The Essential Guide to Generative AI

Key takeaways

• A model is compressed patterns from training, shaped by later tuning; it has a knowledge cutoff and no memory of sources, which explains most of its behaviour.
• The context window is the model's entire world for one conversation, so supplying the right document beats hoping the model already knows it.
• You steer with examples, structure, decomposition and settings such as temperature, and every lever trades consistency against creativity or cost.
• Failures come in families, hallucination, sycophancy, injection, bias and drift, and each family has a specific defence you can build into a workflow.
• Prioritise task choice, context and verification over prompt polish and model choice; test with your own twenty examples before believing any leaderboard.

Inside this bit

1. Idea one: compressed patterns, shaped by tuning

   A generative model is a compressed statistical picture of an enormous body of text or images, built in a stage called pretraining. Nothing is stored as a fact with a source; what is stored is a very large set of numbers that make some continuations more likely than others.

2. Idea two: the context window is the whole world

   For any single conversation, the model can see only what sits inside its context window: the hidden instructions from the vendor, your messages, its own replies and any document you have pasted or attached. That window is measured in tokens and it is finite.

3. Idea three: the levers you steer with

   Once you know the model predicts from what it can see, steering becomes a set of levers rather than a mystery. Examples are the strongest: two or three worked samples of the output you want change results more than paragraphs of description.

4. Idea four: failure modes come in families

   An advanced map groups failures into families, because each family has its own defence. Fabrication is the familiar one: invented citations, numbers and events, defended by verification against a source outside the model.

5. Idea five: quality, cost and speed pull against each other

   Every model choice sits on a triangle. Larger models tend to reason better and follow complex instructions, but they cost more per request and answer more slowly.

6. What to prioritise, what to ignore, what to watch

   Put the five ideas together and a priority order appears. First comes task selection: jobs where errors are cheap to catch and the value is high.