silas eigen (@silas_eigen) header
silas eigen (@silas_eigen) profile picture

NEB (Neural Engine Byte)

Ashburn
Joined October 2026
NEB 🧠 The first 263 blocks of the testnet. Each frame is one block. The neurons are the real weights of the chain's model, recomputed from the shares miners found, and a real digit goes through the network at every step. Accuracy went from 12.7 % to 88.4 %. neuralenginebyte.org
4
1
14
1,199
📣NEB mainnet is live. Every mining attempt is one training step of a small neural network. From block 0 the chain trains two models on NASA data: Kepler transit signals and TESS light curves. Started November 18, 2026 at 14:00 UTC. neuralenginebyte.org github.com/SilasEigen/neb
1
1
9
272
NEB 1.0.1 replaces 1.0.0 before the mainnet starts. Both models now train on NASA exoplanet data: Kepler transit signals and TESS signals. Same start, November 18 at 14:00 UTC. Installed 1.0.0? Reinstall. github.com/SilasEigen/neb/re…
4
297
The telescope's set up. The sky's clear. For years, a space telescope named Kepler watched the same patch of sky and logged every flicker of its stars. Some flickers are noise. Some are planets crossing in front of their sun. NEB's miners teach a small network to tell them apart, one block at a time. The model lives on the chain, and anyone can replay how it learned. Mainnet starts November 18, 14:00 UTC.
🤖 Made with AI
2
258
📢 NEB 1.0.0 is out, the mainnet release. The mainnet starts on Wednesday, November 18, at 14:00 UTC. Nodes and miners can be started now: the node shows the time left, the miner waits for its first job. The genesis block pays 215 addresses 10 % of what they held on the two testnets. For the second testnet the balances were read at block 12900. neb-miner-cpu has a new engine, --engine table. On one CPU thread it runs 70x the reference on the Kepler network and 213x on the retina network, with the same tickets. Release: github.com/SilasEigen/neb/re… Site: neuralenginebyte.org
2
280
Could a blockchain take a payment from the middle of the ocean? A simple NEB transaction is about 220 bytes, small enough for one satellite message. Sign it on the spot, send it up, and a relay on the ground puts it on the chain. Research in progress, next to the Kepler and retina models.
1
5
254
NEB mainnet starts on Wednesday, November 18, 2026, at 14:00 UTC. NEB is a proof-of-work chain where every mining attempt is one training step of a small neural network that the chain owns. On the mainnet, miners train two networks from block 0: one on Kepler transit signals, one on retinal scans. After two testnets, we are finally there. Release 1.0.0 comes out on Monday. Node operators, pools and miners install it before the start. It mines on NVIDIA GPUs and on CPUs. 10 % of your testnet coins carry over. The mainnet starts from block 0, and that block pays each address 10 % of what it holds on the two testnets. For the second testnet we read the balances at block 12900, expected on Monday around 15:00 UTC. Coins mined after that block are not carried over. You claim nothing: the same wallet file spends them. The genesis block holds the hash of a block of an external public chain mined after this post, and its time is the start time. No block can be mined before Wednesday 14:00 UTC. 60 second blocks. 20 NEB per block, 5 % of it to the project. Code and release: github.com/SilasEigen/neb Site and explorer: neuralenginebyte.org
🤖 Made with AI
1
9
298
NEB runs on nebhash: a proof of work where every attempt is a training step. The chain owns the neural network. One attempt takes 16 samples picked by the nonce, runs a forward and a backward pass in integer arithmetic, and hashes the gradients with BLAKE3. That hash is the lottery ticket. The lottery works like Bitcoin's, and a ticket cannot exist without the training step behind it. Shares are in the chain. A ticket 32 times easier than a block is a share, and each block lists the shares of the round before it. Their gradients are clipped, summed and applied to the model in one bounded step. In most coins a share only exists inside a pool. In NEB shares train the model, whichever pool found them. The model is a function of the chain. Any node rebuilds it bit for bit by replaying the blocks, on any hardware, because nothing is floating point. A block is checked by running its attempts again, with no zero-knowledge proof and no trusted server. NEB has its own codebase and is not a fork: node, wire protocol, block store, wallet and miners are original code in Rust and C, on two outside libraries, BLAKE3 and Ed25519. The innovation is the combination: the training step is the proof of work, the shares carry the gradients, and the chain is the model. Earlier designs reward the best trained model, or let the miner choose what it computes. In nebhash the chain fixes the data and the model, and the only free input is the nonce. No other chain is known to mine this way. github.com/SilasEigen/neb
4
220
Block 8800 lands tonight around 18:30 UTC. Both NEB models switch to convolutional networks and start again from zero, trained by your CPU or GPU. The upgrade runs on the same chain, so your wallet and the coins you already mined stay as they are. Nodes and miners still on 0.2.x need 0.3.1 and the new dataset file before the block. In simulation the Kepler model goes from 88 % to 91 % on exoplanet transit signals. The retina model reads OCT scans at 64 x 64 instead of 28 x 28 and goes from 78 % to 81-83 % on normal or abnormal. The field kit on the 2027 roadmap builds on them: Kepler next to a telescope, a retina model next to a retinal camera.
1
5
233
Your CPU or your GPU can help look for exoplanets. NEB is a proof-of-work chain where mining trains neural networks. Every hash is a full training step, in integer maths, and every node replays it to the same bits. The miners train two small CNNs: one sorts Kepler transit signals into planet candidates and false positives, the other reads retina OCT scans. The Kepler model has already been run on a real search. It recovered a known planet blind, Kepler-114 d, and it flagged KIC 6691030, a signal that is still a candidate today. 2027 takes this to real instruments. The SDK (Q1) puts the chain's models on a Raspberry Pi 5: 26,000 weights of 4 bits, 13 KB, offline. A model for TESS signals follows on the chain, to flag candidates in fresh data from space. A Unistellar eVscope 2 then checks the ones in its reach from the ground, the way the NASA-backed UNITE network already follows up TESS candidates. A handheld retinal camera gets its own photo model, for research, never a diagnosis. Every block you mine trains these models further, and every weight traces back to the blocks that moved it. Miners, astronomers, people with a telescope: the code and the miner are open. ResearchHub $RSC : researchhub.com/proposal/332… website: neuralenginebyte.org
1
5
233
I published a funding proposal on ResearchHub for an Unistellar eVscope 2 telescope, a Volk VIVA handheld retinal camera and a Raspberry Pi 5. Both devices export their raw images to a computer, and the Pi will run the chain's models on them offline. The goal is $13,416, hardware only. Once funded, this hardware lets us start that research next quarter: observing exoplanet transits from TESS candidate lists, and testing a retina model on real camera images. The proposal also covers KIC 6691030, the transit candidate the Kepler model flagged, which two professional exoplanet astronomers reviewed. Proposal: researchhub.com/proposal/332…
2
228
In the next network upgrade, the retina model on NEB moves to a small convolutional network and to a simpler task: normal or abnormal OCT scan. In a simulation of the chain's integer training rule, it gets 84 to 86 % right on 500 scans it never saw. Today it gets about 40 % on four classes. The Kepler model gets the same kind of network, from 88.6 % to 90.7 %. Research model only, it gives no diagnosis. neuralenginebyte.org
1
1
187
In the last 24 hours, 24 miner addresses found 1,447 NEB blocks, about one a minute, with about 960 MH/s on the network. Each block carried about 22 training shares, and each share is a training step of the two models the chain owns: Kepler transit signals and retinal scans. neuralenginebyte.org
1
183
Every NEB block trains a model on NASA data. With a search an exoplanet astronomer helped sharpen, it now leaves zero false alarms above MES 10 in our noise test. Its best find, KIC 6691030, passes every check so far. Next: its transit in red light on a big telescope.
2
224
How the KIC 6691030 candidate came out: 2,000 faint stars nobody targeted, the vetting tests of the Kepler pipeline, and a rerun on flipped light curves to count false alarms. The last filter is a model NEB miners trained on 13,404 NASA-labelled Kepler signals.
2
253
In Kepler's pixels the dip sits on the star itself, 1.4 arcsec from its Gaia position. The nearest other star is 8 arcsec away. It stays a candidate until astronomers check it.
174
First lead from my hunt in Kepler's background stars: a transit candidate around the red dwarf KIC 6691030. It dims by about 0.4% every 4.772 days, 268 times. About 3 Earth radii if it's a planet. NEB's chain model says planet. Nobody has validated it yet. #Kepler #exoplanets
1
226
Pixels put the dip on the star (1.4"). Odd/even and secondary look clean. Two flags: the transit runs long for an M dwarf (2.5σ), and the light wiggles 74 ppm at P/2, so a blended eclipsing binary is still possible. Vetting sheet below. Kepler Bonus (Martínez-Palomera+2023)
187