You’re in a meeting and someone says “we need to make this more agentic.” Or you’re scrolling your feed and a founder is bragging about “vibe coding” a whole app over a weekend.
Or a friend mentions their favorite chatbot “hallucinated” a fake statistic and you just nod along like you know exactly what that means.
we have all been there.
In this guide, we’re finally explaining the AI terms you keep hearing, in plain English, with zero assumed computer science background.
Here’s what you’ll walk away with: clear definitions for the AI terms that show up constantly in tech news, product launches, and casual conversation right now, from the basics like “large language model” all the way up to newer buzzwords like “agentic AI” and “vibe coding.” We’ll go roughly in order from foundational concepts to the more advanced stuff, so feel free to skim to whatever term brought you here.
Let’s Start With the Big One: What’s an LLM?
Large language model, or LLM, is the term underneath almost everything else on this list. An LLM is a type of AI system trained on enormous amounts of text so it can recognize patterns in language and generate new text that sounds human. IBM describes it as an AI system capable of understanding and generating human language by processing vast amounts of text data, over billions or trillions of words pulled from books, websites, and articles.
Under the hood, LLMs run on a type of neural network architecture called a transformer, which is very good at figuring out how words in a sentence relate to each other, even when those words are far apart. Google’s own developer materials explain that a language model is essentially predicting the probability of what word (or word-fragment) comes next in a sequence, over and over, which is how it strings together coherent paragraphs.
ChatGPT, Claude, Gemini, Microsoft Copilot, and Meta AI are all examples of products built on top of LLMs, according to the University of Arizona’s library guide on the topic. When people say “the AI,” they usually mean one of these.
Foundation Models: The Bigger Umbrella
You’ll also hear “foundation model,” and it’s easy to assume that’s just a fancier word for LLM. It’s actually a broader category. A foundation model is a large-scale model trained on broad, diverse data, often using self-supervised learning, so that it can later be adapted to lots of different tasks through fine-tuning or prompting, rather than being built to do just one thing.
LLMs are a type of foundation model, but foundation models aren’t limited to text. Some are built for images, some for audio, and increasingly the big-name models handle several types of input at once. That brings us to the next term.
Multimodal: When AI Isn’t Just Reading Text
Multimodal simply means a model can work with more than one type of input : text, images, audio, even video, rather than being limited to plain text. TechTarget notes that today’s leading foundation models are largely multimodal because they combine language, audio, and vision capabilities in a single system. This is why you can now drop a photo into a chatbot and ask it what’s in the picture, or hand it an audio clip and ask for a summary.
Tokens and the Context Window: AI’s Working Memory
Here’s one that trips people up constantly: tokens and context windows. A token is the smallest chunk of text a model actually processes, sometimes a whole word, sometimes just part of one. NVIDIA explains that to understand a full prompt, an AI model has to process a whole sequence of these tokens at once, and every model has a limit on how many it can handle in a single go.
That limit is the context window, essentially the model’s short-term working memory. Decagon’s glossary puts it well: the context window is the amount of information a model can consider in one request, and it sets a hard ceiling on how much of your conversation, documents, or instructions it can actually “see” at once. This is why a chatbot might seem to forget something you said earlier in a long conversation, it’s not being forgetful on purpose, it’s just run out of room to hold everything in view. Worth noting too: a bigger context window isn’t automatically better. Very long contexts can suffer from what Anthropic’s own documentation calls “context rot,” where accuracy and recall actually degrade as the amount of stuffed-in text grows.
Fine-Tuning: Turning a Generalist Into a Specialist
Fine-tuning is what happens after a model has already been broadly trained. It’s the process of taking that general-purpose model and further training it on a narrower, more specific dataset so it gets better at a particular job – say, medical terminology, legal writing, or customer support for one specific company. As one industry explainer puts it, pre-training gives a model its general capability, while fine-tuning shapes its behavior for a specific domain. It’s the difference between a well-read generalist and a specialist who’s read everything in one particular field, closely.
Hallucinations: When AI Confidently Makes Stuff Up
This is probably the AI term with the most real-world consequences for regular users. A hallucination is when an AI model generates information that sounds completely plausible but is actually false, fabricated, or unsupported by any real source. Stanford’s Human-Centered AI institute defines it as instances where a system produces incorrect, misleading, or entirely made-up content while presenting it as fact.
The tricky part is the confidence. Grammarly’s explainer notes that AI tools generate this false information while appearing just as certain as when they’re correct, which makes hallucinations genuinely hard to catch without double-checking. It happens because, as Decagon explains, language models don’t store facts the way a database does, they’re predicting the next most statistically likely token based on patterns, not pulling up a verified record. That’s a fundamentally different process than “looking something up,” and it’s why fact-checking AI-generated claims (especially dates, statistics, and citations) is always worth the extra thirty seconds.
Prompt Engineering: A Quick Refresher
We’ve covered this one in depth in a previous piece, but the short version: prompt engineering is the practice of writing and refining your input to an AI model so it reliably gives you the output you actually want. Same model, same capabilities but a clearer, more structured prompt gets a noticeably better result than a vague one. If a chatbot’s answer feels flat or generic, the prompt is usually the first thing worth adjusting.
Chain of Thought and Reasoning Models: Watching AI “Show Its Work”
You may have noticed some AI tools now display a little “thinking” process before they answer. That’s chain-of-thought, or CoT – a technique, as NVIDIA’s glossary describes it, that helps a model reason more accurately by getting it to show its work, the same way a teacher asks a student to write out the steps of a math problem instead of just blurting the answer.
This has evolved into an entire category called reasoning models, which build in extra “thinking” steps automatically rather than needing to be specifically prompted for them. IBM notes that this step-by-step approach breaks complex problems into smaller, sequential pieces before arriving at a final answer, which tends to produce more reliable results on multi-step tasks like math, logic, or planning. If you’ve seen a chatbot pause and display a “thinking” summary before responding, that’s this in action.
RAG: Giving AI a Library Card
Retrieval-augmented generation, or RAG, solves a specific problem: LLMs only know what was in their training data, which has a cutoff date and doesn’t include your company’s internal documents or this morning’s news. RAG fixes that by connecting the model to an external, searchable source of information before it answers. NVIDIA describes it as a technique for enhancing the accuracy and reliability of generative AI by grounding it in facts fetched from outside sources, rather than relying purely on what it memorized during training.
In practice, when you use a chatbot’s “search the web” or “search my files” feature, that’s RAG doing the work behind the scenes, the system pulls in relevant, current information, feeds it to the model alongside your question, and the model uses that fresh material to write its answer. It’s also one of the more effective tools for reducing hallucinations, since the model has something concrete to point to instead of guessing.
AI Agents and Agentic AI: From Chatting to Actually Doing Things
This is the term you’re probably hearing the most right now. An AI agent is a piece of software built to carry out a specific task somewhat independently, checking your calendar, booking a reservation, filing a support ticket. Agentic AI describes the broader shift toward systems that can set sub-goals, plan multiple steps, and take action with limited hand-holding, rather than just answering a single question and stopping.
Red Hat draws a useful distinction: an AI agent is a noun, a specific thing you build or deploy while “agentic” is more of a descriptor, like saying a piece of software has become more capable of acting on its own. Google Cloud’s definition frames it as AI focused on autonomous decision-making and action, versus traditional AI that mostly responds to direct commands. In practice, that means an agentic system doesn’t just tell you the weather is bad for your flight, it might actually go check alternate flights, compare prices, and flag the best option, chaining several steps together toward a goal you gave it.
MCP: The Universal Connector
Model Context Protocol, or MCP, is a newer term that’s become common shorthand in the developer world, and increasingly outside it. It’s an open standard, originally introduced by Anthropic, that gives AI assistants a consistent way to connect to outside systems : your files, a calendar, a database, a search tool without needing a custom-built integration for every single one.
The analogy nearly everyone uses is a good one: MCP is like a USB-C port for AI applications. Just as USB-C gave every device a single standard way to plug in, MCP gives AI models a single standard way to plug into outside tools and data, instead of every company building its own one-off connection. This is part of what makes agentic AI practical, an agent can only take real-world action if it has a reliable way to actually reach the tools it needs.
Guardrails and Alignment: Keeping AI in Its Lane
As AI systems get more capable and more autonomous, “guardrails” and “alignment” come up a lot, and they’re related but not identical. Alignment generally refers to training a model so its behavior matches human intent and values in the first place. Guardrails are more like an external safety net layered on top, IBM describes them as the policies, controls, and technical measures that sit between an AI model and whatever it’s connected to, catching problems the model itself might miss.
Think of it like the difference between raising someone with good judgment (alignment) and still putting a fence around the pool anyway (guardrails). Both matter, and most production AI systems today, as one guide on the topic notes, lean on both, a well-behaved model plus hard external checks — because relying on either one alone tends to leave gaps.
Vibe Coding: Building Software by Describing the Vibe
This one’s newer and it’s spread fast. Vibe coding is the practice of using an AI system to generate computer code from natural-language descriptions rather than writing every line by hand. Merriam-Webster actually added it to the dictionary as a real term, which tells you something about how quickly it caught on.
The phrase was coined by AI researcher Andrej Karpathy in a 2025 post, where he described essentially forgetting the code even exists and just describing what he wanted, letting the AI handle the implementation. It’s genuinely useful for prototypes, personal projects, and getting an idea off the ground fast but it comes with real trade-offs. Because the person “driving” often isn’t reading every line the AI produces, vibe-coded projects can accumulate bugs, security gaps, or code nobody fully understands, which matters a lot more once something moves from a weekend project to a real product with real users.
A Few More Worth Knowing
A handful of other terms worth having in your back pocket, quickly:
- Parameters: the internal values a model adjusts during training that shape how it makes predictions. More parameters generally (though not always) means a more capable model.
- Inference: what’s happening when a trained model is actually generating a response to your prompt, as opposed to the earlier training process.
- Open-source vs. closed model: whether a model’s underlying code and weights are publicly available for anyone to inspect, run, or modify (open) or kept private by the company that built it (closed).
- Embeddings : a way of converting words, images, or other data into numerical form so a computer can measure how similar or related different pieces of content are to each other. This is part of what makes RAG’s search step work.
Quick Answers
Is “AI agent” the same thing as “agentic AI”? Not quite. An agent is a specific piece of software; agentic AI describes the broader behavior of systems that can plan and act with more independence.
Why do AI tools hallucinate instead of just saying “I don’t know”? Because they’re built to predict likely-sounding text based on patterns, not to check facts against a verified database, so a fluent, wrong answer and a fluent, right answer can feel identical to the model as it’s generating them.
Do I need to understand all of this to use AI tools well? No. But knowing this handful of terms makes it a lot easier to understand what a new AI feature actually does before you rely on it for something that matters.
The Bottom Line
None of these terms are as complicated as they sound once you strip away the jargon.
An LLM predicts text based on patterns.
A context window is how much it can “remember” at once.
RAG gives it a way to check outside sources.
Agentic AI lets it take multiple steps toward a goal instead of just answering once. And guardrails are there to keep all of it pointed in a safe, useful direction.
The AI conversation moves fast, but the underlying ideas are more graspable than the headlines make them seem and now you’ve got the vocabulary to keep up.
