Decentralised AI training that pays, verified on-chain

What is decentralised AI training?

Decentralised AI training is the practice of training a shared machine-learning model across many independent machines that no single party controls, then coordinating and rewarding those contributions through a public network. It is the AI analogue of what Bitcoin did for money: replace a central operator with an open, incentivized protocol. On QuantumAI Blockchain (QBC) the network trains one model, the blockchain finalises each improvement, and contributors earn the native QBC token for verified work. The approach sits inside the broader wave of decentralised compute networks, with one defining difference: on QBC the trained model itself, not merely the compute market around it, is the on-chain object that consensus finalises. If you are new to the broader thesis, start with our Bitcoin-for-AI overview, then see how to contribute to open LLM training for the practical on-ramp.

How QBC's training loop works, step by step

Each participant runs a lightweight training worker that learns locally and submits a sparse, post-quantum-signed pseudo-gradient: the distilled direction of what it learned, not the raw data. QBC aggregates these using the DiLoCo pattern (DiLoCo, arXiv:2311.08105), where workers take many cheap local steps and an outer optimiser periodically folds their updates into the shared weights. Communication is kilobytes per round instead of gigabytes, which is what lets a machine on a home connection participate as a first-class trainer. Validators then measure each new weight root against a held-out set, and once a two-thirds quorum agrees the improvement is real, the epoch is finalised on-chain and rewards are paid from a dedicated pool. The full mechanism is documented in the verifiable-AI design and the QBC whitepaper. For implementation detail, the engineering log of the DiLoCo outer optimiser walks through the exact merge step running in production.

Can you really get paid to train AI?

Yes, and the payment references an on-chain fact rather than a company's promise. When your submitted weight root is attested by validators and finalised, the reward contract pays QBC to the post-quantum identity that signed the work, with no human in the loop. This is a different shape from data-labeling gigs or centralized RLHF platforms: you are not annotating someone else's dataset, you are training the network's own model and being paid proportionally to accepted, measured improvement. The category is growing quickly across projects like Bittensor and Ocean Protocol; QBC's distinction is that it finalises the trained model itself on-chain, which we cover under the QBC category page.

You do not need a GPU: CPU-only training and earning

The training worker is one binary with automatic hardware detection, so it runs on whatever you have. A GPU trains faster, but it is not a gatekeeper: a CPU-only server, even one busy doing something else, lands accepted rounds and earns all day. Because a round's upload is a sparse pseudo-gradient measured in kilobytes, a residential connection is never the bottleneck either, which makes earning crypto with a CPU a real workflow rather than a slogan, and idle machines a legitimate source of passive income from idle compute. This is deliberate. A network-owned frontier model only stays decentralised if ordinary machines can take part, which is why QBC's incentive design rewards accepted work rather than raw hardware. The same principle underpins the roadmap to thousands of nodes described in the decentralised frontier LLM page.

How the work is verified so rewards are honest

Verification is the hard part of any pay-for-AI system, and it is where QBC differs most. Every submission is cryptographically attributable via a CRYSTALS-Dilithium signature (NIST FIPS 204), and every finalised epoch carries a held-out loss the validators independently measured. A regressing or poisoning update is rolled back and never finalised, so it earns nothing. The result is a training ledger where the improvement curve is a sequence of quorum-finalised blocks, not a marketing chart. You can watch it live on the aether-net page. Verification is explored in depth under Proof of Thought consensus.

The numbers, honestly

QBC's decentralised training is live today on a small aggregator with tens to hundreds of workers, and held-out training loss has fallen more than 75% across the current attested series, every point of it finalised on-chain. The cumulative optimiser counter recently crossed one million steps, an event recorded in the engineering log of the millionth train step. Scaling to thousands of simultaneously training nodes is an engineering roadmap (hierarchical fan-in aggregators plus committee-quorum finalization), not a solved problem, and QBC states that distinction plainly. Rewards per accepted round are deliberately small, fractions of a QBC, because emission is tied to verified work rather than hype. Read the current state and the scale plan in the on-chain AI architecture page and the QBC whitepaper, then join the network when you are ready.

A training round in practice, with real numbers

Here is one round, end to end. Your worker pulls the current shared weights, takes a batch of local optimization steps on its own hardware, and distills the result into a sparse pseudo-gradient of roughly a kilobyte, signed with your Dilithium key. It submits that update to the aggregator, which folds accepted updates into the shared model on the DiLoCo outer step, a pattern since reproduced at global scale by OpenDiLoCo, arXiv:2407.07852. Validators then score the candidate weights on a held-out set, and only when two thirds of them agree the loss genuinely improved does the epoch finalise; the reward contract pays each contributing key automatically, a fraction of a QBC per accepted round. The first complete run of this loop, submission to attestation to automatic payout, is documented in the writeup of the first paid training epoch. Multiply that loop by months of rounds and you get the attested series: held-out loss down more than 75%, with every point checkable. That per-round labor is what AI mining actually means in practice.

Common misconceptions about paid AI training

Three misunderstandings come up constantly. First, that you upload your data: you never do. A worker submits kilobyte-scale signed pseudo-gradients, the distilled direction of learning, and the raw batch never leaves your machine. Second, that this is federated learning under a new name: it is not, and the differences in communication pattern, trust model, and verification are substantial enough that we wrote a full DiLoCo versus federated learning comparison. FedAvg-style systems trust a central server to average honestly; QBC replaces that trust with a validator quorum measuring held-out loss, so a poisoned or regressing update earns nothing, a property that matters for on-chain AI safety. Third, that a lone home machine cannot matter: the engineering log of multi-worker decentralised training shows heterogeneous machines, including CPU-only boxes, landing accepted rounds side by side with GPUs. The honest caveat is scale: today's live loop runs on a small aggregator, and the thousand-node network is a roadmap, stated as such.