Week 1 has two 80-minute meetings. Today builds the course map and the LLM foundation.
Meeting 1 of 2 · Start here
Use AI with judgment—not just confidence.
This week’s destinationBuild a useful mental model of an LLM, separate plausible language from evidence, and decide where human approval belongs.
Choose Quick Review or Extended Review if useful. Neither route is graded. If you complete one, submit its PDF in eCampus for formative feedback; choosing no homework carries no penalty.
Certification: preparation 2%, scheduled exam attempt or approved equivalent 3%, reflection 2%. External pass/fail does not change those points. eCampus will announce each quiz’s coverage, date, and weight before administration.
CapabilityWhat is it doing?
ContextWhat can it see?
EvidenceWhat supports it?
AuthorityWhat may it do?
These in-class activities are ungraded. Follow eCampus grading milestones; no exam registration or payment in this lesson.
The course map · Four checkpoints
Four questions will follow us all semester.
Why this map mattersWhenever AI produces an answer or takes an action, these four questions help us slow down and locate the real decision.
Before we open the map: Which checkpoint do people most often forget? Choose one, then we will examine all four in order.
Class prediction
What is it doing?
Identify whether AI is generating text, retrieving information, using a tool, or taking an action.
UpcomingWhat supports the answer?
A confident response is not automatically verified. Look for appropriate sources, tests, and evidence.
UpcomingWhat may it do?
Capability is not permission. Define privacy boundaries, allowed tools, and where human approval is required.
UpcomingWho remains responsible?
Using AI does not transfer responsibility from the student, teacher, developer, company, or university to the machine.
UpcomingKeep the map: What is it doing? What supports the answer? What may it do? Who remains responsible? Week 1 introduces the map; later weeks apply it to GitHub, agents, multimodal systems, ethics, deployment, and projects.
Your prediction has no right answer; it gives us something to revisit.
Discussion · Future work
Map the task. Do not predict the whole job.
What students will doChoose a familiar role, separate one piece of work into tasks, then identify where AI may help and where a person must remain responsible.
Step 1 of 3 · Try first6-minute discussion
Choose one role and one real task inside that role.
Work alone for 90 seconds. Break the task into smaller steps. Mark where AI could assist, where evidence must be checked, and where a person should decide.
GatherAI may organize supplied material.
DraftAI may propose a structure.
VerifyA person checks claims and sources.
Decide + communicateA responsible person owns the judgment and consequence.
No labor-market forecast is required. Analyze a task you understand.
Running example · Fictional
What should happen before anyone presses Send?
What students will doStart with the request alone. Propose a safe process, then compare it with four stages revealed one at a time.
Your turn · Process firstNothing has been sent
What should happen between this request and pressing Send?
Think for 30 seconds. Name the first step your process should take. The four course stages are hidden until the class commits to an idea.
?Stage 1 hidden
?Stage 2 hidden
?Stage 3 hidden
?Stage 4 hidden
The sequence will be Context → Draft → Check → Act.
Challenge 1 · Context → Draft
What is the minimum sufficient context?
Connection + taskThe previous slide ended Context → Draft → Check → Act. Now test that first connection: inspect three fictional packets, commit to one, and compare usefulness, relevance, and privacy.
Context Without OversharingFixed course examples · No live AI · Nothing saved or graded
1 · Inspect
Open Packet A, B, or C.2 · Choose
Select the minimum sufficient packet.3 · Commit
Reveal only that fixed response.4 · Compare
Evaluate all three after commitment.
Open Packet A, B, or C.2 · Choose
Select the minimum sufficient packet.3 · Commit
Reveal only that fixed response.4 · Compare
Evaluate all three after commitment.
Information this packet gives the AI
Simulated AI draft · Hidden until commitment
Inspect the packet, then commit before the course-authored draft appears.
Page-memory only · Refresh clears activity state
Evidence card · Page 1
Write your minimum-context prompt and one sentence explaining what you intentionally omitted—and why.
First principles · From draft to model
An LLM builds text one token at a time.
ConnectionThe last activity changed the context given to a draft. Now we examine the simplified generation loop that turns visible context into a continuation.
Step 1 of 5 · Visible sequenceSimplified generation loop
The model receives the token sequence visible in its context.
It does not begin with a verified answer. It begins with tokens representing the visible instructions and text.
1 · Read sequenceUse visible tokens.
2 · ScorePossible next tokens.
3 · SelectApply a selection rule.
4 · AppendAdd one token.
5 · RepeatContinue the sequence.
Follow the loop, then identify the important step it does not automatically contain.
Challenge 2 · Human prediction game
Same unfinished email. Three different contexts.
What students will doVote on the most plausible continuation in three rounds. The unfinished email and answer choices stay fixed; only visible context changes.
Round 1 of 3Bars count people—not model probabilities or truth
Raise hands for the continuation that feels most plausible. The instructor records the class distribution with the + buttons.
No answer is automatically true. This activity measures how context changes human plausibility judgments.
Worked case · Two judges
A sentence can sound right and still fail the evidence test.
What students will doJudge the same fictional sentence twice. Commit to each decision before the course explanation appears.
Test 1 of 3 · Judge 1Commit before reveal
Does it pass Judge 1—does it sound like a possible continuation of the email?
Judge 1 · Language fit: Is the sentence fluent and plausible as the next part of an academic email? This judge does not check truth.
Choose PASS or FAIL before revealing the course analysis.
Challenge 3 · Exact tokenizer example
Predict the boundaries before the tokenizer reveals them.
What students will doClick where token boundaries might occur, lock the class prediction, then compare it with one verified tokenizer result.
Step 1 of 3 · Predict boundariesFixed verified example · No live tokenizer call
Token: a tokenizer-specific piece of input/output. Token ID: that piece’s vocabulary index—not its probability, importance, or truth. The symbol ␠ makes a leading space visible.
Gold = your predicted boundaryGreen = matchRose = extra predictionCyan = missed exact boundary
Exact token pieces and IDs
Predicted token count: 1. Click boundary slots before locking.
Challenge 4 · 4 explicit rounds
Same question. Different visible context.
What students will doJudge one fixed response at a time as USE, ASK, or REJECT. Classify the response as written, then reveal the evidence analysis.
Round 1 of 4Fixed course-authored response · No live AI
Response to classifyChoose the label that best fits this response as written. A rejected response might still be replaced by a request for better evidence.
Context check: Is it available, relevant, authoritative, and current?
Watch · Clip A · 0:00–1:28
Use the clip to test the model we just built.
Before playbackListen for three things: what is predicted, why “token” is more precise than “word,” and what repeats to produce a response.
Large Language Models explained briefly
Play the clip with captions. If playback fails or external media is unavailable, use the static fallback below and answer the same three checks.
Static fallback
An LLM receives a visible token sequence, scores possible next tokens, selects one, appends it, and repeats. This loop can produce useful language without automatically verifying every claim.
Source: 3Blue1Brown, full 7:58 video ↗. Captions available. No autoplay. Opening the clip contacts YouTube/Google; no student responses are sent.
After the clip · Check 1 of 3Commit before answer
1 · What does the model predict?
2 · Why is “token” more precise than “word”?
3 · What repeats to produce a response?
The narrator may say “word”; this course uses the more precise term “token.”
Meeting 1 close · 3 phases
Build your evidence card—then explain it.
Submission statusUse Evidence Card pages 1–2. It is ungraded, not submitted, and remains with you. The anonymous question uses a separate slip.
Phase 1 of 3 · Individual write · 4 min
Phase 2 of 3 · Partner explain · 3 min
Phase 3 of 3 · Anonymous question · 3 min
Phase 1 · Evidence Card pages 1–2
Write your two-sentence mental model.
- What does an LLM do?
- Why can useful, fluent language still need evidence?
- Complete: “A token is ___; a tool is ___.”
- For the email task, check: available, relevant, authoritative, current.
Self-check: Did you separate model generation from evidence checking and external tools?
Keep your evidence card. Nothing is collected or graded.
Meeting 2 · Retrieval practice · 3 rounds
Complete one idea at a time.
What students will doComplete the current sentence aloud or on paper, commit as a class, then reveal and explain the course model before the next sentence appears.
From data to behavior · 3 stages
Training answers two different questions.
Why this follows Meeting 1We know an LLM predicts tokens. Now we ask how it learns broad sequence patterns and why an assistant behaves differently from a raw base model.
Base model → assistant → product · 4 stages
Build the layers, then test the boundary.
What students will doReveal one layer at a time, identify what each layer changes, watch an optional short clip or read its equivalent summary, then complete one sentence about behavior and truth.
Worked example · 4 rounds · Fictional policy
Judge claims separately—not the paragraph by vibe.
What students will doSplit one polished sentence into checkable claims, classify each claim against the supplied syllabus excerpt, then repair the answer.
Claim and evidence lab · 3 phases · 4 classification rounds
Classify, correct, then prioritize.
What students will produceOn Evidence Card pages 3–4: a four-row ledger, a corrected summary, and one reason for checking a particular claim first.
Phase 1 · Classify · 4 rounds
Phase 2 · Correct
Phase 3 · Prioritize
Bridge · 4 distinctions
A response becomes a system when more parts can participate.
Why this bridge mattersThe evidence check told us whether text was supported. The next question is what components can retrieve, repeat, prepare, or act—and who grants authority.
System builder · 3 rounds
Build the system before we name its parts.
What students will doRead one fictional system, select every component the evidence supports, lock the map, then compare each label with the exact phrase that supports it.
Component tray · Select all that the description supports
Round 1 of 3 · Choose components before revealing the evidence map.
Permission map · 3 stages
Choose the least authority the task needs.
What students will doOpen the five authority levels, choose the least sufficient level for the fictional missed-class task, commit, then compare with the course boundary.
Fictional task
A system may retrieve the approved syllabus, cite it, and draft a missed-class email. Nothing should leave the student’s account without the student reviewing the exact message.
Operational control · 3 rounds
Proceed, ask, or block?
Repeat the decision ruleRead only the current fictional case, choose one control, commit, then compare the recommendation with the visible facts. All three cases appear together only after Round 3.
Week 1 close · 3 phases
Revise, explain, then leave one question.
Submission statusYour Evidence Card is ungraded, not submitted, and stays with you. Only the separate anonymous question slip is collected.
Phase 1 of 3 · Revise · 4 min
Phase 2 of 3 · Explain · 3 min
Phase 3 of 3 · Question · 3 min
Week 1 classroom deck: 22 screens across two 80-minute meetings.