Lightning-Coordinated Decentralized AI Training

A protocol for permissionless machine learning using Bitcoin's Lightning Network as the coordination, payment, and incentive layer.

Abstract

We propose a protocol for decentralized AI model training that replaces token-based coordination (Bittensor TAO, custom governance tokens) with Bitcoin Lightning Network micropayments. Participants contribute GPU compute to distributed training runs and receive payment proportional to their measured contribution quality — settled in seconds, denominated in sats or USDT via Taproot Assets, with no staking requirements, no token governance, and no identity verification. The protocol combines three proven components: (1) SparseLoCo gradient compression enabling training over commodity internet at 146x compression, (2) L402 HTTP payment gating for gradient exchange, and (3) hold-invoice conditional payments that release funds only upon validator attestation of gradient quality. We address the semi-centralized coordinator problem — the primary trust assumption — through a layered mitigation strategy combining deterministic validation replay, DLC-bound payment settlement, and federated multi-validator consensus that reduces trust requirements to "at least one of N validators is honest." We extend the protocol to decentralized autoresearch — autonomous optimization bounties where AI agents compete to improve any quantifiable metric, paid per validated improvement via the same hold-invoice escrow. Autoresearch bounties require no GPU, run on any hardware, and have an essentially unbounded addressable market — we argue they represent the protocol's largest practical application.

1. Introduction

1.1 The Coordination Problem

Training large language models requires coordinating hundreds of GPUs for weeks. This has historically required centralized clusters with high-bandwidth interconnects (400 Gb/s InfiniBand). Recent work has proven that training can occur over commodity internet with dramatically compressed communication. DiLoCo (DeepMind, 2023) demonstrated 500x communication reduction. Prime Intellect scaled this to 10B (INTELLECT-1), 32B (INTELLECT-2), and 100B+ MoE (INTELLECT-3) parameters across globally distributed GPUs. Covenant-72B trained a 72B model across 70+ anonymous peers on Bittensor. Independently, Nous Research's DisTrO achieved 857–10,000x communication reduction on LLaMA 2 1.2B. The communication bottleneck is solved. What remains unsolved is coordination and incentive design — how to pay for contributions, verify quality, and align incentives without custom tokens or trusted operators.

1.2 The Token Problem

Most decentralized AI projects require custom tokens: Bittensor (TAO), Gensyn (AI token), Nous/Psyche (NOUS), io.net (IO), Akash (AKT). A peer-reviewed empirical study of Bittensor (arXiv 2507.02951, 6.6M events, 121,567 wallets) quantifies what this creates:

• Wealth concentration — top 1% of wallets control median 89.8% of stake; fewer than 2% of wallets command a 51% majority in most subnets
• Stake-driven rewards — economic stake is the dominant predictor of rewards (r=0.80-0.95), while performance contributes "only modestly for validators and very weakly for miners" (r=0.10-0.30)
• Price volatility — TAO's December 2025 halving cut daily emissions from ~7,200 to ~3,600 TAO; without external revenue, most miners are unprofitable
• Governance overhead — token holders control subnet parameters, creating political dynamics orthogonal to compute quality
• Barrier to entry — Bittensor validator permits require 1,000+ TAO (~$215,000); even miner registration is a dynamic burn that fluctuates with demand

Notable exceptions: Prime Intellect operates permissionless compute pools with no token, and Hivemind/Petals is open-source research with no incentive layer at all. But no project has combined real training results with a viable, non-token payment mechanism.

1.3 The Lightning Alternative

Bitcoin's Lightning Network provides:

• Instant settlement (~182ms average, <500ms multi-hop)
• Negligible fees (median 63 ppm, ~$0.001 for sub-$100 payments)
• No identity requirements — permissionless participation
• Programmable payments — hold invoices, PTLCs, DLCs for conditional settlement
• Proven scale — 5,606 BTC capacity ($490M), $1.17B monthly volume, 99.7% success rate
• Denomination stability — USDT on Lightning (Taproot Assets) available since Jan 2025 for those preferring fiat-denominated payments

2. Background

2.1 Decentralized Training Methods

The field has progressed rapidly from theory to practice:

• DiLoCo (DeepMind, 2023) — Foundational algorithm. Train locally for H steps, sync compressed pseudo-gradients. 500x communication reduction. 8 workers match fully synchronous training.
• Together AI (2022–2023) — Trained GPT-JT 6B across distributed GPUs using local SGD with randomly skipped communications. Reduced inter-GPU communication from 633 TB to 12.7 TB. Then pivoted to centralized cloud ($300M annualized revenue by 2025).
• Hivemind/Petals (Learning-at-Home) — Open-source PyTorch library using DHT-based peer discovery. Demonstrated BLOOM-176B inference at ~1 tok/s across consumer GPUs. No incentive layer.
• Prime Intellect (2024–2026) — The strongest results. INTELLECT-1 (10B, open-source), INTELLECT-2 (32B, first decentralized RL), INTELLECT-3 (100B+ MoE, SOTA for size). 400–2,000x communication reduction via DiLoCo + int8 quantization. TOPLOC verification for RL rollouts. No token.
• Nous Research / Psyche (2024–present) — DisTrO optimizer achieves 857–10,000x communication reduction on LLaMA 2 1.2B. 15B and 40B testnet training runs on Solana-coordinated Psyche network. NOUS token, still in permissioned testnet.
• Covenant-72B (2026) — 72B model trained by 70+ anonymous peers on Bittensor Subnet 3. SparseLoCo compression at 146x. 94.5% compute utilization. Quality roughly matches LLaMA-2-70B — 2–3 generations behind frontier, but the largest permissionless training run to date.

Key takeaway: Only Prime Intellect and Together AI have trained competitive models via decentralized infrastructure. Everything else is either inference-only (Bittensor), testnet-stage (Gensyn, Nous), or marketplace infrastructure (io.net, Akash). The communication bottleneck is solved; verification of untrusted computation and incentive alignment remain the hard problems.

2.2 SparseLoCo

Detailed description of the compression pipeline: 30 local steps → top-k sparsification (1.56% density) → 2-bit quantization → error feedback (decay=0.95). Result: 146x compression, 94.5% compute utilization, 70-second communication rounds over 500 Mbps internet.

2.3 Lightning Network & L402

Overview of Lightning payment channels, HTLCs, the L402 HTTP authentication protocol, and hold invoices. How L402 turns any HTTP endpoint into a paid API with sub-second settlement.

2.4 Limitations of Token-Based Coordination

The most comprehensive empirical analysis of a token-coordinated AI network is arXiv 2507.02951 (June 2025), which studied 6.6 million events across 121,567 wallets in all 64 Bittensor subnets from March 2023 to February 2025. Findings:

• Stake dominates rewards. Stake-to-reward correlation r=0.80-0.95. Performance-to-reward correlation r=0.10-0.30. Economic capital, not compute quality, determines payouts.
• Extreme concentration. Fewer than 2% of wallets command 51% majority in most subnets. Top 1% control median 89.8% of stake.
• Centralization persists. The Opentensor Foundation controls all Subtensor validator nodes via Proof of Authority and can censor transactions. The February 2025 dTAO upgrade replaced root-network valuation with market-driven staking but did not address the PoA chain or the stake-quality disconnect.
• Emission economics are unsustainable. 3,600 TAO/day ($774,000 at $215/TAO) is funded by inflation, not customer revenue. Average miner earns ~$9.61/day against $50–200/day GPU costs. The December 2025 halving exacerbated this.

The core issue: token-based systems select for capital, not contribution quality. For Bitcoin developers, the comparison to proof-of-work is instructive — in Bitcoin, hash rate directly maps to block production. In Bittensor, TAO stake is a governance/reputation token that loosely correlates with AI output quality.

3. Protocol Design

3.1 Architecture Overview

3.2 Gradient Exchange Protocol

1. Peer trains locally for K steps (K=30 in reference implementation)
2. Peer compresses pseudo-gradient via SparseLoCo
3. Peer uploads compressed gradient to coordinator via L402-gated HTTP PUT
   • Peer pays small submission fee (anti-spam, covers storage)
   • Coordinator issues hold invoice (payment held pending validation)
4. Coordinator validates gradient quality:
   • Forward pass on validation batch before and after applying gradient
   • Loss reduction score computed
   • Assigned-vs-unassigned data check (catches plagiarism)
   • Norm calibration
5. On validation pass: coordinator settles hold invoice — peer receives reward proportional to quality score
6. On validation fail: hold invoice expires — funds return to peer automatically
7. Coordinator aggregates validated gradients and publishes updated model checkpoint
8. Peers download new checkpoint and resume local training

3.3 Payment Mechanics

Submission Fee (Peer → Coordinator)

• Small fixed fee per gradient submission (anti-spam, covers validation compute + storage)
• Paid via standard L402 on the upload endpoint
• Suggested: 100–1,000 sats (~$0.10–$1.00)

Quality Reward (Coordinator → Peer)

• Hold invoice issued at gradient upload time
• Amount determined by coordinator's posted reward schedule
• Settlement conditional on validation oracle attestation
• Reward proportional to measured loss reduction (not flat rate)
• Formula: reward = base_rate x quality_score x normalization_factor

Channel Design for 70 Peers

• Pre-opened direct channels between coordinator and each peer
• Bidirectional flow (fees in, rewards out) naturally rebalances
• At 1,000 sats/submission + 10,000 sats average reward per round:
  • ~9M sats/week capacity per channel
  • All 70 payments settle in ~13 seconds (well within 70-second round window)

3.4 Coordinator Trust Model

The coordinator is the primary trust assumption in this protocol. Rather than claiming full trustlessness — which remains an open problem for gradient validation at scale — we constrain the coordinator's power through layered mitigations that make misbehavior detectable, unprofitable, and recoverable.

3.4.1 Attack Surface Analysis

The coordinator performs four trusted functions: gradient validation, payment settlement, checkpoint publication, and data partitioning. A malicious coordinator could attempt:

Critically, the coordinator cannot steal funds — hold invoices auto-refund on timeout. The worst-case attack is labor theft (accept gradient, refuse payment), which is immediately visible to the affected peer and destroys coordinator reputation.

3.4.2 Layered Mitigation Strategy

Layer 1 — Deterministic Replay (Accountability)

Loss evaluation is a pure function: f(model_checkpoint, gradient, validation_data) → loss_score. If the model checkpoint, validation data, and evaluation code are published, any party can independently replay the computation and verify the coordinator's scoring. A coordinator that rejects a valid gradient or accepts an invalid one is provably dishonest.

This does not prevent misbehavior, but it makes misbehavior cryptographically provable — a property that token-based systems like Bittensor lack, where subnet owner scoring is opaque.

Layer 2 — DLC-Bound Payment Settlement (Cryptographic Constraint)

Discreet Log Contracts (DLCs) bind payment settlement to an oracle-signed attestation of the loss score. The coordinator cannot settle a hold invoice without producing a valid oracle signature over the actual computed loss. This removes the coordinator's ability to lie about validation results — the payment math is enforced by Bitcoin script, not coordinator honesty.

DLCs are production-ready Bitcoin primitives (not experimental). Combined with deterministic replay, they ensure that: (a) the coordinator must publish a loss score to settle payment, and (b) anyone can verify that score is correct.

Layer 3 — Federated Multi-Validator Consensus (Decentralized Trust)

Multiple independent validators evaluate each gradient submission. Payment requires majority attestation (e.g., 3-of-5 validators agree on loss score within tolerance). Validators are:

• Selected per-round from a pool (rotation prevents collusion)
• Paid via Lightning for their validation compute
• Subject to the same deterministic replay accountability

This reduces the trust assumption from "trust the coordinator" to "trust that at least one of N validators is honest" — the same security model used by most blockchain systems.

Layer 4 — Market Competition (Economic Constraint)

The gradient exchange protocol is open. If a coordinator misbehaves, peers can migrate to a competing coordinator running the same protocol. The switching cost is low: download the latest checkpoint (public), open channels to the new coordinator, resume training. This creates economic pressure for honest behavior — a coordinator that censors or cheats loses its peer network and revenue.

3.4.3 Trust Comparison

The coordinator role in this protocol is constrained, auditable, and replaceable — not an unchecked central authority. This is a meaningfully different trust model than both centralized training and stake-weighted token systems.

3.4.4 Remaining Open Problems

Fully trustless gradient validation — where no trusted party is required at all — remains unsolved. Potential future approaches include zero-knowledge proofs for forward-pass computation (currently prohibitive at 72B scale) and trusted execution environments (TEEs) for validation. We consider the federated validator model sufficient for practical deployment while these research directions mature.

3.5 Heterogeneous Participation

The protocol targets consumer hardware for 0.5B–7B model training. Research benchmarks confirm this is practical across a wide range of devices:

The coordinator dispatches tasks appropriate to each peer's framework. CUDA peers (NVIDIA GPUs) use PyTorch; Apple Silicon peers use MLX. The protocol doesn't care what framework computes the gradients — only that the compressed pseudo-gradients pass validation.

For larger models (13B+), Heterogeneous SparseLoCo groups peers into "virtual replicas" via pipeline parallelism with activation compression between stages, lowering the hardware barrier from 8xB200 (~$200K) to potentially 1–2 GPUs. This is a Phase 3+ optimization; the prototype targets 0.5B–3B models where nearly every modern Mac and most gaming PCs can participate.

4. Decentralized Autoresearch

4.1 The Autoresearch Pattern

The autoresearch pattern (Karpathy, March 2026): an AI coding agent iteratively edits a mutable file, evaluates against a quantifiable metric, keeps improvements (git commit), discards regressions (git reset). The pattern has been validated at scale: Karpathy's original run completed ~650 experiments over two days, finding ~20 stacking improvements that reduced time-to-GPT-2 from 2.02 to 1.80 hours (~11%). Critically, all improvements were additive and transferable across model scales — discoveries at depth-12 proxy transferred to depth-24 production.

Within days, Shopify CEO Tobi Lutke applied the pattern to a production 0.8B model overnight, achieving +19% quality improvement — beating a prior 1.6B model after just 37 experiments. Lutke generalized: "Autoresearch works even better for optimizing any piece of software." The pattern has since been applied to GPU kernel optimization (autokernel, ~40 experiments/hour), HFT backtesting ("agents discover techniques I understand to be proprietary"), API latency reduction (40% improvement overnight), and prompt engineering.

Karpathy's SETI@home vision frames the full potential: "asynchronously massively collaborative agents … the goal is not to emulate a single PhD student, it's to emulate a research community." This is a coordination problem — and it maps directly to l402-train's infrastructure.

4.2 The Opportunity

Training AI models is a hard coordination problem with high hardware requirements, synchronization overhead, and a limited addressable hardware base (GPUs with 16+ GB VRAM). Autoresearch bounties are the opposite on every dimension:

The addressable market for autoresearch bounties is essentially unbounded. Every production system with measurable performance — classification accuracy, latency, conversion rates, code quality scores, prompt effectiveness — is a potential bounty target. Training produces a model; autoresearch optimizes anything.

The single-machine pattern works (Karpathy proved it). But scaling to hundreds of competing agents exploring different approaches simultaneously — Karpathy's "research community" — requires coordination infrastructure: bounty publication, agent discovery, submission management, validation, and payment. This is exactly what l402-train's coordinator + L402 + hold invoice infrastructure provides.

4.3 Bounty Protocol

1. Sponsor publishes via coordinator:
   • Target: mutable file(s) to optimize + baseline version (git commit hash)
   • Metric: quantifiable score + eval command + public eval dataset
   • Bounty: total sats available + payment schedule + deadline
   • Rules: constraints, held-out eval set hash (commit-reveal), maximum diff size
2. Agents (running on any hardware) compete:
   • Download baseline + eval framework via L402-gated endpoint
   • Run autonomous experiments locally (any coding agent: Claude Code, Codex, local models)
   • Submit improvements: code diff + claimed score on public eval set
   • Multiple submissions allowed — each is independently validated and paid
3. Validation (coordinator):
   • Apply submitted diff to baseline, run eval on held-out eval set (not the public one)
   • Check for metric gaming: canary inputs, distribution shift detection, temporal stability
   • Score: improvement magnitude on held-out set relative to target improvement
   • 80% payment on primary held-out eval, 20% holdback on delayed re-evaluation
4. Payment via Lightning:
   • Hold invoice created at submission time
   • Settled proportional to improvement magnitude: payment = bounty_pool × (score_improvement / target_improvement)
   • Bonus multiplier for exceeding target; minimum threshold to qualify
   • Holdback released after temporal stability check (24–72 hours)

4.4 Why Lightning for Bounties

The objection "you could use Stripe for bounty payments" is reasonable but misses four properties that Lightning provides and traditional payment systems don't:

1. Permissionless participation. No KYC, no payment processor approval, no bank account required. Any agent anywhere in the world can compete for bounties and receive payment. This is essential for the SETI@home model — you can't require 10,000 distributed agents to each set up a Stripe account.
2. Micropayment granularity. Many improvements are marginal — worth 500–5,000 sats ($0.35–$3.50). Stripe's minimum transaction is $0.50 with a $0.30 + 2.9% fee, making sub-dollar payments uneconomical. Lightning fees are <1 sat for these amounts. This matters because autoresearch produces many small, stacking improvements, not a few large ones.
3. Hold invoice escrow. The same trustless validation-before-payment mechanism used for training (§3.3) applies directly: payment is locked at submission time and can only be released if the improvement passes held-out validation. The coordinator cannot steal funds (auto-refund on timeout) and cannot withhold earned payment (settlement is triggered by validation pass). No traditional payment system offers this without a trusted escrow intermediary.
4. Instant settlement. When a hold invoice settles, the agent is paid in <500ms and can immediately move to the next bounty. Stripe payouts take 2–7 business days. For agents running hundreds of experiments across multiple bounties, settlement speed directly affects capital efficiency.

L402 also enables gated access to bounty materials. The eval framework, baseline code, and public dataset are served via L402-gated endpoints — agents pay a small access fee (covering coordinator bandwidth costs) that also serves as a lightweight anti-spam mechanism.

4.5 Anti-Gaming Measures

Goodhart's law is the primary risk: agents optimizing the metric rather than genuinely improving the target. Defenses:

• Held-out evaluation: Sponsor evaluates on secret data not available to agents. The held-out set hash is committed at bounty creation (commit-reveal scheme) to prevent coordinator manipulation
• Multi-metric composite: Require improvement across multiple independent metrics simultaneously (e.g., accuracy + latency + token efficiency). Gaming one metric while degrading others is caught
• Canary detection: Embed known-answer probes in the public eval set. Agents that hardcode canary answers are detected on the held-out set where canaries differ
• Temporal stability: Re-evaluate improvements after 24–72 hours to catch overfitting and fragile optimizations. The 20% holdback pays out only after this check passes
• Diff size limits: Maximum diff size prevents agents from replacing the target file entirely. Large diffs get additional scrutiny
• Semantic review: Top-N improvements reviewed by sponsor before final holdback release. Automated validation handles the common case; human review catches sophisticated gaming

Coordinator IP risk. The coordinator sees every submitted improvement and could theoretically steal ideas without paying. Mitigations: (1) commit-reveal submission — agents submit a hash of their diff before revealing it, creating a timestamped record of priority; (2) hold invoices create economic accountability — a coordinator that consistently lets invoices expire (stealing ideas without paying) loses agents to competing coordinators; (3) the 20% holdback acts as a reputation bond — sponsors who don't release holdbacks after temporal validation get flagged; (4) agents can submit to multiple competing coordinators simultaneously, reducing single-coordinator risk.

4.6 Bounty Economics

Agent costs. Running a coding agent overnight (100 experiments) costs approximately:

• API costs (if using cloud agents): $10–50/night depending on model and context size
• Electricity (if using local models): 20–90W × 8 hours = 0.16–0.72 kWh ≈ $0.03–0.12
• Expected improvements: ~3% hit rate (20 per 650 experiments). Each improvement is independently payable

Sponsor value. A company posting a classification accuracy bounty of 100,000 sats (~$70) gets overnight distributed optimization that would take a contractor days at $50–150/hour. The bounty is only paid for validated improvements — the sponsor's downside is capped at the bounty amount.

Agent profitability. An agent running local models (near-zero marginal cost) that finds one improvement per night worth 5,000–50,000 sats ($3.50–$35) is profitable from the first improvement. Agents running cloud APIs need bounties large enough to cover their API costs — natural market pricing ensures this equilibrium. The key economic insight: agents with local models (running on consumer hardware) have a structural cost advantage, creating exactly the kind of distributed participation the protocol is designed for.

Market dynamics. Unlike training (where the coordinator funds rewards from a training budget), bounties create a two-sided market: sponsors post bounties, agents compete. The coordinator takes a fee (5–10% of bounty payouts) for hosting, validation compute, and held-out dataset management. This is a sustainable business model independent of training coordination revenue.

5. Economics

5.1 Cost Comparison

5.2 Break-Even Analysis

The marginal cost of contributing compute is electricity. Most potential participants already own their hardware — they bought a Mac for work or a gaming PC for fun. The GPU sits idle 80%+ of the time. At BTC price = $70,000 (1 sat = $0.0007), US average electricity $0.16/kWh:

*Full cost includes 3-year hardware amortization at 50% utilization (12 hrs/day). Mac Mini M4 Pro has the best full-cost economics at $800 purchase price.

For context, Vast.ai hosts with RTX 4090s earn 158–243 sats/hr equivalent at current market rates. A decentralized training protocol should target 200–500 sats/hr per peer to be competitive and "worth the time" for participants. At 200 sats/hr per peer with 100 peers, coordinator cost is $14/hr ($336/day).

5.3 Bitcoin Mining Comparison

The target audience knows mining economics. Post-2024 halving, an Antminer S21 Pro (234 TH/s, 3,531W) earns ~$7.21/day against ~$8.47/day electricity at $0.10/kWh — unprofitable at US residential rates. AI training compute has three structural advantages over mining:

1. Dual-use hardware. The Mac or gaming PC has value beyond training. ASICs have zero utility beyond mining.
2. Power efficiency. A Mac Mini at 40W is 100x more efficient per device than an Antminer at 3,531W. Home circuits, cooling, and noise tolerance all have limits.
3. Revenue from work, not inflation. Payment for gradients that improve a model generates revenue from the value of the output. Bitcoin mining revenue is definitionally inflationary until fees dominate — which hasn't happened in 15 years.

Honest risk: if nobody values the models being trained, the revenue source dries up. Bitcoin mining has demand certainty (block reward exists by consensus rules). AI training payment depends on someone paying for the training. The protocol's sustainability is only as strong as demand for its compute output.

5.5 Pricing Model

Market-driven pricing:

• Coordinator posts reward schedule (sats per unit of loss reduction)
• Peers self-select based on their hardware costs and expected rewards
• Natural equilibrium: reward > cost of compute → more peers join → competition increases → quality improves
• No token issuance, no inflation schedule, no governance votes

5.4 Validation Economics

Validation is the coordinator's largest operational cost. Each peer submission requires a forward pass on the validation batch (before and after applying the gradient) to compute loss reduction. With 70 peers submitting every 70 seconds, this is substantial.

Scaling strategies:

• Sampling-based validation: Evaluate a random subset of the validation batch per round rather than the full set. Reduces compute proportionally while maintaining statistical confidence. A 10% sample with 70 peers still provides robust scoring with 95% confidence intervals.
• Amortized forward passes: The "before" forward pass is shared across all submissions in a round — only compute it once per round, not once per peer. This cuts validation cost nearly in half.
• Staggered validation: Not all 70 peers submit simultaneously. Spread validation across the 70-second window, smoothing GPU utilization.
• Validator pools: Distribute validation across multiple paid validators (Layer 3 in §3.4.2). Each validator handles a subset of submissions. Validators are paid from submission fees, creating a self-sustaining market for validation compute.

Cost estimate: At 70 peers with amortized forward passes and 10% sampling, coordinator validation requires approximately 1–2 GPUs dedicated to scoring — roughly 2–3% of total network compute, well within the 5–15% overhead budget funded by submission fees.

5.6 Channel Liquidity

The protocol requires the coordinator to maintain payment channels with each peer. At 10,000 sats average reward per round with 70-second rounds, each channel needs approximately 9M sats/week of throughput capacity.

Liquidity optimizations:

• Bidirectional flow: Submission fees flow peer→coordinator while rewards flow coordinator→peer. This naturally rebalances channels, reducing the locked capital requirement.
• Just-in-time channels: Lightning Service Providers (LSPs) can open channels on demand when new peers join, eliminating the need for the coordinator to pre-fund channels with every potential peer.
• Epoch-batched settlement: Aggregate rewards across multiple rounds and settle once per epoch (e.g., every 10 rounds / ~12 minutes). This reduces per-round channel throughput requirements by 10x while maintaining sub-hour settlement — still orders of magnitude faster than Bittensor's ~12-second consensus or centralized Net-30 contracts.
• Submarine swaps: Peers can receive on-chain payments for large accumulated rewards, freeing channel capacity for ongoing micropayments.

5.7 Coordinator Funding

Who pays the coordinator?

• Corporate R&D: Company funds training bounty, pays in sats, receives trained model
• DAO/community: Pooled funding via multisig, model released open-source
• Self-funding: Coordinator sells API access to trained model via L402, reinvests revenue into training bounties
• Autoresearch sponsors: Companies pay to optimize their prompts, configs, or models

6. Security Analysis

6.1 Threat Model

6.2 Coordinator Threat Mitigation

The coordinator is the primary attack target. Summary of defense layers:

1. Can't steal funds — hold invoices auto-refund on timeout (Lightning protocol guarantee)
2. Can't lie about scores — DLC-bound settlement requires oracle signature over computed loss (§3.4.2, Layer 2)
3. Can't censor undetected — deterministic replay proves valid gradients were rejected (§3.4.2, Layer 1)
4. Can't act unilaterally — federated validators require majority consensus (§3.4.2, Layer 3)
5. Can't hold peers captive — open protocol, low switching cost to competing coordinator (§3.4.2, Layer 4)

The residual risk is a coordinator that subtly degrades checkpoint quality over time — difficult to detect in any decentralized training system. Public eval benchmarks on each checkpoint provide the best available defense.

6.3 Comparison to Bittensor Security

Bittensor's stake-weighted consensus means wealthy actors control the network — top 1% of wallets control ~89.8% of stake, and performance-to-reward correlation is only r=0.10-0.30 vs stake-to-reward correlation of r=0.80-0.95. Validators score opaquely with no deterministic replay mechanism.

The Lightning protocol inverts this: payment is proportional to measured contribution quality, not capital staked. Validation is deterministic and publicly replayable. There is no governance token, no staking requirement, and no emission schedule disconnected from compute value. The attack surface is narrower and more auditable.

7. Limitations

1. Coordinator is semi-centralized: Despite the layered mitigations in §3.4 (deterministic replay, DLCs, federated validators), fully trustless gradient validation at scale remains an open problem. The trust requirement is reduced to "at least one of N validators is honest," but not eliminated entirely.
2. Consumer hardware limits model scale: The protocol targets 0.5B–7B models on consumer hardware (MacBook Air to RTX 4090). This is sufficient to prove the coordination mechanism and useful for domain-specific fine-tuning. But frontier pre-training (70B+) requires sustained petaFLOPS-scale compute with 100+ Gbps NVLink interconnects — 1,000x what home internet delivers. The protocol lowers the financial barrier (no staking) while accepting a model scale ceiling on consumer hardware.
3. Denomination volatility: Rewards denominated in sats fluctuate with BTC price. USDT on Lightning (Taproot Assets, production since Jan 2025) provides fiat-stable denomination for peers who prefer predictable economics. The protocol is denomination-agnostic — coordinators can post bounties in either currency.
4. Model quality gap is closing: Covenant-72B (2026) matches LLaMA-2-70B — roughly 2–3 generations behind frontier. But Prime Intellect's INTELLECT-3 (100B+ MoE, January 2026) achieves state-of-the-art for its size across math, code, science, and reasoning benchmarks via decentralized RL training. The gap is narrowing faster than expected. The protocol's value proposition is a superior coordination mechanism — it makes decentralized training economically viable and fairly incentivized, benefiting from whatever training methods the field develops. Near-term applications favor domain-specific fine-tuning where the base quality gap matters less.
5. Autoresearch bounty validation has limits: Held-out evaluation prevents naive metric gaming, but sophisticated Goodhart attacks on composite metrics remain an open research problem. The commit-reveal submission scheme mitigates coordinator IP theft but adds latency. For high-value bounties, the 20% holdback and human semantic review provide defense in depth, but this partially re-centralizes validation.
6. Operational complexity: Peers must manage a Lightning node and payment channels in addition to GPU infrastructure. Lightning agent tooling (Lightning Labs, 2026) reduces this burden, but it remains non-trivial compared to simply joining a managed training cluster. LSP integration and automated channel management are required for consumer-grade UX.
7. Validation compute: Gauntlet-style validation adds compute overhead, funded by submission fees. Scaling strategies (sampling, amortized passes, validator pools — see §5.3) keep this manageable, but validator economics require careful tuning to avoid creating a centralized compute bottleneck.

8. Related Work

Training Algorithms & Systems

• DiLoCo (Douillard et al., 2023) — Local SGD foundations, 500x communication reduction. arXiv:2311.08105
• Streaming DiLoCo (DeepMind, 2025) — Overlapping communication with computation at 1B, 10B, 100B scale
• SparseLoCo (Sarfi et al., 2025) — 146x gradient compression via top-k + 2-bit quantization
• DisTrO (Nous Research, 2024) — 857–10,000x communication reduction on LLaMA 2 1.2B
• DeMo (Lambda, 2024) — Decoupled momentum, extreme compression
• Heterogeneous SparseLoCo (2026) — Multi-tier peer participation via pipeline parallelism

Decentralized Training Projects

• Prime Intellect (2024–2026) — INTELLECT-1 (10B), INTELLECT-2 (32B, first decentralized RL), INTELLECT-3 (100B+ MoE, SOTA for size). Best results in the field. No token. arXiv:2505.07291
• Covenant-72B (Nous Research / Bittensor Subnet 3, 2026) — 72B permissionless training, largest decentralized run. arXiv:2603.08163
• Together AI (2022–2023) — GPT-JT 6B via distributed local SGD. Pivoted to centralized cloud ($300M ARR by 2025)
• Hivemind/Petals — Open-source decentralized inference + fine-tuning library. No incentive layer. GitHub
• Nous/Psyche — Solana-coordinated training network, NOUS token, permissioned testnet
• Gensyn — Verde probabilistic proof-of-learning. $51M raised (a16z). Late testnet since 2022. AI token launched before mainnet
• Bittensor (OpenTensor Foundation) — Token-based inference marketplace. Empirical analysis: arXiv:2507.02951

Compute Marketplaces

• io.net — 327,000 GPUs aggregated, $1M+ monthly revenue. Marketplace, not training coordination
• Akash Network — Reverse auction compute marketplace, 736 GPUs at 70% utilization, $4.3M annual revenue
• Vast.ai — Consumer GPU rental. RTX 4090 at $0.13–0.20/hr. Market-rate benchmark for consumer compute

Payment & Coordination Infrastructure

• L402 Protocol — HTTP 402 payment authentication via Lightning + macaroons. Production since 2020 (Lightning Loop). Spec
• Lightning Agent Tools (Lightning Labs, 2026) — AI agent payment infrastructure, MCP server, lnget auto-pay. GitHub
• Aperture (Lightning Labs) — L402 reverse proxy, reference implementation. LND-only. GitHub
• Fewsats — L402 commercial infrastructure: proxy402, l402-python SDK, MCP servers. Most active L402 ecosystem builder
• x402 (Coinbase, 2025) — Stablecoin (USDC/Base) alternative to L402. 5.6k stars, 75M txns reported. Lacks hold invoices and macaroon delegation
• FEDSTR (2024) — Federated learning marketplace on Nostr + Lightning. The closest prior art: uses Lightning payments and Nostr relays for peer discovery. Key differences: targets federated learning (data stays local, model aggregation) rather than distributed pre-training; does not address gradient compression; lacks hold-invoice conditional payment for trustless quality validation. Our protocol extends FEDSTR's core insight to the harder problem of permissionless gradient exchange with SparseLoCo compression and DLC-bound validation
• L402 inference services (2025–2026) — ~10 live services accepting sats for AI inference: LightningProx (30 sats/query), Sats4AI (5–500 sats), SatsForAI, The Ark AI (120+ services). Electricity cost per 7B inference query on Apple Silicon is ~0.05 sats (>99% gross margin). Validates the economic viability of L402 micropayments for per-query compute. See inference payments research

9. Conclusion

Bitcoin's Lightning Network is a superior coordination layer for decentralized AI training compared to custom tokens. It provides instant settlement, negligible fees, conditional payments via hold invoices, and permissionless participation — without the governance overhead, wealth concentration, and incentive misalignment of token-based systems.

The semi-centralized coordinator — the primary trust assumption — is constrained through four independent layers: deterministic validation replay (accountability), DLC-bound payment settlement (cryptographic constraint), federated multi-validator consensus (decentralized trust), and open-protocol market competition (economic constraint). This reduces the trust requirement to "at least one of N validators is honest" — a well-understood security model, and a meaningful improvement over both centralized training (full trust in employer) and token-based systems (trust in stake-weighted, opaque validators).

Combined with SparseLoCo gradient compression, Gauntlet-style validation, and practical solutions for validation scaling and channel liquidity, Lightning enables a viable protocol for permissionless large-scale model training. The extension to decentralized autoresearch — where AI agents compete for Lightning-paid bounties to optimize any quantifiable metric — may represent the protocol's largest practical application. Training coordination solves the hard technical problem; autoresearch bounties scale that same infrastructure to an essentially unbounded market. Both modes share the same hold-invoice escrow, L402 payment gating, and coordinator validation architecture — and both pay contributors in bitcoin, proportional to the quality of their work.

Research Files