SEPTEMBER 2026 · THE TOOL DESKPractical technology. No theatre.

Practical guide · verified against the real thing

How large language models actually work (in plain English, no maths)

In one line: They are not databases and not minds — they are extremely good guessers of the next word. Understanding that one idea explains both their brilliance and their nonsense.

Most confusion about AI — both the fear and the overselling — comes from not knowing what the thing actually is. A large language model is not a database it looks things up in, and not a mind that understands you. It is a system trained to do one thing extraordinarily well: given the words so far, predict the most likely next word. Everything it does, good and bad, follows from that single mechanic.

Next-word prediction, at enormous scale

During training the model reads a vast amount of text and learns statistical patterns — which words tend to follow which, in which contexts. Ask it something and it generates an answer one token at a time, each chosen because it is likely to follow the last. There is no lookup table of facts and no belief underneath; there is a deeply learned sense of what plausible text looks like. That is why it can write a poem it has never seen and also state a falsehood with total confidence — both are just "what would likely come next."

Why this explains the brilliance and the nonsense

The same mechanic that makes a model fluent and useful makes it unreliable as a source of fact. It is optimised for plausibility, not truth, so when it does not "know" something it will still produce something that fits the pattern — the failure explored in why AI hallucinates. It also explains why phrasing matters so much: you are steering a predictor, which is the whole subject of writing better prompts.

What it is genuinely good at, and what it is not

Because it models language, it excels at anything linguistic: drafting, summarising, rephrasing, explaining, brainstorming, translating, and turning messy text into structure. It is weak at precise recall of specific facts, arithmetic it cannot reason through, and knowing the limits of its own knowledge. Judging a tool against what it actually is — a powerful language engine, not an oracle — is the same scepticism the desk applies in useful vs hype, and it is what lets you use these tools well instead of being impressed or burned by them.

Next

Related on this desk.