Generative AI & Large Language Models (LLMs)

Generative AI is the branch of AI that creates new content — text, images, audio — instead of just analyzing what already exists. This tutorial explains what it is and zooms into the technology behind tools like ChatGPT and Claude: Large Language Models.

Table of Contents

• What Is Generative AI?
• What Generative AI Can Create
• What Are Large Language Models?
• How an LLM "Thinks"
• LLMs You'll Actually Use
• The Hallucination Problem

What Is Generative AI?

Generative AI is artificial intelligence that produces new content rather than sorting or summarizing existing content. You give it a description and it creates something that didn't exist before.

Older AI classified (is this email spam? is this photo a cat?). Generative AI generates — it writes a report, drafts an email, builds a formula, or draws an image from a text description.

Why it matters: this is the type of AI you'll actually talk to in Excel — the assistant that turns "make me a quarterly summary" into a finished piece of work.

Example — the difference:

Explanation: same input type, totally different job. One analyzes what's there; the other creates something new.

What Generative AI Can Create

Generative AI is not one tool — it's a family of tools, each trained to create one kind of content:

• Text — reports, emails, explanations (LLMs like ChatGPT, Claude, Gemini)
• Images — logos, illustrations, product mockups (DALL·E, Midjourney)
• Audio — voices, music, podcast narration
• Video — short clips and animations
• Code and formulas — scripts, and yes, Excel formulas

Example in the Excel context:

You describe your goal — "I need a formula that calculates the growth rate between this year and last year" — and the tool writes the formula. That's generative AI creating something new from your description.

What Are Large Language Models?

A Large Language Model (LLM) is a generative AI trained on enormous amounts of written text — books, articles, websites, documentation. From that, it learns how language works: how words relate, what answers usually follow which questions, and what a sensible response looks like.

"Large" refers to the size of the model — billions of parameters (the internal numbers the model tunes while learning). "Language model" means it models language: it has built a statistical map of how humans write.

Why it's useful: LLMs are the engine inside the chat assistants you'll use with Excel. When you understand what they're really doing, their strengths and weird mistakes both make sense.

Common misconception: the model isn't "trained on your workbook" and it isn't searching the internet live. It's applying patterns it learned during training.

How an LLM "Thinks"

Here's the single most useful mental model for an LLM: it predicts the next word (or token) in a sequence, one step at a time.

Give it "The cat sat on the," and it picks the most likely continuation — "mat" — based on everything it learned. It repeats this until the answer is complete.

Example — a tiny version:

Input:    "To calculate profit, subtract costs from"
Predicted: "revenue"  →  "To calculate profit, subtract costs from revenue."

Explanation: that's it — but done billions of times, with a context window that lets it keep track of what you said earlier in the conversation. When you ask an Excel question, it's not "looking up" the answer; it's generating the most probable useful response given your question and its training.

Why this explains the weird behavior: because it's predicting, it can sound confident even when wrong, and it tends to guess rather than say "I don't know."

LLMs You'll Actually Use

You won't build an LLM — you'll use one. The common ones:

Best Practice: the exact brand matters less than the skill of prompting. All of them do the same core job, and the techniques in the next tutorial work across all of them.

The Hallucination Problem

Because an LLM generates the most plausible answer — not a verified one — it sometimes produces fluent nonsense. This is called a hallucination: a confident-sounding answer that's wrong.

Common Mistakes:

• Believing a hallucinated formula will "obviously" be wrong — sometimes it looks perfect and still fails on real data.
• Asking for a specific fact (a tax rate, a date, a function name) and assuming it's accurate.
• Not testing the AI's output before using it on real work.

Example — in Excel:

Ask an LLM for "the function to find the second-smallest value," and it may give a plausible but wrong function. The fix is the same as with any tool: verify. In the spreadsheet, that means pasting the formula into a test cell, not your live report.