---
id: con-large-language-model
schema: concept/1
layer: garden
status: established
name: Large language model
aka: ["LLM", "the model", "foundation model", "GPT", "frontier model"]
summary: >
  The engine behind AI chat tools: a system trained on vast amounts of text to predict what words
  should come next, which turns out to be enough to write, summarise, translate and reason in useful ways.
first_seen: "2020"
relations:
  builds_on: [con-generative-ai]
  evolved_from: []
  related: []
  parallels:
    - idea: "predictive text on a phone keyboard"
      field: "everyday technology"
      note: "The same next-word guessing, trained on vastly more text and much better at keeping a thread."
implementations:
  - { provider: openai, name: "GPT models" }
  - { provider: anthropic, name: "Claude models" }
  - { provider: google, name: "Gemini models" }
  - { provider: meta, name: "Llama models" }
  - { provider: xai, name: "Grok models" }
  - { provider: huggingface, name: "Open models on HuggingChat", resource: res-huggingface-huggingchat }
competencies: [cmp-check-the-result]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "Knowing the engine predicts likely text explains hallucination, sycophancy and why context matters."
---

A large language model, or LLM, is the part of an AI tool that produces the words. It was trained by reading a huge amount of text and learning, over and over, to guess the next word. Trained on enough text, that guessing becomes good enough to draft, summarise and translate.

Two facts about LLMs explain a lot of everyday behaviour. First, they produce what is likely, which is not always what is true. Second, they only know what was in their training material plus whatever you put in front of them in the conversation. They do not look things up unless the product gives them a search tool.

Chat apps such as ChatGPT, Claude and Gemini wrap a model in an interface, plus extras like file uploads, search and memory. The app and the model are different things, and one app can offer several models.

## Example

You ask for "the latest figures" on your industry. The model gives tidy numbers with no source. Knowing it predicts likely text, you turn on web search or paste in the actual report, and ask again.

## How it connects

- It is the engine inside most **generative AI** tools.
- Its guessing nature is the root of **hallucination** and **sycophancy**.
- Different models suit different jobs, which is why **model choice** matters.

## Try it

Write down what you expect your AI tool to get wrong about your company. Then ask it: "What do you know about my company, and how confident are you?" Check its answer against what you know.