What are LLMs and how do they actually work, explained simply?

The honest middle version: An LLM is a next-word predictor trained at absurd scale. It read most of the internet and learned, statistically, which token (word fragment) tends to follow a given sequence. When it 'answers' you, it's repeatedly asking 'given everything so far, what's the most plausible next token?' — thousands of times per response. The non-obvious part: to predict the next word REALLY well across all of human text, the model was forced to internally learn grammar, facts, reasoning patterns, coding conventions — because you can't predict the next token of a physics explanation without compressing some physics. That's why 'fancy autocomplete' undersells it. Chat behavior comes from a second phase: fine-tuning on human-rated conversations so it prefers helpful, instruction-following continuations. And hallucinations make sense in this frame — it always produces a PLAUSIBLE next token, and plausible isn't the same as true.