How do large language models actually 'think' when they answer a question?

They don't think the way you do. Not even close. A large language model, or LLM, is more like an incredibly fast pattern-matching machine. When you type a question, it doesn't pause to ponder or understand. It breaks your words into tiny pieces called tokens and then predicts, one token at a time, what the most likely next token should be. That's it. A really fancy autocomplete. Imagine you're typing 'The capital of France is...' on your phone. Your phone's keyboard suggests 'Paris' because it's seen that pattern millions of times. An LLM does the same thing, but with a much bigger memory and a lot more math. It's been trained on a massive chunk of the internet, so it's learned that when a sentence starts a certain way, specific words usually follow. When you ask it to explain quantum physics, it's not accessing a hidden library of concepts. It's generating a sequence of words that, based on its training, sounds like a valid explanation of quantum physics. This is why they can sometimes sound incredibly smart and then fail at a simple logic puzzle. They're masters of language, not masters of thought. A useful tip: because of this, the more specific and structured your prompt is, the better the output. If you just say 'write a story,' the model has to guess from a billion possible next tokens. If you say 'write a 200-word mystery story set in a lighthouse, in the style of Agatha Christie,' you've massively narrowed down the probable token paths. You're not asking it to think harder. You're giving its pattern-matching engine a much clearer map to follow.