From information to tokens
Explore training data, tokens, representations, prompts, and context as parts of the generation process.
Unit overview
Unit 2 explains generative AI at a conceptual level without requiring programming or advanced mathematics. You will examine how training data, tokens, prompts, context, probabilistic generation, retrieval, tools, multimodal processing, interfaces, and provider-controlled settings contribute to a complete AI system. You will also distinguish the underlying model from the application used to access it and examine why plausible output may differ across systems, versions, sessions, and configurations.
From probabilities to outputs
Understand next-token generation conceptually and why outputs can vary or contain unsupported content.
From model to system
Distinguish models from applications, retrieval, tools, memory, multimodal functions, interfaces, and governance.
Central idea: A generative AI system produces
context-sensitive output by modelling patterns and probabilities.
Coherent output does not demonstrate human-like understanding, source
awareness, truthfulness, or accountable judgment.
Unit learning outcomes
By the end of Unit 2, you should be able to:
- Define training data, token, prompt, context, context window, model, application, retrieval, tool, and multimodal AI.
- Explain conceptually how a language model generates one token after another using conditional probabilities.
- Distinguish model training from the generation of an output during use.
- Explain why the same or similar prompt can produce different outputs across runs, systems, versions, or configurations.
- Distinguish the underlying model from the interface, application, retrieval system, memory, tools, and provider controls around it.
- Explain how a context window limits the information available during a particular generation process.
- Distinguish information contained in model parameters from information supplied through prompts, retrieval, files, tools, or connected services.
- Identify opportunities and risks created by retrieval, tool use, multimodal processing, and broader AI ecosystems.
- Explain why probabilistic generation can produce fluent but unsupported, outdated, incomplete, or contextually unsuitable output.
- Construct and use a system map when deciding what evidence, information boundaries, review, and human control are required.
Unit 2 lessons
Complete these lessons in order. The filenames continue the course-wide lesson sequence established in Unit 1.
2.1 Training Data, Tokens, Prompts, and Context
Examine the information and representations that shape generation, distinguish training from use, and explore how text is processed as tokens.
2.2 Probabilistic Generation and Context Windows
Build a conceptual next-token model, examine variability, and assess how context limits and competing instructions influence an output.
2.3 Retrieval, Tools, Memory, and Multimodal AI
Distinguish model generation from retrieved sources, connected tools, retained information, and the processing of text, images, audio, video, and documents.
2.4 From Model to AI System
Map the complete ecosystem from user and interface to model, sources, tools, provider settings, outputs, review, decisions, and accountability.
File-link convention: Keep this homepage and
t-01.html through t-04.html in the same
Moodle folder or content package.
Embedded Unit 2 study guide
This sequence supports asynchronous study from late Week 1 into Week 2. Follow official Moodle dates if they differ.
-
Retrieve
Explain fluency and reliability before consulting notes. -
Model generation
Study tokens, prompts, probabilities, and context. -
Map the system
Add retrieval, tools, memory, modalities, and interfaces. -
Transfer
Revise your explanation and apply the system map.
Suggested study sessions
Total: 6–8 hours-
Session 1: Retrieval and reading, 60–75 minutes.
Recall Unit 1, preview Chapter 2, and identify unfamiliar terminology without relying on AI-generated definitions. -
Session 2: Lesson 1, 75–90 minutes.
Map training data, tokens, prompts, context, and output. Distinguish training from generation during use. -
Session 3: Lesson 2, 75–90 minutes.
Use a simple worked example to explain token-by-token generation, variability, and context-window limits. -
Session 4: Lesson 3, 75–90 minutes.
Compare a model-only response with a retrieval- or tool-supported workflow and identify additional information and permission risks. -
Session 5: Lesson 4, 90–120 minutes.
Construct a complete AI-system map and assign human verification, decision, and stop responsibilities. -
Session 6: Consolidation, 45–60 minutes.
Revise your opening explanation, complete the self-check, save evidence, and write your own Assignment 1 notes.
Evidence to retain
- Initial and revised explanation of fluency without understanding
- Training-versus-generation comparison
- Token and context-window exercise
- Retrieval and tool boundary analysis
- Multimodal information inventory
- Complete AI-system map
- Assignment 1 reflection notes
Show a system-mapping checklist
Identify the human purpose, user, interface, model, prompt, supplied context, retrieval source, connected tools, memory or logs, modalities, provider controls, output, reviewer, decision-maker, action authority, correction process, and stop authority. Include only components demonstrated or supplied in the case.
Assignment 1 continues in Unit 2
Assignment 1: AI Learning and Reflection Journal is worth 15%. Preserve evidence that shows how your explanation of generative AI changed after examining tokens, probabilities, context, retrieval, tools, and complete systems.
Official Moodle due date and time: ______________________________
Evidence to preserve
- Your initial explanation before Chapter 2
- Your corrected explanation after Lesson 4
- One misconception you identified
- One system-level risk that was invisible when viewing only the model
- One new verification or information boundary
Reflection boundaries
- Write the reflection yourself.
- Do not present anthropomorphic language as a technical explanation.
- Distinguish demonstrated facts from analogy and interpretation.
- Use the textbook terminology where supplied.
Show Unit 2 reflection-planning questions
- How did your mental model of generative AI change?
- Why can probabilistic generation produce persuasive but false content?
- Which system components affect an output beyond the underlying model?
- How do retrieval and tools create both opportunities and additional risks?
- Which claim about AI “understanding” would you now revise?
Use these questions to plan your own reflection, not as model responses.
Required learning activities
| Activity | Purpose | Expected product | Estimated time |
|---|---|---|---|
| Training versus use classification | Distinguish model development from prompting and generation during use. | Classification with explanation | 25–35 minutes |
| Token and probability walkthrough | Explain generation conceptually without advanced mathematics. | Annotated sequence | 35–45 minutes |
| Context-window experiment | Identify relevant, missing, competing, and excessive context. | Context analysis | 30–45 minutes |
| Retrieval-versus-generation comparison | Distinguish generated content from retrieved evidence and source citation. | Comparison table | 35–50 minutes |
| Tool and permission map | Identify what a system may read, retrieve, draft, modify, send, or control. | Bounded permission record | 30–45 minutes |
| AI ecosystem map | Connect people, models, applications, sources, tools, outputs, and authority. | System map with text alternative | 45–60 minutes |
Participation options
Option A: Approved AI system
Compare low-risk outputs using an approved system and only synthetic, public, supplied, or authorized information.
Option B: Supplied records
Use supplied prompts, outputs, token illustrations, retrieval records, tool calls, and system descriptions.
Option C: Non-AI analysis
Complete the conceptual modelling and system-mapping activities without operating an AI system.
All options assess the same conceptual and responsible-use outcomes.
Responsible-use boundaries and support
Boundaries for Unit 2
- Do not upload confidential, personal, proprietary, privileged, security-sensitive, or culturally governed information.
- Do not connect email, files, calendars, payments, cameras, microphones, location, or external tools.
- Do not test security weaknesses or prompt injection against real systems.
- Do not infer product features that have not been verified.
- Do not equate prediction of tokens with human understanding or accountable judgment.
Support and accessibility
- Use supplied examples if tokenization, visual maps, live systems, or multimodal features are inaccessible.
- Provide a text description for every system diagram.
- Contact the course coordinator for conceptual or assignment questions.
- Use technical or accessibility support when needed.