If you read AI news regularly, you have run into a wall of jargon: parameters, tokens, context windows, RAG, agents. Most glossaries either oversimplify or drown you in math. This one is built for a specific purpose — so you can judge whether an AI headline is meaningful or marketing.
Each entry gives a definition, an analogy, and the reason the term shows up in coverage. Read it end to end, or jump to what you need.
Quick reference
| Term | One-line definition | Why it matters |
|---|---|---|
| LLM | A large model trained to read and write text | The engine behind every chatbot |
| Parameters | Tunable values inside the model | A rough proxy for size and cost |
| Tokens | The smallest chunks a model processes | How pricing and limits are measured |
| Context window | How much it holds at once | Determines long-document ability |
| Fine-tuning | Adapting a trained model to a task | How most company AI is built |
| RAG | Retrieve documents, then answer | Adds freshness and sources |
| Hallucination | Confident, fluent, wrong | The core reliability problem |
| Multimodal | Handles text, images and audio | Now standard in flagship models |
| Reasoning model | Works step by step before answering | Better at math and code, slower |
| AI agent | Plans and executes multi-step tasks | The industry current focus |
| Open weights | Model files published for download | Lets companies self-host |
| On-device AI | Runs locally on your phone or laptop | Privacy and offline use |
The basics
LLM (Large Language Model)
A model trained on enormous amounts of text to predict what comes next. That single training objective — guess the next chunk of text — turns out to produce translation, summarization and coding ability as side effects.
Common misconception. An LLM does not look up facts in a stored database. It generates statistically likely text. That distinction explains almost every weakness discussed below.
Parameters
Think of them as tuning screws inside the model. Training nudges each one slightly until output improves. A 70B model has 70 billion of them.
Read headlines carefully. More parameters is not automatically better. Larger models cost more to run and respond more slowly, and a well-trained small model routinely beats a poorly trained large one. An announcement of a bigger model tells you very little on its own.
Tokens
The unit a model reads and writes in — roughly a word or word fragment in English. Tokens matter because pricing is per token and limits are counted in tokens. When you see an 80 percent price cut or a one-million-token context, this is the unit being discussed.
Context window
The maximum amount of text a model can consider at once. Picture a desk: a wider desk lets you spread out more documents, but it does not guarantee you read them carefully. Models with very large context windows still miss details buried in the middle of long inputs.
How models are built and improved
Pre-training and fine-tuning
Pre-training builds general language ability from web-scale text. It is extraordinarily expensive, and only a handful of organizations do it. Fine-tuning takes a finished model and adapts it to a narrower job — legal documents, customer support, medical summaries.
Most enterprise AI announcements are fine-tuning stories, not new models.
RAG (Retrieval-Augmented Generation)
Before answering, the system searches a document store or the web, then writes its answer using what it found. This addresses two weaknesses at once: the model not knowing recent events, and the model inventing details.
Its limit: if retrieval surfaces the wrong document, the answer is confidently wrong with a citation attached. RAG improves grounding; it does not guarantee truth.
Hallucination
When a model states something false with complete confidence. Because it optimizes for plausible text rather than verified fact, saying nothing is known is often a less likely continuation than a fluent invention.
The dangerous cases are fabricated citations, statistics and legal references. They arrive in the correct format, which is exactly why people accept them.
Treat any factual claim from a model as unverified until you check the primary source. This applies doubly to numbers, dates, names and quotations.
Terms you will see this year
Multimodal
Able to work with images, audio and video alongside text. Once a headline feature, now close to table stakes for flagship models.
Reasoning models
Trained to work through a problem in steps before producing an answer. Substantially better on math, logic and programming, at the cost of latency and price. Many products now route easy questions to a fast model and hard ones to a reasoning model.
AI agents
Systems that take a goal and carry out multiple steps on their own: searching, writing files, calling other software. This is where most current investment is going.
Reading tip. Agent demos are unusually persuasive and unusually unrepresentative. Success rates in messy real conditions are typically far below what a rehearsed demo shows. Always ask whether the capability is generally available.
Open source versus open weights
Publishing the model file lets anyone run it on their own hardware, which appeals to organizations that cannot send data to a third party. However, training data and code are rarely released, so open weights is the more accurate term. Licenses vary, and downloadable does not mean free for commercial use.
On-device AI
Models that run on your phone or laptop instead of a remote server. The appeal is privacy and offline availability; the tradeoff is that on-device models are smaller and less capable than cloud ones.
Three questions this vocabulary lets you ask
Once the terms are familiar, most overhyped coverage collapses under three questions.
- Is it shipped, or announced? A demo, a paper and a waitlist are not a product.
- Who measured that number? Vendor-run benchmarks are marketing until independently reproduced.
- Compared to what? Twice as fast against a two-year-old model means very little.
We update this glossary as new terms become common in coverage. If something is unclear or missing, tell us through our contact page and we will address it.
작성자 소개
