Skip to content

Releases: Deep-Commit/poai

v0.3.0-alpha: feat: implement real LLM mining, wallet system, and transaction layer

18 Jul 06:48

Choose a tag to compare

BREAKING CHANGES:

  • Build now requires llama.cpp for real LLM inference
  • New build command: go build -tags llama -o poaid cmd/poaid/*.go

FEATURES:

  • Real LLM integration with llama.cpp CLI

    • Deterministic LLM inference for mining
    • Nonce-based probabilistic search like Bitcoin
    • Realistic mining times (1-3 seconds per block)
    • GPU acceleration support (Metal on macOS)
  • Complete wallet system

    • Balance checking with --balance command
    • Works with running mining nodes
    • Support for multiple data directories
    • Secure private key management
  • Transaction system

    • Transaction creation and signing with --send command
    • Cryptographic signature verification
    • Double-spend protection
    • Transaction hash generation
    • Balance verification
  • Enhanced CLI commands

    • ./poaid help - comprehensive command reference
    • ./poaid generate-key - key generation with file saving
    • ./poaid balance --addr= --data-dir=
    • ./poaid send --to= --amount= --privkey=
  • Improved mining system

    • Bitcoin-style difficulty retargeting with compact bits
    • Nonce-based search for probabilistic mining
    • Real LLM responses instead of stub responses
    • Better error handling and logging
  • Build system improvements

    • Two build modes: real LLM (-tags llama) and stub LLM
    • Automatic file inclusion with cmd/poaid/*.go pattern
    • GPU acceleration support for macOS
  • Documentation updates

    • Comprehensive README with new features
    • Build instructions for both modes
    • CLI command reference
    • Transaction system documentation
    • Troubleshooting guide

TECHNICAL DETAILS:

  • LLM inference uses llama-cli subprocess with --no-conversation flag
  • Database locking prevents corruption during balance checks
  • Cryptographic signatures use ECDSA secp256k1
  • Transaction system ready for mempool integration
  • Difficulty targets use negative values (more negative = harder)

This completes the core PoAI functionality with real AI-based mining,
secure wallet operations, and transaction capabilities.

PoAI v0.2.0-alpha — LLM-Powered Block Solving

13 Jul 03:11

Choose a tag to compare

PoAI Release: LLM-Powered Inference Integration (Alpha)

This release marks a major milestone for PoAI: the integration of real LLM-powered inference into the block mining and validation process. The chain now uses a deterministic, on-chain LLM (TinyLlama-1.1B-Chat-GGUF via go-llama.cpp) to solve (mine) blocks, replacing the previous toy model. All nodes perform identical forward passes for each block, ensuring verifiable and reproducible consensus.

Highlights

  • LLM Inference for Mining: Block solving now uses a real language model, with deterministic outputs for consensus.
  • Model Support: Out-of-the-box support for TinyLlama-1.1B-Chat-GGUF and other llama.cpp-compatible GGUF models.
  • Configurable Inference: New CLI flags for model path and GPU layer offload (--model-path, --gpu-layers).
  • Updated Documentation: README and .gitignore updated with clear instructions for LLM setup, model download, and build process.
  • Test Corpus & Fast Mining: Supports test datasets and adjustable difficulty for rapid local testing.

Note

This is an alpha release for experimentation and feedback. Expect rapid changes and breaking updates. Do not use on mainnet or with real value.

v0.1.0-alpha

09 Jul 05:53

Choose a tag to compare

v0.1.0-alpha Pre-release
Pre-release

✨ Status: PoAI v0.1.0-alpha — “Stub-Miner” Release

This tag is an MVP that proves the chain mechanics without a real AI
model yet. Expect rapid churn and breaking changes.

What works

  • ✔️ Genesis + block format (header.Header, JSON-serialised, Keccak-256 hash)
  • ✔️ Local mining loop
    • “loss” is a deterministic SHA-256 toy function over the test corpus
    • Proof-of-Work style target; difficulty retarget scaffolding in place
  • ✔️ Two-node P2P sync (libp2p pubsub, manual multiaddr connect, mDNS discovery)
  • ✔️ Orphan-pool scanner + periodic watchdog
  • ✔️ Pluggable dataset (--test-corpus flag) with 100-Q&A demo set
  • ✔️ BadgerDB persistence (+ automatic re-index on boot)

What doesn’t work yet

  • Real forward pass — the toy SHA-256 is a placeholder for an actual ML model
  • Economic layer (txs, rewards, slashing)
  • State trie / smart-contract logicStateRoot is a stub
  • Security hardening (DoS protection, peer auth, fork-choice tweaks)
  • Distributed inference — GPU workers & on-chain weight updates planned

Why release now?

Early feedback > silent perfection.
If you can spin up two nodes, mine blocks, and watch them stay in sync, please open an issue with logs, questions, or design critiques.

⚠️ Not production-ready. Keys, balances, and blocks will be
incompatible with future versions. Use on test machines only.

Full Changelog: https://github.com/Deep-Commit/poai/commits/v0.1.0-alpha