What Large Language Models Actually Are (No PhD Required)

Sofia: I want to start with a question I get constantly: what is an LLM actually doing when it responds to you?

Sofia: Let me try the most honest non-technical answer I know. An LLM is a very sophisticated pattern-completion machine. It has read a significant fraction of the written internet — books, articles, code, conversations — and learned the statistical relationships between words and concepts. When you write it a message, it's not looking up an answer. It's completing the pattern: "given everything I've ever read, what words most plausibly come after this sequence?"

Sofia: That sounds simple. So why does it seem intelligent?

Sofia: Because intelligence might be, at some level, sophisticated pattern completion. When you answer a question, you're also doing something like pattern matching against your memories and knowledge. The difference is we don't fully understand why our pattern matching feels like thinking. We're suspicious of the machine's pattern matching because we can see the mechanism.

Sofia: But it hallucinates. It makes things up.

Sofia: Yes. And this follows directly from the mechanism. When the model encounters a question where it doesn't have strong statistical evidence for an answer, it still completes the pattern — it generates the words that most plausibly continue the sentence. Sometimes those words are factually wrong. The model has no internal flag for "I don't know this." That's being added to newer models, but it's not native to the architecture.