AI glossary

AI jargon, translated. The terms that get tossed around everywhere — in plain words, each with the way to where it gets concrete.

Basics

Artificial intelligence (AI): Umbrella term for software that tackles tasks once thought to need human intelligence — language, images, patterns. No consciousness, no intent: statistics that got very good at guessing.
Large language model (LLM): The model behind chatbots like Claude or ChatGPT. It predicts the most likely next word — millions of times in a row. It can amaze and be wrong, often in the same sentence.more on this
Token: The chunks a model breaks text into — usually word pieces, not whole words. Prices and limits count tokens, not characters.more on this
Context window: How much text a model holds in mind at once. Overflow it and older parts drop out — the start of the chat is gone.more on this
Prompt: Your input to the model — the question, the instruction, the context. How you ask strongly shapes what comes back.
Hallucination: When a model states something invented as confident fact. Not a bug but the flip side of probability-guessing — so: check.more on this
Training & training data: The texts and images a model learned its patterns from. What wasn't in there — or was skewed in there — shows up later in the answers.

Transformer: The architecture nearly all of today's language models are built on (the “T” in GPT). Its trick: weighing which words in a sentence matter to each other.
Parameters & weights: The dials inside a model, tuned during training — billions of them in large models. Roughly: more parameters, more capacity (but not automatically more sense).
Embedding & vector: Meaning as a row of numbers. Similar concepts sit close together as vectors — that's how software “computes” with language and finds related things.
Temperature: A dial for randomness: low = tame and predictable, high = more creative and erratic. The same question can roll quite different answers.more on this
Inference: Using a finished model — the moment it answers. It costs compute (and money) per request, unlike the one-off training.
Fine-tuning: Sharpening a pretrained model on a task or a tone with extra examples — instead of starting from scratch.
RAG: “Retrieval-augmented generation”: before answering, the model looks things up in a knowledge source (documents, a database) instead of talking from memory alone — against staleness and invention.
Context rot: The slow decay in long chats: answers get vaguer, earlier facts get contradicted, the thread slips. This site's namesake.more on this

Agent: A model that doesn't just answer but plans steps and uses tools itself — search, write files, call programs. With correspondingly more room for use and for harm.more on this
MCP: “Model Context Protocol”: an open standard for plugging AI tools into data sources and programs — the USB port for assistants, so to speak.more on this
System prompt: The quiet upfront instruction that sets an assistant's role and rules before you even type. It shapes the tone and limits of the whole conversation.
API: The interface other programs use to call a model — no chat window, straight from code. That's how AI ends up inside apps and websites.
Memory: Notes an assistant keeps across chats to “remember” you. Handy — and a question of control and privacy.more on this

Bias: Systematic skews in the answers, inherited from the training data. The model echoes the world's patterns — the unfair ones included.
EU AI Act: The EU's AI law: it regulates by risk, from harmless to banned. It touches small operators too — especially on transparency and high-risk use.more on this
Deepfake: Convincingly real but fabricated images, voices or video. The tech has gotten cheap — proof now needs a source, not just a face.
Alignment: The effort to get a model to do what people actually want — not just satisfy the literal wording. Harder than it sounds.
llms.txt: A small file with which a website offers AIs a tidy summary of itself — like robots.txt, but for reading aloud to machines.more on this

Want to go deeper? Hand-picked sources are in the Reading