# AI in Plain English

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Everything AI - beyond the hype](https://dylit.info/pr/everything-ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [AI in Plain English](https://dylit.info/ch/ai-in-plain-english/6ab3536999638e119f592c02)

Six Words You Need, Defined Once There is a particular kind of meeting where someone says "we should fine-tune a model on our own data" and four people nod, and at least two of them are nodding at a phrase they could not define. No shame in it. The vocabulary arrived fast, mostly from engineering, and nobody circulated a glossary. Six terms cover most of what you will hear. Here they are, defined once, in the way they actually get used. Model The trained system itself. The thing that does the guessing. When people say "which model are you using," they mean which of these trained systems, since they differ in capability, cost and speed. A useful mental image is a vast collection of learned patterns, frozen in place. The model is not a program someone wrote line by line. It is the result of a training process, and even the people who built it cannot point to where inside it any particular piece of knowledge sits. Training The process that produced the model. Enormously expensive, done once, finished long before the tool reached you. The important consequence: the model does not change while you use it. It is not learning from your conversations. If it gave you something wrong this morning and you corrected it, that correction is gone the moment the conversation ends. Most of the "it's getting smarter as I use it" feeling is you getting better at asking. Inference What happens each time you send a message. The model takes what it learned in training and applies it to your particular request. Cheap and fast compared to training. This gap is the reason these tools are affordable at all. The expensive part happened once, years of compute condensed into a file, and every individual use draws on it for a fraction of a cent. Generative AI AI built to produce new content, text, images, audio, code, rather than to classify or score things that already exist. The distinction matters more than it sounds. Older AI mostly answered questions with a fixed set of possible answers: is this transaction fraudulent, is this email spam, which of these customers will churn. Generative systems produce open-ended output, which is why they are useful for drafting and also why the output needs reviewing in a way a fraud score does not. Prompt Everything you type. Instructions and context together. The word makes it sound like a magic phrase, which has caused a lot of unnecessary mystique. A prompt is closer to a brief. If you handed the same text to a capable new colleague with no other information, could they produce what you want? That question tells you almost everything about whether your prompt is any good. Token A chunk of text, roughly three quarters of a word. Models read and write in these units rather than in whole words. You will meet tokens in two places. Pricing, because most tools bill by tokens consumed and produced. And limits, because every model has a ceiling on how many it can hold in view at once. When a tool says it supports a context of some large number, that number is in tokens, and it covers your input, any documents you attached, and the reply combined. Two more worth knowing Context window: The total amount the model can see at one time. Everything outside it does not exist as far as the model is concerned, including the earlier parts of a long conversation once the window fills up. Hallucination:  Content the model produced that is simply false, stated with the same confidence as everything else. A made-up statistic, a plausible citation for a paper that was never written. The word makes it sound like a rare malfunction. It is a normal consequence of how the system works. Why this is enough You do not need to know about attention heads or embedding dimensions to use these tools well, any more than you need to understand a codec to run a video call.