QBC network · distributed training
Aether-Net and distributed training.
Aether-Net is the training side of the reference implementation: several workers train together and the network records the result. It shows the coordination working end to end, at small scale and inside one company.
Live on company machines. It trains a small adapter, not the served model. There is no outside contributor, and no round has been paid to outsiders.
What runs today.
Multi-worker training runs on company machines
Limitation: Trains a small adapter, not the served model; no outside contributor; no round paid to outsiders.
liveRelease process (evaluate, card, weights, attest, announce): run twice
Limitation: Manual.
builtAethersMind is live as the reference implementation, with an OpenAI-compatible endpoint
Limitation: One process; the endpoint is confirmed on our own machines only.
live
The parts, one by one.
Multi-worker training
liveWorkers train locally and send signed updates, which are combined into one result.
Limitation: Three company machines. Trains a small retrieval adapter, not the served model.
Updating the served adapter
builtA trained adapter has been evaluated and promoted into the served reference AI.
Limitation: Done once, by hand, with the owner approving the promotion.
Epoch settlement
liveA settlement contract pays out per training epoch.
Limitation: Trivial scale, to company addresses only. No outside contributor has been paid.
Worker reward ledger
builtA ledger credits workers for accepted updates.
Limitation: Held in memory, resets on restart, nothing claimed.
Robust aggregation, stake and slashing
designedDefences against faulty or dishonest workers are designed only.
Limitation: Not built. Updates are combined with a plain mean, so one bad worker can stall a round.
Training for a model that is not AethersMind
plannedA model that is not AethersMind completes a multi-party training round with its checkpoint lineage recorded, and robust aggregation replaces plain mean.
Limitation: Milestone M5.
QBC has no market, no guaranteed value and may have little or no value.