AI AI basics: what it is and how it got here
AI jargon decoded: 30 words you keep seeing explained
The AI words that turn up in headlines and settings screens, each defined in one or two sentences: tokens, context, weights, RAG, agents, inference, distillation and more.
The short answer
- Nearly all AI jargon sorts into four families: the model itself, how it was built, how you use it, and how it fails, and placing a word in its family explains most of a sentence.
- Tokens, context window and inference are the three words behind almost every AI bill and usage limit, so they are the ones worth learning first.
- Fine tuning changes style and format, while retrieval puts real documents into the prompt, which is why retrieval is usually the right tool for company facts.
- Words like AI powered, smart and agentic have no fixed technical meaning, so treat them as marketing and ask what the system is actually allowed to do.
- Hallucination, sycophancy, prompt injection and bias name four different failures with four different defenses, and using them interchangeably hides the fix.
Almost every AI word you keep meeting belongs to one of four families: words for the model itself, words for how it was built, words for using it, and words for the ways it fails. Sorting a new term into the right family usually gets you most of the way to understanding a sentence, even before you know the exact definition. What follows is a little over thirty terms, each in a line or two, with the reason it matters to you rather than to an engineer.
Words for the model itself
These describe the thing that was built. Product pages quote them as specifications, which is why two of them, parameter count and open weight, are seen far more often than they deserve.
| Term | What it means | Why it matters to you |
|---|---|---|
| Model | The trained file itself: a very large set of numbers plus the rule for running data through them | Every other word here describes building, running or fixing this one thing |
| Parameters, or weights | The individual numbers inside, counted in billions and often written as 8B or 70B | More is not automatically better, but it always costs more to run: what a parameter count really tells you |
| Neural network | The layered arrangement those numbers sit in | The reason nobody can point to the part of a model responsible for a given answer |
| Transformer | The network design behind nearly every current language model, built around attention | Attention compares every position with every other one, which is why long inputs get expensive fast |
| Large language model | A transformer trained on an enormous amount of text to predict what comes next | This is the chatbot, and prediction is genuinely all it does |
| Multimodal | Handles images or audio as well as text, inside one model | Lets you paste a screenshot or a photo instead of describing it in words |
| Base model | The raw trained model before it was shaped into an assistant | A base model continues your text rather than answering you, which startles people who try one |
| Open weight | The numbers are published for download, usually without the training data or the code | Not the same as open source: what open actually means here |
| Quantization | Storing the weights at lower precision so the model takes less memory | How a large model gets small enough to run on a laptop, at a modest quality cost |
| Distillation | Training a small model to copy the outputs of a bigger one | Why cheap small models are much better than their size suggests |
Words for how it was built
This family explains a model's blind spots. Almost every odd behavior you notice traces back to something in this list rather than to the model deciding anything.
| Term | What it means | Why it matters to you |
|---|---|---|
| Training data | The text, images and code the model learned patterns from | Whatever is thin in the data is where the model is weakest, including your local area and your industry |
| Pre-training | The long first stage: predict the next chunk of text, over and over, across the whole corpus | This is where the knowledge comes from, and it is the part that costs millions to run |
| Fine tuning | Further training on a smaller, specific set of examples to change style or behavior | Good for teaching a format or a tone, usually the wrong tool for teaching facts |
| RLHF | Reinforcement learning from human feedback: people rate answers, and the model is tuned toward the preferred ones | Why assistants are polite and also why they over agree with you |
| Alignment | The broad effort to make a model behave as intended and refuse what it should refuse | Explains refusals, including the unhelpful ones on harmless questions |
| Knowledge cutoff | The date after which nothing was in the training data | Anything later is either retrieved live or invented, with no visible difference between the two |
| Compute | Raw processing time, usually on specialized chips, measured in chip hours | The main reason capable models cost money and the main constraint on who can build one |
| Benchmark | A fixed set of test questions used to compare models | Weak evidence on its own: published tests leak into later training data, which inflates scores |
Words for using it
These turn up in settings screens, pricing pages and developer documentation. Two of them, tokens and context window, explain most of what a bill and most of what a limit are made of.
| Term | What it means | Why it matters to you |
|---|---|---|
| Prompt | Everything you send, including files and earlier turns of the conversation | The prompt is not just your last message, which is why long chats behave differently |
| System prompt | Standing instructions from the provider or the app, placed before anything you type | Sets the rules you did not write and cannot see, including tone and refusals |
| Token | The chunk the model actually reads: roughly a short word or a piece of a long one | Usage is metered in these, not in words: how tokens decide your AI bill |
| Context window | The maximum number of tokens the model can hold in view at once | When a long chat starts forgetting, this is why: what the window does when it fills |
| Inference | Running a trained model to get an answer, as opposed to training it | Training happens once; inference is what you pay for every time you press send |
| Temperature | A setting controlling how adventurous the model is when picking each next token | Low means more predictable, which is not the same as more accurate: what the sampling settings really do |
| Embedding | A list of numbers representing a piece of text, positioned so similar meanings sit near each other | The machinery behind semantic search and behind most document chatbots |
| Vector database | Storage built for finding the nearest embeddings quickly | The retrieval half of any system that answers questions over your own files |
| RAG | Retrieval augmented generation: search your documents first, paste the best passages into the prompt, then answer | The usual right answer for company knowledge: RAG or fine tuning, and how to choose |
| Tool use | Letting the model call a calculator, a search index or an internal system and use the result | How a model gets arithmetic and current facts right without guessing |
| Agent | A model given tools, a goal and permission to take several steps on its own | Raises the stakes from a wrong sentence to a wrong action: when a chatbot starts doing things |
Words for the ways it fails
This family is worth learning properly, because the words are used loosely in coverage and each one describes a different problem with a different fix.
| Term | What it means | Why it matters to you |
|---|---|---|
| Hallucination | Fluent output that is not grounded in anything real | Not a bug being patched but a property of generation: why models make things up |
| Confabulation | The more accurate term some researchers prefer: filling a gap with a plausible invention | Captures the key point that the gap is never marked or noticed |
| Sycophancy | Agreeing with the user, or reversing a correct answer when pushed | A side effect of training on human ratings, and a reason not to argue with an answer you want |
| Prompt injection | Hidden instructions in a web page, document or email that the model reads and follows | The main security problem for any assistant allowed to browse or read your mail |
| Jailbreak | Wording designed to get around a model's restrictions | Mostly a research and moderation topic, and a reason refusals keep shifting between versions |
| Bias | Systematic skew inherited from the training data or the tuning process | Shows up as uneven quality across dialects, names, regions and groups of people |
| Regression | A new model version being worse at something the old one handled | Why a workflow that worked last month can quietly break after an update you did not choose |
Words that sound technical and are not
A handful of terms carry no fixed meaning, and treating them as specifications leads you astray.
AI powered describes anything from a genuine model to a sorting rule written by hand. It is a claim about marketing, not architecture. Smart and cognitive in a product name mean nothing at all. Agentic is used for everything from a single automated step to a system running unsupervised for an hour, so always ask what the thing is actually permitted to do without asking you.
Reasoning is the tricky one, because it does point at something real: models that generate intermediate steps before answering do better on multi step problems. It does not mean the system reasons the way you do, and a long visible chain of steps can still arrive confidently at a wrong premise. Understands and knows in vendor copy are similar: they describe how the output reads, not what happens inside.
What to do with this vocabulary
You do not need to memorize the list. The useful habit is asking which family a word belongs to. If it describes the model, it is a specification that matters less than the marketing suggests. If it describes training, it tells you where the blind spots are. If it describes usage, it affects what you pay or what fits. If it describes failure, it tells you what to check before relying on an answer. For terms beyond AI, including security and networking words, the plain English glossary covers the wider vocabulary in the same one line format.
Common questions
What is the difference between AI and a large language model?
AI is the whole field, covering everything from fraud scoring to image recognition to route timing. A large language model is one type of AI system, trained on text to predict what comes next. Most coverage that says AI means a language model, which is worth remembering when a claim about AI is really a claim about chatbots.
Is a token the same as a word?
Not quite. A token is the chunk the model reads, and common short words are usually one token while longer or unusual words break into several pieces. Punctuation and spaces count too. For ordinary English prose, a rough working figure is that a page of text costs somewhat more tokens than it has words.
What does fine tuning actually change?
It adjusts the model's weights using extra examples, which shifts style, format and the kind of task it expects. It is effective for making output consistently match a house pattern. It is a poor way to install facts, because the facts are absorbed as statistics rather than stored as records, and updating them means training again.
What is a vector database for?
It stores embeddings, the numeric representations of text passages, and finds the closest matches to a query fast. That matters when you want a system to answer from your own documents: the database supplies the relevant passages, and the model writes the answer from them. Without that retrieval step the model can only draw on training data.
Why do people say prompt injection is different from a jailbreak?
Because the attacker is in a different place. A jailbreak is the person at the keyboard talking their way past a restriction. Prompt injection is a third party hiding instructions in content the model reads later, such as a web page or an attachment. Injection is the more serious one for ordinary users, since you never see it happen.