Generative AI fundamentals: what the model can and cannot know

Put the terminology in context

Artificial intelligence is a broad field concerned with systems performing tasks associated with intelligence. Machine learning fits patterns from data rather than requiring a hand-written rule for every case. Deep learning uses multilayer neural networks. Generative AI produces content such as text, images, or audio from learned patterns and supplied inputs.

These terms are related, but the neat staircase often drawn in roadmaps hides exceptions. Transformers are an architecture, not a synonym for all generative AI. Large language models are one important family of generative systems. A conventional classifier and a text generator may both help the same product: one routes a support ticket while the other drafts a reply.

Follow a request through inference

For a text model, input is divided into tokens and represented numerically. The model processes the context and predicts a distribution over possible next tokens. A decoding procedure selects output tokens until generation stops. Training previously adjusted the model's parameters; a normal inference request does not retrain those parameters.

Supplying a policy in a prompt therefore changes the information available for this response. It does not permanently teach the base model your company policy. Conversation history, stored application memory, and training are three different mechanisms. Keeping them separate will save you confusion when you debug a follow-up question.

Fluent language is not evidence

Northstar's fictional return window is 30 days. If you ask a model without providing that policy, it may produce a plausible window from patterns learned elsewhere. Even a confident answer is not proof. Grounding means supplying relevant evidence and checking that the answer follows from it.

Hallucination is a useful shorthand for unsupported or fabricated content, but name the actual failure in your bug report: wrong number, invented source, unsupported eligibility claim, or false action confirmation. Those labels point to different fixes. A missing policy may need retrieval; an incorrect total may need a calculator; an invented confirmation needs execution-result validation.

Match the task to the right mechanism

Use generation for rewriting a supplied notice, explaining a policy, or drafting a response. Use ordinary code for exact totals, identifier validation, permissions, and state transitions. Use retrieval for approved documents. Use an authenticated tool for a current order status. Many useful systems combine all four.

Temperature affects sampling when the model supports that setting. It is not an accuracy knob: a low value does not turn a false premise into a true one. Context windows limit how much information a model can process in a request, but a larger window does not ensure it uses every detail correctly. Test the behavior you need instead of treating a parameter as a guarantee.

References and further reading

Use these primary references for deeper study and current API details. Examples in this lesson use fictional Northstar data.

• Research paper — 1706.03762
• Research paper — 2005.11401