COMP 3270 / Module 1 / Unit 2 / Consolidation

Unit 2 Consolidation · Week 2 · Chapter 2

Unit 2: How Generative AI Works — Consolidation

Synthesise everything from Unit 2 — training data, parameters, tokens, probability, prompts, context, retrieval, tools, multimodal AI, and human‑governed systems.

Estimated total time: 3–4 hours Concludes Unit 2 Prepares for Unit 3

Table of Contents

Overview

You have completed all four lessons of Unit 2. Now it is time to consolidate your learning. This page brings together the following cumulative components:

Work through these activities in sequence. They are formative and ungraded, but they will prepare you for later assessments, especially Assignment 1 and Unit 3.

Purpose: To integrate and apply everything you have learned in Unit 2: how generative models are built, how they generate output, how prompts and context shape results, and how retrieval, tools, and multimodal capabilities extend the system.
Responsible‑use reminder: Do not enter personal, confidential, proprietary, privileged, or security‑sensitive information into any public or unapproved AI system. Use only synthetic, public, or generalised information for all activities.

A. Unit 2 Consolidation Exercise

This exercise integrates all the concepts from Unit 2. You will analyse a synthetic AI‑supported process and answer questions that span the entire unit.

Synthetic scenario: AI‑assisted research summarisation

A graduate student is using an AI system to help summarise research articles for a literature review. The system works as follows:

  • User prompt: "Summarise the key findings of the attached paper on renewable energy policy."
  • Supplied context: The student uploads a PDF of a recent journal article (text and figures).
  • Model: A large language model with a context window of 8,000 tokens.
  • Retrieval: The system also searches an internal repository of related papers and retrieves two additional abstracts.
  • Tool: A citation formatter that generates a reference in APA style.
  • Multimodal: The system processes a chart from the PDF and includes a description of it in the summary.
  • Generated output: A 300‑word summary with a chart description and a formatted citation.
  • Human review: The student reviews the summary, checks the citation, and verifies the chart description against the original figure.

Your task: Answer the following questions based on this scenario.

Questions
  1. Identify the human purpose.
  2. Identify the model.
  3. Identify the application.
  4. Identify the complete AI system.
  5. Distinguish training from inference in this scenario.
  6. Identify possible tokens conceptually for the prompt.
  7. Explain how next‑token generation would work for this task.
  8. Identify the context window's role — what might be included in the context?
  9. Distinguish supplied context from verified evidence.
  10. Identify retrieval in the scenario.
  11. Identify tool use in the scenario.
  12. Identify multimodal processing in the scenario.
  13. Identify three possible failure points.
  14. Identify two human‑review points.
  15. Connect one technical capability to an opportunity.
  16. Connect the same capability to a corresponding risk.
  17. Identify one question for Unit 3.

Estimated time: 30–45 minutes.

Response template: You may use bullet points or paragraphs. Provide clear answers for each question.

Show self‑assessment checklist
  • All 17 questions are answered.
  • Answers are consistent with Unit 2 concepts.
  • Failure points and review points are realistic.
  • Opportunity‑risk pair is well‑justified.
Show Hint 1: Identifying components

Look for the human user, the core computational component, the user‑facing product, and the broader operational arrangement including retrieval, tools, and human roles.

Show Hint 2: Training vs. inference

Training is the creation of the model (parameters adjusted on data). Inference is the processing of the prompt and context to produce the summary during the student's interaction.

Show Hint 3: Failure points

Consider: the retrieval might pull irrelevant abstracts, the model might misinterpret the chart, the citation tool might format incorrectly, the context window might truncate the PDF.

Show solution design and illustrative partial answers

Sample answers (partial):

  1. Human purpose: To obtain a concise summary of a research paper for a literature review.
  2. Model: The large language model that processes the input and generates text.
  3. Application: The AI summarisation tool (the student's interface).
  4. System: The complete setup: model, application, retrieval component, citation tool, context window, human reviewer (student).
  5. Training vs. inference: The model was trained earlier on a large corpus. During the student's interaction, it performs inference — processing the prompt and context to generate the summary. The model's parameters are not updated.
  6. Possible tokens: The prompt might be tokenized as [“Summarise”] [“ the”] [“ key”] [“ findings”] [“ of”] … (conceptual).
  7. Next‑token generation: The model predicts tokens one by one based on the prompt, context, and previously generated tokens.
  8. Context window: Includes the prompt, the PDF text, the retrieved abstracts, and the chart description (if converted to text). Must fit within 8,000 tokens.
  9. Supplied context vs. verified evidence: The PDF and abstracts are supplied context; the student must verify the summary and chart description against the original sources.
  10. Retrieval: The system searches for and returns two additional abstracts.
  11. Tool: The citation formatter.
  12. Multimodal: The chart from the PDF is processed and described.
  13. Failure points: (1) Retrieval pulls outdated abstracts; (2) Chart description misinterprets data; (3) Citation tool misformats.
  14. Human review: The student checks the summary for accuracy and verifies the citation.
  15. Opportunity: Rapid synthesis of multiple sources.
  16. Risk: Misinterpretation of the chart leading to a false claim.
  17. Question for Unit 3: How do I assess whether the AI summary is sufficiently accurate for my literature review?

B. Unit 2 Cumulative Homework Set

This set of 14 problems covers the entire unit. They are formative and ungraded, but are excellent preparation for the cumulative project and Assignment 1.

Problem 1: AI and Machine‑Learning Relationships

Task: Draw a diagram or write a hierarchy showing the relationships among AI, ML, deep learning, generative models, foundation models, and LLMs. Indicate at least two non‑interchangeability relationships.

Show solution design

AI is the broadest; ML is a subset; DL is a subset of ML; generative models are a type of model often built with DL; foundation models are large generative models; LLMs are a type of foundation model for language. Non‑interchangeability: not all ML is generative; not all generative models are LLMs.

Problem 2: Training Data and Parameters

Task: Explain what parameters represent, and why they are not stored documents. Use an analogy and explain its limitations.

Show sample answer

Parameters are numerical values encoding patterns. Analogy: a map's coordinates and connection strengths. Limitation: the map is static and interpretable; parameters are high‑dimensional and not directly readable.

Problem 3: Training vs. Inference

Task: A user interacts with a chatbot. The chatbot was trained last year. Explain why the chatbot's knowledge is not automatically updated by the user's questions.

Show sample answer

During inference, the model's parameters are frozen. The user's questions are processed using the existing patterns; the model does not learn from the interaction. Updates require retraining or fine‑tuning.

Problem 4: Tokenization

Task: Give an example of a word that might be split into multiple tokens, and explain why tokenization varies across systems.

Show sample answer

"unbelievable" could be split as "un", "believ", "able". Tokenization varies because different tokenizers use different vocabularies and rules.

Problem 5: Next‑Token Prediction

Task: Explain why the model does not plan the entire response in advance, using an example of a prompt and two possible first tokens.

Show sample answer

The model generates one token at a time. Example: prompt "The capital of France is". Possible first tokens: "Paris" or "a". Choosing "Paris" leads to a factual answer; choosing "a" might lead to a different sentence. The model doesn't know the final outcome until it finishes.

Problem 6: Probability and Variability

Task: Why can the same prompt produce different outputs, and why does agreement across multiple runs not constitute independent verification?

Show sample answer

Generation is probabilistic; sampling leads to variability. Repeated outputs come from the same model and are not independent; they may all contain the same error. Verification requires external sources.

Problem 7: Prompts and Context

Task: Distinguish between a system instruction, a user instruction, and supplied context in a prompt. Provide an example.

Show sample answer

System instruction: "You are a helpful assistant." User instruction: "Summarise this article." Supplied context: The article text. They are all part of the input but serve different purposes.

Problem 8: Context Windows

Task: What happens when the context window is exceeded? Why is a larger window not always better?

Show sample answer

When exceeded, older tokens may be dropped. A larger window may not guarantee better output because the model may not attend evenly to all parts; quality and relevance of context matter more.

Problem 9: Retrieval

Task: Explain what retrieval‑augmented generation is and list two ways retrieval can fail.

Show sample answer

RAG retrieves external information to ground generation. Failures: retrieving outdated documents, misinterpreting the source, or retrieving irrelevant content.

Problem 10: Tools

Task: Give an example of an external tool that could be used with an AI system, and explain one risk it introduces.

Show sample answer

Example: a tool that sends emails. Risk: the system could send inappropriate messages without human approval.

Problem 11: Multimodal AI

Task: Why is multimodal coherence not a guarantee of correctness? Provide an example of a possible error.

Show sample answer

Coherence may be coincidental or based on common patterns. Error example: a system describes a bar chart as showing a 50% increase, but the actual increase is 20% due to misreading the scale.

Problem 12: Model vs. System

Task: Distinguish an AI model from an AI system. Why is governance of the system more complex than governance of the model alone?

Show sample answer

A model is the trained component; a system includes the model, interface, retrieval, tools, human roles, etc. Governance must address all components, not just the model's technical performance.

Problem 13: Human Oversight

Task: Identify at least two points in an AI‑supported workflow where human review is essential and explain why.

Show sample answer

After generation: to check for factual errors. Before acting on the output: to ensure decisions are appropriate and comply with policies.

Problem 14: Opportunity‑Risk Connection

Task: Pick one technical characteristic (e.g., probabilistic generation, retrieval) and describe an opportunity and a corresponding risk.

Show sample answer

Characteristic: retrieval. Opportunity: access to current information. Risk: reliance on outdated or incorrect sources.

C. Unit 2 Cumulative Project

Project Title: How a Generative AI System Produces and Extends an Output
  • Purpose: To synthesise all Unit 2 concepts by analysing a generative AI system from training to output, including human oversight.
  • Learning outcomes: All Unit 2 outcomes.
  • Estimated effort: 2–3 hours.

Instructions: Select one bounded, low‑risk application from:

  • Learning (e.g., a study assistant).
  • Professional practice (e.g., draft writing).
  • Communication (e.g., email drafting).
  • Creative work (e.g., image generation).
  • Organisational activity (e.g., policy summarisation).
  • Public information (e.g., health advisory).
  • Everyday life (e.g., travel planning).

Develop a structured analysis with the following 27 sections:

  1. Application and context
  2. Human purpose
  3. Intended user
  4. AI model type, if known
  5. AI application
  6. Simplified AI ecosystem
  7. Training‑data influence at a conceptual level
  8. Training‑versus‑inference distinction
  9. Prompt or input
  10. Conceptual tokenization example
  11. Next‑token generation explanation
  12. Probability and variability
  13. Context supplied
  14. Context‑window consideration
  15. Retrieval source, if applicable
  16. External tool, if applicable
  17. Multimodal component, if applicable
  18. Generated output
  19. Three possible technical failure points
  20. Claims requiring verification
  21. Human reviewer
  22. Final human authority
  23. Opportunity
  24. Corresponding risk
  25. Preliminary information boundary
  26. One question for Unit 3
  27. Assignment 1 reflection connection

Suggested length: Approximately 1,000–1,250 words or an equivalent structured submission.

If your selected application does not use retrieval, tools, or multimodal input, state that clearly and explain how adding them would change capability and risk.

Privacy and academic‑integrity boundaries:
  • Use only synthetic, public, or generalised information.
  • Do not use personal, confidential, or proprietary content.
  • Do not ask the AI to perform external actions or generate personal data.
Show self‑review checklist
  • All 27 sections are addressed.
  • Tokenization is conceptual, not exact.
  • Failure points are realistic and connected to technical aspects.
  • Opportunity‑risk pair is well‑justified.
  • Human oversight is clearly defined.
Show Hint 1: Choosing an application

Pick something you are familiar with, e.g., using AI to summarise course notes or to draft emails. Avoid high‑stakes applications (medical, legal).

Show Hint 2: Organising the sections

Group sections logically: first describe the context and purpose, then the technical components (model, tokens, generation, context), then retrieval/tools/multimodal if applicable, then the output and human review, and finally opportunity/risk and reflection.

Show Hint 3: Failure points

Think about each component: model (hallucination), retrieval (wrong source), context (overload), tools (error), human (misses error).

Show solution architecture and partial outline

Solution architecture: Use a clear heading for each section. For each technical component, explain it conceptually and tie it to the specific application.

Illustrative outline (based on a different application — AI for travel planning):

  • 1. Application: AI travel assistant for itinerary suggestions.
  • 2. Human purpose: Plan a 5‑day trip to a European city.
  • 3. Intended user: A budget‑conscious traveller.
  • 4. Model: LLM.
  • 5. Application: Chatbot interface.
  • 6. Ecosystem: Travel booking APIs, reviews, maps, etc.
  • … etc.

Note: This is just an outline; your project should be specific to your chosen application and include all 27 sections.

D. Unit 2 Cumulative Self‑Check Quiz

This quiz contains 16 questions covering all Unit 2 topics. It is ungraded and includes explanatory feedback.

Question 1

Which of the following is a subset of machine learning?

Select one:
Show answer and explanation

Correct answer: B. Deep learning.

Explanation: Rule‑based systems are not ML; AI is broader than ML.

Question 2

What is the role of training data?

Select one:
Show answer and explanation

Correct answer: A.

Explanation: Training data is used only during training, not during inference.

Question 3

Model parameters are best described as:

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Parameters are not stored documents or explicit rules.

Question 4

During inference, the model's parameters are:

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Inference does not update parameters.

Question 5

A token is:

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Tokens are not always words.

Question 6

How does a language model generate extended text?

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Generation is iterative.

Question 7

Probability in token prediction indicates:

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Probability is a statistical measure, not a measure of truth.

Question 8

Why might the same prompt produce different outputs?

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Variability is normal.

Question 9

What is the context window?

Select one:
Show answer and explanation

Correct answer: B.

Explanation: It limits the input.

Question 10

Which of the following is an example of conflicting context?

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Conflicting information can confuse the model.

Question 11

Retrieval‑augmented generation (RAG) is:

Select one:
Show answer and explanation

Correct answer: B.

Explanation: RAG combines retrieval and generation.

Question 12

Which of the following is a potential retrieval failure?

Select one:
Show answer and explanation

Correct answer: B.

Explanation: That is a retrieval‑specific failure.

Question 13

What does multimodal AI refer to?

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Multimodal includes text, images, audio, etc.

Question 14

An AI application differs from an AI model in that:

Select one:
Show answer and explanation

Correct answer: A.

Explanation: The model is just the computational core; the application is the product.

Question 15

Why is human oversight essential in an AI system?

Select one:
Show answer and explanation

Correct answer: B.

Explanation: Humans are accountable for decisions and must verify outputs.

Question 16

Which of the following does generated content not establish?

Select one:
Show answer and explanation

Correct answer: D.

Explanation: Generated content does not establish any of these human qualities.

E. Assignment 1 Journal Guide

This section helps you prepare your Assignment 1: AI Learning and Reflection Journal entry for Unit 2. You will reflect on your learning journey and how your mental model has evolved.

Unit 2 journal topics

Choose from the following topics to guide your journal entry. You do not need to address all of them; select the ones most relevant to your learning.

  • Mental model shift: How has your mental model of generative AI changed after completing Unit 2? What did you previously assume about training data, parameters, tokens, or generation that has been corrected?
  • Training data: What did you previously assume about training data? What have you learned about its limitations (gaps, errors, imbalances, currency)?
  • Tokens and prediction: How does understanding tokenization and next‑token prediction change the way you view AI‑generated text?
  • Probability vs. truth: Why is it important to distinguish between what is linguistically likely and what is factually correct?
  • Output variability: How does the probabilistic nature of generation affect your trust in AI outputs? What does variability tell you (and not tell you)?
  • Prompts and context: How do prompts, instructions, and context influence generation? Why is context not the same as evidence?
  • Retrieval and tools: How do retrieval and tools expand capabilities and risks? Give an example.
  • Model vs. system: Why is it important to distinguish the model from the application, system, and ecosystem?
  • Technical limitation and trust: Which technical limitation (e.g., training data currency, context window, probabilistic errors) most affects your trust in AI? Why?
  • Opportunity‑risk pair: Identify one technical characteristic (e.g., pattern learning, retrieval) and explain one opportunity and one corresponding risk.
  • Question for Unit 3: What question do you want Unit 3 to address about responsible AI use?
Suggested organisation
  • Introduction: Briefly state what you learned in Unit 2.
  • Body: Address 2–3 topics in depth. Provide specific examples from your learning (e.g., from activities, scenarios, or your own exploration).
  • Conclusion: Summarise your key takeaway and state your question for Unit 3.
Evidence to retain
  • Your responses to exercises, homework problems, and mini‑projects.
  • Notes from the lessons.
  • Any AI‑generated outputs you used (with appropriate documentation).
AI‑interaction documentation reminder

If you used an AI tool to assist your learning, document your use according to course policy (e.g., include prompts and outputs).

Academic‑integrity and privacy reminders:
  • Your journal should be your own work.
  • Do not disclose personal, confidential, or proprietary information.
  • Use general, synthetic, or public examples.
Show self‑review checklist
  • I have addressed at least two topics from the list.
  • I have included specific examples from my learning.
  • I have articulated a clear change in my mental model.
  • I have connected technical concepts to opportunities and risks.
  • I have formulated a question for Unit 3.
  • I have adhered to privacy and academic‑integrity guidelines.

Note: Do not provide personal model responses; these are your own reflections.

Looking Ahead to Unit 3

You have now completed Unit 2: How Generative AI Works. You have built a solid conceptual foundation covering training, parameters, tokens, prediction, probability, prompts, context, retrieval, tools, multimodal AI, and the distinction between model, application, system, and ecosystem.

In Unit 3: Effective and Responsible AI Use, you will move from how AI works to whether and how to use it responsibly. You will learn to:

The technical insights you have gained — about probabilistic generation, context limitations, retrieval failures, and the gap between fluency and truth — will be essential as you evaluate AI‑supported activities and determine when and how to intervene.

Key questions for Unit 3:
  • "Should this AI system be used for this purpose, under these conditions, with this evidence, oversight, and accountability?"
  • "What are the risks, and how can they be mitigated?"
  • "What is the human role, and who is accountable?"

Before moving on, ensure you have completed the consolidation exercises and quiz, and that you have reflected on your learning in the Assignment 1 journal guide. Your instructor will provide further guidance on submitting Assignment 1.

Congratulations on completing Unit 2!

You are now well‑prepared for the responsible and effective use of generative AI.

Continue to Unit 3: Effective and Responsible AI Use Return to Unit 2 Home
👤