Day 1. The Foundations of Generative AI

A generative model does not look anything up. It predicts what comes next.

That is the entire mechanism, and almost every strange behaviour you will meet over the next fourteen days follows from it. The model has read an enormous amount of text and images and has learned, in extraordinary statistical detail, what usually follows what. When you type a question, it is not searching for your answer. It is continuing your text in the way its training suggests text like yours is usually continued.

People resist this because the output feels like understanding. Hold both ideas at once: the mechanism is prediction, and the behaviour that emerges from prediction at this scale is genuinely useful. You do not need to resolve that tension. You need to remember which one is true when the model surprises you.

The written lesson. Read it slowly, it saves you later.

The nested circles

Four terms get used interchangeably in the market and they are not the same thing. Artificial intelligence is the whole field, going back to the 1950s. Machine learning is the subset that learns patterns from data rather than following rules a programmer wrote. Deep learning is the subset of machine learning that uses many-layered neural networks, building up from edges to shapes to faces. Generative AI is what those deep networks do when the task is production rather than classification. A classifier looks at a photo and says cat. A generative model is asked for a cat and produces one that has never existed.

Every circle contains the next. When a vendor says we use AI, they have told you almost nothing.

Tokens, and why the model cannot count letters

The model does not read characters or words. It reads tokens, fragments of roughly three to four characters in English. A common word is one token. An unusual name might be three. Arabic and Malayalam typically consume more tokens per word than English, which is why the same paragraph costs more in those languages.

Ask a model how many letter Rs are in a word and it may get it wrong, not because it cannot count, but because it never saw the letters. It saw two or three fragments, and it is answering from the statistics of how such questions are usually answered.

The context window

Everything the model can see at once, your prompt, the uploaded document, the whole conversation so far, plus its own reply, must fit inside a fixed budget of tokens. That budget is the context window. In a long conversation, early material falls out of view. The model has not forgotten in a human sense, that text is simply no longer in front of it.

Working habit: start a fresh chat when the subject changes. Do not treat one thread as a filing cabinet.

Hallucination, properly understood

A hallucination is fluent, confident output that is not true. First, it is not lying. Lying requires knowing the truth and choosing otherwise. The model is producing plausible continuation, and a citation that does not exist looks statistically like a real citation. Second, and this is the part professionals internalise, the model has no confidence signal you can see.

The practical rule for the rest of your career: anything with a number, a name, a date, a law, a price or a citation gets verified outside the model before it goes to a client. Everything else, the structure, the phrasing, the ideas, the first draft, you can move fast on.

Temperature, cutoff and bias

Temperature is the dial that trades reliability against range. Low settings give consistent, conservative output. Higher settings give variety, surprise and more errors. You will meet it again on Day 6 with images and Day 9 with video. Training cutoff means the model's knowledge stops at a date, and asking past it gives either honest refusal or confident invention. Tools with web search bolted on go and look, which is a different mechanism you learn properly on Day 4. Bias lives in the data, not the algorithm. Default outputs skew heavily Western. If you want a majlis and you type living room, you will get a sofa from Ohio. Specificity is the fix, and it is your job, not the model's.

What good looks like today

By the end of Day 1 you should be able to say, to a colleague, in plain language, why the AI made up a source. If you can explain that without using the word lied, you have Day 1.

Two tracks. Pick yours, produce something.