How LLMs work (a deeper explanation, still in plain English)
If you enjoyed Module 1 and want a little more detail, this page is for you. You will learn what 'training' means, why AI can sound confident, and what words like 'tokens' and 'context window' actually mean.
No maths. No coding. Just useful understanding.
Training: how the model learns patterns in language
An LLM is trained on a very large collection of text. During training, it learns patterns: which words often follow other words, how sentences are structured, and how information is typically expressed.
A simple way to imagine it is: the model practises guessing the next piece of text, over and over, millions (or billions) of times. Over time, it becomes very good at producing language that looks like human writing.
Prediction, not 'knowing'
When you ask a question, the model does not 'open a book' in the human sense. It generates an answer by predicting what text would be most appropriate given the conversation.
Why AI can be wrong (even when it sounds certain)
AI can be wrong for several reasons:
- It misunderstood your question (missing context).
- It is guessing details that were not provided.
- It is mixing up similar concepts (especially names, dates, or numbers).
- It is producing a 'plausible-sounding' answer instead of an evidence-based one.
That is why good prompts and verification habits matter. The AI is not offended by follow-up questions - it expects them.
Instruction-following and 'helpfulness'
Modern chat-style AI models are designed to follow instructions. That is why the prompt recipe works so well.
In addition to basic training, models are often tuned to be more helpful in conversation. This can include learning to:
- answer in a friendly tone
- follow directions like 'use dot points' or 'keep it short'
- refuse unsafe requests
- ask clarifying questions
Tokens: the AI's way of counting text
AI systems often measure text in tokens. A token is a chunk of text - sometimes a whole word, sometimes part of a word.
You do not need to count tokens manually. The practical takeaway is: more text means more 'budget' used.
- Long conversations use more tokens.
- Long documents use more tokens.
- Some plans allow a larger context window (more text held in mind).
Context window: why the AI 'forgets'
The context window is the amount of text the model can consider at once when generating the next reply.
If you exceed it, older parts of the conversation may be dropped or summarised. This can look like the AI 'forgetting'.
Temperature: why the same prompt can give different answers
Some models include randomness in how they generate text. This can help creativity and prevent repetitive answers.
That means: if you ask the exact same question twice, you may get slightly different wording or even different suggestions.
Why the AI sometimes refuses
AI tools have safety systems. They may refuse requests that involve harmful instructions, illegal activity, or sensitive personal information.
If it refuses and you think it misunderstood, rephrase your goal in a safe way. Example:
I am not asking for anything unsafe. I want general, high-level information for educational purposes. Please keep it safe and legal.The best mental model for beginners
Here is a calm, practical way to think about chat AI:
- It is a drafting and organising tool.
- It is excellent at writing and rephrasing.
- It is helpful for learning and planning.
- It can be wrong about facts - so you verify what matters.
- You are always in charge of decisions.
Mini exercise: 'teach me gently'
Copy this prompt into ChatGPT or Gemini:
Teach me about [TOPIC] in simple terms. Start with a 5-dot-point overview. Then ask me 3 questions to check my understanding. If I get something wrong, correct me kindly.