What is a large language model and how is it different from a search engine?

Fundamentally different. A search engine finds existing documents and points you at them. A language model generates new text based on patterns it learned. One retrieves, one composes — and that difference explains where each is trustworthy. **Search engine**: maintains an index of web pages. You give it terms, it ranks pages by relevance and authority, and returns links. The content it shows you was written by an identifiable someone, at a knowable time, and you can inspect the source. What it can't do is synthesise — you get ten links and do the integration yourself. **Language model**: has no index and isn't looking anything up. Its 'knowledge' is compressed into parameters during training. When you ask something, it generates a response one token at a time. What it can do is synthesise, adapt to your specific situation, rewrite, translate, summarise and explain at whatever level you ask. What it can't do is guarantee accuracy or show you where anything came from. When to use which: Use **search** for: anything current (news, prices, availability, releases), authoritative sources (official docs, government or medical guidance), verifying a specific fact, finding a real citation, and anything where you need to see who said it. Use **an LLM** for: explaining a concept at your level, summarising text you provide, drafting and rewriting, brainstorming, code generation and debugging, translation, and questions too specific or oddly-shaped to search for — 'given these three constraints, which approach makes sense and why' has no page to find. The genuine difference in failure mode is what matters most: a search engine failing gives you irrelevant results, which is obvious. A language model failing gives you a fluent, confident, wrong answer, which is not obvious at all. That asymmetry is why 'just use it like Google' is risky for factual questions. The blurring: many products now combine both — the model performs a search, reads the results, and writes an answer with citations. That's meaningfully more reliable for factual queries because the output is grounded in retrieved text, and you can click through. When accuracy matters, prefer a mode that shows sources, and actually click them — grounded answers still misread their sources sometimes. The practical habit: ask 'do I need a fact, or do I need thinking done?' Facts go to search or a grounded tool. Thinking, drafting, explaining and transforming go to the model.