What Is an LLM? Large Language Models Explained Without the Hype

Every AI coding tool — Claude Code, Cursor, Lovable, Bolt, v0 — is built on a large language model, or LLM. You don't need to understand the maths to use one well. But a clear picture of what an LLM is explains most of what AI tools are good at, and most of what goes wrong.

The one-sentence version

An LLM is a program that, given some text, predicts what text should come next — trained on so much writing that its predictions capture grammar, facts, reasoning patterns, and code.

How it's made

1. Pre-training. The model is shown an enormous amount of text: books, websites, documentation, and a great deal of public source code. Over and over, it practises predicting the next piece of text, and adjusts billions of internal numbers (its parameters) to get better at it. To predict well, it has to pick up how language works, how facts relate, and how code is structured.

2. Post-training. The raw model is then trained further to be a helpful assistant: to follow instructions, answer questions, refuse harmful requests, and — for coding models — to write, test, and fix code across many steps. This stage involves human feedback and a lot of practice on real tasks.

The result is the model: a very large file of numbers that, given text in, produces text out.

Key terms

Model — the trained LLM itself. Claude Sonnet, Claude Opus, and so on are different models: bigger ones are generally more capable, slower, and more expensive.
Prompt — the text you give it: your question or instructions. (How to Prompt AI Coding Tools.)
Tokens — the chunks of text it reads and writes, and what usage is measured in. (What Are Tokens in AI?.)
Context window — how much text it can consider at once. (What Is a Context Window?.)
Temperature — a setting controlling randomness. Low gives more predictable answers; high gives more varied ones.
Training cutoff — the date its training data ends. It doesn't know about anything after, unless you tell it or it can search.

Why LLMs are good at code

Code is an ideal fit:

There's a huge amount of public code to learn from, often alongside explanations.
Code has strict, learnable patterns.
Code can be checked — run it, test it, see the error — which lets coding agents correct themselves. That feedback loop is what makes tools like Claude Code much more reliable than a chatbot pasting code. (What Is an AI Coding Agent?.)

Why they make things up

An LLM produces plausible text. Usually plausible and true are the same. Sometimes they aren't, and the model states something false with complete confidence — a hallucination.

In coding, that looks like:

Calling a function or setting that doesn't exist in the library you're using.
Using an outdated version of an API it saw a lot of in training.
Importing a package that doesn't exist — which attackers have learned to exploit. (AI Hallucinated a Package.)
Claiming it ran the tests, or that something works, without checking.

The defence isn't to distrust AI entirely. It's to make it verify: run the code, run the tests, check the docs. Agents that can execute code in a real environment catch most of their own hallucinations, because the error message doesn't lie.

What an LLM is not

Not a database. It doesn't look facts up; it reconstructs them. Exact figures, prices, and version numbers may be wrong or out of date.
Not a memory. Each request starts fresh. It "remembers" a conversation only because the whole conversation is sent back each time.
Not deterministic. Ask twice, get two different answers. Useful for ideas; a reason to test for correctness.
Not you. It doesn't know your business, your users, or why you made a decision unless you tell it.

What this means when you build with AI

Give it context. It only knows what's in the prompt and the files it reads. Describe the goal, the constraints, and what "done" looks like.
Let it check its work. Tools that run your code and tests beat tools that only write it.
Verify anything important — security, payments, data handling — yourself or with tests. (Review AI-Generated Pull Requests.)
Keep sessions focused. Quality drops as context fills up.

EasySpawn gives Claude Code a real, persistent environment to run your app, tests, and database — so the model can check its own output against reality instead of guessing. See how it works or join the waitlist.