A large language model is the engine behind tools like Claude, ChatGPT, and Codex. The useful shift is to stop picturing a fact database and start picturing a very well-read assistant.
Think about that assistant across the desk from you. It has read manuals, emails, contracts, code, and a lot of messy human writing. Hand it a note and it doesn’t pull one perfect answer from a filing cabinet; it looks at the words in front of it and asks, “Given everything I’ve seen, what should probably come next?” That sounds small, but most professional work is language work: describing a problem, writing instructions, comparing options, drafting a reply. An LLM is good at that because it learned patterns in language at massive scale.
How it shows up
A large language model isn’t a database of facts. It’s a language engine that reasons over the context you give it.
When you use Claude Code or Codex, the LLM is the language engine inside the harness, reading your prompt, looking at files it can access, deciding what steps make sense, and writing back. The model isn’t the whole system. Model plus context plus tools plus permissions becomes the working agent.
Why you care
Most frustration comes from treating it like a smarter Google: you ask for a perfect fact, get a confident answer, and assume it’s broken when it makes something up. Search finds pages; an LLM generates an answer from patterns, instructions, and context. If you want it grounded, give it the right material or connect it to a retrieval system like RAG. This is where tokens and the context window matter: the assistant only works with what fits on the desk, so leave the important file out and it may still write smoothly while guessing more than you want. The quality of the work usually starts with the quality of what you put on the assistant’s desk.