Glossary

Abliteration: A process that reduces learned refusal behavior in a compatible language model. It does not guarantee truth or safety.

Agent: A goal-oriented AI workflow that can use models, skills, memory, and approved connectors.

Companion: A persistent AI identity with personality, memory, skills, media abilities, and optional onchain identity/economy.

Contribution Pack: A portable dataset package containing records, sources, permissions, hashes, quality, and attribution.

GGUF: A file format commonly used for quantized local language models.

Local AI: AI inference performed on the user's device rather than by a hosted model provider.

Mint: Create an onchain NFT or registered asset through a wallet transaction.

Model token: A fixed-supply token attached to a validated fine-tuned or abliterated model.

Proof of Contribution: BonzAI's system for recording who or what helped create a useful AI asset and routing ownership/revenue accordingly.

Provenance: The traceable origins and parent relationships of data, models, and outputs.

Quantization: Compression that lowers model memory use, with a possible quality tradeoff.

Revenue route: An onchain allocation connecting an asset's income to contributor wallets.

Staking: Locking $BONZAI to create utility weight and demonstrate durable commitment.

Utility weight: Stake adjusted by lock duration and attested contribution score.

WebGPU: Browser access to GPU computing used by supported BonzAI Web/BonzAI+ local models.