AI cheat sheet
AI terms in plain English.
This page turns common AI words into normal language. If a launch post, benchmark chart, or product demo sounds confusing, start here first and then go back to the story.
Short definitions
Why the term matters
Useful for news readers
How to use this page
If a company says a model has a bigger context window, better reasoning, or agent skills, you can scan this page in a minute and know what those words actually mean before deciding whether the claim is impressive.
Model basics
These are the words that show up in almost every AI launch, product page, or news story.
Core term
LLM
A large language model is a system trained on a huge amount of writing so it can predict words and respond like a chat partner.
Why it matters: when a company says it launched a new model, this is usually what it means.
Core term
Token
A token is a small chunk of text. Models read and write tokens, not full paragraphs all at once.
Why it matters: token limits affect how much the model can read, remember, and answer in one go.
Core term
Context window
This is how much text, code, or conversation a model can keep in view at one time before older details start falling out.
Why it matters: a bigger context window can help with long documents, large code files, and longer chats.
Core term
Fine-tuning
Fine-tuning means taking a general model and training it more so it becomes better at one job, one style, or one company’s data.
Why it matters: this is how companies try to make a broad model more useful for their own workflows.
Core term
Inference
Inference is the moment the model does the actual work of answering your prompt after training is already finished.
Why it matters: many pricing and speed claims are really about making inference cheaper or faster.
Core term
Open weights
Open weights means a company shares the model files so other people can run the model themselves, usually with some license rules attached.
Why it matters: open weights can make it easier to self-host, inspect, or customize a model.
Trust and quality words
These are the terms that usually tell you whether a model is trustworthy, tested, or still shaky.
Risk word
Hallucination
This means the model says something that sounds confident but is wrong, made up, or unsupported by evidence.
Why it matters: a polished answer can still be false, so confidence alone should never be treated as proof.
Risk word
Benchmark
A benchmark is a test used to compare models. It can be useful, but it does not always tell you how well a tool works in normal real-world tasks.
Why it matters: a benchmark win can sound huge even when the everyday product barely changes.
Risk word
Safety policy
A safety policy is the rulebook for what a model should refuse, warn about, or handle carefully.
Why it matters: policy changes can affect what a model will answer, who gets access, and how risky the product feels.
Quality word
RAG
RAG stands for retrieval-augmented generation. It means the model looks up outside information first, then answers using that extra material.
Why it matters: this can improve accuracy when the system actually uses fresh or trusted sources.
Quality word
Evaluation
An evaluation is a broader set of tests used to check whether a model is accurate, safe, reliable, or useful for a specific job.
Why it matters: serious teams show evaluations, not just marketing lines and dramatic demo videos.
Quality word
Reasoning
Reasoning usually means the model is better at multi-step thinking, problem solving, or sticking with a harder task without drifting.
Why it matters: this is one of the most common upgrade claims, so it helps to ask what real tasks improved.
Product and rollout words
These are the words that explain what the AI product can actually do for a person or a company.
Product term
Agent
An agent is an AI system that does more than chat. It can plan steps, use tools, and carry out tasks with less hand-holding.
Why it matters: agent claims usually suggest the product can take action instead of only talking.
Product term
Multimodal
Multimodal means the system can work with more than one kind of input or output, like text, image, audio, or video.
Why it matters: multimodal systems can read a photo, hear a question, and answer in text or voice.
Product term
API
An API is the way developers connect their own app or website to an AI model so the AI becomes part of another product.
Why it matters: many model launches matter most because they change what developers can build or afford.
Product term
Latency
Latency is the delay between the moment you ask for something and the moment the system starts answering.
Why it matters: lower latency makes AI feel faster, smoother, and more usable in real products.
Product term
Deployment
Deployment means moving a model from an announcement or test into the real world where people, teams, or customers can actually use it.
Why it matters: the jump from demo to deployment is where many bold claims either hold up or fall apart.
Product term
Copilot
Copilot usually means an AI helper that sits inside another app and tries to assist you while you work.
Why it matters: the word sounds friendly, but the real question is what work it can truly save you.
Quick questions to ask when you read AI news
- Is this a real product, or only a demo?
- Who can use it right now, and who cannot?
- What is actually new, and what is just a new name for an older tool?
- Did the company show proof, customer use, or independent testing?