The working language of new intelligence.

A software actor that can pursue goals across steps and tools.

A recurring process redesigned around AI delegation and review.

How closely a system's actions match operator intent and limits.

The mechanism that lets a model weigh which parts of context matter.

AI-assisted work where the human remains visibly accountable.

The relevant facts, constraints, examples, and history a model needs.

The amount of text, history, and source material a model can see at once.

The line between human-owned work and safely delegated AI work.

A generative model that creates media by denoising toward a target.

A numeric representation of meaning used for search and retrieval.

A repeatable way to judge if AI output is useful and safe.

Adapting a trained model with additional task-specific examples.

Confident AI output that is false, unsupported, or invented.

A design where humans review or steer important system decisions.

A transformer model trained to predict and generate language at scale.

A lightweight way to adapt a model without retraining all weights.

The durable layer that carries useful facts across sessions.

The underlying AI capability used to reason, draft, classify, or act.

A layered statistical system that learns patterns by adjusting weights.

Quavia shorthand for new intelligence as an operating layer.

The rhythm that keeps AI-enabled work useful after the demo.

An instruction package with task, inputs, constraints, and criteria.

Designing instructions, examples, and constraints for better model output.

Compressing model weights so a model can run cheaper or faster.

Retrieval first, generation second, grounded in source material.

Training models with human preference feedback.

Intermediate reasoning steps used to help or inspect model problem-solving.

A generation setting that controls output variety and randomness.

Breaking input into tokens before a model processes it.

AI systems calling software, files, commands, or structured actions.

Reusing general model capability for a more specific task.

The attention-based architecture behind most modern language models.

A database for storing embeddings and finding semantically similar items.

Ranking workflows before building automation or AI support.