Glossary · vaathi.dev

Model: Software trained on data that maps inputs to outputs (text, images, scores). You usually call a model through a product or API rather than training it yourself.
Gen AI: Generative AI — systems that create new content (text, images, audio, code) from a prompt or input, rather than only classifying or scoring existing items.
Label: A human- (or tool-) assigned answer attached to an example — a category, score, corrected text, or preference — used to train, evaluate, or fine-tune models.
Supervised learning: Training a model on examples that already have labels, so it learns to predict similar labels on new inputs.
Token: A chunk of text the model reads or writes — often a word piece, not always a whole word. Longer prompts and answers use more tokens.
Embedding: A list of numbers (a vector) that represents text or other media so similar meanings tend to sit nearby. Used for search, retrieval, and “more like this.”
Transformer: A model architecture that uses attention to weigh which parts of the input matter when predicting the next output. Common backbone for modern language models. Often stacked in many layers; chat products are frequently decoder-only.
Attention: A learned “spotlight” that weighs which tokens (or positions) matter when the model updates a representation or predicts the next token. Central idea behind transformers — compute-heavy at long context.
CPU: Central processing unit — strong at a few complex, branching tasks (running apps, orchestration, preprocessing). Not specialized for huge parallel matrix math the way GPUs/accelerators are.
GPU: Graphics processing unit — strong at many similar simple operations in parallel. Training and large-model inference lean on this shape of work (matrix multiplies). Other accelerators play a similar role.
Hallucination: When a model produces fluent text that is wrong or invented. Treat confident wording as a style, not a guarantee of truth.
Prompt: The instructions and context you give a model. Clear goals, constraints, and format usually beat “magic” one-liners.
Prompt engineering: Deliberately structuring and iterating prompts — goals, context, examples, constraints — so outputs are more useful. Habit and clarity, not secret spells.
BLEU: A classic metric for machine translation that scores overlap between a model’s output and human reference translations. (Sometimes misspelled “BLUE.”) Useful in labs; not a full measure of quality alone.
ROUGE: A family of metrics often used for summarization that score overlap with a human reference summary. Helpful for benchmarks; still incomplete without human judgment.
METEOR: A reference-based text metric that is often a bit more flexible than classic BLEU when matching related wording. Still needs human references and still incomplete alone.
BERTScore: An embedding-based metric that compares meaning similarity between a model output and a reference, rather than counting exact word overlap only.
Perplexity: A measure of how “surprised” a language model is by a piece of text. Useful for model fit; not a direct score of helpful chat answers.
F1: A single number balancing precision and recall for labeled (classification) tasks. Not the right default tool for open-ended free text.
Claude: Anthropic’s family of language models and related products (chat, API, Claude Code). Distinct from cloud platforms like AWS Bedrock, which may host Claude among other models.
Claude Code: Anthropic’s agentic coding tool (CLI and other surfaces) that can read and edit a project. Different from Claude chat. See official docs at code.claude.com for install.
Frontier model: A large general-purpose foundation model that products and agents call via a vendor or cloud API. Catalogs change; the access pattern matters more than any single brand.
Amazon Bedrock: An AWS managed platform for calling foundation models via AWS APIs, permissions, and billing. Not “Claude’s home” — Claude is one possible model choice among others.
API: A programmed interface an application calls to request work (for example, “generate text”) and receive a response — as opposed to a human-facing chat screen.
Agent: A system that uses a model plus tools (search, code, apps) to take multi-step actions. More power, more ways to go wrong — start simple.
Agentic workflow: A goal-driven loop: plan, act with tools, observe results, then loop or finish — unlike a single chat reply. Supervise carefully as a beginner.