The AI Jargon Decoder

Thirty AI terms, each in one sentence, each linked to a fuller explanation. Use it as an index when a word blocks you mid-article.

One sentence per term. Follow the link when you need the real version.

The core objects

Token — the chunk of text a model actually reads, usually a common word or word fragment. → full

Vocabulary — the fixed list of all tokens a model knows, frozen when the model was built. → full

Embedding — a token converted into a list of numbers positioned so that distance means similarity. → full

Vector — an ordered list of numbers; an embedding is one. → full

Parameter — one adjustable number inside a model; "7B" means seven billion of them. → full

LLM — a large language model: a machine trained to predict the next token. → full

Autoregressive generation — producing text one token at a time, each conditioned on all previous. → full

Inference — running a trained model to get output, as opposed to training it. → full

Temperature — a knob controlling randomness in token selection; higher is more varied. → full

Top-p / nucleus sampling — restricting choices to the smallest set of tokens whose probabilities sum past a threshold. → full

Streaming — sending tokens to you as they are generated rather than waiting for the full response. → full

Pretraining — the expensive first stage where a model learns language from enormous text. → full

Base model — a model with pretraining only; it continues text rather than following instructions. → full

Instruction tuning — additional training that turns a base model into something that answers questions. → full

Fine-tuning — further training on your own data to change a model's behavior or style. → full

Training cutoff — the date after which the model has no knowledge.

Transformer — the architecture nearly all current language models use. → full

Self-attention — the mechanism letting each token draw information from other relevant tokens. → full

Attention head — one independent attention computation; models run many in parallel. → full

Positional encoding — how word order gets injected, since attention alone is order-blind. → full

KV cache — stored intermediate values so generating token 100 does not recompute tokens 1–99. → full

Residual connection — a shortcut letting information skip past a layer, making deep networks trainable. → full

Prompt — the text you send in. → full

System prompt — instructions set once at the start, framing the whole conversation. → full

Context window — the total token budget covering prompt, history, documents, and the answer. → full

Few-shot prompting — including examples of what you want instead of describing it. → full

Chain of thought — prompting a model to work step by step before answering. → full

Hallucination — fluent, confident output that is factually wrong. → full

Structured output — constraining a response to a machine-readable format like JSON. → full

RAG — retrieval-augmented generation: fetch relevant documents, then answer using them. → full

Chunking — splitting documents into pieces small enough to retrieve usefully. → full

Vector database — storage optimized for finding nearest neighbors among embeddings. → full

Cosine similarity — measuring how much two vectors point the same direction. → full

Reranking — a second, more accurate pass that reorders initial retrieval results. → full

Agent — a model that can call tools and loop until a task is done. → full

Tool calling / function calling — a model requesting that your code run a specific function. → full

MCP — a standard protocol for exposing tools to models. → full

Eval — a repeatable test measuring whether your AI system actually works. → full

Next: In What Order Should You Learn AI? — if you want the path rather than the index.