How well do you actually understand AI?

Models don't see words, they see tokens: sub-word chunks averaging about four characters in English. It matters because you're billed per token, and because context limits are counted in tokens, not pages.

The dangerous part isn't the error — it's the confidence. A model has no internal signal for “I don't know”, so a fabricated citation reads exactly like a real one. That's why sourcing habits matter more than model choice.

Everything competes for the same window: your instructions, the pasted document, the conversation so far, and the answer. When a long chat starts “forgetting” the beginning, that's the window filling up — not the model losing interest.

RAG is retrieval plus a bigger prompt — nothing about the model changes. That's the appeal: your documents stay yours, updating them is instant, and the model can cite what it was given.

One concrete example beats any amount of adjectives. “Be accurate” gives the model nothing to act on; a sample of the format, length, and tone you want gives it a target to match.

Built-in knowledge is frozen at the cutoff. Most tools paper over this with live web search — which is why the same question can get a current answer in one product and a stale one in another using the same model.

The rule of thumb: the more precise and checkable a claim is, the more likely a fabrication is to look convincing and cause damage. Drafts and brainstorms fail harmlessly; a made-up case number doesn't.

Fine-tuning teaches behaviour and style, not facts. Teams reach for it and then discover their real problem was knowledge — which RAG solves better, faster, and cheaper.

Low temperature makes the model pick the likeliest next token, so output is consistent and repetitive. High temperature loosens that. It does not make anything truer — a confident falsehood at temperature 0 is still a falsehood.

Free and consumer tiers often reserve the right to keep your inputs and train on them, and “delete” rarely means gone from their systems. Treat anything you paste into a public chatbot as potentially public — use an enterprise tier with data controls, or a private deployment, for anything sensitive.