By the end of this lesson you will be able to
- Describe what a large language model does in one sentence
- Explain why AI is fluent about things it does not know
- Predict which kinds of tasks it will handle well before you try them
Almost every confusing thing AI does — the confident wrong answers, the different reply to the same question, the sudden brilliance followed by a basic mistake — comes from one fact about how it works. Get this fact and the rest stops being mysterious.
The one idea
A large language model is a prediction machine. Given some text, it predicts what text most plausibly comes next. That is the entire operation. Everything else — writing an email, summarising a contract, drafting code — is that same prediction running over and over, each new word feeding back in as part of the input for the next one.
Definition Large language model- A system trained on an enormous amount of text to predict what comes next in a sequence. "Large" refers to the size of the model and the training data, not to its intelligence. Full definition
It is tempting to dismiss that as trivial. It is not. Predicting the next word well enough, across billions of examples, requires the system to encode a great deal about grammar, structure, tone, common facts, and the shape of a well-formed argument. Fluency is a real achievement. It is just not the same achievement as understanding.
Prompt “The shipment has been ___”
Medium: the default range. Natural-sounding, still on topic.
Why it always sounds certain
Here is the consequence that catches people out. The model is predicting plausible text, and confident text is more plausible than hedging text — because most of what humans write down is written by people who knew what they were talking about. So the model produces confident-sounding output regardless of whether the underlying claim is right.
This is why "it seemed so sure" is not evidence of anything. The model has no separate confidence signal it is choosing to hide from you. Confidence is a property of the writing style, not a report on the truth of the content.
What it is not
Two mental models. The one on the left causes most of the disappointment.
What people assume
- Looks up facts in a database
- Knows when it does not know something
- Reasons step by step toward a conclusion
- Gives the same answer to the same question
- Remembers your previous conversations
What actually happens
- Generates plausible text from statistical patterns
- Produces confident text either way
- Produces text that resembles reasoning
- Samples from a distribution, so answers vary
- Sees only what is in the current context window
None of the right-hand column makes AI useless. It makes it predictable. Once you know the machine generates rather than retrieves, you know to verify anything factual, to give it the information rather than assume it has it, and to expect variation.
What follows from this
Three practical rules fall straight out of the prediction idea.
Supply the facts, do not request them
Paste the contract, the numbers, the policy. The model is excellent at working with text you give it and unreliable at recalling text it was trained on. This single habit removes most hallucinations.
Verify anything specific
Names, dates, figures, citations, legal claims. Verify these every time in work that matters. Verify prose and structure much less — that is where the model is genuinely strong.
Expect variation, and use it
The same prompt twice gives two answers. That is a feature when you want options and a problem when you want consistency. Ask for several drafts when you are exploring; pin down the format tightly when you need repeatability.
The questions that come up every time we teach this.
Does the AI understand what it is saying?
Not in the way you mean. It has statistical structure that captures a great deal about how concepts relate, which is why it can handle novel phrasing and analogy. But there is no separate layer that checks a claim against reality. Whether that counts as understanding is a genuinely open philosophical question; for practical purposes, treat it as no.
If it is just predicting text, how can it write working code?
Because code is text with unusually strict patterns, and there is an enormous amount of it in the training data. Prediction over highly structured text works very well. This is also why AI code that looks right can fail in subtle ways — plausibility and correctness are not the same test.
Will this change with better models?
Models get better at prediction, which makes the failures rarer and harder to spot. That is an improvement and also a risk. The underlying mechanism has not changed, so the habits in this lesson still apply.
Key takeaways
- A language model predicts likely next text. That is the whole mechanism.
- Fluency and accuracy come from the same process, so confidence tells you nothing about correctness.
- Give it the information rather than asking it to recall the information.
- Verify specifics — names, numbers, dates, citations. Trust structure and prose far more.