For non-engineers who use AI assistants daily and want an accurate mental model of what happens inside one, so they can predict its behavior instead of being surprised by it.
Level
Basic
Study time
8 hours
Lessons
16 in 6 modules
Exam
20 questions
About this course
Most explanations of large language models stop at "it predicts the next word". That is true but not enough to explain why a model can follow instructions, why it forgets the middle of a long document, why the same question gets different answers, or why newer "reasoning" models pause to think. This course opens the box one layer at a time, using worked examples you can try in any chat assistant and the free tokenizer tools vendors publish.
You will follow a piece of text from tokenizer to embedding vectors, through the attention layers of a transformer, and back out as a probability distribution over the next token. You will see how pretraining on internet-scale text differs from the instruction tuning and preference training that turn a raw model into an assistant, how reasoning models spend extra computation before answering, and how context windows, tools and retrieval extend what a model can do. The course closes by tying each mechanism to a behavior you have seen at work. It ends with an automated exam and a Wirewalk Institute certificate of completion and assessment. Product facts are stated as of the review date and the course is re-reviewed on each major Anthropic or OpenAI model release or every six months; errors can be reported through the contact link on this site.
What you will be able to do
Trace how a sentence is split into tokens and explain why token counts drive cost, limits and some odd mistakes such as miscounting letters
Describe embeddings and attention in plain language and use them to explain how a model relates words across a passage
Distinguish pretraining, supervised instruction tuning and preference training, and explain what each contributes to an assistant's behavior
Explain what reasoning models do differently at answer time and when the extra cost and latency are worth it
Explain how temperature and sampling produce varied answers, and when to expect consistency
Describe context windows, tool use and retrieval, and predict where a model will struggle with long or unfamiliar material
Map common surprising behaviors, such as hallucination, sycophancy and knowledge cutoffs, to the mechanism that causes them
Syllabus
6 modules · 16 lessons · about 8 hours
1
Module 1: From Text to Tokens to Vectors
How a model turns your words into numbers it can work with.3 lessons
Tokenizers in Action: Byte-Pair Encoding with Real Examples
Why Models Miscount Letters and Struggle with Some Languages
Embeddings: Meaning as Position in a Space
2
Module 2: Inside the Transformer
Attention and layers explained with analogies and toy examples, not formulas.3 lessons
Attention: How Every Word Looks at Every Other Word
Stacking Layers: From Grammar to Facts to Tasks
Parameters, Model Size and What "70 Billion" Actually Means
3
Module 3: How a Raw Model Becomes an Assistant
The training stages that turn a text predictor into something that follows instructions.3 lessons
Pretraining: Learning from a Large Slice of the Written World
Instruction Tuning and RLHF: Teaching the Model to Help
Reasoning Models: Spending More Computation Before Answering
4
Module 4: Generating an Answer
What happens between pressing Enter and seeing the reply.3 lessons
Probabilities, Temperature and Why Answers Vary
Context Windows: What the Model Can See at Once
System Prompts, Conversation History and Memory Features
5
Module 5: Extending the Model: Tools and Retrieval
How assistants search, run code and read your files without retraining.2 lessons
Tool Calling: How a Model Asks for a Search or a Calculator
Retrieval-Augmented Generation in Plain Language
6
Module 6: Explaining Behavior from Mechanism
Connect what you now know to the surprises you meet at work.2 lessons
Hallucination, Sycophancy and Knowledge Cutoffs, Explained Mechanically
Reading Vendor Claims: Context Length, Benchmarks and Model Cards
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 40 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by OpenAI, Anthropic, Google, Meta, Mistral AI, or DeepSeek. Product and company names are trademarks of their respective owners.
A free, plain-English introduction taught through one realistic company getting ready for its busiest season. Sort the kinds of AI, see why a chat assistant can sound right and be wrong, decide which tasks it fits and how to measure them, catch its three predictable failures, ask well, and decide what never goes into an AI tool.
A plain-language grounding in artificial intelligence: what AI, machine learning and generative AI actually are, how chat assistants produce text, why they sometimes get things confidently wrong, and how to use them safely for real work.
Work through a fictional Ohio learning venue’s campaign using supplied briefs, review logs, scripts and timeline records. Compare manual production and generative approaches, repair concrete defects, and prepare a rights-aware accessible release recommendation.