TessorioLearning Academy

Sign in Sign up free

Tessorio Library · Free

How AI Actually Works: From Neurons to ChatGPT

artificial intelligencemachine learningdigital literacy17 lessons · Adult learner (also suitable for teens, ages 14+); no prior math or coding

Unit 1

Inside a Neural Network: Numbers In, Guess Out

One Neuron: Multiply, Add the Bias, Then ReLU

10:02You will compute a single artificial neuron's output by hand: multiply each input by its weight, add them up, add the bias, and pass the result through ReLU.

Stacking Neurons into Layers: Which Numbers Are Given and Which Are Learned

09:45You will trace a number all the way through a two-layer network by hand and say, for every number involved, whether it came from the data or was learned during training.

Guess, Measure the Error, Nudge the Weights

17:04You will walk through the training loop and justify whether a specific weight should go up or down to shrink a given error, then say what happens on the next pass.

Unit 2

Learning Language by Guessing the Next Word

From Sentence to Training Examples: Tokens and the Hidden Answer Key

15:57You can split a sentence into tokens, slide through it to build next-token training examples, and say exactly how many examples a 12-token sentence gives with no human labeller.

One Token at a Time: The Probability List and the Loop

07:10You can describe how a model turns a prompt into a ranked list of probable next tokens, picks one, glues it on, and runs the whole thing again to build a sentence.

Temperature: Always the Safe Word, or Roll the Dice

07:16You can predict which token gets chosen at temperature 0 versus a high temperature, and explain why the same prompt twice gives two different answers without the model changing its mind or learning from you.

Unit 3

From Raw Model to Helpful Chatbot

Why a Raw Model Answers a Question With More Questions

07:01You'll be able to explain that a pretrained model only continues text, and describe how instruction tuning — training on thousands of written request-and-reply examples — teaches it to answer instead.

Thumbs Up, Thumbs Down, and the Hidden Instructions

10:27You'll be able to describe how human preference ratings and a system prompt shape a model's tone and limits, and decide which of the three stages is responsible for fixing a given misbehaviour.

Unit 4

Peeking Inside: Attention and the Transformer

Why "It" Means the Suitcase: Attention Scores in Plain Words

08:51You'll explain, step by step, how a model decides which earlier words each word should look at, and use that to work out what "it" refers to in an ambiguous sentence.

Every Word Looks at Every Word: Heads and Layers

07:27You'll describe how several attention heads look for different kinds of links and how stacking layers turns those links into a richer meaning for each word.

A 12,000-Token Report in an 8,000-Token Window

07:05You'll predict exactly what a model does when a document is longer than its context window and choose a sensible fix, such as splitting the text into chunks and summarising each.

Unit 5

Why It Makes Things Up and Gets Things Wrong

Confident Fiction: Invented Citations and Yesterday's Facts

07:28You can explain why a next-word predictor produces smooth-sounding fake sources, and tell that kind of made-up detail apart from a fact that was simply true when training stopped.

Fuzzy Questions, Skewed Data, and 12 Per Cent of 250

10:17You can spot when a bad answer came from your own ambiguous wording, from lopsided training data, or from a step-by-step slip like calling 12 per cent of 250 forty instead of thirty.

Unit 6

Prompting Well and Checking the Answer

Four Things Every Weak Prompt Is Missing: Role, Context, Format, Examples

10:16You can turn a vague request like "write about marketing" into a prompt that states a role, the context, the exact output format, and a worked example.

Read the Output, Fix One Lever: Revising a Prompt That Didn't Work

07:03You can diagnose which of the four ingredients a disappointing answer is missing, change just that one thing, and point to the before-and-after differences that prove the rewrite worked.

Check First, Then Decide: Trust, Verify Further, or Discard

10:23You can split an AI answer into separate claims, rank them by how likely and how costly a mistake would be, check the riskiest ones against a source or method independent of the model, and give a reasoned verdict.

Unit 7

Your Own AI Task, Start to Finish

Run a Real Task and Keep a Working Log

10:30You will take one real task of your own through an AI tool from first prompt to finished output, keeping a one-page log of your prompts, your revisions and why you made them, every factual claim you checked and how, and at least one error you caught.