NEB roadmap: from testnet 2 to mainnetOn NEB, one mining attempt is one training step of a neural network that belongs to the chain. Every node replays that training bit for bit from the genesis, so every weight of the model can be traced to the blocks that moved it. The roadmap below takes those ideas as far as they go. The dates are targets: they move when the measurements say so.
Q4 2026: science on the chain, then mainnetTestnet 2 started on November 7. Mining now trains two models on public science data: transit signals from NASA's Kepler mission, sorted into planet candidates and false positives, and retinal OCT scans, sorted into three diseases and normal.
Mainnet launches this quarter, at a time I announce in advance. On launch day, 10 % of the testnet 2 balances go into the mainnet genesis, so testers keep part of what they mined. Keep your wallet file: the coins go to the same lock.
An open source reference pool follows, ready for existing pools to integrate. On NEB a share is a training step, and the pool checks each one.
A replay kit (the independent Python oracle) rebuilds the weights from the blocks and lands on the same hash as the node, and model cards show, for each weight, the blocks that moved it. The Apple GPU miner, a community contribution, gets merged once it passes review, and NEB gets listed on the tracking sites.
Q1 2027: the chain's model in your projectsAn SDK puts the chain's model in your own app or device. The model is 50 KB of integer weights, so it runs offline on a phone or on a microcontroller like the ESP32. Think of a mobile app that reads an OCT scan exported from a device, or a small box that sorts light curves.
You can also try the model in your browser: drop a retina scan or a light curve and the model answers, and any node can check the result.
The models go into the hands of researchers: astronomers and doctors evaluate them on their own data, and the results get published.
Desktop wallet (Windows, Linux, macOS) and mobile wallet.
Q2 2027: models that grow with the labsA first partnership with a lab: a dataset chosen with them becomes a model of the chain.
A network upgrade at an announced height, on the same chain, lets a model grow in place: it gets wider while computing the same function as the block before, then keeps learning. No reset, no lost blocks. The same upgrade brings integer convolutional networks and more data: TESS after Kepler, and retina photos, a first step toward analysis by camera.
The weights hash goes into every block header, so a light client can check a model without replaying the chain. An external security review closes the quarter.
Also planned: per-pool stats on the explorer.
Site:
https://neuralenginebyte.orgHow it works:
https://medium.com/@neuralenginebyte/every-hash-is-a-training-step-the-engineering-behind-neb-6f0a97714ff8Code:
https://github.com/SilasEigen/neb
