1 What an LLM Actually Is Tokens, parameters, and the one sentence that explains everything • Tokenizer lab • Next-word game

The whole secret fits in one sentence. A large language model is a machine that answers one question, billions of times per conversation: "given the text so far, what is likely to come next?" It does not look things up in a database. It does not "know" facts the way you do. It has read an enormous amount of text and compressed the patterns of that text into numbers. When it writes, it is rolling weighted dice over what word-piece should come next, again and again, very fast.

Hold onto that sentence. Every strength and every weakness of these tools falls out of it. Why are they brilliant at drafting emails? Because millions of emails shaped the dice. Why do they sometimes invent a court case or a citation? Because the dice produce plausible text, and plausible is not the same as true.

Token: the unit an LLM reads and writes. Not quite a word, not quite a letter. Common words are one token ("the", "email"). Rarer words get split into pieces ("demystify" might be "dem", "yst", "ify"). A model never sees letters or words, only token numbers.

Parameter (or weight): one of the numbers inside the model that gets adjusted during training. GPT-3 has 175 billion of them. When people say a model "learned" something, they mean these numbers moved. Nothing else is stored.

Training vs inference: training is the once, months-long, massively expensive process of adjusting the weights on trillions of tokens of text. Inference is what happens when you chat: the weights are frozen, and the model just predicts. Your conversation does not reprogram the model in the moment.

Interactive Lab 1

See Like a Model: The Tokenizer

Type any sentence. Watch it break into tokens, the only thing the model ever sees. Try a common word, then a rare one, then "strawberry". This is an approximation of how real tokenizers (like BPE) behave, simplified so you can see the idea.

Why this matters at work: models are priced and limited by tokens, and they can struggle with letter-level questions ("how many r's in strawberry?") because they never see letters, only chunks.

Game 1

Think Like the Machine: Predict the Next Word

For each sentence, pick the word an LLM would rate most likely. You are not guessing the truth. You are guessing the statistics of human text. That distinction is the entire lesson.

Stop and think In the game, "The capital of Jamaica is ___" had one overwhelming winner, but "Our company's biggest asset is our ___" did not. What does that tell you about which questions an LLM answers reliably, before you ever check a fact?

2 The Engine Room: Train a Network by Hand Weights, loss, and gradient descent, learned by moving them yourself • Neural network playground

Feynman said the best way to understand a thing is to build it. You are about to train a neural network, the same species of machine that powers every LLM, just small enough to hold in your head: 2 inputs, 2 hidden neurons, 1 output, 9 numbers total.

The task is real office work: should this support ticket be escalated? The network reads two signals, urgency and customer frustration, each scored 0 to 1. It must output near 1 (escalate) when either signal is high, and near 0 (routine) when both are low. You have four training examples, shown below the diagram.

A neuron is embarrassingly simple: multiply each input by a weight, add the results plus a bias, and squash the sum into a number between 0 and 1. Whatever intelligence emerges comes from stacking thousands of layers of these and finding the right weights.

Loss function: a single number measuring how wrong the network is across all training examples. Here it is mean squared error: for each example, take (prediction minus target), square it, then average. Loss 0 means perfect. Training is nothing more than pushing this number down.

Gradient descent: the algorithm that trains every modern AI. For each weight, ask "if I nudge this up a hair, does the loss go up or down?" Then move every weight a small step downhill. Repeat millions of times. The step size is called the learning rate.

Interactive Lab 2 • The Centerpiece

The Neural Network Playground

Drag the sliders and watch the calculations flow left to right. Line thickness shows weight size; cyan is positive, orange is negative. First try to beat the loss by hand. Then press "Step Downhill" and watch gradient descent do what you were doing, but with calculus. Get the loss below 0.02 to win.

Loss: ?

Loss under 0.02. You just trained a neural network.

Voxel Lab

The Loss Landscape

Zoom out from nine weights to the whole picture. Every possible setting of the weights is a point on a landscape, and the height at that point is the loss. Training is a ball rolling downhill. Click anywhere on the terrain to drop the ball and watch gradient descent hunt for a valley.

No ball on the terrain yet. Click to drop one.

You found the deepest valley: the global minimum.

One catch is visible from up here: the terrain has more than one valley. Drop the ball near the shallow valley and it settles there, stuck at a loss that is low but not the lowest. Real training fights this with momentum, random restarts, and other tricks. The picture to keep: learning is descent across terrain far too large to ever see all at once.

Now scale the picture up. An LLM is this exact machine with two changes: it has billions of weights instead of nine, and its training goal is not "escalate or not" but "predict the next token of internet text". The loss function rewards it for assigning high probability to the token that actually came next. Everything you just felt in your fingers, the nudging, the downhill steps, the loss falling, happened to GPT-4 and Claude too. Just multiplied by a few trillion.

Stop and think When you pressed Auto-Train, the loss sometimes dropped fast and then crawled. Real models cost tens of millions of dollars in computing power partly because of that long crawl. What does that suggest about why companies do not retrain models every week, and why the model's knowledge has a cutoff date?

3 Attention and the Context Window How the model decides what matters • Attention visualizer

Read this sentence: "The trophy would not fit in the suitcase because it was too big." What does "it" refer to? You answered instantly: the trophy. Now change one word: "...because it was too small." Suddenly "it" is the suitcase. You resolved that using meaning, not grammar. For decades, computers could not do this.

The breakthrough was attention, introduced in a 2017 paper from Google titled "Attention Is All You Need". The idea: when processing each token, let the model look back at every other token and assign each one a relevance score. High score, strong influence. The architecture built on this trick is the transformer, the T in GPT.

Interactive Lab 3

Watch Attention Decide

Click any word to see where its attention flows. Then flip "big" to "small" and click "it" again. Watch the arcs move to a different noun even though the grammar is identical. Arc thickness shows attention strength.

Click a word above.

Context window: the maximum number of tokens the model can pay attention to at once: your conversation, your pasted documents, and its own replies combined. Modern models handle from tens of thousands to over a million tokens. Anything that scrolls out of the window may as well never have been said.

What this means at work

Paste the source, not a description of it. The model can attend to what is in the window, so give it the contract, the report, the email thread itself.
Long chats drift. In a marathon conversation, early instructions can fall out of the window or get diluted. Restate what matters or start fresh.
Put instructions near the task. Attention finds them more reliably when your requirements sit next to the content they govern.

Stop and think A colleague complains: "I told the chatbot our formatting rules an hour ago and it stopped following them." Using the words "context window" and "attention", explain to them what likely happened, in two sentences.

4 Temperature: Why the Same Question Gets Different Answers Probability, sampling, and controlling randomness • Temperature lab

Remember: the model produces a probability for every possible next token. It then has to pick one. Always picking the single most likely token makes output repetitive and flat. So instead, the model samples: it rolls dice weighted by those probabilities. Temperature is the knob that reshapes the dice.

Low temperature sharpens the distribution: the favorite almost always wins, output becomes consistent and predictable. High temperature flattens it: underdogs win often, output becomes varied, creative, and occasionally unhinged. Neither is "better". They are settings for different jobs.

Interactive Lab 4

The Temperature Knob

The model must complete: "Our quarterly results were ___". Move the slider and watch the probabilities reshape. Then press Sample a few times at temperature 0.2, and again at 1.8. Feel the difference you would feel in a real chat.

Temperature 0.8

Samples: none yet

Work settings, plain rules

Contracts, data extraction, policy answers, code: you want low temperature behavior. Consistency beats flair. Many tools let you set this; in a chat interface, asking for "precise, deterministic" phrasing and using the same wording each time helps.
Brainstorming, naming, marketing angles, first drafts: higher temperature behavior is the point. Ask for 20 options, not 3. Variety is what you are paying for.
Never judge a model on one roll. One bad answer is a sample, not a verdict. Regenerate before you conclude.

Top-p (nucleus) sampling: a sibling knob. Instead of reshaping all probabilities, it cuts the long tail: sample only from the smallest set of words whose probabilities add up to p (say 90%). You will see it next to temperature in API tools.

5 Prompting That Works Role, context, task, format, examples • Prompt duel game

Prompting is the same skill as briefing a smart new hire on their first day: they are capable, they know nothing about your situation, and they will fill every gap you leave with a guess. The quality of your output is mostly the quality of your brief.

The five-part brief

Role: who should the model be? "You are a CFO reviewing a budget" activates different patterns than "you are a copywriter".
Context: the situation and the raw material. Paste the actual document, data, or thread.
Task: one specific verb. Summarize, rank, rewrite, extract, critique. Not "help me with".
Format: exactly what the output should look like: five bullets, a table, 120 words, JSON.
Examples: one or two samples of what "good" looks like. This is called few-shot prompting, and it is the single most powerful trick on this list.

Why bother? In the Harvard/BCG field experiment, consultants using GPT-4 finished 12.2% more tasks, 25.1% faster, at 40% higher rated quality, on tasks within the model's competence. The gains went to people who directed the tool well. Dell'Acqua et al., "Navigating the Jagged Technological Frontier", HBS Working Paper (2023)

Game 2

Prompt Duel: Pick the Stronger Brief

Four rounds. Each shows two prompts aimed at the same job. Pick the one that will reliably produce better output, then read why.

Practical patterns to steal

Meeting notes: "Here is a raw transcript. Produce: decisions made, action items with owner and deadline, open questions. Flag anything ambiguous rather than guessing."
Email: "Rewrite my draft below. Keep my facts, cut it to 120 words, professional but warm, end with one clear ask."
Analysis: "Here is a CSV of monthly sales. List the three biggest changes versus the prior year, with the numbers. Then list what data you would need to explain each change. Do not speculate beyond the data."
Learning: "Explain X like I run a business and have no math background. Then give me the one equation that matters and walk through it once."

Stop and think "Do not speculate beyond the data" appears in the analysis pattern above. Given what you learned in Module 1 about what an LLM fundamentally does, why is that line doing so much work?

6 Hallucinations: When Plausible Beats True Why models invent facts and how to catch them • Spot-the-fabrication game

A hallucination is the machine doing exactly what it was built to do: produce the most plausible next token. When the truth is well represented in training data, plausible and true coincide. When it is not, the model produces something that sounds right, with the same confident tone either way. The model has no internal signal for "I am making this up".

This is why the danger zone is specifics: citations, case numbers, statistics, URLs, prices, names, dates. These look like facts, pattern-match like facts, and are the easiest things for the model to fabricate fluently.

Two cases every professional should know. In Mata v. Avianca (2023), New York lawyers filed a brief containing six court cases that ChatGPT invented. The court fined them $5,000 and the story went global. In 2024, a Canadian tribunal ruled Air Canada liable after its website chatbot told a grieving passenger a bereavement discount could be claimed after the flight. The airline argued the chatbot was "responsible for its own actions". The tribunal disagreed: your AI's words are your company's words. Mata v. Avianca, S.D.N.Y. (2023); Moffatt v. Air Canada, BC Civil Resolution Tribunal (2024)

Game 3

Spot the Fabrication

Each round shows an AI answer containing three claims. One is the kind most likely to be fabricated. Click the claim you would verify first. Train your eye for the danger zone.

The verification rule

Scale your checking to stakes times specificity:

Low stakes, low specificity (brainstorm, tone rewrite): read it and go.
Any citation, number, quote, or legal/medical/financial claim that leaves your desk: verify against a primary source. Not another chatbot, the source.
High stakes (court filings, financial reports, public statements): the LLM drafts, a named human owns every fact. Write that into your process, not your hopes.

Stop and think Why is asking the same model "are you sure?" not verification? Answer using the word "plausible".

7 Safe and Secure: LLMs With Your Company's Data Data leakage, tool tiers, prompt injection • Safe-to-paste game

In 2023, Samsung engineers pasted proprietary source code and internal meeting notes into ChatGPT to get quick help. The data left the building, and Samsung banned generative AI tools company-wide. The lesson: treat a prompt as a data transfer, because that is exactly what it is.

Know your tool tier

Consumer tools (free chatbots, default settings): your inputs may be stored and used to improve the service. Assume anything you paste could be seen by the provider. Public or non-sensitive content only.
Business/enterprise tiers: contracts typically promise no training on your data, plus admin controls and retention settings. This is what your confidential work belongs in, if your company has approved it.
Your company's own deployment (private or API-based): the strongest option, governed by your own security team.

The exact policies differ by vendor and change over time, which is the point: an AI Boss checks the current data terms of the specific tool, and asks IT which tier the company has, before pasting anything sensitive.

Game 4

Safe to Paste?

You are using a free consumer chatbot with default settings. For each item, decide: safe to paste, or not. Assume no special agreement with the vendor.

Prompt injection: an attack where malicious instructions are hidden inside content the model reads (an email, a webpage, a resume), hijacking it into ignoring your instructions. It is ranked the #1 risk for LLM applications by OWASP. The defense mindset: anything the model reads can try to steer it, so limit what a model connected to real systems is allowed to do. Course 03 drills this with a full lab.

The AI Boss data rules

Strip secrets before you paste: names, keys, credentials, personal data. The redacted version usually gets an equally good answer.
Match the sensitivity of the data to the tier of the tool. When unsure, ask before you paste, not after.
Never paste other people's personal information into a consumer tool. Their privacy is not yours to spend.
Follow your company's AI policy. If there is none, propose one. That is what a boss does.

LLMs for Work

What You Will Be Able to Do

What the Research Says

The Seven Modules