AI 101: Learn About AI | AI at The U | University of Miami

Learn About AI

AI 101

The key ideas behind today's AI tools in plain language, what's fact and what's fiction, and where you'll meet each one at the University of Miami.

The Foundations

Model (the large language model, or LLM)

A model is the engine. A large language model is a very large pattern-matcher that was trained on huge amounts of text and learned, in effect, to predict the next word given everything that came before. That sounds almost too simple to be useful, but when you scale it up, that single skill turns into the ability to answer questions, draft documents, summarize, translate, and write code. Picture a brilliant graduate who has read nearly everything in the library, but whose reading stopped the day they walked at commencement: fluent and well-read, but frozen at a point in time.

Two facts cut through most of the confusion. First, the model's knowledge is frozen at the moment its training ended, so on its own it does not know today's news or anything about your private files. Second, it does not "look things up" or reason like a person; it generates the most statistically plausible continuation of the text. That is why it can sound completely confident and still be wrong.

Fact vs. fiction

A model is not a database and it is not conscious. It has no goals, opinions, or awareness. It is a tool that produces text, and everything else in this guide is really about how we feed it better input and connect it to the outside world.

Where you'll see it at UM

Every tool we offer sits on top of a model or multiple. Microsoft 365 Copilot in Chat and your Office apps (where you can pick from the latest OpenAI and Anthropic models), and Google Gemini are each a friendly front end on one of these engines.

Tokens and the context window

Models do not read letters or whole words; they read tokens, which are small chunks of text (a short word is often one token, a longer word might be two or three). The context window is the total amount of text, measured in tokens, that the model can hold in view at once. It includes your instructions, any documents you paste in, and the conversation so far. Think of it as the desk in a closed-book exam: only what you can fit on the desk is in front of you, and anything you set on the floor is out of reach until you put it back on the desk.

Fact vs. fiction

A bigger context window is not the same as memory. Once something scrolls out of the window, or you start a new chat, it is gone unless a system deliberately stores and re-supplies it. The model is not quietly remembering your last session.

Prompt (and the system prompt)

A prompt is simply the input you give the model: your question or instruction. "Prompt engineering," stripped of its mystique, just means writing clearer instructions and giving useful examples. Being specific about the task, the audience, and the format you want will get you most of the way there.

A system prompt is a special set of instructions placed above your conversation that sets the ground rules: the assistant's role, tone, what it should and should not do, and how to format answers. You usually do not see it, but it is why a coding assistant behaves differently from a customer-service bot built on the same underlying model. Think of the system prompt as the course syllabus that governs the whole term, and your own prompt as the specific assignment you turn in that day.

Fact vs. fiction

There are no magic words. A clever phrase will not unlock hidden intelligence, and you do not need to be a programmer to write a good prompt. Clear beats clever almost every time.

Randomness (why the same prompt gives different answers)

At each step, the model has a range of plausible next words to choose from. A setting often called temperature controls how much it favors the safest choice versus a less likely one. Higher settings produce more varied, creative output; lower settings produce more focused, repeatable output. This is why asking the same question twice can give you two different answers.

Fact vs. fiction

Variation is not the model "thinking for itself" or being moody. It is a deliberate dial. If you need consistent, predictable output, that is a setting someone can turn down.


Grounding and Customization

Hallucination

A hallucination is when the model states something false as if it were true. Because it generates plausible text rather than retrieving verified facts, when it does not know an answer it will often produce a confident guess that looks right. This is not lying, which implies intent. It is the predictable result of a system built to sound fluent.

Fact vs. fiction

Hallucination has been reduced, not solved. The two most common fixes are grounding the model in real source documents (RAG, below) and keeping a human in the loop for anything that matters. Treat unverified output as a confident draft, not a citation.

RAG (retrieval-augmented generation)

RAG is the "open book exam" approach. Before the model answers, a search step pulls the most relevant passages from a specific, trusted knowledge base (your documents, help center, or database) and hands them to the model along with your question. The model then answers based on those passages, ideally with citations you can check. The model itself is not changed; you have simply given it the right pages to read for this one question.

Diagram of RAG as an open-book exam: your question, search trusted sources, relevant pages retrieved, model answers with citations.
Figure 1. RAG as an open-book exam: the system searches trusted sources, hands the model the relevant pages, and the model answers from those pages with citations you can check.

This is the workhorse behind most useful enterprise AI. It keeps answers current and grounded, and when something is wrong you fix the source document rather than retraining anything.

Fact vs. fiction

RAG does not teach the model anything permanently, and it does not guarantee correctness. If the retrieval step surfaces the wrong passage, or your source document is wrong, the answer will be too. Good RAG lives or dies on the quality of the underlying content and the search.

Where you'll see it at UM

When you use the licensed Microsoft 365 Copilot inside Word, Outlook, or Teams and it answers using your own emails, meetings, and files, that grounding in your work data is RAG at work.

Fine-tuning

Fine-tuning actually adjusts the model itself by continuing its training on a focused set of examples. It is best for changing how the model behaves: adopting a house style, following a consistent format, or getting fluent in the vocabulary of a specialized domain. If RAG is an open-book exam, fine-tuning is sending a capable generalist through a specialized residency or graduate program: it does not hand them a book to read during the test, it changes the instincts they bring to every case.

One thing worth understanding is that fine-tuning often narrows the space of outputs rather than broadening it. Imagine you fine-tune a model on a yes-or-no classification dataset drawn from our health system: thousands of patient portal messages, each paired with a single label of whether the message describes a symptom that needs urgent clinical attention. One row might be "I've had chest pain and shortness of breath since this morning," labeled "yes," and the next might be "Can I get a copy of my vaccination records?" labeled "no." Feed the model enough examples like these and what you get back is very good at exactly one thing: answering "yes" or "no" about urgency. If you later ask that same fine-tuned model to write you a research paper or a bedtime story, it will still only say yes or no, because that is the narrow behavior you trained into it and it never saw an essay or a story to learn from. Fine-tuning specializes a model, and specialization means giving something up.

Fact vs. fiction

Fine-tuning is usually the wrong first tool for "make it know our facts." It is comparatively expensive, it goes stale the moment your information changes, and it does not add a reliable memory of specific facts. For keeping knowledge current, RAG is almost always the better and cheaper choice. Reach for fine-tuning to change style and behavior, not to load in a knowledge base.

Diagram comparing RAG, which gives the model the right pages to read, with fine-tuning, which changes how it writes and behaves.
Figure 2. Two different jobs: RAG supplies the right pages to read and keeps facts current, while fine-tuning changes how the model writes and behaves.

Embeddings and the vector database

These are the plumbing behind RAG, worth knowing because the terms come up constantly. An embedding turns a piece of text into a list of numbers that captures its meaning, so that passages about similar topics end up with similar numbers. A vector database stores these numbers and, given your question, quickly finds the passages closest in meaning. It works like a library catalog organized by subject rather than by title: you can find the relevant material even when you do not know the exact words in the book. That is how a system retrieves "the relevant pages" even when your wording does not match the document word for word.

Fact vs. fiction

A vector database is not intelligent and it is not a brain. It is a specialized search index that matches by meaning instead of exact keywords. Useful, but ordinary.

Personalization and long-term memory

Earlier we noted that a bigger context window is not the same as memory, and that a model does not quietly remember your last session. Long-term memory is how modern applications work around that. The model itself still forgets, but the application wrapped around it keeps a running profile of you and feeds the relevant parts back in at the start of each new conversation. That is why a tool can seem to "know you" over time even though the underlying model has not changed.

This usually happens in two ways. First, the app saves specific facts and preferences, either because you asked it to ("remember that I prefer bullet points") or because it noticed something worth keeping. Second, it summarizes your past conversations and stores those summaries, so it can bring up an earlier discussion when you return to the same topic. Both are kept in a separate profile, not baked into the model, and the app pulls in whatever is relevant to your current question. Think of it as an academic advisor who keeps a file of notes on you: they do not memorize every word you have ever said, but they jot down what matters, glance at the file before each meeting, and pick up where you left off instead of starting from scratch.

Diagram of AI memory as a loop: you chat, it notes what matters, saves it to your profile, and recalls it next time.
Figure 3. Long-term memory as a loop: the assistant notes what matters from your chats, saves it to a profile, and recalls the relevant parts in later conversations so you do not start from scratch.
Fact vs. fiction

Memory is not the model learning about you or being retrained. It is a stored profile that gets read back in, which means it is selective, sometimes wrong, and fully under your control. You can view what a tool has saved, correct or delete it, and use a temporary or private chat when you want a session that neither reads nor updates your profile. It is also worth remembering that these notes are stored somewhere, so treat what you tell a tool to remember the same way you would treat any saved record.

Where you'll see it at UM

Microsoft 365 Copilot has a memory and personalization feature that saves memories, draws on your chat history, and applies any custom instructions you set. For UM accounts these memories are stored in your own profile and follow the same protections as your other work data. You can review and delete them in your Copilot settings, and a temporary chat skips memory entirely. Personal tools such as ChatGPT have their own memory settings that work in a similar way.


Agents and Their Parts

This is the area with the most hype, so it is worth going slowly. The key move is to see that an agent is built out of a few plain parts, and each part is limited in a way you can point to.

Tool

A tool is a single, specific function you let the model call to do something it cannot do on its own. A tool might search the web, look up a passage from your files (a RAG tool), fetch today's exchange rate, or write one record into a database. Each tool has a defined input, a defined output, and a narrow job. Tools are the hands of the system; the model is the part that decides which hand to use.

Fact vs. fiction

A tool is a specific, permissioned action, not a blank check. "Take over my computer" is not a tool. "Write a row to this one table" is. A system can only do what its tools allow, and each tool is something a person deliberately built and connected.

Agent

An agent is just a brain plus a set of tools. The brain is a model with some instructions, and the tools are the specific actions it is allowed to take. Once the brain and tools are configured, they define and limit what the agent can do: it can only read and act within its instructions and the tools it has been given.

A concrete example makes this clear. Suppose you configure an agent to analyze your grant submissions and give it access to a set of your publications. If you later ask it to do your taxes and file them with the IRS, it will not be able to. It was never instructed to do that, and it has no tool to do it: no access to your financial records and no ability to submit anything to the IRS. The agent is bounded by its design, not by good intentions.

Diagram of an AI agent: an LLM brain with instructions and a few tools, unable to file taxes because it has no tool for that.
Figure 4. An agent is a brain plus a bounded set of tools. It can act only within its instructions and the tools it was given, which is why a grant-analysis agent cannot file your taxes.

What makes an agent feel different from a plain chatbot is that it can take a goal, break it into steps, use its tools, check the results, and keep going until the task is done, rather than answering in a single shot. That loop is genuinely useful, but it runs entirely inside the fence built from its instructions and tools.

Fact vs. fiction

An agent is not autonomous in the science-fiction sense. It has no goals of its own, it cannot grant itself new powers, and it cannot wander outside the tools and permissions it was given. "Agentic" describes the multi-step working style, not independence. If an agent should not be able to do something, the fix is simple: do not give it the tool or the access.

Where you'll see it at UM

Copilot Cowork, the newer feature inside Microsoft 365 Copilot, is an agent that plans and carries out multi-step tasks while checking in for your approval. Our advanced users build their own agents in Copilot Studio.

MCP (Model Context Protocol)

MCP is a standard way to plug tools and data sources into an AI system. The common analogy is a USB-C port for AI: instead of building a custom, one-off connection between every model and every tool, each side implements the MCP standard once and they can all talk to each other. Introduced by Anthropic in late 2024, it has since been adopted across many major AI platforms.

Before MCP, connecting five models to five tools could mean building twenty-five separate integrations. With MCP, each model and each tool connects to the shared standard once, which cuts the integration work dramatically. That is the whole idea.

Diagram of MCP as a USB-C port for AI: tangled custom connections without a standard, one shared standard with MCP.
Figure 5. MCP as a shared standard: instead of a tangle of custom connections, each model and each tool plugs into one common port, like USB-C for AI.
Fact vs. fiction

MCP is plumbing, not intelligence. It standardizes how tools are described and called; it does not make the model smarter or more capable on its own. It also does not remove the need for permissions. A tool exposed over MCP is still just a tool, with the same limits and the same need to be granted access.

Where you'll see it at UM

In Copilot Studio you can connect an agent to an MCP server to give it new tools, alongside a library of ready-made connectors. The tools appear automatically once the server is connected.

Skill

A skill is the written know-how that tells an agent how to do a task well, or how to use a tool correctly. Building a tool and handing it to an agent is only half the job. The agent also needs to know when to reach for that tool, what good input looks like, how to interpret what comes back, and what the common pitfalls are. A skill captures all of that: the step-by-step instructions, the examples, the templates, and the context that turn a raw capability into reliable, correct work.

Think of it as the documentation or the standard operating procedure that sits next to a piece of lab equipment. The instrument alone does not guarantee good results; the written protocol next to it, explaining how and when to use it, is what makes the work repeatable and safe. In short, a tool is what an agent can do, and a skill is the instructions for doing it the right way, in the right situations.

Fact vs. fiction

A skill is not extra intelligence or a smarter model. From the model's point of view it often ends up looking like a well-organized set of instructions with some attached examples and reference files, and that is exactly the point: careful documentation is what makes a tool usable in the right contexts. Two cautions worth keeping in mind. The term means slightly different things from one platform to the next, so do not assume they are identical. And a skill is only as good as the instructions in it; vague or wrong guidance produces vague or wrong work, just like a poorly written protocol.

Where you'll see it at UM

Copilot Cowork uses skills, which are reusable sets of instructions, so a recurring task like drafting a newsletter follows your structure and tone every time.

Guardrails

Guardrails are the rules and checks placed around a model or agent to keep it inside acceptable bounds: filtering unsafe requests, blocking certain topics, validating output before it is used, or requiring a human to approve a sensitive action. They are part of the fence that, together with instructions and limited tools, defines what the system is allowed to do.

Fact vs. fiction

Guardrails reduce risk; they do not make a system foolproof. They are one layer of defense, which is exactly why the more reliable protections are structural: limited tools, limited access, and a human in the loop for anything that carries real consequences.


Myths vs. Reality, At a Glance

What the hype saysWhat is actually true
The AI "knows everything" and is always up to date.A model's knowledge is frozen at training time. Current or private information has to be supplied through RAG or tools.
Agents are autonomous and can do anything.An agent can only use the tools and access it was given. Remove the tool and the capability is gone.
Fine-tuning is how you teach an AI your company's facts.Fine-tuning changes style and behavior. For keeping facts current, RAG is usually cheaper and better.
If it sounds confident, it is correct.Fluency is not accuracy. Models can hallucinate, so verify anything that matters.
MCP, connectors, and skills are advanced intelligence.They are plumbing and packaging: standards, bridges, and organized know-how, not extra brainpower.
There are secret prompts that unlock hidden power.Clear, specific instructions do the real work. There are no magic words.

The Bottom Line

Once you strip away the marketing, generative AI is a set of understandable parts. A model predicts text. A prompt tells it what you want. RAG hands it the right documents to read, and fine-tuning adjusts how it writes. Tools let it act, an agent strings those actions together toward a goal, and MCP, connectors, and skills are the plumbing and packaging that make all of it easier to build. Every one of these parts is bounded by design.


Level Up Your AI Knowledge

  • AI for Everyone

    AI for Everyone (AI4E) is the University of Miami's enterprise AI learning program, designed to build AI literacy and practical skills across our institution. Grounded in the principle of digitally enabled, human led, AI4E provides a progressive learning journey that empowers faculty and staff to use AI responsibly, work more effectively, and unlock more time for innovation, collaboration, and high-value decision-making.

    All UM faculty and staff are invited to begin their AI4E learning journey today. Explore the available Learning Levels to discover the knowledge and skills you will build as you progress.

  • University of Miami Libraries

    The University of Miami Libraries' Artificial Intelligence Research Guide is made for students, staff, and faculty at UM to support your AI research by highlighting resources from the UM Libraries and beyond. Whether you're new to AI or more experienced, librarians are here to help; plus, you can contact a Subject Librarian for personalized assistance.

  • LinkedIn Learning

    Explore a wide range of self-paced courses for every level of experience designed to help you understand and apply AI in your work and daily life. These courses are available at no cost to the University of Miami community. Sign in with your UM credentials to get started.

  • O'Reilly

    Dive into a robust collection of AI and machine learning resources curated by industry experts. O'Reilly provides in-depth, technical content to help you build and advance your AI skills, available at no cost to the University of Miami community. Sign in with your UM credentials to get started.

Back to top ↑
Top