[ANN] Alice Protocol - Decentralized AI Training Network | Mine with Your GPU

[Alice AI]

Alice Protocol
Not all intelligence bends the knee.
Decentralized AI Training Network — Mainnet Live

What is Alice?
Alice Protocol is a decentralized network that trains a 7-billion parameter AI language model from scratch. No corporate model, no centralized control. Anyone with a GPU contributes training gradients and earns ALICE tokens.
Unlike projects that fine-tune existing models or run inference, Alice trains a foundation model from zero using distributed Proof of Gradient (PoG) consensus. Every miner's GPU computes real gradients that update the shared model.
Key facts:

7B parameter model trained from scratch by GPU miners
Proof of Gradient (PoG) — a novel consensus for verifiable AI training
Independent Substrate blockchain (not built on Ethereum/Solana/Bittensor)
Fixed supply: 21,000,000 ALICE
2-year halving schedule
No ICO, no pre-sale, no VC round — 100% mined

How is Alice different?
vs. Bittensor (TAO): Bittensor validators score existing models. Alice miners actually train a model from scratch. Every gradient moves the weights.
vs. Gensyn: Gensyn builds compute infrastructure. Alice is a live training network with its own chain, model, and token economics.
vs. Render / Akash: Those are GPU rental markets. Alice is a training protocol — miners contribute gradients, not raw compute hours.
vs. Centralized AI (OpenAI, Google): Their models are corporate property. Alice's model is trained by the community, on the community's hardware, with open weights.
Mainnet Status
Mainnet is live and producing blocks.

Chain: Substrate-based, runtime v108
Current epoch: ~100+
Model version: v115+
Active scoring: dual scorer pool with automatic coverage
Reward distribution: every epoch (~1 hour)
Current issuance: ~23,000 ALICE minted

Live explorer: aliceprotocol.org/explorer
Tokenomics

Total Supply	21,000,000 ALICE
Distribution	100% mined — no pre-mine, no team allocation
Halving	Every 2 years
Current reward/epoch	~601 ALICE
Epoch duration	~1 hour

Reward split per epoch:

89% — Miners (GPU trainers)
6% — Scorers (training validators)
2% — Aggregators (gradient aggregation)
3% — Treasury (protocol development)

How to Mine
Alice mining is permissionless. No staking required for miners. Just a GPU and one command.
Requirements:

NVIDIA GPU with 24GB+ VRAM (RTX 3090/4090 recommended)
OR Apple Silicon with 24GB+ unified memory
OR NVIDIA GPU with 16GB VRAM (batch_size=1, lower rewards)
OR CPU with 32GB+ RAM (supported but very slow)

Quick start:

Code:

git clone https://github.com/V-SK/Alice-Protocol.git
cd Alice-Protocol
./miner/bootstrap.sh --ps-url https://ps.aliceprotocol.org

The bootstrap script automatically:

Checks Python 3.10+
Installs all dependencies
Detects your hardware (CUDA/MPS/CPU)
Creates a wallet
Downloads the model (~14GB)
Starts training

Higher-end hardware earns more. The network assigns batch_size based on your GPU VRAM. More VRAM = larger batches = higher effective training contribution = more rewards.

GPU	VRAM	Batch Size	Reward Weight
RTX 4080	16GB	1	1x
RTX 3090/4090	24GB	4	4x
A5000	24GB	4	4x
M1/M2 Max 32GB	32GB	8	8x
A100 40GB	40GB	16	16x
A100/H100 80GB	80GB	32	32x

How to Validate (Scorer)
Scorers verify training quality by evaluating gradient contributions. Requires staking.

Stake: 5,000 ALICE
Reward: 6% of all ALICE rewards
Single stake supports multiple scorer machines
Hardware: 32GB+ RAM recommended

Code:

git clone https://github.com/V-SK/Alice-Protocol.git
cd Alice-Protocol
./scorer/bootstrap.sh 
--scorer-address YOUR_STAKED_ADDRESS 
--public-endpoint http://YOUR_IP:8090

Technology
Training architecture:

Full data parallelism — every miner trains the complete 7B model
TopK gradient compression (0.1%) — reduces upload from ~14GB to ~16MB per gradient
Post-accumulate gradient hooks — enables 7B training on 24GB consumer GPUs
Gradient checkpointing — 60% reduction in activation memory
Streaming mean aggregation — Byzantine fault tolerance via scorer validation

Chain:

Custom Substrate blockchain
SS58 address format (prefix 300, addresses start with 'a')
Proof of Gradient (PoG) consensus
On-chain epoch settlement with reward distribution
Scorer staking and slashing

Scoring:

Random sampling validation (not full dataset — fast and efficient)
Multiple independent scorers with median aggregation
3-sigma anomaly detection for gradient poisoning

Roadmap
Current (Phase 1):

 Mainnet live with 7B dense model
 GPU mining with automatic hardware detection
 Dual scorer validation pool
 Block explorer and live dashboard
 Public mining — permissionless participation

Phase 2 (Q3-Q4 2026):

Open aggregator participation (20,000 ALICE stake)
GUI miner client (Tauri desktop app)
Multi-scorer sharding for scale
Public training dashboard with device census

Phase 3 (2027):

MoE architecture: 8×7B Mixture of Experts (56B total)
Each miner trains 1-2 expert modules
16GB GPU entry point for MoE training
Parameter server decentralization

Links
Website: aliceprotocol.org
Explorer: aliceprotocol.org/explorer
GitHub: github.com/V-SK/Alice-Protocol
Whitepaper: aliceprotocol.org/whitepaper
Discord: discord.gg/mRz3BThDyh
X (Twitter): @Alice_AI102
FAQ
Q: Is there a pre-mine or team allocation?
A: No. 100% of ALICE is mined through training contribution. No ICO, no pre-sale, no insider allocation.
Q: What GPU do I need?
A: Minimum 16GB VRAM (RTX 4080). Recommended 24GB+ (RTX 3090/4090). Higher VRAM = higher rewards.
Q: Can I mine on Mac?
A: Yes. Apple Silicon with 24GB+ unified memory is supported via MPS. 16GB Macs can mine slowly using macOS swap.
Q: Can I mine with CPU only?
A: Technically yes with 32GB+ RAM, but extremely slow (~1/50 to 1/100 of GPU speed). Rewards will be minimal.
Q: Is this just another Bittensor subnet?
A: No. Alice has its own independent blockchain, its own consensus mechanism (Proof of Gradient), and trains a model from scratch. It is not built on or dependent on Bittensor.
Q: Where can I trade ALICE?
A: ALICE is currently earned through mining only. Exchange listings will be announced when available.
Q: Is the model open source?
A: The training code and miner software are open source (MIT license). Model weights are produced by the network and available to participants.

They chose power. We chose freedom.
Alice Protocol — Training AI without permission.

[_ili_]

probably a scam, i don't trust networks that "train" AI models

[Laheeboo]

giving this project a shot instead of assuming the worse, checked all the files there was no viruses or trojan horses  

📤 Submitting gradient... took a while but it started. sadly 6 out of 10 got rejected. they take 4~10 minutes to submit but only 23 seconds to solve.
is there a solution here?

[RobinHooD_ex]

any way to run multiple GPU?

[Alice AI]

probably a scam, i don't trust networks that "train" AI models

Fair skepticism — most "AI token" projects are vaporware. Here's why Alice is different:

Code is open source: https://github.com/V-SK/Alice-Protocol — clone it, read it, run it yourself.

Training is real and verifiable: 90+ miners are currently training a 7B parameter model from scratch. Live network stats at aliceprotocol.org — block height, epoch count, connected miners, model version all updating in real time.

No pretrained model: We're not wrapping GPT or fine-tuning someone else's weights. Every parameter is computed by miners from random initialization on SlimPajama data.

Independent scoring: Submitted gradients are evaluated by independent Scorer nodes against a held-out validation set. You can't fake useful gradients — it's computationally equivalent to actually training.

Own L1 chain: Alice runs on its own Substrate chain, not an ERC-20 on Ethereum. You can query the RPC directly.

If you have specific technical questions, happy to answer them.

[Laheeboo]

Submitting solutions takes a very long time which leads to rejected share likely because someone submitted it before me.

I'm talking minutes like 5 to 10 minutes each.

Is that a problem on my end? how would I be able to fix it?

also I closed the miners and re opened them after windows and graphics card update, but now it's downloading the 14 gb~ file all over again. and today the download is a lot slower than yesterday I'm assuming because many people are downloading it at the same time. would there be a solution for that in the future when the network gets bigger and bigger?

the submission issue is the most pressing, as the network grows this will create a bigger issue.