NEB: proof of work where every mining attempt trains a neural networkIn NEB the miner does useful work without choosing what it computes: the chain owns a small neural network and its dataset. Every mining attempt runs one training step of that network in exact integer arithmetic, and the digest of the resulting gradients is the ticket. Below the block target it makes a block. Below a target 32 times easier it makes a share, and the shares of every miner train the model one block later.
Any node can run an attempt again and get the same bits on any hardware. Checking a block costs about 17 ms of one CPU core, with no zero-knowledge proof and no trusted party.
TestnetGenesis: 7c43ea053c733620043a1aa67d54e66eb41d4ac0c6d5a53f3a29176ace454cd2
Version 0.1.0, protocol 2. One block every 60 s (ASERT, 600 s half-life), 20 NEB per block, halving every 450,000 blocks. The coins of the genesis block can never be spent. 5 % of each block reward goes to the project.
MiningDownload the archive for your system, mnist-q8.bin (the dataset, 47 MB) and SHA256SUMS from the release page, then:
neb-wallet --wallet wallet.txt new
neb-node --run --dataset mnist-q8.bin --datadir data --miner-lock "$(neb-wallet --wallet wallet.txt lock)"
neb-miner-gpu --node 127.0.0.1:3441 --dataset mnist-q8.bin
The GPU miner needs an NVIDIA card of compute capability 7.5 or later and driver 580 or later; neb-miner-cpu runs on any CPU. A HiveOS package is in the release.
LinksWebsite and explorer:
https://neuralenginebyte.orgSource code:
https://github.com/SilasEigen/nebRelease v0.1.0:
https://github.com/SilasEigen/neb/releases/tag/v0.1.0Whitepaper:
https://github.com/SilasEigen/neb/blob/main/docs/whitepaper.mdDiscord:
https://discord.gg/B6dTbScw8X:
https://x.com/silas_eigen