Frequently Asked Questions

General

What is NeuroShard?

NeuroShard is a decentralized protocol for training Large Language Models (LLMs). Instead of a company training a model in a data center, NeuroShard coordinates thousands of individual nodes to collectively train a shared model called NeuroLLM.

How is this different from other decentralized AI projects?

Feature	NeuroShard	Bittensor	Gensyn
Model ownership	Community trains from scratch	Pre-trained models	User-provided
Architecture	Dynamic, grows with network	Fixed	Fixed
Verification	Proof of Neural Work	Output validation	Compute verification
Training	Full distributed training	Inference/fine-tune	Verification

Is NeuroLLM any good?

NeuroLLM starts from random weights and improves over time as the network trains it. Early on, output quality will be low. As the network grows and more training happens, quality improves. This is intentional — we refuse to bootstrap from a corporate model.

Why train from scratch instead of fine-tuning an existing model?

1. True Ownership: Models trained by corporations have unknown biases, filters, and training data
2. Decentralization: Starting from scratch means no single entity ever controlled the model
3. Transparency: Every gradient, every training round is recorded on the Intelligence Ledger
4. Philosophy: We're building collective intelligence, not redistributing corporate AI

NEURO Token

What is NEURO?

NEURO is the utility token of NeuroShard. It's earned by contributing compute (training) and spent to use inference. NEURO is not a cryptocurrency investment — it's a utility token for network participation.

How do I earn NEURO?

1. Training: 0.0005 NEURO per batch (~43 NEURO/day active)
2. Uptime: 0.0001 NEURO per minute (~0.14 NEURO/day idle)
3. Data: 0.00001 NEURO per sample served
4. Inference: Market-based (supply/demand)

Was there a pre-mine or ICO?

No. The Genesis Block starts with total_minted = 0.0. All NEURO must be earned through verified Proof of Neural Work. Even the project creators must run nodes to earn NEURO.

What's the total supply?

There's no hard cap. NEURO is minted through work. However:

• 5% of all spending is burned (deflationary)
• Rewards have diminishing returns
• Circulating supply = minted - burned

Running a Node

What hardware do I need?

Minimum:

• 2GB RAM
• Internet connection
• Any CPU

Recommended:

• 8GB+ RAM
• NVIDIA GPU with CUDA or Apple Silicon
• Stable internet

Can I run multiple nodes?

Yes! You can run multiple nodes with the same wallet token:

• Each node gets a unique network identity
• All earnings go to the same wallet
• More nodes = more earnings

Do I need to stake to run a node?

No for basic participation. You can run a Driver or Worker without staking.

Yes (100 NEURO) to become a Validator who verifies proofs from other nodes.

How much can I earn?

Node Type	Memory	Daily Earnings (Active)
Raspberry Pi	2GB	~10-20 NEURO
Laptop	8GB	~40-60 NEURO
Gaming PC	16GB	~80-120 NEURO
Server	64GB+	~200-400 NEURO

Earnings depend on network activity and model quality.

Why is my node holding all layers?

In early network stages (few nodes), each node holds many layers. This is temporary. As more nodes join, layers are redistributed across the network.

Can I run inference without training?

Yes:

neuroshard --token YOUR_TOKEN --no-training

Technical

What model architecture is NeuroLLM?

NeuroLLM is a modern transformer with:

• RMSNorm (stable distributed training)
• Rotary Position Embeddings (RoPE)
• Grouped Query Attention (GQA)
• SwiGLU activation

Architecture scales dynamically with network capacity.

How does DiLoCo work?

DiLoCo (Distributed Low-Communication) is our training protocol:

1. Train locally for 500 steps
2. Compute pseudo-gradient (what we learned)
3. Sync with peers (rarely!)
4. Apply outer optimizer with Nesterov momentum

This reduces communication by 500×.

How are gradients aggregated?

We use Byzantine-tolerant aggregation:

• Trimmed Mean: Remove outliers, then average
• Krum: Select gradient closest to majority
• Coordinate Median: Median of each parameter

This protects against malicious nodes.

What is Proof of Neural Work?

PoNW is our consensus mechanism. It validates that nodes are doing real neural network computation (training/inference). Unlike Bitcoin's PoW which wastes energy on hashing, PoNW produces useful AI training.

Each proof includes:

• Work metrics (tokens, batches)
• ECDSA signature
• Timestamp and nonce
• Role and layer information

How is the network secured?

1. ECDSA Signatures: Proofs are cryptographically signed
2. Robust Aggregation: Malicious gradients are rejected
3. Validator Stakes: Validators stake 100 NEURO (slashed for bad behavior)
4. Rate Limiting: Prevents reward inflation
5. Slashing: Fraud results in token loss

Troubleshooting

My node says "No GPU detected"

Install CUDA-enabled PyTorch:

pip install torch --index-url https://download.pytorch.org/whl/cu121

I'm not earning any NEURO

1. Ensure training is enabled (no --no-training flag)
2. Wait for Genesis data to load (30-60 seconds)
3. Check your balance at the Ledger Explorer

Training loss isn't decreasing

Normal in early network stages. The model improves as:

• More nodes join
• More data is processed
• More training rounds complete

Can I see my earnings history?

Yes, visit the Ledger Explorer and search for your wallet address.

Security

Is my wallet token safe?

Your wallet token should be treated like a private key:

• Never share it
• Don't post it in logs or screenshots
• Store it securely

Can someone steal my NEURO?

Only if they have your wallet token. Keep it secret.

What happens if I lose my token?

Your NEURO is unrecoverable without your token. Back it up securely.

Can nodes poison the model?

Our defenses make this difficult:

1. Robust aggregation rejects outlier gradients
2. Validators verify proof plausibility
3. Slashing punishes detected fraud

Future

Will there be a token launch/ICO?

No. NEURO is only earned through work. There's no sale, no pre-mine, no investor allocation.

Will the model be open-sourced?

The model is inherently open — it exists across the network. Checkpoints can be assembled from shards and are community property.

How will governance work?

NeuroDAO (coming) will enable token holders to vote on:

• Reward rates
• Slashing conditions
• Protocol parameters
• Architecture changes (tokenizer, attention mechanism)

What NeuroDAO will not control:

• Model size (automatic based on network capacity)
• Who can participate (open to all)

See Also

• Introduction
• How It Works
• NEURO Token
• Troubleshooting