How AI Works
A working mental model of what modern AI systems actually do.
Move past the buzzwords to how the systems really work. Explain a concept in your own words and a tutor probes exactly where your model breaks, in plain language, no code required.
Fluent language and checked facts are different things
A language model generates text in small pieces called tokens, using patterns learned during training and the context supplied with your question. That can produce useful explanations, but generating an answer does not automatically check each claim against a reliable source.
Think of an assistant asked about a company's refund policy. It might produce a familiar-sounding answer without having that company's policy. Giving it the actual document supplies evidence to work from; checking that its answer follows the document is a further step.
Read the worked example →Understand the concepts
- AI hallucinations: why confident answers can be wrong
Understand AI hallucinations with a worked example: why fluent answers can be false, how retrieval helps, and how to check the evidence behind a claim.
- How do LLMs work? Next-token prediction explained
Follow a worked example of next-token prediction. Learn what tokens, training and context do, and why generating an answer is different from looking it up.
- How do AI agents work? Tools, feedback and the next step
Follow an AI agent through a practical task. Understand the model, tools and feedback loop, how agents differ from workflows, and when they should stop.
- Context window vs memory: why AI forgets your conversation
Separate an AI's context window, saved memory and trained knowledge. Follow a worked example to understand why a chatbot can remember one detail and miss another.
- What are AI evals? Testing answers that can change each time
Understand AI evals with a fictional support assistant: test cases, grading rules, useful paraphrases, critical failures and why an average score is not enough.
What you’ll learn
- Why a model can sound confident and still be wrong
- What context windows and memory really are
- How retrieval, fine-tuning and prompting solve different problems
- What turns a model into an agent, and why that is riskier
Concepts you’ll explore
Try a question
An AI gives a confident, detailed answer that turns out to be wrong. How can it sound convincing without knowing whether the answer is true?
Answer this question →One free answer, instant feedback from an AI tutor, no account needed to start.