Large Language Models (LLMs)
The engine behind ChatGPT, Claude, and Gemini — explained simply
Contents
What is an LLM?
A Large Language Model is an AI system trained on vast amounts of text — books, websites, code, and more — to understand and generate human language. Think of it as a very well-read assistant that has processed more text than any human could read in a thousand lifetimes.
The "large" refers to the number of parameters (essentially, learned connections) — models like GPT-4 have hundreds of billions of them. These parameters capture patterns in language: grammar, facts, reasoning styles, and even nuance.
How does it actually work?
LLMs work by predicting the next word (technically, "token") in a sequence. When you type "The capital of India is", the model calculates probabilities for every possible next word and picks the most likely one — "New Delhi".
This sounds simple, but doing it billions of times across complex text requires understanding context, logic, and language at a deep level. The magic is in scale: train on enough data with enough parameters, and this next-word prediction starts looking a lot like reasoning.
Key terms to know
Tokens: The chunks text is broken into (roughly ¾ of a word). "ChatGPT" = 2 tokens. This matters because models have a token limit (context window).
Context window: How much text the model can "see" at once. GPT-4 handles 128K tokens; Claude 3 handles 200K — roughly 150,000 words.
Temperature: Controls randomness. Low temperature (0.1) = consistent, predictable. High temperature (0.9) = creative, varied.
Parameters: The learned values inside the model. More parameters ≠ always better, but generally enables more capability.
Popular LLMs and their strengths
GPT-4o (OpenAI): Best all-rounder. Great at coding, multimodal (text + images + voice), widely integrated.
Claude 3.5 Sonnet (Anthropic): Best for long documents and nuanced writing. Follows complex instructions carefully.
Gemini 2.0 (Google): Fastest and free. Best when using Google Workspace. Strong at real-time information.
Llama 3 (Meta): Open-source. Free to run locally. Great for developers who want privacy or customisation.
Mistral: Lightweight, fast, open-source. Good for self-hosting on modest hardware.
What LLMs can and cannot do
Can do: Write, summarise, translate, answer questions, generate code, analyse documents, have conversations, reason through problems.
Cannot do (reliably): Access the internet (unless tool-enabled), perform precise maths, remember past conversations (without memory features), guarantee factual accuracy.
The most important thing to understand: LLMs can be confidently wrong. Always verify important facts from authoritative sources.