Large Language Models (LLMs)

KEY TAKEAWAY

An LLM is an AI that learned language by reading an enormous chunk of the internet — and can now write, reason and converse.

💡 EVERYDAY ANALOGY — BEFORE THE TECH

📚

An LLM is like someone who read the entire internet and has exceptional pattern memory

Imagine someone who read every book, article, Wikipedia page and forum post ever written. They didn't memorise everything word for word, but they absorbed patterns: how arguments are structured, how questions are answered, how stories flow. When you ask them something, they generate an answer based on all those patterns. That's an LLM.

👁️ Visual Explanation

Diagram: Text input → tokenization → LLM processes context → predicts next token → generates response

🌍 In Plain English

LLMs like ChatGPT, Claude and Gemini were trained on trillions of words. They learned the statistical patterns of language so deeply that they can write, explain, reason, translate, summarise and code. They predict the most likely next word given everything said so far.

Real-World Example

When you ask Claude "explain quantum physics to a 10-year-old", it draws on patterns from countless explanations, children's books and science articles to construct a new explanation tailored to your request.

🎓 For Students

Ask an LLM to explain a difficult concept at three different levels — basic, intermediate and expert — until the explanation clicks for you.

💼 For Professionals

Legal professionals use LLMs to summarise 200-page contracts into 2-page summaries, flagging the most important clauses.

💬 Example Prompt to Try

"Explain the French Revolution as if you're telling a 15-year-old who's never studied history."

🔧 Technical DepthOptional — expand when ready

LLMs are transformer-based neural networks with billions of parameters. Training involves predicting masked tokens across massive datasets. The key innovation — the attention mechanism — lets the model weigh which words in context are most relevant for each next-word prediction. GPT-4 has an estimated 1 trillion parameters; Claude 3.5 was trained on similar scales.

Key Terms

ParametersThe millions or billions of numerical values that store everything the model "knows"

Context windowHow much text an LLM can "see" at once — its working memory

TransformerThe neural network architecture that powers modern LLMs, invented at Google in 2017

⚠️ Common Misconception

LLMs don't "know" facts the way you know your own name. They learned patterns. They can generate confident-sounding wrong answers — called hallucinations.