Generative AI and Language Model Foundations

Review questions

Integrated Summary: Generative AI and Language Model Foundations

This subject establishes the foundation needed before detailed prompt-engineering techniques. It starts with a simple question: what does generative AI actually do, and what should remain a human responsibility? Generative AI produces new content from learned patterns, available inputs, and context. The output may be text, image, audio, or code. Producing content, however, does not make that content true or a correct decision. The first distinction was therefore between “producing an output” and “being able to rely on that output.”

We also separated the model from the application. The model generates a response; the application adds standing instructions, files, tools, memory, and controls. You cannot infer an application’s full capability from a model name alone. We also distinguished generation, extraction, classification, and retrieval. A workflow can combine several of them, but verification differs: extraction is compared with a source, retrieval is checked for coverage and currency, generated content is reviewed, and final decisions may require an authorized person.

The foundation then moves to how a language model handles text. Input is converted into tokens, processing units that do not necessarily equal one word. What the model can use in a request is the context passed by the application: instructions, relevant conversation history, examples, retrieved passages, or tool results. A document existing somewhere does not mean the model can see it. If you request “the latest policy” without supplying the current version or a tool that retrieves it, the problem is not merely prompt wording; the source is absent from context.

Generation builds a response progressively from possible continuations. The teaching example showed how context changes the distribution of plausible next steps. Those probabilities are not the probability that each statement is true; they reflect how a continuation fits what came before under the model and generation process. Wording can therefore vary across runs. Historically, the Transformer architecture made attention central to modeling relationships within a sequence, but that historical foundation does not mean every current model is an exact copy of the 2017 paper.

Next, we separated capability from reliability. A model can draft, summarize, extract, and propose, yet still confabulate or operate on missing, stale, or conflicting context. NIST uses confabulation for erroneous or false content stated confidently in a way that may mislead users. Saying “do not make anything up” is not enough. When facts matter, we need a source or external test: recompute a percentage, locate the original quotation, run the code, identify the active version, or request an evidence location and inspect it.

Verification changes with claim type. An official name comes from an approved source, a date from a current record, a percentage from original data plus recomputation, and a causal claim requires stronger evidence than two variables moving together. This approach makes uncertainty visible instead of converting it into confident prose. When two versions of a policy conflict, do not silently merge them; resolve source authority or disclose the conflict.

Before optimizing a prompt, screen the task. Ask about consequence of error, verifiability, reversibility, data sensitivity, and source availability. An internal heading draft is not the same as a final financial decision. Summarizing a supplied document is not the same as answering a current fact with no source access. Human review is useful only when the reviewer has enough information, time, and authority to change the outcome before text becomes action. An “approve all” button after a stream of one hundred cases is not serious verification.

The practical workflow for the rest of the program is: define the task and success criterion; prepare the minimum necessary data and context; request an output in a checkable form; verify claims and structure; expose errors and uncertainty; obtain authorized approval when consequences matter; then act and record what is necessary. This subject does not provide a magic prompt. It provides a way of thinking: defined task, appropriate context, realistic expectations of model capability, verification proportionate to risk, and clear human responsibility. The next subject turns these principles into a testable prompt specification.

Review questions

Test your understanding with these questions. Marking and saved results are available with a free account.

Question 1Which description best matches generative AI? A system that creates new content from learned patterns, inputs, and context A database that always returns one stored sentence for the same question A search engine that can only display existing links Software that can produce only numbers and probabilities
Question 2Why is this a suitable starting task? Because any model-generated summary is automatically correct Because the source exists, the output is defined, and it can be checked against the original Because summarization needs neither context nor review Because the model can approve decisions on behalf of the team
Question 3Which pair represents “retrieval followed by generation” in one application? Change font size, then save the file Sort files alphabetically, then delete temporary copies Find an approved policy, then draft an explanation grounded in it Add four numbers, then display the total
Question 4What is the best next step? Accept it because fluency is sufficient evidence of correctness Repeat the request until the same sentence appears twice, then accept it Make the tone formal because that increases factual accuracy Treat it as an unverified draft and obtain evidence before using it
Question 5Which statement about a token is most accurate? It is a processing unit that may be a word, part of a word, or another symbol It is always exactly one whole word in every language and model It is another term for a whole page of text It is a confidence score that a sentence is true
Question 6What is the main limitation? Language models are incapable of summarization The model lacks the required source in context to establish the latest version A summary needs no source when it is short Making the prompt longer automatically fetches the file
Question 7In a simplified example, what does a next-token probability distribution primarily represent? The probability that every generated sentence is a verified fact The share of internet data the model read just now How plausible different continuations are given the preceding context during generation A quality score for the organization using the model
Question 8What is the main historical relevance of Attention Is All You Need in this lesson? It proved all future models must use an identical architecture It converted language models into literal databases It removed the need for context in text generation It introduced the Transformer architecture based primarily on attention mechanisms for sequence modeling
Question 9What does confabulation mean in this lesson? Erroneous or false content stated confidently and potentially misleading the user Any small wording difference between two runs A model refusing to perform a disallowed request A correct summary of a long document into a short paragraph
Question 10What is the strongest next step? Accept it because a specific number looks professional Flag it as unsupported and find its source or recompute it before use Round it to 20% so it is easier to explain Hide the percentage inside a longer paragraph
Question 11What is the best handling? Automatically merge the two versions into a compromise policy Use the longer version because it is probably newer Resolve the active version from an approved source or explicitly surface the conflict before summarizing Ask for more confident wording to hide uncertainty
Question 12Why is “do not make anything up” alone insufficient for a high-consequence task? Because instructions never affect model outputs Because high-consequence outputs must always be longer Because all errors are caused only by text formatting Because instructions do not eliminate error, so source-based or external verification is still needed
Question 13Which task is most suitable for direct generative experimentation with a simple check? Turn non-sensitive general notes into revision cards and compare them with the source Make an irreversible final financial decision automatically Infer missing personal data from a short description Report a current fact without a source or access tool
Question 14When is “human review” a meaningful control point? When someone clicks approve-all without reading When the reviewer has the information, time, and authority to reject or change the output before action When review occurs after an irreversible action has already happened When the reviewer’s name merely appears in the interface
Question 15What should happen before drafting the follow-up message? Choose a plausible nearby date so the table is complete Keep the date because extraction is faster than review Mark the due date as absent from the source and block it from entering the message until resolved Remove the minutes from context after generation
Question 16Which sequence best reflects the responsible workflow taught in the subject? Automatic action → generate explanation → look for a source only if challenged Send all data → generate → default approval → optional later review Choose a model → repeat until it looks good → execute without a success criterion Define task → prepare minimum necessary context → produce checkable output → verify → authorized approval → act