How to use
- Speaker · VO — host voice from the deck speaker notes.
- Figure — diagram / press image notes when present.
- Facilitation — deeper Ocean do / don’t / recovery craft.
- Project the student deck; keep this on a second screen or print.
- Through-line: When answers are free, what still costs a human?
Opening
Cold open · five marks
Speaker · VO
Cold open — the front page. Do not introduce yourself yet. Let the question sit for five seconds of silence, then say: 'That is the only question this whole talk is about.' The index on the right is the only agenda the room gets; do not read it aloud.
Figure
FIGURE — A broadsheet front page. The thick-over-thin rule pair with the dateline running between them is the one place this design system prints rules; everything else is separated by whitespace. The headline is set as three process plates (cyan, rose, yellow) slightly out of register over a paper-white union, so it reads as ink on newsprint rather than type on a screen. The right column is the edition index: the five acts as small plate numerals. Do not read the index aloud.
Speaker · VO
Five marks, one ink, one square field — the same press screen in the first and the last, which is the whole point: nothing in the world changes over the next hour, only what you can see in it. Do not explain each mark; let the room read them. If anyone asks: the second mark is a plate split on its own axis, half inked; the fourth is the four stations we walk in Act IV.
Figure
FIGURE — Five marks, one ink, one 230px square field, one element changing between them. I Fog: the deck’s halftone press screen with no image under it — loud, no shape yet. II Mirror: a disc split on its own axis, half inked, the mirror plane drawn in cyan. III Ocean: seven hairlines of unequal length (the wide net) with one accent dot — the single idea landed from it. IV Voyage: one line, four stations, the second in accent; this is the same four-station track drawn in full on slide 53, so the mark and the diagram rhyme when you get there. V Return: mark I again, completely unchanged, with a boundary drawn round it and one accent dot at the centre — you, standing in the same fog. The device is that the first and last marks are the same screen: nothing in the world changes over the hour, only the boundary, and you are inside it. Let the room read the marks; do not narrate them one by one.
Act I · Fog
The world outside the model
Speaker · VO
Act opener. Say: this act is the only part of the talk that is about the world instead of the machine. It exists so nobody in the room mistakes AI for software news.
Speaker · VO
This board stays up. Ask the room to point at the one force they have personally felt this year. Do not explain all eight — three named out loud is enough, the board does the rest.
Figure
FIGURE — Eight press clippings pinned as a board, each over the one line that says what the force is: the forces are evidence, not assertions, and the room should recognise at least two sources without help. The same eight hard-wired board clippings return later, where the caption changes from what the force is to the mechanism under it — the two boards are meant to be photographed and compared. CAPTURE BRIEF, one per well — 01 a DRAM or HBM price-spike story, chart or headline with a number in it (trade press 2025–26). 02 an export-control or fab-capacity ruling; the drier and more official-looking the better. 03 the vendor’s own model-deprecation page showing a retirement date — the access statement itself, not coverage of it. 04 a launch post claiming a benchmark win, cropped so the number is the loudest thing in the clip. 05 a local paper on a data centre’s power or cooling draw — local, not national, named town if possible. 06 a story about a role being automated, ideally close to your students’ intended field. 07 availability or pricing shown by country or tier; a map or a table reads best. 08 a prompt-injection or over-permissioned-agent post-mortem, not a warning piece. Board wells are hard-wired (board-clip-01…08.jpg). Crop intent: one fact legible from the back of the room; outlet name and date in the crop for attribution. Clippings print through the newsprint dot-screen this system reserves for interface imagery — it keeps small type readable where four misregistered plates would destroy it. Even so, capture headlines, numbers and charts rather than paragraphs. The wells crop to fill: after dropping a clipping, double-click it to drag and scale the crop until the headline sits inside the frame. SOURCES for the crops, verified July 2026 — 01 tomshardware.com, memory prices climbing through Q3 2026 (TrendForce survey). 02 Federal Register document 2026-00789, the BIS final rule effective 15 January 2026 amending 15 CFR 742/744/748; the Council on Foreign Relations commentary of 14 January 2026 is the readable companion. 03 platform.claude.com/docs/en/about-claude/model-deprecations — Sonnet 4 and Opus 4 notified 14 April 2026, retired 15 June 2026: an access statement, not coverage of one. 04 OpenAI’s Frontier Evals team stopped reporting SWE-bench Verified on 23 February 2026 after auditing 138 of the hardest tasks and finding 59% with materially flawed tests. 05 cbsnews.com/chicago — Naperville City Council voted down a proposed data centre after months of resident opposition over water and utility costs. 06 hrdive.com on the ResumeTemplates Class of 2026 report, 17 July 2026: 48% of hiring managers would rather invest in AI tools than hire and train a recent graduate. 08 helpnetsecurity.com, 11 June 2026, on OWASP GenAI’s State of Agentic AI Security 2.01 — the LiteLLM backdoor sat on PyPI about three hours in March 2026 and was downloaded roughly 47,000 times. Well 07 has no verified source line yet: capture the provider’s own country-availability page and let the page be the citation. Every clipping is someone else’s copyright — keep the outlet name and date inside the crop, and use the clipping as evidence under discussion, not as decoration.
Speaker · VO
Four dated facts, so this is reporting and not vibes. The retirement row is the one to dwell on: a model you built on can be withdrawn on a published date, and the notice period is the whole of your warning. Sources: Anthropic agentic-coding reporting 2026; memory market reporting 2025-26 (TrendForce surveys, via trade press); Anthropic’s own model-deprecations page — Sonnet 4 and Opus 4 notified 14 April 2026, retired 15 June 2026, Mythos Preview retired 30 June 2026; production-agent security guidance, 2026.
Figure
FIGURE — Four dated facts as a stack of rows, each led by the number that carries it (0–20%, HBM, 2026·06·12, stop). Reporting layout, not slideware: the figure on the left, the claim in the middle, the source in the notes. The June 2026 row is the one in accent because it is the only one that is a lifecycle decision rather than a market or research finding. Every figure on this sheet is checkable; nothing here is inferred.
06Tie the elephant to the tree
Deck
Speaker · VO
The story first, the slide second. Tell it plainly: at Foxconn, H. L. Cheung — ex Apple headquarters — told our team, 'you need to tie the elephant to the tree.' Then unpack it in front of them: the elephants are the masters we just named. You cannot lead an elephant around on a leash the way you walk a puppy — a body of thought that took thirty years to earn does not follow you into a prompt. So you do not move it. You tie it to something that will not move: AI. Then say the thesis out loud — a real leader's thinking is timeless, and the only honest way to learn AI is to go back and understand the people who thought hardest about building things, then watch how well their thinking survives the new medium. That survival is the reason we bothered naming them at all: a track record of influence that large is evidence there is wisdom in it we can use. Three knots on this board, each one a master's idea tied to a place AI is weak. The rope is the same every time — 'this matters for AI because…' — and that is the gate: if you cannot tie the knot, the idea stays on the shore, on Act III's ocean wall, and you do not teach it here.
Figure
FIGURE — The metaphor drawn, not illustrated. Left: the trunk, a single misregistered black bar with three collars — AI, the one thing in the room heavy enough to tie to. Right: three masses rendered as dense process screens rather than pictures, because an elephant here is weight, not an animal; each is dated and named so the mass is a real body of work. The ropes sag under load and each ends in a collar on the trunk; the italic caption names the rope as the sentence every beat must finish. Bottom: one mass with no collar and a dashed rope trailing off the sheet — the untied case, which is where the old 'gate' rule now lives. Point at the ropes, never at the boxes.
Act II · Mirror
Ten mechanisms from inside
Speaker · VO
Act II is the heart of the talk. Ten mechanisms. Each one is a diagram, not a claim — students should be able to redraw it from memory. Warn them: after this act, 'the AI understands' becomes a sentence they cannot say without wincing.
Speaker · VO
Ask the room to count the r's in strawberry before you reveal the split. Then show that the model was never handed letters. Design consequence: every prompt is silently re-cut into someone else's alphabet before the model reads it at all. Source: subword tokenization (BPE-style) mechanics.
Figure
FIGURE — A side-by-side of two alphabets. Left: ten letter cells, the three r’s in accent, with the count 3 beneath — what you were handed. Right: three token tiles (str · aw · berry) with a question mark beneath — what the model was handed. The letters are physically inside the tiles and never inspected separately; that is the whole mechanism. Ask the room to count the r’s before you reveal the right-hand side.
Speaker · VO
Draw the parallelogram on the whiteboard if you can. Key line: this is literal vector arithmetic on the model's internal coordinates, not a metaphor. Then the honest limit — clean on hand-picked examples, messier the further you push it. The gap: the model has a consistent direction for a concept; you have a lived encounter with it.
Figure
FIGURE — A parallelogram plot on percentage coordinates: man, king, woman, queen, with the two displacement arrows drawn as one SVG so they stay parallel at any size. The point of the drawing is that the two arrows are the same arrow — one direction, reused. The three cards on the right are the real geometry, the honest limit, and the gap. Say the limit out loud: this is clean on hand-picked examples and messier the further you push it.
Speaker · VO
Point at the photo. You saw a dog before you finished the sentence. The model got a grid of numbered tiles plus position tags, and reasons over them like words in a sentence. That is why image models are confidently wrong about spatial detail — left-of, behind, how many.
Figure
FIGURE — A real photograph with a 14×14 hairline grid laid over it, beside the same image as the model receives it: numbered tiles, each a vector plus a position tag. The grid is the argument — you saw a dog before you finished the sentence; the model got 196 numbered squares in a known order and reasons over them like words. That is why image models are confidently wrong about left-of, behind, and how many.
Speaker · VO
This is the one slide in Act II where humans are winning. Feature discovery, then circuit tracing in 2025 — not just which features fire but the sequence of features triggering features. The Golden Gate demo: researchers forced one feature far past normal firing and the model became obsessed, at one point claiming to be the bridge. Same technique now checks pre-release models for deceptive internal features.
Figure
FIGURE — Three stages across the top (feature found, circuit traced, used before shipping), then the intervention itself: two bars showing one feature at normal firing and at ten times normal, with the feature-to-feature chain beneath. The photograph on the right is the concept that feature fires for. This is the one slide in Act II where the humans are winning — say so.
Speaker · VO
A human learns hot by touching a stove once. The model inherits the description, not the encounter — that is the symbol grounding problem. World-model research (JEPA-style) tries to close the gap and researchers are candid that it does not, by itself. Design mistake to avoid: treating confident language as verified experience.
Figure
FIGURE — A two-column comparison with the gap named down the middle. Left: one photograph and the numeral 1 — one contact, once, at about four years old. Right: four lines of text the model was trained on and the numeral 0 — no contacts, ever. The asymmetry between the two numerals is the mechanism; everything else on the slide is caption.
Speaker · VO
The classic panda-to-gibbon case. The noise exploits a narrow pattern the model over-relies on: invisible to your eye, decisive to the model. Then the hard consequence — if a system can be confidently fooled by input a human would not notice, model confidence alone is not evidence. Same asymmetry underlies prompt injection and jailbreaks.
Figure
FIGURE — The classic panda-to-gibbon case as an equation: photograph, plus a field of calculated noise printed in the three process inks, equals a second photograph. Under each, a confidence bar — 57.7% and 99.3%. Read the two bars, not the words: the model got more confident as it got more wrong.
Speaker · VO
Everything the model knows about this chat lives in the context window and is wiped when the session ends. Inside the window it feels like memory — total recall, instantly. Next session: zero idea yesterday happened. This is why agent designers build external memory instead of trusting the model to remember. State design fixes a structural limit, not a nicety.
Figure
FIGURE — Two context windows side by side. Left: bars filled in accent, tapering — full recall of this conversation. Right: the same six slots, all dashed and empty. The arrow between them is labelled with the only thing that happened: the session ended. Nothing was deleted; there was never anywhere for it to persist.
Speaker · VO
Every training example has a right next word to predict; abstaining was almost never the rewarded move. So in the gray zone, the finish-the-sentence habit wins. This is not a random glitch — it is the same mechanism that makes the model fluent and correct most of the time. The real fix is evidence checks, abstention thresholds and human gates, not asking it to be more careful.
Figure
FIGURE — Three reward bars: fluent and correct (full), fluent and wrong (most of the way), and “I don’t know” (almost nothing). The gap between the second and third bar is the entire hallucination mechanism — abstaining was never the rewarded move. The panel on the right is the real fix, with the thing that does not work greyed at the bottom.
Speaker · VO
A teenager's values come from years of family, friction and consequence. A model's come from two engineered layers: rater preference, and in some labs a written constitution drawing on documents like the UDHR. Judgment did not leave — it moved upstream. So 'the AI agrees with me' is a claim about raters and a document, not about lived ethical experience.
Figure
FIGURE — Eighteen years against two layers. Left: four bars stepping through family, friction, consequence, their own — time and answerability. Right: two labelled boxes, RLHF and a written constitution. The visual asymmetry is the argument: judgment did not disappear, it moved upstream to a rating process and a document.
17The mirror, held up
Deck
Speaker · VO
Close Act II slowly. Read the list. Then: that gap is not a flaw to patch — it is the reason judgment, memory and lived responsibility stay the expensive human part of the system. The Chinese line is now earned by evidence rather than asserted.
Figure
FIGURE — No figure: this is a full-bleed statement page, deliberately the quietest slide in Act II after ten diagrams. Read it slowly and stop talking at the end.
Act III · Ocean
Fifteen builders · four architectures · conflict
Speaker · VO
Act III opener. Fifteen real thinkers nested under four architectures. Every face still gets a named stop; the architectures are how you remember them. Slow down here — the act only works if the room is allowed to admire someone before it hears the rest. Say the governance line once, then stop defending it: nobody is banned here, and nothing is hidden. Do not preview the flaws on this slide.
Facilitation
18 · Act opener
Say: Fifteen builders under four architectures. You will love some before you learn the rest — that is the point.
Don’t: Preview Torvalds 2018, DHH 2021, or Dijkstra’s cruelty here.
Beat: Slow. Let the room settle into Ocean.
Speaker · VO
Hold this board up before a single name is spoken. The ocean stays wide on purpose — nobody is banned here. What decides is not my taste: it is two published documents, and I want the room to see that they are older and larger than this course. Point at the mesh and read the documents out loud: the UNESCO AI competency frameworks and the UNESCO ethics recommendation adopted by acclamation by all 193 member states in November 2021. Then point at what rests on the mesh and at what went through it. Say the honest part: the three ideas held here are held because they survive a floor written by people who have never met us; the two maxims that fell through are not evil, they are simply industry sayings that cannot survive contact with a room of fourteen-year-olds. If a student thinks something was thrown back unfairly, the mesh is public — go and read it and come back and argue with the document, not with me.
Figure
FIGURE — The net drawn as a net. Top: the ocean as a light process screen, everything still in the water. Middle: the mesh, a crosshatch band across the full sheet, labelled with the two real documents so the floor has authorship. Resting on the mesh, three dated masses in the deck's own press screen — the same mass mark used for the elephants in Act I, so the eye connects them. Below, two smaller marks with dashed trails, already falling back. Nothing here is a step in a process; it is one physical event, so the board can be photographed and understood without me.
Speaker · VO
Read only the top row first. Let them fall for the four ideas — properly, out loud, no hedging, no “but”. Give it time; the love is the mechanism, because an idea you love is the only kind that can change your shape. Then read the magenta line and say nothing after it. Do not rescue anyone. Do not rescue yourself. The bottom row is the last thing you read, and you read it flat. Two things not to do. Do not explain the diagram — if you explain it, nothing rewires. Do not put a thinker here you were only lukewarm about; the room can smell self-protection, so at least one of these must still be someone you love. The claim, if you are asked: an idea’s truth and its author’s decency are independent variables, and holding both at once is a skill, not a disclaimer. “Just be sceptical” is the coward’s version — it keeps you safe and teaches you nothing.
Figure
FIGURE — Four bonds running down the sheet: thick full-strength ink where the love was whole, cut by a magenta rule, then continuing thinner to a small square. Nothing reaches zero — the residue is the point, and falling out is never symmetrical with falling in. Armstrong’s bond is not cut: a cyan ring where the others take the slash, and below it the line forks into two, because a disputed technical position is not a disgrace.
Facilitation
20 · Ocean is wide (falling in / falling out)
Do: Top row with full love. Magenta line. Silence. Bottom row flat.
Don’t: Explain the diagram. Don’t protect yourself with a lukewarm example.
Claim if asked: An idea’s truth and an author’s decency are independent variables.
Speaker · VO
The published frameworks: UNESCO’s AI competency frameworks for students and teachers (2024), and whatever policy your own school has already written. Respect these first — this finale is a premium complement, never a substitute for school frameworks. Say the last line out loud: the assessment rule is theirs, and softening it is not ours to do.
Figure
FIGURE — Left: the published frameworks drawn as the ground rule in full-strength ink, with three posts rising from it (people first, process over output, staged by age) and the finale above them as a dashed cyan plate carried by the same posts — complement, not substitute, argued by the drawing rather than asserted in a bullet. Right: the assessment rule as a submission gate — raw AI output stops at the rule in magenta, AI as one source among others passes through in black and continues off the sheet, because the accountability travels with the student.
Speaker · VO
Adopted by acclamation by all 193 member states, November 2021. Human rights and dignity is the cornerstone — every other value sits under it. Note the environment clause: cost, energy and water are ethics questions here, not footnotes, which ties straight back to the fog board.
Figure
FIGURE — The document drawn as masonry rather than a rack: the three other values stand as columns, each on a short leg, and they all land on one full-width bar of solid ink. Under the bar, the cornerstone — human rights and human dignity — carries the whole sheet. The adoption facts are demoted to a dateline, because the structure is the argument, not the numbers. FIGURE — How the artwork translates the UNESCO Recommendation into an abstract composition. Central radiant silhouette (human rights and dignity): the illuminated figure at the centre is the literal and metaphorical cornerstone; everything stems from and wraps around it, reinforcing that human dignity anchors all technological progress. Interconnected geometric grid (peaceful and just societies): concentric rings, glowing nodes and structural lines radiate outward — not cold isolated technology but the social fabric, how data, systems and people live together in a connected, balanced ecosystem. Mosaic of varied patterns (diversity and inclusiveness): diverse geometric tiles, varied colour blocks and multi-faceted shapes sit inside the circular networks, standing for the breadth of human culture, perspectives and datasets — an ethical system is built on a multiplicity of voices, not a monolithic one. Flowing blue and green currents (ecosystem and environmental ethics): fluid organic waves with leaf motifs and water ripples sweep the lower half and sides, physically intertwined with the digital grid, bringing water, energy and land out of the background and into the core AI conversation. Warm golden light at the centre against cool teal and green makes the point that ethical AI is not about halting progress but harmonising technology, humanity and nature.
Speaker · VO
Do not define the piles. Show them one idea at a time and let the room hear the sentence that goes with each. Pile one, the deepfaked classmate's voice: this is not a debate, it is a rock, and I name it and set it down in front of everyone so they know rocks exist and what one looks like. Pile two is the one that earns their trust — Dijkstra's full formal-verification argument is true, it is one of the great arguments in the field, and it is not for this room this year. Say the throw-back sentence here, verbatim, so they hear that being released is not being dismissed. Pile three is Torvalds' linked list, which lands and gets taught tomorrow. The point of doing it with real ideas is that a student can disagree with a specific placement — and that is exactly what I want them to do.
Figure
FIGURE — Three piles, each carried by one real named idea instead of a definition, with the pile mark in the top corner: a rough block for rocks, an open ring for the release, a solid accent square for the landing. The verdict line under each is the sentence spoken out loud. Same three-column rhythm as the flaws board, so the two read as a pair.
24Where you may argue
Deck
Speaker · VO
One axis, and I want them to watch two real decisions land on it in opposite places. Start on the left, where I have the least authority and they have the most: Rust or C for this build — argue with me, bring evidence, I will change the plan in front of you, and that is not me being generous, it is the only way the small end of this axis works. Then walk to the right and drop the second pin: a student wants to feed a classmate's essay to a model to grade it. That one is not mine to decide and it is not theirs to argue, because the floor was set by UNESCO and by the school before any of us walked in. Say the shape out loud once — as the stakes rise my authority narrows to nothing, and as the room gets smaller their right to challenge widens to everything — and then never say it again; the two pins do the teaching. If they only remember one thing: the place where you may grill me hardest is the place where I am most likely to be wrong.
Figure
FIGURE — Two opposing wedges around one axis. Above the line, the authority that overrules thickens as stakes rise toward E; below the line, the room's right to challenge thickens as stakes fall toward M. The wedges are drawn, not labelled with numbers, because the taper IS the claim. Two dashed pins carry real decisions and land in opposite zones, which is what stops this being an org chart of my own process.
25Five of six is a fail
Deck
Speaker · VO
Two ideas, one inspection, and the room learns the gate by watching it fail. Take the top row first: Kapur's productive failure collects all six stamps, and I want them to see me actually check the boring ones — who is affected, can we ask why, does it admit what the machine cannot do. Then the bottom row, which is the whole point of the board: peer-grading a classmate's essay with a model collects three stamps cleanly and dies at the fourth. It does not get a score of three out of six. It does not get a discussion. It stops, because dignity and law are a floor and not a weighting. Say the sentence: five out of six is a fail, and it is a fail even when the idea is popular, even when it would save me an evening of marking. The last stamp is the one from Act I — if the sentence will not finish, the rope will not tie.
Figure
FIGURE — The gate as a rubber-stamp inspection rather than a rubric. Six stamps per row, named once in the header so the rows stay quiet. Row one collects all six in accent; row two stops dead at the fourth, which is drawn as a struck ring with the two remaining stamps left as empty dashed outlines — the visible emptiness is the argument. Verdicts sit at the right where a docket stamp would go.
Speaker · VO
Run one real case through the gauges in front of the room so the system is not theoretical. Verdict: split landing. The linked-list taste lesson lands and is taught. The mailing-list conduct is named, not modelled, with the 2018 apology included as part of the honest record.
Figure
FIGURE — One real case run down the four gauges on the left, with the verdict on the right: a split landing. Green mark for what is taught, accent mark for what is named but not modelled. The shape of the verdict — two marks, not one — is the thing to photograph.
Speaker · VO
Teach the exact sentence. Use it instead of 'banned' or 'wrong' when a real, true idea surfaces that is not for this age or this room yet. It teaches release without teaching denial — the idea still exists, still has a name, and can be picked back up later.
Figure
FIGURE — A statement page carrying the exact sentence to say, with three chips beneath: not banned, not wrong, released with its name still on it. Teach the sentence verbatim.
Speaker · VO
Say almost nothing here. Read the left column, then the right column, and let the room notice that they are the same four things. Nobody designed this twice — the shape shows up in the classroom and in the datacentre because both are systems that take in more than they can trust, hold it against something written down, sort it, and keep an escalation path for the calls that are too expensive to make locally. The one row worth stopping on is the last: the escalation policy is the boring, unglamorous part of every production agent, and it is the same question as who in this room gets the final call. If a student can draw the left column from memory, they can read an architecture diagram at work in ten years.
Figure
FIGURE — Two columns, four rows, hairline connectors and no arrows, no numbering, and nothing labelled 'pattern' — the identical geometry is the entire argument. Left column set in the deck's serif at reading weight; right column the same weight, so neither side looks like the explanation of the other.
Facilitation
28 · Shape repeats
Do almost nothing. Left column, right column. Pause on escalation = who gets the final call.
Speaker · VO
Internal teaching roster — faces credited on Sources. This is the roster, not the syllabus. Fifteen named stops, nested under four architectures — Loop, Representation, Temperament, Stack. We do not tour them as a celebrity census. We learn them as a system: each stop earns its place by a mechanism you can reuse when the names fade. Point at the board: four rows, fifteen faces, one question underneath. Then step into Architecture A.
Speaker · VO
Architecture opener. Read principle once. Name carriers. Do not teach masters yet.
Facilitation
30 · Opener A
Read once: Wrongness, contained and corrected, beats instruction alone.
Name carriers: Kapur, Pocock, Karpathy, Gagné.
Don’t teach them yet.
Speaker · VO
Manu Kapur, ETH / ex-NIE. Two-phase Productive Failure — order is the mechanism; contrast is the lesson. Honest limit: STEM, secondary+, fragile without Phase 2. AI fold — do NOT stop at gradient descent: training failure productive; deployment failure silent (no gradient, no memory); system failure designed vs undesigned. Student still has a feedback loop; model does not. Optional one-liners: Transformer, Anthropic, DeepSeek as named-failure inventions.
Facilitation
31 · Kapur — Productive Failure
Mechanism: Phase 1 attempt (will be wrong) → Phase 2 canonical + explicit contrast. Order is the design.
Honest limit: STEM / secondary+; fragile if Phase 2 is skipped.
AI fold (hit this):
- Training failure = productive (gradient).
- Deployment failure = silent (no gradient, no memory).
- System failure = designed (Erlang) vs undesigned (poisoned memory).
Don’t stop at: “Gradient descent is productive failure.” That’s the shallow mapping.
Optional depth (if time): Transformer / Anthropic / DeepSeek as named-failure inventions — one sentence each, not a history lesson.
Student move: “So I should always struggle first?” → Only when transfer matters and failure is safe. Unsafe struggle is not Kapur.
Speaker · VO
Grill-Me: one question at a time + suggested answer until tree closed. Backstory 30s: voice coach → XState → TypeScript → synthesis. Model trained to fill gaps not ask — Grill-Me is hard override. Link to plate next: Pocock is Kapur for domains where student cannot productively fail alone.
Facilitation
32 · Pocock — Grill-Me
Mechanism: One question at a time + suggested answer → accept / revise / reject until tree closed → then build.
Backstory (30 sec): Voice coach → XState → TypeScript → synthesis. He learned to hear the gap between what people say and what they need.
Honest limit: Friction; he narrowed Grill-Me himself.
AI fold: Model is trained to fill gaps, not ask. Grill-Me is a hard override.
Student move: “Can’t I just prompt better?” → Better prompts help; closed trees help more when stakes are real.
Speaker · VO
HIT HARD. Do not rush. Do not skip.
Facilitation
33 · Kapur × Pocock plate — **HIT HARD**
This is the structural gold of the whole Masterclass.
| | Kapur | Pocock |
|---|---|---|
| Who authors wrongness | Student | Teacher / interrogator |
| Domain | Constrained (STEM) | Open (requirements) |
| Phase 1 | Self-generated wrong answer | Manufactured suggestion to evaluate |
Say:
Two implementations of the same principle. AI as Grill-Me: the model generates plausible wrongness; you correct it; the contrast is the spec. Without Kapur, Grill-Me is a trick. Without Pocock, Kapur is only a classroom technique.
Don’t rush. Don’t skip under time pressure. Cut trivia before you cut this plate.
Speaker · VO
1.0/2.0/3.0 and leaky abstractions. Tiny loop before library miracle is scar tissue, not cute method — do not full-biography. Bridge: living proof architectures were earned under pressure. Full letter stays in teacher notes / Bonus, not read aloud unless dessert.
Facilitation
34 · Karpathy
Mechanism: Software 1.0 / 2.0 / 3.0; leaky abstractions; tiny loop before library miracle.
Human line (if room can hold it): Tiny loop is scar tissue, not a cute method. Tesla pressure, then teaching as the path he chose.
Don’t: Turn into celebrity bio.
Don’t: Read the full “letter” — that stays in your notes / Bonus.
Bridge: He is evidence the architectures were earned under pressure — not Pedagogy peer #3 for its own sake.
Speaker · VO
Gift → catalysts → talent. Benchmark=gift, product=talent. One breath each: GPT-4 RLHF months, Claude constitution, DeepSeek-R1 pure RL. Then admission plate next — do not skip.
Facilitation
35 · Gagné
Mechanism: Gift → catalysts → talent. Nothing automatic. Benchmark = gift; product = talent.
AI examples (one breath each): GPT-4 RLHF months; Claude constitution; DeepSeek-R1 pure RL. Same gift class, different catalysts.
Honest limit: Hard to operationalize socioaffective gift — why we avoid single IQ gates.
Speaker · VO
Bernard speaks this. Do not paraphrase into comfort.
Facilitation
36 · Gagné admission — **SPOKEN, NOT PARAPHRASED**
Say roughly:
I am giving you a spark. The catalysis that turns gift into talent happens after you leave this room. Practice, feedback, time, community — Gagné’s model requires them. A lecture can start that. It cannot be that.
Don’t: Soften into pep talk. Don’t: Apologize for the course. Do: Own the limit. You’re modelling honesty about tools — including this one.
Speaker · VO
Architecture opener. Read principle once. Name carriers. Do not teach masters yet.
Facilitation
37 · Opener B
Principle: Design the shape of the information; then behavior has somewhere to stand.
Speaker · VO
A+DS=Programs. Transcript vs designed record. Lead into quote plate. Honest limit: Pascal lost to C for systems — purity had a bill. Structure at scale: tensors/embeddings/attention still representation discipline.
Facilitation
38 · Wirth
Mechanism: Algorithms + Data Structures = Programs. Transcript vs designed record {goal, step, results, risks}.
Honest limit: Pascal lost to C for systems work — purity had a bill.
AI fold: Agents drift because they re-infer state from prose.
Speaker · VO
Leave on screen. Silence. Do not explain over it.
Facilitation
39 · Quote plate — **LEAVE ON SCREEN**
The fix is not a better prompt. It’s a better representation.
Silence. Count to five in your head. Then one line if needed: “Lost-in-the-middle is often a representation failure in a model costume.”
Speaker · VO
Linked-list good taste: special case disappears when representation changes. Split landing already in Governance — one sentence only on 2018. Invisible foundation 45s if time: models run on Linux+Git. Prompt-patch = bad list; state slots = pointer-to-pointer.
Facilitation
40 · Torvalds
Mechanism: Good taste = delete the special case by changing representation (linked list).
Split landing already practiced in Governance — don’t re-litigate 2018; one sentence: taste ≠ kindness.
Invisible foundation (if 45 sec): Every model run sits on Linux + Git. He never built a model; models run on his work.
AI fold: Prompt-patching is the bad linked list. State slots are the pointer-to-pointer.
Speaker · VO
Café 1959 sixty seconds: no pen, twenty minutes, shortest path; fame surprised him. Invariant + local steps. Agent tool-jumping is goto. Think before you prompt. Named cruelty on BASIC — split landing, rigor ≠ kindness.
Facilitation
41 · Dijkstra
Story (60 sec): Café, 1959, no pen, no paper, twenty minutes — shortest path. Fame surprised him; he kept writing the EWDs.
Mechanism: Invariant + local steps; program readable without running.
AI fold: Agent tool-jumping is goto in new clothes. Café method: think before you prompt.
Honest limit: Named cruelty about BASIC students — rigor and kindness are separable. Split landing.
Speaker · VO
Full 97/3 quote — truncation inverts him. Prompt often 97%; architecture 3%. 2023 confident nonsense → 2025 revised in public. Correction is practice. Eval literacy: design your measurement.
Facilitation
42 · Knuth
Full quote: Premature optimization… yet the critical 3%. Popular truncation inverts him.
AI fold: Prompt ≈ 97%; state, schemas, hard stops ≈ 3%.
Human line: 2023 confident nonsense → 2025 revised in public. Correction is a practice.
Student move: “So never optimize prompts?” → Optimize after you measure; don’t live in prompt cosplay.
Speaker · VO
Simple ≠ easy. Token abundance makes architecture invisible. Three questions before prompt: concerns, ownership, never-happen. Then generate. Free tokens, expensive confusion.
Facilitation
43 · Hickey
Mechanism: Simple (one fold, structural) ≠ easy (near you, subjective). Complect = braid.
AI age: Token abundance makes architecture invisible — worse than hardware abundance.
Take-away (on slide energy): Before prompt — (1) name concerns (2) who owns what (3) what must never happen. Then generate.
Student move: “But generating is free.” → Free tokens, expensive confusion.
Speaker · VO
Architecture opener. Read principle once. Name carriers. Do not teach masters yet.
Facilitation
44 · Opener C
What you optimize for becomes culture.
Speaker · VO
Least MY surprise after learning Ruby well — not universal least surprise. Track record vs fluent care. Sets up five voices + happiness question.
Facilitation
45 · Matz
Correction everyone skips: Least MY surprise — after you learn Ruby well — not universal least surprise.
MINASWAN: Temperament became norm.
AI fold setup: Thirty years of choices vs one fluent sentence of “I want to help.”
Speaker · VO
One pass. Then Matz question. No debate yet. These five are outside voices, not part of the fifteen — contemporary critics of AI’s social effects, brought in to set up Matz. Say so if anyone asks: no numbered stop, no portrait, deliberately. Turkle (MIT, on simulated intimacy), Newport (Georgetown, on convenience and capability), Harris (attention economy), Mitchell (Santa Fe, on RLHF and sycophancy), Mollick (Wharton, on upskilling versus deskilling).
Facilitation
46 · Five voices plate
One pass, no debate yet: Turkle · Newport · Harris · Mitchell · Mollick
Then Matz question:
The model makes you happy. What did you have to BECOME to earn that happiness?
If answer is “nothing”: “That’s not a feature. That’s a warning.”
Don’t open a therapy circle. One beat. Move.
Speaker · VO
Defaults are product ethics. Ghost setting is the policy that ships. Rails declined; philosophy reincarnated in agent frameworks. 2021 ban named not modelled.
Facilitation
47 · DHH
Mechanism: Convention over configuration; defaults = policy; ghost setting ships.
Reincarnation line: Rails declined; the philosophy is in every agent framework.
Honest limit: 2021 speech ban / departures — named, not modelled. Architecture lesson ≠ workplace edict.
Student move: “So defaults are evil?” → Defaults are powerful. Ask who chose them and for what.
Speaker · VO
Architecture opener. Read principle once. Name carriers. Do not teach masters yet.
Facilitation
48 · Opener D
Refuse above the model. Trust below it. Supervise beside it.
Speaker · VO
Two Zen lines. Interpreter refuses; model guesses. Nobody quotes him; everyone implements him. Bridge to sandwich plate after Ritchie.
Facilitation
49 · Guido
Two Zen lines: Special cases aren’t special enough… / In the face of ambiguity, refuse to guess.
AI fold: Interpreter stops; model continues. Hallucination is this design difference at scale.
Line: Nobody quotes him. Everyone implements him (schemas, hard stops, CAI honesty).
Speaker · VO
Total trust by design; bill arrived (memory vulns, NSA/CISA). llama.cpp chose small C on hot path. Model may be safest layer; foundation is not. Then sandwich plate.
Facilitation
50 · Ritchie
Design: Total trust — no bounds check by choice.
Bill: ~70% memory vulns; Heartbleed; NSA/CISA 2025.
AI fold: llama.cpp / hot path still chose small C. Model may be safest layer; C underneath is not.
Don’t C-shame students who like systems — teach the bill.
Speaker · VO
Draw the sandwich. Rust = compile-time refuse; Armstrong = runtime recover.
Facilitation
51 · Guido × Ritchie sandwich — **DRAW IT**
TOP Guido schemas / hard stops / abstention
MIDDLE Model constrained above, runs on trust below
BOTTOM Ritchie CUDA / drivers / C
Say: Nobody quotes them. Everyone implements them. They are the water. Then: Rust = refuse at compile time; Armstrong = recover at runtime.
Speaker · VO
Elevator spark 20s. Borrow checker. Open question: borrow checker for agent teams — leave hanging.
Facilitation
52 · Rust
Elevator spark if 20 sec. Borrow checker: many readers XOR one writer.
Open question: What would a borrow checker for agent teams check? Leave hanging.
Speaker · VO
Crash was never the problem — nobody noticed. Crash small, isolate, restart clean. Let it crash ≠ don't handle errors; separate happy path from recovery.
Facilitation
53 · Armstrong
Quote: Problem isn’t the crash — it’s that nobody noticed and nobody did anything.
Three rules for agents: crash small · isolate · restart clean.
Don’t: “Let it crash” = “don’t handle errors.” Separate happy path from recovery.
Speaker · VO
Four architectures is the memory structure. Then Conflict Map.
Facilitation
54 · Converged
Four boxes. One question:
Whose judgment is this, actually — and which architecture is under stress?
Point at the four names. Do not re-list all fifteen.
Speaker · VO
Do not fake harmony. Exercise always: (1) name a specific AI system (2) pick one axis (3) finger on a pole — not the middle (4) name which architecture is under stress — Loop, Representation, Temperament, or Stack. Poles 04–06 are vault-fed: Guess vs refuse, Prompt vs representation, Earned joy vs given comfort.
Facilitation
55 · Overview
Say: Do not fake harmony. These people did not agree.
Exercise (always):
- Name a specific AI system (ChatGPT homework help / school chatbot / agent that can email).
- Pick one axis.
- Put a finger on a pole — not the middle.
- Name which architecture is under stress: Loop / Representation / Temperament / Stack.
Speaker · VO
Fluency is not correctness. Both poles can be right same week at different layers — plan easy, gate irreversible tools on correct. Trap: the model answered fast so it must be right. Architecture B owns the right pole when cost of wrong is real.
Figure
FIGURE — hard-wired tension-01.jpg concept illustration of the axis.
Facilitation
56 · Easy vs correct
Key: Fluency ≠ correctness. Same person can pick both poles same week at different layers.
Trap: “The model answered fast so it must be right.”
Speaker · VO
Metal → control (Ritchie, Torvalds, CUDA). Orchestration → safety (Rust, Armstrong). Trap: fast-kernel culture applied to agent shared memory. Architecture D is this axis as a sandwich. Point at the layer before you pick the pole.
Figure
FIGURE — hard-wired tension-02.jpg concept illustration of the axis.
Facilitation
57 · Control vs safety
Key: Metal → control. Orchestration → safety.
Trap: Fast CUDA culture applied to agent shared memory.
58Never fail vs fail small
Deck
Speaker · VO
99.9% is a never-fail-flavoured claim; the 0.1% still needs a circuit breaker. Kapur failure lives in people (lesson). Armstrong failure lives in systems (incident design). Same word, two architectures. Aim high in lab; ship containment anyway.
Figure
FIGURE — hard-wired tension-03.jpg concept illustration of the axis.
Facilitation
58 · Never-fail vs fail-small
Key: Kapur failure = lesson (person). Armstrong failure = incident design (system).
Trap: “99.9% accurate” without a circuit breaker for the 0.1%.
Speaker · VO
PRIORITY if short on time. Left: model native — complete and continue. Right: Guido + Grill-Me + schemas + hard stops. Exercise: name a tool call in your life — guess or stop? Line: nobody quotes Guido, everyone implements him. Jailbreak = guess pole attacking refuse pole.
Figure
FIGURE — hard-wired tension-04.jpg concept illustration of the axis.
Facilitation
59 · Guess vs refuse — **priority if short on time**
Left: Model native — complete and continue.
Right: Guido + Grill-Me + schema + hard stop.
Exercise: “Name a tool call in your life. Guess or stop?”
Line: Jailbreak = guess pole attacking refuse pole.
60Prompt vs representation
Deck
Speaker · VO
PRIORITY if short on time. Left: another sentence (97%). Right: better state/structure (3%). Exercise: failing agent — next move sentence or slot? Echo Wirth quote plate. Transcript-as-memory is representation failure in a prompting costume. Architecture B owns this axis.
Figure
FIGURE — hard-wired tension-05.jpg concept illustration of the axis.
Facilitation
60 · Prompt vs representation — **priority if short on time**
Left: Another sentence (97%).
Right: Better state / structure (3%).
Exercise: Failing agent — next move: sentence or slot?
Echo plate 40.
61Earned joy vs given comfort
Deck
Speaker · VO
PRIORITY if short on time. Matz earned craft joy vs LLM given comfort. End on: what did you have to BECOME to earn that happiness? Do not rescue the room. If silence: hold the question, no forced out-loud answer. Do not moralize; invite judgment. Architecture C.
Figure
FIGURE — hard-wired tension-06.jpg concept illustration of the axis.
Facilitation
61 · Earned joy vs given comfort — **priority if short on time**
End on Matz question. Do not rescue.
If silence: “Hold the question. You don’t have to answer out loud.”
Don’t moralize addiction; invite judgment.
Act IV · Voyage
Habits that install judgment
Speaker · VO
Act IV is craft applied — the six modules compressed into habits. Nothing here is new theory; it is the masters' rules turned into things a student can do on Monday.
Speaker · VO
Do not narrate all twelve cells. Point at the second column and say: that is the same move in all three rows — a wrong answer, produced on purpose, early. Then the payoff line at the foot. Kapur's productive failure, gradient descent, and agent debugging are one loop; skipping the attempt teaches performance, not transfer. In training the wrong loss, the leaky split and the silent bug are the real curriculum. In agents, prompt patches do not fix a bad representation.
Figure
FIGURE — One path, walked by three different things. Four stations across the top (predict, attempt, observe, revise) and three lanes beneath: a student, a model in training, an agent in production. Same four dots, same spacing, same order in every lane; only the words change. The second column is the only one in colour — bigger dot, accent type — because the attempt that goes wrong is the same move in all three rows. Each lane loops back with a hairline arc, and the reason is written under the arc in italic, different per lane. Point at the second column and say one sentence: that is the instrument. Do not narrate all twelve cells.
Speaker · VO
2026 practice: specs with goals, boundaries, tests and ownership outperform one-line vibe prompts. Agents fail first on underspecification, not on capability. Grill-Me is how you get here: one question at a time, recommended answer, accept-revise-reject, until you can restate the whole tree. Only then stamp it closed and implement.
Figure
FIGURE — The six fields of a spec as six cells, with owner in accent because it is the field teams leave blank. The GRILL-CLOSED stamp at the foot is a gate: nothing gets implemented above it.
Speaker · VO
AIMA and CS188: define Performance, Environment, Actuators, Sensors — then say how state updates. A tool-agent without a performance test is chat with side effects. Make students write PEAS for something concrete before they touch a framework.
Figure
FIGURE — Four columns, each led by its letter at display size, because the letters are the mnemonic students will carry out of the room. The line at the foot is the whole slide: a tool-agent without a performance test is chat with side effects.
Speaker · VO
Expand states until a goal test succeeds. Modern agent planners still sit on this skeleton. Breadth-first: layer by layer, often shorter paths, frontier memory grows fast. Depth-first: dive and backtrack, low memory, can wander far from optimal. The trade-off is completeness versus memory, and path quality versus thrift.
Figure
FIGURE — The two search strategies drawn as dot trees. Breadth-first: rows of dots widening layer by layer, with the unexplored frontier dashed — you can see the memory cost growing. Depth-first: a single diagonal chain with unexplored siblings dashed — cheap memory, no guarantee of a good path. The dashes are the trade-off.
Speaker · VO
The Wirth fold, drawn. Left is what most agents actually carry: a chat log. Point at it and say the honest thing — there is no goal in there, there is only prose that a goal can be re-inferred from, freshly, every single turn, and re-inference is where drift comes from. Right is the same agent with five named slots. Nothing on the right is smarter; it is only decided. Then the smell: if your fix for drift is another paragraph of prompt, you are patching a representation problem with glue. Change the representation. And the test they can apply without me: draw your state on a whiteboard with no code on it. If you cannot, you do not have state, you have a transcript.
Figure
FIGURE — Two plates. Left, a field of hairline rules standing in for a transcript, with three questions written across it in italic — goal, step, risk — because those are the things the model has to go fishing for in prose. Right, the same information as five ruled slots with registration ticks at the corners, drawn as empty fields waiting to be filled. The argument is entirely in the difference between a texture and a rack.
Speaker · VO
Constraints drawn as an aperture stack, because that is physically what they are. Start at the source and say it plainly: on day one an agent's reach is everything you gave the process — every API, every file, every credential you happened to have in the environment. Then walk the four bars left to right. The allow-list is not a policy document, it is a hole; the schema is a narrower hole, because a typed call can be refused by a machine and a paragraph can only be refused by a human who is paying attention. The action gate is the one people fake — 'please confirm before deleting' inside a prompt is a suggestion to a text generator, whereas a deterministic approve-or-deny outside the model is a control. And the last bar is the Torvalds move applied to permissions: do not check for the bad combination, make it unrepresentable. Finish on the far right — three actions get through, and every one of them is one you chose. Alignment by construction means the narrowing is in the plumbing, not in the model's good manners. Add the standing rule: every byte of external text is untrusted input, because an adversarial example never looks adversarial to you.
Figure
FIGURE — A beam narrowed four times. Nineteen rays leave the source; each ink bar has a smaller aperture than the last, and a blocked ray simply stops with a dot where it hit — no crosses, no red, nothing dramatic, because in production nothing dramatic happens either. Only three rays reach the right-hand edge, drawn heavier and in cyan, and they are named. The bars carry the ink weight of press furniture on purpose.
69Scores are not truth
Deck
Speaker · VO
One number on one line, and everything a number cannot tell you drawn around it. Put a finger on the hollow ring first: that is what we believed before the model said anything, and if nobody in the room can state the prior out loud then the score has nothing to update. The arrow is the observation, and it is the noisy part — same input twice, different number, and the arrow's length is a guess about how much to move. The filled dot is the posterior, which is the only thing the model actually gave us. Now the two wedges, which is the part that changes behaviour: the same 0.62 is a trivial number when a wrong yes costs a wasted minute and an unacceptable number when it costs a person's place in a class, so the threshold is not a property of the model, it is a property of what breaks. Then the shaded band. Every serious system has one and almost no student project does: a range where the honest output is not an answer. Three exits, and name them as exits — refuse, escalate to a human, or go and gather more data. Close on the sentence: a confident wrong answer is worse than an abstention, and only one of those two is a design choice you can make today.
Figure
FIGURE — A score line, drawn as an instrument rather than a chart. Prior as a hollow ring, evidence as the arrow that moves it, posterior as the filled dot, and the two costs of being wrong as opposing wedges above and below the axis — the wedges taper, so the eye reads that the threshold slides with the stakes. The abstention band is the shaded interval across the middle with three exit paths dropping out of it, and it is deliberately the widest single element on the sheet.
Speaker · VO
The Karpathy fold, and the figure is the whole argument, so resist explaining it. On the right is what a student's first contact with AI usually looks like: three lines of import and a demo that works. It is not fake and it is not bad — it is illegible. On the left is a ruler with four graduations, and the ruler is the small loop you built yourself: you chose the features, so you know what the system can see; you built the dumb baseline, so you know what the easy half of the problem was worth; you then swapped in the real library and can name what it automated and what it hid; and because you own all three, the fourth graduation exists at all — you can judge an evaluation number instead of reciting it. Say the closing line and stop: the demo is only as legible as the smallest loop you own. If you have no loop, you are not evaluating a system, you are admiring one.
Figure
FIGURE — A measuring instrument laid against an opaque mass. Left, a ruled scale whose four graduations are the four steps, drawn as a real rule with major and minor ticks. Right, the demo as a dense process screen with no internal structure at all — nothing to read, on purpose. The eye supplies the argument: without the rule, the mass has no size.
Speaker · VO
Cost drawn as the staircase it actually is. Walk it from the bottom, where storage is wide and cheap and slow, up to the top step, which is narrow and gold and where every token you generate has to live for the instant it is generated. That top step is high-bandwidth memory, and its scarcity is not a software problem — it is a small number of fabs, a queue, and industrial policy, arriving on your invoice as a price per thousand tokens. This is Wirth's Law with a receipt attached: the software got heavier faster than the memory got cheaper, which is the same board we opened the deck on, and the sentence to say out loud is that the fog in Act I is priced in DRAM. Then point at the ground the whole staircase is standing on, because students never think about it: inference sits on a real grid and real cooling water, in a real place, usually a place that did not get asked. UNESCO put that in an ethics recommendation rather than an engineering one for a reason. Last, the tick on the right — your evaluation budget is finite too, so you cannot test the universe; design the cheap check that catches the failure you actually fear.
Figure
FIGURE — A memory hierarchy as a real staircase: three steps whose widths encode capacity per dollar and whose heights encode speed, the top step in process gold because that is the scarce one. Underneath, a single heavy ground rule labelled with the grid and the cooling water — the whole diagram rests on it, which is the argument. A finite eval budget is a short tick at the right, deliberately small.
72Defaults are ethics
Deck
Speaker · VO
Four switches, and every one of them is already thrown. That is the board. Do not read the labels out; ask the room a question instead — who moved these? Because somebody did, and it was not the user. The model default decides whether every answer is fast and cheap or slow and strong, and whether the conversation ever leaves the building. The tools default decides what the thing can reach on a bad day, and least privilege is a default, not a feature you add later. The tone default answers 'who is the human on the other side' before that human has said a word — a fourteen-year-old and a surgeon are not the same audience, and shipping one voice for both is a decision about who you were actually building for. And the safety default prices somebody's risk. Someone chose to refuse or to answer, and the honest question is whose risk was acceptable and who was in the room when that was decided. Then the line at the foot: the ghost handle is the position nobody will ever move it to. That is what a default means. It is not a setting, it is a policy that ships.
Figure
FIGURE — Four toggle tracks, each with the handle already parked and a dashed ghost handle at the position it will never reach. The tracks are drawn plainly, the ghosts are the whole figure — the emptiness at the far end is the argument. The safety track is in accent because it is the one where somebody else decided whose risk was acceptable.
Speaker · VO
The Erlang fold, drawn as the shape Erlang actually gave us: a supervisor above, workers below, and a restart. Say the uncomfortable part first — a multi-agent system is flaky by default, not by accident, because you have chained several non-deterministic things together and every link has its own way of going wrong. The move is not to prevent all six of these; it is to decide, in advance and in code, what happens when each one fires. Walk the branches: a timeout retries or switches model or abandons the plan; malformed output is re-asked with a schema and quarantined, never parsed hopefully; a failing tool trips a breaker and the system degrades instead of thrashing; external text is untrusted input every single time; and collapsing confidence escalates to a person, which is only possible if you built the escalation path when you were calm. Then stop at the sixth, which is the one people forget, and it is drawn differently on purpose: the crashed worker is trying to write its bad result into long-term memory, and the bar across that arrow is the entire lesson. A crash is cheap and recoverable. A poisoned memory is neither, because it will be politely recalled as fact for weeks. Close on the quotation: the problem was never that the process crashed.
Figure
FIGURE — A supervision tree. One supervisor, six workers, hairline branches, and the fifth worker struck through as crashed with a restart arc curving back to the supervisor — crash small, isolate, restart. The sixth branch does not end in a worker; it ends at the long-term memory drum with a heavy bar across the write, because that is the failure that does not announce itself.
Speaker · VO
The last board of Act IV, and it closes the loop the deck opened with. Four rules, one per master, and each one is drawn rather than quoted, because a rule you can picture is a rule you can use on Monday. Wirth is a rack of named slots — decide the parts and the data before you write any glue. Torvalds is a branch with the special case struck off, and the mark matters: he did not handle the edge case better, he deleted the condition by changing the shape of the data. Matz is the smooth path drawn over the staircase — the staircase works, and no human enjoys walking it, and enjoyment is not decoration when a person has to use the thing every day. DHH is the vertical slice through every layer, thin and complete and opinionated, shipped, as against a wide shallow tour of five frameworks that runs nothing. Then say the title out loud, because it is the actual thesis of the whole talk: install, do not idolise. Every one of these people has something in their record I would not defend, which is exactly why we tie the rule to the trunk and leave the person on the shore. If a student walks out with one of these four as a habit, the talk worked. One is enough.
Figure
FIGURE — Four marks, each a small diagram of its own rule rather than an icon: named slots with registration ticks, a branch with the special case struck out, a smooth curve drawn over a staircase, and a thin vertical slice through four layers. All four in the same 220-unit frame at the same ink weight, so they read as one printed sheet of marks.
Act V · Return
Witness · close · weather
Speaker · VO
Act V. Short. Do not add new material here. Go back to the fog board, then hand the room the closing question and let each person speak one witness line.
Speaker · VO
Same eight forces as the fog board; each well is a takeaway illustration (takeaway-01..08). Walk the mechanisms out loud.
77What done looks like
Deck
Speaker · VO
Ten checks. Read them as a self-test, not a rubric — ask students to silently count how many they can do right now, then say the number to nobody but themselves.
Figure
FIGURE — Ten checks in two columns, numbered, no boxes. Ask the room to count silently how many they can already do and tell nobody the number. It is a self-test, not a rubric.
Speaker · VO
Slow down. Say the three lines. Then: name one learning, one change to how you use AI, one direction. Host names one thing learned from the room — that part is not optional.
Figure
FIGURE — Three chips for the three things each person names, plus a fourth in reverse colour for the one the host owes the room. The fourth chip is not decoration; do not skip it.
Speaker · VO
The back page. Leave it up while the room empties — do not talk over it. The end mark is the printer's -30-, the old wire-service sign for end of story.
Figure
FIGURE — The front page inverted: the masthead furniture moves to the foot (thin-over-thick with the dateline between), the logo sits at the bottom, and the registration targets move to the top corners. The end mark is the printer’s –30–, the old wire-service sign for end of story, set as a plate numeral.
Speaker · VO
Addendum, rail 1 of 3. This plate is the swappable one: the shape is global weather, then the documents your own room answers to, then the conclusion. Teaching a class in Malaysia or Italy? Keep slide 1 and 3 almost as they are and replace rail 2 with that country's own instruments, quoted from the issuing body rather than from commentary. Read the four dates out loud and let the room notice they are all recent. Nothing on this sheet is a prediction: every item is a published document with a date, and any student can go and read all four. What you want them to feel is that the floor under this subject is being poured right now, by rooms they are not in — which is exactly why the why-question is the only durable skill.
Figure
FIGURE — One timeline, four instruments. Names and issuing bodies above the rule, dates and what each one actually binds below it, so the eye can compare like with like. UNESCO's tick is in accent because it is the only one all the others sit inside.
81China — AI as infrastructure
Deck
Speaker · VO
Addendum, rail 2 of 3 — the local plate, and the one to swap when this class travels. Every figure here is from the issuing body: the State Council's Opinions on deeply implementing the “AI Plus” action for the 2027, 2030 and 2035 targets, and MIIT's AI+ICT implementation plan for 2026-2028 for the network target. The three instruments along the bottom are all from 2026, and all three are worth naming out loud, because they are the answer to a student who thinks governance is something that happens elsewhere: an ethics-review measure that covers research as well as public services, agent guidelines issued jointly with CAC and NDRC, and a three-year plan that treats AI as utility infrastructure rather than as a product. Do not editorialise the targets. Read them, note that they are dated and published, and let the room draw the obvious conclusion: the ambition is stated in public, so the work can be checked in public.
Figure
FIGURE — A rising staircase of published targets, each step a dated slab of ink, with the three 2026 instruments listed as a dateline band underneath. Ascending because the commitments are staged in time, not because anything here is a forecast.
Speaker · VO
Addendum, rail 3 of 3. Say this one slowly and then stop talking. The argument: the floors move, the vendors change, the targets get revised — and none of that touches the question a student has to be able to answer, which is why this, and who is accountable for it. That is the whole reason the course is built on judgment rather than on tools. Keep this sheet exactly as it is when the class travels; it is the only part of the addendum that is not local.
Figure
FIGURE — A statement page. One rule, one line of ink, no diagram: after two sheets of other people's documents, the room should be looking at nothing but the sentence.
Reference
How to run it · sources
83Reference — how to run it
Deck
Speaker · VO
Reference card. Not projected during the talk.
84Reference — Acts I & II
Deck
Speaker · VO
Reference card. The full text behind Act I and Act II, for prep.
85Reference — Act III
Deck
Speaker · VO
Reference only. Ocean + Conflict Map.
86Reference — Acts IV & V
Deck
Speaker · VO
Reference card. Module map and the journey table.
Speaker · VO
Sources and image credits for internal teaching distribution. Roster portraits hard-wired under /portraits/ with credits listed on this slide. Credit sources; re-check rights before any public or commercial use. Diagrams are original to this deck.
Ocean facilitation pack
Deep Act III facilitation on live sheet numbers.
Deck: EN sheets 18–61 (Ocean through Conflict Map)
Shape: Governance → 4 architectures with 15 named stops nested → Converged → Conflict Map
Memory structure students should leave with: Loop · Representation · Temperament · Stack
Not the memory structure: fifteen résumés in equal depth
0. Your posture for the whole act
- Fall first, limit second. Admiration is allowed. Flaws are named, not banned, not glamorized.
- Architecture is how they remember; faces are how they care. Every master still gets a stop.
- Do not fake harmony. Conflict Map is the honesty test after the roster.
- Gagné admission is non-negotiable. Say it out loud. Do not soften it into “and of course you’ll practice later.”
- Throw-back sentence is verbatim:
> “Not wrong — not for this landing. Ocean still has it. We name it, measure it, release it.”
Timing (full Ocean, intensive)
| Block | Sheets | Time | Notes |
|---|---|---|---|
| Governance | 18–28 | 25–35 min | Do not rush throw-back |
| A · Loop | 30–36 | 25–30 min | Protect synergy plate 34 |
| B · Representation | 37–43 | 20–25 min | Protect quote plate 40 |
| C · Temperament | 44–47 | 12–15 min | Protect Matz question |
| D · Stack | 48–53 | 15–18 min | Sandwich plate 51 |
| Converged + Conflict | 54–61 | 20–25 min | One full worked axis minimum |
| Total | | ~2–2.5 h | Compressed: see §8 |
If you only have 60–75 min for Ocean: Governance skim (throw-back + gauges) → A full → Wirth quote + Stack sandwich → Conflict priority axes only → Converged Q.
1. Governance (18–28)
19–22 · Mesh
Say: The floor is older and larger than this course — EDB / school frameworks + UNESCO.
Don’t: Turn this into a compliance lecture. One clear read of the rails is enough.
Link forward: Environment / energy on UNESCO board → fog board, not a footnote.
23–27 · Deck, authority, gauges, worked example, throw-back
Worked example (26): Torvalds 2018 — split landing. Taste lesson lands; conduct is named, not modelled.
27: Teach the throw-back sentence verbatim. Practice once with a volunteer idea.
7. Conflict Map (55–61) — facilitation script
How to run each tension (56–61)
- 20 sec poles + anchors
- 20 sec two real systems on the line
- 30 sec heuristic
- 30 sec AI mirror
- Optional 60 sec: one student places their system
Close Conflict
None of the six settle permanently. Re-decide per system, per layer. Bridge to Voyage: habits install architectures as practice.
8. Compressed runs
75-minute Ocean
18 (30s) → 27 throw-back → 28 → A 30–36 full → 39 quote → 51 sandwich → 54 → 59, 60, 61 → stop
45-minute Ocean (exception only)
27 → 33 synergy → 36 admission → 39 → 51 → 55 overview + 60 only → Matz question spoken without full C
What you may cut first
Double-takes, full café story, Rust ecosystem, five-voice names (keep question), DHH cloud numbers
What you may not cut
33 synergy · 36 admission · 39 quote · throw-back script · at least one of 59/60/61
9. Common student moves → recovery lines
| Student move | Recovery |
|---|---|
| “So which master is right?” | None permanently. Pick a system and a layer. |
| “AI will make this obsolete.” | Then judgment gets more expensive, not less. That’s the through-line. |
| “I already use ChatGPT well.” | Show me your failure log. Name the gap. (Kapur.) |
| “This is just philosophy.” | Plate 39 and 50 are engineering. Representation and stack are load-bearing. |
| Defends idol / attacks person | Governance gauges. Split landing. Idea ≠ conduct. |
| Wants more biography | “Vault has depth. This room installs architecture.” |
| Uncomfortable at Matz question | Good. Don’t rescue. Temperament axis is working. |
| “Just add it to the prompt” | Sheet 61. Say the Wirth line again. |
10. Teacher self-check (before you walk in)
- ☐ I can draw Kapur×Pocock table from memory
- ☐ I can say Gagné admission without apologizing
- ☐ I can recite throw-back verbatim
- ☐ I can draw Guido/Ritchie sandwich
- ☐ I know my compressed cut list
- ☐ I will not fake harmony on Conflict Map
- ☐ I know which plate I protect if time dies (33, then 39, then 36)
11. Map to Voyage (Act IV) — say once at Conflict close
| Architecture | Voyage habit |
|---|---|
| Loop | Productive failure cycle · Grill-Me / spec before tools |
| Representation | State design · PEAS · constraints · features first · eval literacy |
| Temperament | Defaults are ethics · cost as constraint · taste rules |
| Stack | Supervise failure · hard stops · refuse-to-guess |
“Ocean named the architectures. Voyage installs them as habits.”
12. After-class (for you, not students)
- Vault deep-dives:
projects/iits/masterclass-*.md - lock:
masterclass-ocean--locked.md - Conflict bodies:
masterclass-conflict-map-.md - Karpathy letter: speaker-only unless dessert explicitly offered
If a student asks for “all fifteen in depth”: point to full spine / PDF Bonus path, not a promise to re-run Ocean as biography night.