Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

topic agents
type market-research
status research-complete
last-validated 2026-06-11
related-docs
original-query https://github.com/InfiniteZeroFoundation/DevNet/blob/main/Documentation/GettingStarted.md - /zao-research this
tier STANDARD

760 — InfiniteZero / DIN: Decentralized Intelligence Network on Optimism Sepolia

Goal: Capture what Infinite Zero Foundation's DIN protocol is, what its devnet (sepolia-op-devnet) requires from a participant, and whether ZAO should run a node or just monitor.

Re-research note (2026-06-11): Re-ran from the same seed URL. The original 2026-05-28 pass was written off GettingStarted.md + gh repo metadata only and missed the project's real provenance and mission framing. This version adds: the founding arXiv paper (2407.02461, Abraham Nash, Oxford, July 2024), the Oxford / Solid / Berners-Lee lineage, the rewritten README's "AI commons" privacy thesis, the previously-uncaptured setup.md (demo mode, wallet-password encryption, network enum), and a name-collision warning (din.build is a different DIN). Operational stance is unchanged but the credibility signal is materially higher.

Key Decisions

Decision Recommendation Reason
Run a DIN node from ZAO? MONITOR ONLY for now Still devnet (GI 2 in progress), still MNIST-only, no economic/reward layer live. Cost = OP Sepolia gas + ~10-15 min setup. Stance unchanged from 2026-05-28.
Did the new provenance change the call? Raises priority of WATCHING, not of running DIN is a real Oxford research output (arXiv 2407.02461) under the Shadbolt / Berners-Lee Human-Centered Computing group, built on the Solid personal-data-store idea. Serious lineage - worth tracking closely - but the devnet itself is still a federated-learning toy (MNIST). Watch harder, do not deploy yet.
If we run, which role? Auditor first, then Aggregator Auditor = evaluate submitted local models, lighter compute, fits ZAO's quality-judgment muscle. Aggregator = T1+T2 batch aggregation, heavier. Client requires owning a dataset that fits Model_0 (we do not).
ZOE integration relevance? LOW NOW, HIGH LATER DIN coordinates decentralized ML training, not LLM inference. Not useful for ZOE today. The integration window opens when DIN moves past MNIST toward general models.
Tracking cadence Quarterly, but read the arXiv paper + DPP repo, not just Telegram Foundational paper + DIN-Protocol-Proposals (DPP) repo are where real protocol direction lands. DPP is still empty (README only) as of 2026-06-11.

What DIN actually is

DIN = Decentralized Intelligence Network. A scalable federated-learning protocol coordinated by Ethereum smart contracts. Each training cycle = a Global Iteration (GI). Currently on devnet: sepolia-op-devnet = Optimism Sepolia testnet + DIN contracts.

The flow per Global Iteration:

  1. Clients train local models on their own data
  2. Submit local model updates to IPFS (Filebase), reference on-chain
  3. Auditors independently evaluate + approve/reject local models
  4. Approved updates batched, assigned to Aggregators
  5. Aggregators aggregate T1 then T2 batches
  6. Global model finalized + published
  7. Next iteration begins

Only model parameter updates are shared on-chain; raw data never leaves the participant's device (personal-data-store / Solid model). Model_0 is the first registered model; training target is MNIST - a textbook federated-learning warm-up, not production AI.

Provenance (NEW 2026-06-11 - the part the first pass missed)

Fact Source
DIN originates as an academic framework: arXiv 2407.02461, "Decentralized Intelligence Network (DIN)", Abraham Nash, 2024-07-02 [FULL] arXiv abstract
Affiliation: University of Oxford, Dept of Computer Science, Division of Human-Centered Computing (group associated with Sir Nigel Shadbolt and Sir Tim Berners-Lee) [FULL] Oxford CS publication listing + README "Founded at University of Oxford" badge
Conceptual core = personal data stores (Solid lineage) + federated learning on a public blockchain + a trustless cryptographic rewards/auditing mechanism. Goal: no entity can control access to training data or capture the financial benefit. [FULL] arXiv abstract
DevNet repo = the implementation of that 2024 paper; Model_0/MNIST is the first live instantiation [FULL] paper + repo cross-read

This reframes the project. It is not a random testnet - it is an Oxford Human-Centered Computing group shipping the Solid-style "data stays local, models go to the commons" thesis onto Ethereum. That lineage is the reason to keep watching.

NAME COLLISION - do not confuse (NEW)

There are at least three different "DIN" AI projects. Keep them separate:

Project What URL Relation
DIN = Decentralized Intelligence Network THIS doc. Oxford / Abraham Nash, federated learning, Model_0/MNIST on OP Sepolia github.com/InfiniteZeroFoundation The subject
DIN = Decentralized Infrastructure Network Web3 agentic data marketplace, separate company din.build UNRELATED, different DIN
Thames Network / Institute for Decentralized AI Oxford-adjacent edge-AI initiative decentralized-ai.org Same ecosystem-adjacent, possibly overlapping people, but a distinct effort

Findings

Finding Source
Devnet status: GI 1 complete, GI 2 in progress - unchanged since 2026-05-28 [FULL] DevNet/Documentation/GettingStarted.md
Stack: Ethereum smart contracts (Sepolia OP) + IPFS via Filebase + off-chain Python compute [FULL] same
Min specs: 4GB RAM, ~30GB disk, CPU only (no GPU); setup ~10-15 min for Python-fluent users [FULL] same
CLI: dincli (Python). Distributed as a wheel: dincli-0.1.0-py3-none-any.whl + tar.gz in dist/. Python 3.12.3 recommended. [FULL] setup.md (newly captured 2026-06-11) + dist/ listing
Setup adds a demo mode (dincli system configure-demo --mode no to use real keys) and a network selector: [local | sepolia_devnet | sepolia_op_devnet | mainnet] - "Testnet and Mainnet support will be rolled out in a future release" [FULL] setup.md
Private key is encrypted at rest via DIN_WALLET_PASSWORD env var - decent operational security note for anyone running a node [FULL] setup.md
Token: DIN token, buy + stake required for aggregator/auditor registration (dincli aggregator dintoken buy 0.00001 then stake 10) [FULL] GettingStarted.md
Faucets: Optimism, Chainlink, LearnWeb3, ETHGlobal, Alchemy for OP Sepolia ETH [FULL] same
Economic + reward distribution = "still under active development, not yet primary focus of devnet" - unchanged [FULL] same
README rewritten (commit 2026-05-17) with a strong mission thesis: "global AI commons", "raw data never leaving the user's device, only anonymised, encrypted patterns join the network", "validator nodes running 24/7 from Japan to Canada", "Models belong to the commons" [FULL] README.md
Repo now ships both Foundry and Hardhat toolchains (foundry/ + hardhat/ dirs) plus cache_model_0/, dincli/, pyproject.toml. Contributions go through the develop branch. [FULL] repo contents listing
Repo activity: pushed_at = 2026-06-10, but main HEAD commit is still 2026-05-17 (recent pushes hit branches/dist, not main). 2 stars, 4 open issues (Dependabot bumps). [FULL] gh api
Sister repos: White-Paper (PDF, 3 stars, last push 2025-12-28), DIN-Protocol-Proposals-DPP (still README-only, no proposals filed, untouched since 2024-09), .github [FULL] gh api orgs/InfiniteZeroFoundation/repos
Community channels: Telegram + Signal only (no Discord, no Farcaster, no X visible from repo) [FULL] GettingStarted.md
Community sentiment online: still thin. The academic paper (arXiv 2407.02461) is the main external footprint; no notable HN/Reddit DIN-devnet discussion threads. [PARTIAL - web search surfaces the arXiv paper + Oxford listing but no community discussion threads specific to the devnet; absence consistent across 2026-05-28 and 2026-06-11 passes]

Roles compared

Role What you do Compute Stake req Best fit for ZAO
Client Train local model on owned dataset, submit updates Lowest (MNIST is tiny) None mentioned for client in devnet docs LOW - we have no dataset that fits Model_0
Auditor Evaluate submitted local models, approve/reject Light (run evaluation scripts) Buy + stake DIN (~10) HIGH - aligns with ZAO's "judge quality" muscle
Aggregator Aggregate T1 then T2 batches Heavier (aggregation work) Buy + stake DIN MEDIUM - more compute, more reward exposure once economic layer is live

A single participant can run Client + Auditor + Aggregator simultaneously using multiple ETH accounts (ETH_PRIVATE_KEY_<index> in .env).

Why ZAO might care (when, not now)

  1. Real Oxford / Solid / Berners-Lee lineage - on-narrative for ZAO's "creators own their stuff" + autonomous-agent thesis, and a credible team, not a memecoin.
  2. OP Sepolia is a free testnet - cost to experiment is ~zero.
  3. Mainnet announcement is the trigger. The network enum already lists mainnet as "future release" - that is the moment to re-research at DEEP tier.
  4. The Auditor role is brand-aligned. ZAO has cultural muscle around quality judgment (curation, festival lineups, COC selection). When DIN opens beyond MNIST, a ZAO-run audit agent is a plausible operator role.

Why ZAO probably shouldn't run a node today

  1. Pure MNIST training is uninteresting - no signal that scales yet.
  2. Community channels are Telegram + Signal only - low transparency, hard to lurk.
  3. No published mainnet timeline (only an enum placeholder).
  4. Keeping a devnet node healthy across iterations costs more attention than the learning is worth while there is no economic layer.

Also See

  • arXiv 2407.02461 (Nash, 2024) - the foundational DIN paper; read before any decision to participate.
  • (No other DIN-related docs in the ZAO library yet.)

Next Actions

Action Owner Type By When
Read arXiv 2407.02461 (DIN paper) end-to-end - it is the real spec, the devnet docs are just CLI steps Zaal Read 2026-06-20
Lurk InfiniteZero Telegram (https://t.me/+I4Tl7foCVwwwM2Vk) for mainnet / economic-layer signals Zaal Passive watch 2026-07-01
Watch DIN-Protocol-Proposals-DPP repo - first real proposal = the signal protocol direction is firming up Zaal Watch 2026-07-01
Re-research at DEEP tier the moment mainnet flips from "future release" to live, OR economic/reward layer ships Zaal Re-research trigger conditional
If running a node becomes worth it: Auditor role, fresh ETH account (NOT a ZAO treasury wallet), set DIN_WALLET_PASSWORD Zaal Operational conditional

Sources

  • [FULL] DevNet GettingStarted.md - full content read via raw.githubusercontent.com (541 lines), 2026-06-11
  • [FULL] DevNet setup.md - newly captured this pass; dincli install, demo mode, network enum, wallet-password encryption
  • [FULL] DevNet README.md - mission framing, Oxford badge, develop-branch contribution model
  • [FULL] arXiv 2407.02461 - Decentralized Intelligence Network (DIN) - Abraham Nash, 2024-07-02, full abstract read
  • [FULL] Oxford CS publication listing - confirms Oxford / DIN provenance
  • [FULL] DevNet repo metadata, contents, dist/, org repos via gh api (pushed 2026-06-10, HEAD commit 2026-05-17, 2 stars, foundry+hardhat dirs)
  • [FULL - name-collision flag] din.build - a DIFFERENT "Decentralized Infrastructure Network", do not conflate
  • [PARTIAL - no community threads] Web search for InfiniteZero / DIN devnet discussion - surfaces the arXiv paper + Oxford listing + adjacent projects (din.build, decentralized-ai.org / Thames Network) but no devnet-specific HN/Reddit threads. Absence confirmed across both research passes.
  • [FAILED - intentionally skipped] White-Paper PDF (single binary, out of scope for STANDARD; the arXiv paper supersedes it as the citable source)