● chatpt glossary · 30 terms
chatpt glossary: the words that get in the way
Every definition is one or two sentences plus the reason a normal person should care. No circular glossaries, no “see also: paradigm”.
Agent
A mode where the model can take several steps and use tools — browse, run code, edit files — instead of only writing a reply.
Why it matters Agents need supervision: a multi-step run can go wrong in step three and still look confident at the end.
AI assistant
A product built on top of a model: the chat window, memory, file handling, and settings that turn a raw model into something usable.
Why it matters When people compare "ChatGPT vs Claude", they are comparing assistants, not just models.
API
An interface that lets software call a model directly, usually billed per unit of text processed.
Why it matters A chat subscription and an API key are separate products with separate billing.
Benchmark
A standard test used to compare models on a task, such as answering exam questions or writing code.
Why it matters Benchmarks measure the test, not your job. Distrust a score presented without the task.
Chain of thought
A model working through intermediate steps before giving an answer, rather than producing the answer in one jump.
Why it matters You rarely see the full reasoning, and a tidy explanation is not proof that the reasoning was sound.
Context window
The amount of text a model can consider at once — your prompt, the conversation so far, and any files you attach.
Why it matters It explains why a long thread starts "forgetting" the beginning: the oldest turns fall outside the window.
Custom instructions
Standing preferences you set once so they apply to every chat, such as your role, audience, and formatting rules.
Why it matters They remove the need to re-paste your context in every conversation — and they apply even when they are wrong for the task.
Deep research
A multi-step mode that reads across multiple sources and produces a longer, cited report.
Why it matters It is a research assistant, not a fact-checker: open the citations before you trust the synthesis.
Fine-tuning
Further training a model on a specific dataset so it behaves differently for a narrow task.
Why it matters Most "custom AI" products are prompting and retrieval, not fine-tuning.
GPT
The family name for OpenAI's generative pre-trained transformer models. "GPT" is a brand and an architecture label, not a synonym for all AI.
Why it matters Other companies make large language models too; "GPT" is not the generic word for them.
Hallucination
A confident, fluent statement that is not true — an invented citation, a wrong date, a person who does not exist.
Why it matters It is a normal failure mode, not a bug you can prompt away. Verification is part of using the tool.
Inference
The act of running a trained model to produce an answer.
Why it matters Inference costs money and time; that is why limits exist and why longer inputs are slower.
Knowledge cutoff
The point after which a model has no built-in knowledge, unless it can search or you provide the material.
Why it matters Ask for current facts only when the tool can actually look them up, and check the sources.
Large language model (LLM)
A model trained on large amounts of text to predict and generate language, which is why it can write, summarize, and explain.
Why it matters It predicts plausible text; it does not look up truth. That single sentence explains most of its behaviour.
Latency
The delay between sending a request and seeing the first part of the answer.
Why it matters Long prompts, long files, and reasoning modes all add latency before the first token appears.
Memory
A feature that lets the assistant save facts about you across conversations, which you can review and delete.
Why it matters Memory is not a document store, and it does not replace pasting the source material.
Model
The trained system that produces the text. The assistant is the product around it.
Why it matters Two products can use the same model and feel completely different because of their prompts, tools, and interface.
Multimodal
A model that can work with more than text — images, audio, or files.
Why it matters A photo is often a better prompt than a description, but check what it read before trusting the answer.
Open-weight model
A model whose trained parameters are published, so it can be downloaded and run on your own hardware.
Why it matters The offline and private option, with trade-offs in size, speed, and quality.
Parameter
A number inside a model that is adjusted during training. Model size is often described in billions of parameters.
Why it matters More parameters is not automatically better for your task; it usually means more compute.
Persona
A role given to the model in a prompt, such as "you are a copy editor".
Why it matters A concrete role with a standard works; a pile of grand titles mostly adds noise.
Prompt
The instruction and context you give the model. In a chat, every earlier turn is part of the prompt too.
Why it matters Most bad answers are bad briefs. The prompt is the part you control.
Prompt injection
Text hidden in a document or web page that tries to override your instructions when the model reads it.
Why it matters Treat any document you did not write as untrusted input, especially if the model can act on it.
Retrieval-augmented generation (RAG)
A pattern where the system first retrieves relevant documents, then asks the model to answer using them.
Why it matters It is how many "chat with your docs" tools work, and why citations matter more than fluent prose.
Reasoning model
A model or mode that spends more compute working through a problem before answering.
Why it matters Useful for multi-step problems, unnecessary for rewriting an email. It is slower and often costs more.
System prompt
Instructions set by the product or developer, above your message, that shape how the assistant behaves.
Why it matters You usually cannot see it, which is why two apps with the same model can answer differently.
Temperature
A setting that controls randomness in output — lower is more predictable, higher is more varied.
Why it matters Most chat products hide it. When you need consistency, ask for a fixed format rather than a lower temperature.
Token
A chunk of text the model processes — roughly a word piece. Limits and pricing are usually counted in tokens.
Why it matters Long documents cost more and hit limits sooner than their word count suggests.
Training data
The material a model learned from before it was deployed.
Why it matters It shapes the model's defaults, including its biases and its gaps. It is not a live database.
Zero-shot / few-shot
Zero-shot means asking without examples; few-shot means giving two or three examples of the output you want.
Why it matters When a format is hard to describe, two examples are worth a paragraph of instructions.
Missing a term? Send it in.