Generate
Produce a statistically plausible continuation from tokens, context and decoding settings.
SRAI Book 4 · Chapter 1 · PU-B04-C01
Generative AI and LLM Foundations
Understand how language models represent text, predict successive tokens and use attention, while applying evidence, evaluation and governance controls for responsible institutional use.
01 / LEARNING OUTCOMES
Understand how language models generate content and how evidence makes their use defensible.
- Distinguish generative modelling from discriminative prediction.
- Explain tokenization, context and autoregressive next-token probability.
- Compare greedy, temperature, top-k and top-p decoding policies.
- Interpret a transparent scaled dot-product attention calculation.
- Separate fluency, groundedness, correctness, safety and decision fitness.
- Construct a reproducible evidence chain for accountable institutional use.
02 / GENERATIVE-AI EVIDENCE CHAIN
Generate, ground, verify and authorize.
Ground
Connect material claims to authorized, current and traceable evidence.
Verify
Test correctness, limitations, safety and fitness for the intended workflow.
Authorize
Retain accountable human approval, escalation paths and an auditable record.
03 / CONTROLLED LABORATORY
A miniature language pipeline makes the essential mechanisms visible.
The approved notebook constructs a tokenizer, estimates next-token probabilities, compares decoding policies and performs a transparent self-attention calculation.
Grounding exercises then show why linguistic fluency cannot establish factual authority, safety or institutional authorization.
VERIFIED NOTEBOOK SCOPE
| Total cells | 41 |
| Code / Markdown cells | 21 / 20 |
| Reproducibility seed | 42 |
| Evidence-chain stages | 6 |
| Decision controls | 8 |
| Validated environments | VS Code / Colab |
A language model generates a probable continuation.
Token probability and fluent wording do not prove that a statement is correct, current, supported or suitable for a consequential decision.
Evidence must come from outside fluency.
Authorized sources, claim-level verification, safeguards and accountable human review determine what the generated output may responsibly support.
04 / SRAI DECISION STANDARD
Generated content earns trust through evidence and accountable use.
- The intended and excluded uses are documented.
- Model, prompt, context and decoding settings are recorded.
- Material claims are grounded in authorized and current sources.
- Fluency, groundedness, correctness, safety and decision fitness are evaluated separately.
- Privacy, security, failure and escalation controls are operational.
- An accountable human authorizes consequential use.
CONTROLLED RESOURCES
Read, reproduce, practise and review.
READ
View the complete GitHub production unit ↗
Audited Chapter 1
Controlled publication edition.
PDF ↗ REPRODUCECanonical notebook
Executable and independently verified.
IPYNB ↗ PRACTISEStudent assessment
Progress toward independent application.
PDF ↗ REVIEWSolutions and marking guide
Review the explained answers, assessment criteria and marking guidance.
PDF ↗ RUNOpen in Google Colab
Run the approved notebook and reproduce the lesson experiments.
COLAB ↗ DOWNLOADComplete release package
Lesson, notebook, assessment, executive, presentation and validation assets.
ZIP ↓ APPLYExecutive brief
Connect analytical controls to decisions.
PDF ↗