The technical LLM definition: a large language model is a Transformer neural network trained on broad text corpora to model the probability of the next token given everything before it. At sufficient scale, Brown et al. (2020) showed that this single objective yields models able to perform many tasks they were never explicitly trained on, simply by being prompted — the in-context learning behaviour that defines the modern LLM.
ChatGPT, Claude, Gemini and Grok are all built on LLMs. They differ in training data, scale, alignment and tooling, which is why their answers to the same question can diverge.
What is an LLM? The meaning in plain terms
In everyday terms, the LLM meaning is simple: a large language model is a program that has read an enormous amount of text and learned the statistical patterns of language. "Large" refers both to that training data and to the model's size — modern LLMs have billions of parameters or more, the adjustable weights that store what was learned. Most labs no longer publish exact sizes for their flagship models.
When you ask an LLM a question, it does not look the answer up. It writes a reply one token at a time, each time choosing a likely continuation of everything so far. That is why LLMs are fluent and flexible, and also why they can sound confident while being wrong.
How a large language model is trained
Training usually happens in stages, each building on the last:
- 1Pre-training. Next-token prediction on huge text collections teaches the model language, facts and patterns of reasoning.
- 2Assistant training. Instruction tuning — a form of fine-tuning on example conversations — turns the raw predictor into an assistant that follows requests.
- 3
What LLMs can and can't do
LLMs handle a wide range of language work:
- Drafting, editing and rewriting text in any tone
- Summarising long documents and translating between languages
- Answering questions and explaining concepts step by step
- Writing, reviewing and explaining code
- Working through multi-step reasoning problems
Their limits matter just as much: they can hallucinate facts, their built-in knowledge stops at a training cutoff, and they can only consider as much text as fits in their context window. Retrieval (RAG), web search and cross-checking answers against a second model help with all three.
Why it matters at MultipleChat
No single LLM is best at everything. MultipleChat runs ChatGPT, Claude, Gemini, Grok and Perplexity on the same prompt side by side, so you can compare their answers and let one model's strengths cover another's blind spots. See how the comparison works.