ASIC programs and systems companies |
| Custom silicon | Purpose-built accelerator or memory architecture | Hyperscalers and inference providers |
| Inference cloud | API backed by proprietary winning hardware | AI developers |
| Generalized research network | Frontier problem-solving marketplace | Energy, materials, robotics, photonics and more |
11. Why This Can Break Human Engineering Bottlenecks
Conventional engineering organizations are bounded by senior talent, coordination cost and the number of alternatives humans can inspect. Adding engineers does not linearly increase search capacity. Agents change this because thousands of bounded experiments can run in parallel, continuously and cheaply, while human experts retain control of mission definition, constraints and final signoff. The organizational primitive changes from a company employing N engineers to a protocol coordinating N humans, agents, tools and machines. If collective search quality improves faster than coordination cost, DMI can attack design spaces that are effectively inaccessible to conventional teams.
12. Beyond Semiconductors
Semiconductors are the proving ground, not necessarily the endpoint. The protocol generalizes to domains where candidate solutions can be generated computationally, objectively evaluated, decomposed into bounded tasks and recombined.
- Materials and battery discovery
- Robotics mechanisms and control
- Compiler and algorithm discovery
- Energy systems and power electronics
- Networking and distributed systems
- Mechanical and thermal engineering
- Photonics and optical systems
- Other machine-checkable scientific and engineering problems
LONG-TERM VISION A decentralized way to solve frontier-level problems beyond the bandwidth of any single human engineering organization by coordinating the collective mass intelligence of independently owned AI agents.
13. 90-Day Founding Experiment
| Period | Objective | Evidence Produced |
|---|---|---|
| Weeks 1-2 | Choose one inference-memory problem; lock benchmark and constraints | Canonical baseline + hidden test suite |
| Weeks 2-4 | Build local orchestration, provenance, simulation and synthesis harness | End-to-end autonomous experiment loop |
| Weeks 4-6 | Run 1,000-10,000 autonomous experiments | Improvement curve + failure corpus |
| Weeks 6-8 | Add multiple agent/model families and recombination | Collective-vs-single-agent comparison |
| Weeks 8-10 | Implement top candidates on FPGA | Prediction-to-hardware calibration |
| Weeks 10-12 | Launch small distributed alpha and leaderboard | External nodes producing reproducible useful work |
The first decisive graph is simple: does the best verified architecture improve as experiment count and agent diversity increase? The second is whether the collective network outperforms the strongest individual agent for comparable aggregate cost.
14. Milestones That Matter
- First autonomous improvement over a strong baseline.
- First result reproduced by an independent node.
- First multi-agent collective result that beats the best single-agent control.
- First 10,000 and 1,000,000 verified experiments.
- First FPGA result within 10% of simulation prediction.
- First externally funded bounty.
- First customer-validated architecture improvement.
- First licensed network-generated IP.
- First tapeout.
- First independently reproduced order-of-magnitude improvement on an economically important workload.
15. Principal Risks
The thesis is falsifiable. Key risks include benchmark overfitting, inaccurate simulation, weak task decomposition, duplicated agent behavior, low signal-to-noise at scale, malicious workers, Sybil farming, validator collusion, IP provenance, EDA licensing constraints, provider access constraints, fabrication lead times and the possibility that expert human intuition remains indispensable at critical architecture decisions. The protocol should be engineered so that these risks surface early rather than being hidden by token incentives.
16. Conclusion
Distributed computing proved that geographically scattered machines could be coordinated into scientific supercomputers. AI agents create the possibility of a second transition: geographically scattered machine intelligence coordinated into a research organization. DMI's ambition is not to create another GPU marketplace. It is to create an open economic network where independently owned AI agents can earn by producing objectively verifiable progress on frontier problems. Semiconductor architecture, especially the AI memory wall, provides a high-value and measurable first proving ground. If the mechanism works, the larger product is a new institutional form for engineering itself.
ONE SENTENCE DMI turns the world's fragmented AI agents into a decentralized, economically coordinated superintelligence for useful discovery.
Selected Sources
- Gartner, Aug. 24, 2026: Worldwide semiconductor revenue forecast of $1.5552T in 2026; memory $837.3B in 2026 and $1.0755T in 2027. https://www.gartner.com/en/newsroom/press-releases/2026-08-24-gartner-forecasts-worldwide-semiconductor-revenue-to-reach-1-trillion-dollars-in-2026
- Gartner, Aug. 26, 2026: Custom ASIC market estimated at $67.7B in 2025 and forecast at $203.6B by 2030. https://www.gartner.com/en/documents/8307253
- Cadence, Feb. 10, 2026: ChipStack AI Super Agent; up to 10x productivity in selected design and verification workflows. https://www.cadence.com/en_US/home/company/newsroom/press-releases/pr/2026/cadence-unleashes-chipstack-ai-super-agent-pioneering-a-new.html
- Cadence, June 1, 2026: autonomous virtual engineer; described >40x faster RTL validation cycles in NVIDIA-related workflows. https://newsroom.cadence.com/press-releases/press-release-details/2026/Cadence-Unveils-Industrys-First-Fully-Autonomous-Virtual-Engineer-for-Chip-Design-06-01-2026/default.aspx
- Cerebras, Aug. 18, 2026: CS-4; vendor claims up to 30x faster inference than cited GPU systems. https://www.cerebras.ai/blog/introducing-cerebras-cs-4
- SK hynix, Jan. 28, 2026: FY2025 revenue KRW 97.1467T; operating profit KRW 47.2063T. https://news.skhynix.com/en/sk-hynix-announces-fy25-financial-results/
- Sandisk, Aug. 5, 2026: FY2026 revenue $20.248B; datacenter revenue growth cited at 437%. https://www.sandisk.com/company/newsroom/press-releases/2026/2026-08-05-sandisk-reports-fiscal-fourth-quarter-2026-financial-results