The AI
dictionary,
in English.
87 AI terms explained the way a smart friend would, no jargon, no textbook definitions. Bookmark it, share it with the person who keeps asking what an LLM is.
A
- Agent / agentic AI
- An AI that can take multiple steps on its own toward a goal, like searching the web, running code, or using other tools, instead of just answering one message and stopping.
- AGI (Artificial General Intelligence)
- The hypothetical point where AI can handle any intellectual task a human can, not just the specific things it was trained for. It does not exist yet, no matter how confident the headlines sound. Worth knowing because companies use the term to raise money and stir hype.
- AI (Artificial Intelligence)
- The big umbrella category. Any computer system trained to recognize patterns and make predictions, not a robot brain and not magic.
- AI detector
- A tool that claims to tell whether text was written by AI. Treat the results as a guess, not a verdict: they flag plenty of human writing as AI and miss plenty of AI writing, which is why students keep getting falsely accused by them.
- AI wrapper
- An app built as a thin layer on top of someone else's AI model, adding a nicer interface or a specific workflow but no AI of its own. Nothing wrong with wrappers, most AI apps are one. The value is in the workflow, not the engine.
- Algorithm
- A set of steps a computer follows to get something done, like a recipe. On social media, "the algorithm" means the recipe that decides which posts you see next, based on what you watched, liked, and skipped.
- API
- A way for one piece of software to talk to another. In AI, it's what lets an app or a piece of code send a request straight to a model and get an answer back, with no chat window involved.
- Artifacts
- A Claude feature that opens a separate window for things like documents, code, or visuals it creates, so they don't get buried in the chat and you can keep working on them.
- Automation (vs AI)
- Automation follows fixed rules and breaks when something unexpected happens. AI adapts to input it's never seen before.
B
- Benchmark
- A standardized test used to compare AI models, like an SAT for machines. Useful for rough comparisons, but companies teach to the test, so a great benchmark score does not always mean great real-world answers.
- Bias (AI bias)
- When an AI's answers unfairly favor or disfavor something, usually because of patterns baked into its training data. Different problem than hallucination: hallucination is about being wrong, bias is about being unfairly skewed.
- Bot
- Any program that acts on its own, from harmless (an auto-reply) to shady (fake social media accounts). Chatbots are bots you talk to. Not every bot is AI, most just follow simple scripts.
- Browser extension
- A small program you add to your web browser that gives it a new ability, like blocking ads, saving passwords, or flagging fake websites. It rides along while you browse. Only install ones from official stores, extensions can see what you do in the browser.
C
- Chain of thought
- When an AI works through a problem step by step before giving its final answer, like showing your work in math class. Models that do this tend to make fewer reasoning mistakes, which is the idea behind reasoning models.
- Chatbot
- Any program you talk to through a chat window. Not every chatbot runs on an LLM, older chatbots just followed scripted rules.
- ChatGPT vs LLM (the app vs the engine)
- ChatGPT is the app (the interface you type into). GPT-4 (or whichever model) is the LLM underneath it doing the actual work. Same idea with Claude: Claude is the app, the model underneath is the LLM.
- Claude Code
- Anthropic's command-line tool that lets Claude read, write, and run code directly in your own project files, not just chat about code.
- Claude Project
- A dedicated workspace in Claude that remembers your context (goals, files, running history) across sessions, so you don't have to re-explain yourself every time.
- CLI (command-line interface)
- A text-only way to control a program by typing commands instead of clicking buttons. It's the plain screen full of text that looks like the 'terminal' or that old-school black window. Claude Code runs in a CLI. Looks intimidating at first, but once it clicks you can move faster than clicking around ever would.
- Cloud (the cloud)
- Someone else's computers, reached over the internet. When an app stores your files or runs AI "in the cloud," it means the work happens in a big data center somewhere instead of on your device.
- Context (in a prompt)
- The background info you give an AI: who you are, what you're trying to do, what your constraints are. The difference between a generic answer and a personal one.
- Context window
- The total amount of text (measured in tokens) an AI can hold in its memory at once during a conversation, including everything you've typed plus everything it's replied. Analogy: a whiteboard of a fixed size. Once it's full, writing more means erasing the oldest lines at the top.
- Custom GPT
- A personalized version of ChatGPT that anyone can build without code: you give it instructions, maybe some reference files, and it behaves like a specialist. People share them free in the GPT store. Analogy: the same employee handed a different job description.
- Cybersecurity
- The practice of protecting computers, networks, and accounts from people who shouldn't have access to them, everything from a strong password to how a company defends against hackers.
D
- Database
- Organized storage for information a program needs to keep track of, like your account details, past conversations, or a company's customer list. Analogy: a giant, searchable filing cabinet a computer can search through in a fraction of a second.
- Deep learning
- The technique behind modern AI: neural networks stacked many layers deep, learning patterns from massive amounts of data. When people say AI "trained on the internet," deep learning is how it digested it.
- Deepfake
- AI-generated video or audio that convincingly imitates a real person saying or doing something they never did. Getting cheaper and easier every year, which is exactly why a weird urgent video call from a "family member" deserves a callback on their real number.
- Deploy / deployment
- Pushing software from your computer to the internet so real people can actually use it. "Deployed to production" just means it is live now. The moment of truth for anything you build.
- Distillation
- Training a smaller, cheaper model to imitate a bigger one, so you keep most of the quality at a fraction of the cost and speed. It's how the fast, affordable models get so good. Analogy: an experienced tradesman training an apprentice to handle the routine jobs the same way he would, so he doesn't have to do all of them himself.
E
- Embeddings
- A way of turning words, images, or other data into lists of numbers that capture their meaning, so a computer can measure how similar two things are. It's what makes search and RAG possible behind the scenes.
- Encryption
- Scrambling information so only someone with the right key can read it. It is why a padlock shows in your browser bar, why your messages can be private, and why stolen data is sometimes useless to the thief.
F
- Fable
- One of Claude's models, part of the same family as Sonnet and Opus, each built with a different balance of speed, cost, and capability.
- Fine-tuning
- Taking an already-trained AI model and training it further on a smaller, focused set of examples so it gets better at one specific kind of task. Analogy: a general doctor doing a specialty residency to become a surgeon.
- Folder
- A container for organizing files on a computer, exactly like a physical folder for organizing papers. Nothing fancy, just a way to group related things together so they're easy to find later.
- Foundation model
- A large, general-purpose AI model trained on broad data that can be adapted for many different tasks, rather than built to do just one job.
- Frontend vs backend
- The frontend is the part of an app you see and click. The backend is the machinery behind it: the server, the database, the logic. Analogy: the dining room versus the kitchen. Both have to work for dinner to show up.
G
- Generative AI
- AI that creates new content, like text, images, or audio, rather than just analyzing or sorting existing data.
- GPT (the acronym)
- Generative Pre-trained Transformer. Generative: it creates text. Pre-trained: it learned from huge amounts of data before you ever met it. Transformer: the neural network design that made modern AI possible. ChatGPT is literally "chat" plus that acronym.
- GPU (graphics card)
- The computer chip that powers AI. Originally built to draw video game graphics, it turned out to be perfect for the math AI needs, doing thousands of small calculations at once. The global scramble for GPUs is why AI companies spend billions on chips.
- Guardrails
- The built-in limits and safety rules that keep an AI from doing or saying things it shouldn't.
H
- Haiku
- Claude's fastest and most affordable model, built for simple, high-volume tasks where speed matters more than deep reasoning.
- Hallucination
- When an AI states something wrong with total confidence. Every AI tool does it sometimes. Doesn't make AI useless, just means anything that actually matters needs a quick check.
- Harness
- Everything wrapped around an AI model that turns it into an actual product: the tools it's allowed to call, the instructions it starts with, how it reads your files, what it can and can't touch. The model is the engine, the harness is the rest of the car. Two apps running the exact same model can feel completely different because their harnesses are different.
I
- Inference
- The moment an already-trained AI actually generates a response to your prompt. Training happens once, ahead of time. Inference happens every single time you use it.
J
- Jailbreak
- An attempt to trick an AI into ignoring its safety guidelines. Worth knowing the term exists, not something to try.
- JSON (JavaScript Object Notation)
- No, not a guy named Jason. JSON is just a tidy way to write down information so both people and computers can read it. It pairs labels with values, like "name": "Izic" or "followers": 4, wrapped in curly braces. It's how apps and AI tools quietly pass data back and forth behind the scenes: your settings, an API's answer, a list of results. If you've ever peeked under the hood of an app and seen text full of curly braces and quotes, that was JSON. Analogy: a fill-in-the-blank form where every blank is clearly labeled, so nothing gets mixed up.
L
- Latency
- How long it takes an AI to respond after you send a message.
- LLM (Large Language Model)
- A specific type of AI trained on huge amounts of text so it can understand what you say and write back in plain, natural language. Every LLM is AI, but not every AI is an LLM (same as every car is a vehicle, but not every vehicle is a car).
- Loop (agent loop)
- The cycle an AI agent runs in when it works on its own: think, take an action, look at what came back, decide the next step, repeat, until the job is done or you stop it. It's the difference between an AI that answers one question and an AI that actually gets something finished.
- LoRA (Low-Rank Adaptation)
- A cheap, lightweight way to fine-tune a model. Instead of retraining the whole thing, which costs a fortune, you train a small add-on that adjusts its behavior. Analogy: a clip-on lens for the camera you already own instead of buying a new camera.
M
- Machine Learning (ML)
- A method inside AI where systems learn from data instead of being explicitly programmed with rules.
- Markdown (.md file)
- A plain text file format that uses simple symbols, like # for a heading or * for a bullet point, to add structure without needing a full word processor. Easy for both humans and AI to read and write, which is why AI tools often store notes and instructions in .md files behind the scenes.
- MCP (Model Context Protocol)
- A standard that lets AI models like Claude connect to outside tools and data sources (Canva, Gmail, a database) in a consistent way, instead of every connection being custom-built.
- Memory (AI memory via folders and files)
- How some AI tools remember things across separate conversations, since each conversation normally starts fresh. Instead of holding everything in its head, the AI saves notes to files (often Markdown files) inside folders, then reads those files back in at the start of a new conversation. Analogy: leaving yourself sticky notes in labeled folders, so future-you doesn't have to start from scratch every time.
- Model size (small vs large)
- Bigger models generally know more and reason better, but are slower and more expensive to run. Smaller models are faster and cheaper but less capable. Bigger isn't automatically better, it depends on the task.
- Multimodal
- An AI that can work with more than just text, for example reading images, listening to audio, or understanding video, not just typing back and forth.
N
- Neural network
- The general design most modern AI is built on, loosely inspired by how neurons in a brain connect and pass signals to each other. It's the underlying structure, not the AI itself.
O
- OCR (optical character recognition)
- Technology that reads text out of images, like pulling the numbers off a photographed receipt or making an old scanned document searchable. Modern AI models do this automatically when you upload a picture.
- Open-source model
- An AI model whose files are released publicly so anyone can download, run, and modify it, instead of only renting access through a company's app. Free and private, but you supply the computer power.
- Opus
- Claude's most capable model, built for the hardest, highest-stakes tasks. Slower and pricier than Sonnet, but reasons more deeply and handles more complex work.
P
- Parameters
- The internal settings inside an AI model that get adjusted during training. More parameters generally means a bigger, more capable model.
- Phishing
- A scam message, email, text, or fake website designed to trick you into handing over passwords or payment info by pretending to be someone you trust, like your bank. The tell is urgency: "act now or lose your account." Slow down and go to the real site yourself.
- Prompt
- Simply the message you type to an AI. That's it, no more complicated than that.
- Prompt engineering
- The skill of writing clear, specific prompts to get better results from an AI. Basically, getting good at asking the right way.
- Prompt injection
- When someone hides instructions inside text an AI is going to read, like a customer message, a webpage, or even characters smuggled inside an emoji, hoping the AI treats those hidden words as new orders instead of just content. It's how attackers try to make a chatbot reveal its instructions or say things it shouldn't. Related to a jailbreak, but sneakier: a jailbreak is someone directly asking the AI to break its rules, prompt injection is slipping the trick into something the AI reads along the way.
R
- RAG (retrieval-augmented generation)
- A technique where an AI looks up relevant information from an outside source, like your documents or the internet, before answering, so it can respond with current, specific facts instead of relying only on what it memorized during training.
- Rate limit
- A cap on how many requests you can send a service in a given window of time. Hit it and you wait. It is why an AI tool sometimes says "try again later" even though nothing is broken.
- Reasoning model
- An AI model built to think before it answers, spending extra time working through the problem step by step instead of replying instantly. Slower and pricier, but noticeably better at math, logic, and multi-step problems.
- Repository (repo)
- The folder where a software project lives, along with the full history of every change ever made to it. Usually hosted on a site like GitHub. When someone says "check the repo," they mean look at the project's files.
- RLHF (Reinforcement Learning from Human Feedback)
- A training step where real people rate an AI's answers and the model learns to give more of what they preferred. It's a big reason modern chatbots feel helpful and know when to refuse, instead of just autocompleting whatever comes next.
S
- Scraping (web scraping)
- Using a program to automatically collect information from websites, like pulling every price from a store or every question from a forum. It is how a lot of AI training data was gathered, and a big reason AI and copyright keep ending up in court.
- Server
- A computer (or program running on one) that sits somewhere else and responds when your device asks it for something, like a webpage, a file, or an AI's answer. Analogy: a waiter who takes your order to the kitchen and brings back what you asked for.
- Sonnet
- One of Claude's models, built as a strong all-around balance of speed, cost, and capability. Good for most everyday tasks without the wait or price tag of the largest model.
- Streaming (AI responses)
- When an AI's answer appears word by word instead of all at once. The model is not typing for effect, it genuinely generates one token at a time, and streaming shows you the tokens as they come instead of making you wait for the whole thing.
- Synthetic data
- Training data generated by AI instead of collected from the real world. Used when real data is scarce, private, or expensive. Powerful, but risky in excess: models trained mostly on AI output can drift away from reality.
- System prompt
- Instructions given to an AI behind the scenes, before you type anything, that shape how it behaves for the whole conversation. Different from a regular prompt, which is what you type.
T
- Temperature (AI setting)
- A setting that controls how random or predictable an AI's answers are. Low temperature gives safer, more repetitive answers. High temperature gives more creative, surprising ones.
- Text-to-speech / speech-to-text
- Text-to-speech turns writing into a spoken voice. Speech-to-text turns your voice into writing (also called transcription). Together they are how you can talk to an AI and have it talk back.
- Token / tokens
- The small chunks of text an AI reads and writes one at a time, roughly a piece of a word rather than a whole word. It's the actual unit AI companies use to measure how much text a conversation costs or how much it can handle at once. A single word often splits into more than one token: "hamburger" typically breaks down into three, ham, bur, and ger.
- Tool use
- When an AI can call outside tools mid-conversation, like running a web search or a calculation, instead of only using what it already knows.
- Training data
- The huge collection of text, images, and other material an AI model learned from before it was ever released. It's where the AI's knowledge comes from, and also where its blind spots and biases come from.
- Two-factor authentication (2FA)
- A second lock on your account beyond the password, like a code from an app or a physical security key. Even if a scammer steals your password, they hit a wall. The single highest-value security habit for the effort involved.
V
- Vector database
- A database built for storing embeddings (meaning turned into numbers) so an AI can instantly find the most similar or relevant pieces of information. It is the filing system that makes RAG fast. Analogy: a library organized by what books mean, not by their titles.
- Vibe coding
- Building software by describing what you want to an AI in plain English and letting it write the code, checking whether it works instead of reading every line. How non-programmers ship real apps now. The skill shifts from writing code to describing clearly and testing honestly.
- Voice cloning
- AI that copies a specific real person's voice from a short sample. Legit uses exist (audiobooks, accessibility), but it is also how the "grandkid in trouble, wire money" phone scam got an upgrade. Families should agree on a code word.
- VPN (virtual private network)
- A service that routes your internet traffic through an encrypted tunnel, hiding it from the Wi-Fi you are on and masking your location. Useful on public Wi-Fi. Not the magic invisibility cloak the YouTube ads sell, sites can still track you in other ways.
Z
- Zero-shot / Few-shot
- Zero-shot means asking an AI to do a task with no examples given. Few-shot means giving it a couple of examples first to show exactly what you want.
Want the words to make sense in practice?
One practical AI tip a week, written for beginners. Same plain English as this page.