How Large Language Models Actually Work — In Plain English

Type: Explainer / Trust Builder

Let's start with the honest version of what a large language model is.

It's a very, very good autocomplete.

That's it. That's the core idea. If you've ever had your phone suggest the next word in a text message, you've seen the basic principle. Large language models — the technology behind ChatGPT, Claude, Gemini, and others — are doing exactly the same thing. Just at a scale and sophistication that makes the result feel like magic.

Here's how it works.

Step 1: Training on Enormous Amounts of Text

Before a language model can do anything useful, it has to learn from data. A lot of data. We're talking about a significant portion of the written internet — books, articles, code, forums, Wikipedia, academic papers — billions and billions of words.

The model reads all of this (in a very technical sense of 'reads') and learns patterns. Not rules. Patterns.

It doesn't learn 'in English, sentences follow Subject-Verb-Object order.' It learns that after the word 'the', certain types of words are much more likely to appear. After 'the capital of France is', the word 'Paris' is overwhelmingly probable. After 'she opened the', the model has learned that 'door', 'window', 'letter', and 'box' are all plausible — and it's learned the contexts in which each is more or less likely.

This happens billions of times across billions of examples. The model adjusts its internal weights — think of them as a vast web of numerical settings — until it gets better and better at predicting what comes next.

Step 2: The Scale Changes Everything

Here's where it gets interesting.

Researchers discovered something surprising: when you scale this process up — more data, more computing power, more parameters — the model doesn't just get better at predicting text. It starts developing what look like emergent capabilities.

It starts being able to reason through problems step by step. It starts translating between languages it was never explicitly taught to translate. It starts writing code, solving maths problems, summarising documents, and explaining concepts — all from the same underlying mechanism of predicting what comes next.

Nobody fully understands why this happens. It's one of the genuinely mysterious things about this technology. But it's reliably reproducible: past a certain scale, the models become dramatically more capable.

Step 3: The Model Generates, Not Searches

This is the most important thing to understand, and the source of most confusion about how these systems work.

When you ask ChatGPT a question, it does not search a database for the answer. It generates a response, word by word, based on patterns learned during training. It produces the text that its model predicts would be a good continuation of the conversation.

This is why it can write poetry, explain quantum physics in plain English, and draft a cover letter. It's pattern-matching and completion at extraordinary scale — not retrieval.

It's also why it sometimes confidently says things that are wrong. More on that in a future article.

What LLMs Are Good At (And Where They Fall Apart)

Good at: Writing and editing. Summarising long documents. Explaining concepts. Brainstorming. Translating. Writing and debugging code. Answering questions where the answer exists in their training data.

Struggle with: Anything requiring real-time information (unless connected to search). Precise arithmetic. Consistent logical reasoning over many steps. Knowing what they don't know. Being factually reliable on obscure topics.

The key insight is that these are not databases. They're not search engines. They're sophisticated pattern-completion systems that happen to have absorbed enough human knowledge to be genuinely useful — and genuinely dangerous when misapplied.

The Common Myths, Quickly Debunked

Myth: LLMs understand language the way humans do. Reality: They process statistical patterns. Whether that counts as 'understanding' is a philosophical debate, not a technical one.

Myth: They're just copying and pasting from the internet. Reality: They're generating novel text based on learned patterns. The output is genuinely new, even if the training data was borrowed.

Myth: Bigger always means better. Reality: Efficiency matters too. Some smaller, well-tuned models outperform larger ones on specific tasks.

Next week, we're going to get practical: real people using AI to claw back hours from their week. But first — does this explanation make sense? What's still confusing? Tell me in the comments.