School of Engineering · Playlist

AI: From Zero to Takeoff

From inspectable models to evidence about progress, prompting and choosing tools.

Understanding Artificial Intelligence
Types of AI
How a Language Model Works
The Pace of AI Progress
AI Winters: Boom and Bust
Prompting Strategies
AI Tool Guide

Artificial Intelligence/Start here

Understanding Artificial Intelligence

Start with a game small enough to inspect completely, then train a language model, test forecasts against evidence, and choose prompts and tools.

AI: From Zero to Takeoff

Start with a game small enough to inspect completely. Then train a language model, test forecasts against evidence, and open classroom activities built for specific disciplines.

These lessons are by Bryant Harrison, Murray State University. Each interactive activity includes its original Guide, presenter notes and printable guide. Measurements, scripted examples and sourced claims retain their original labels and snapshot dates.

Start the learning track Print or save the complete track

What AI is

Two demonstrations make the machinery visible. The first compares three ways to choose a move on the same tic-tac-toe board. The second builds from a small language model to reasoning and tool use.

Types of AI
How a Language Model Works

Artificial intelligence: A broad label for machines performing tasks that call for judgment, prediction, planning, perception, or language.
Machine learning: A part of AI in which behavior improves from data or experience. Neural networks are one family of machine-learning models.
Generative AI: Systems that produce new text, images, audio, or other content. Modern examples often use neural networks, but “generative” describes what a system does.
Models and agents: A language model predicts sequences. Reasoning training can improve multi-step work; an agent adds tools, memory, and a loop around a model.

These labels overlap; they are not one universal ladder. Rules-based systems, learned tables, neural networks, generative systems, reasoning methods, and agents describe different mechanisms or behaviors. Ask what was learned, what was written by people, and what the system can actually do.

Where AI is going

One demo tests your intuition against measured progress. The other tests confident predictions against seventy years of history. Read together, they support urgency with skepticism.

The Pace of AI Progress
AI Winters: Boom and Bust

Use the dates, sources, measurement conditions and qualifications inside each activity when weighing its evidence.

How to use AI

Practice writing a clear request, then choose the right tool for the job. These two activities show how your instructions and your choice of tools shape what AI can help you do.

Prompting Strategies
AI Tool Guide

Three ideas to carry forward

Ask what kind of system you are looking at. Human-written rules, learned values, neural networks, language models, and agents behave differently.
Match the tool to the job and the review you can provide. Choose a workflow whose results you can inspect.
Bring your discipline’s standards. A useful classroom tool starts from the learning goal, exposes its assumptions, and makes evidence available for challenge.

Explore the classroom example projects. Each includes a guide to its model, assumptions, and teaching use. Start with a learning goal and inspect the evidence before adapting a project for your class.

Artificial Intelligence/What AI is

Types of AI

Where does an AI's skill come from? Three architectures on one tic-tac-toe board, the methods that train them, and the model types the world sorts by.

Explore Types of AI; Interactive

Explore Types of AI

Compare written rules, a learned value table and a neural network on one tic-tac-toe board, then inspect the network's actual weights. Separate three questions people mix up: how a model is built, how it was trained, and what job it does — the last one linked to Hugging Face's own catalogue of model types.

Open Types of AI full screen Printable guide · Guide PDF

The activity runs in your browser without an AI service. Use its Guide button for instructions and Settings for presenter notes, presentation mode and reset. Open it full screen when you want more room.

Learning path

Symbolic AI (GOFAI). Eight rules a person wrote. It never learns, and an exhaustive search still finds no losing line.
Value table. One score per position, learned from game outcomes. Train it until the search can call it unbeatable.
Neural Network. Shared weights fitted to the table's frozen scores, with error measured on positions kept out of training.
Other architectures & training methods. The wider map: trees, ensembles, SVMs and the rest, and which of 1a–1c used which training method.
Model types. What a model does rather than how it is built — Hugging Face's own by-task catalogue, with a link to real models of each type.

Presenter notes and learning guide

Murray State University · Presenter guide

Types of AI

Demo 1 of 17 · Part one · What AI is · 20 minutes

You need: one browser and the demo’s index.html; it works offline. Start from a fresh page or press Settings → Reset. Settings → Presentation mode scales the page for a projector; Open Presenter Notes shows this guide in the app.

Learning goals

Each stage’s Takeaway, as the learner reads it:

1a. This opponent’s skill came from a person: changing its rules changes its play, and playing more games changes nothing.
1b. Reinforcement learning stores experience as a separate score for each position; only an exhaustive search over every game line, not a good record, earns the word unbeatable.
1c. Shared weights give the network an answer for a board it never practised on — which a table cannot do — and training on other boards makes that answer better. The error readings say how close the answer is to the teacher’s, not how well it plays.
1d. Architecture (how a model is built) and training method (how it learns) are separate choices: the same architecture can be trained more than one way, and the same method can train more than one architecture.
Model types. A model type names the job — what goes in and what comes out — not the architecture inside. One neural-network family, the transformer, serves text, vision, audio, video and multimodal types; generating images, audio and video also leans on a different family, diffusion models, which Hugging Face keeps in a library of its own.

Where this sits on the path

Demo 1 of 17 opens Part one. Nothing comes before it. It hands forward one idea to Demo 2 (How a Language Model Works): a model is a representation that can be trained, and architecture, training method and model type are three separate questions to ask of any AI.

One question for the session

Where does an AI’s skill come from? Three models play the same board and each got its skill a different way. The top lens sorts models two ways: Model architecture (how it is built and trained — 1a to 1d) and Model types (what it does).

“Architecture is how it is built. Training method is how it learns. Type is the job. Keep asking which one someone means.”

1a · Symbolic AI (GOFAI) — 4 minutes

Question

Can a program play well without ever learning?

Predict prompt

Ask the room to press a prediction: with only the First 2 rules on, can you beat it? Take a show of hands before anyone plays.

Activity

0:00 Press First 2 rules. Play one game; invite the room to find the fork. The Notice line and the panel name the rule that answered each move.
1:30 Press All 8 rules and play again. Point at the rule that fires each turn.
3:00 Open Details: the proof line reports the exhaustive search over every game playable against the ladder.

Expected observation

Two rules lose to a fork; eight rules have no losing line. The opponent explains every move by naming a rule, and nothing about it changes however many games are played.

Check answer: a person wrote every rule and its priority; playing did not change them. The wrong option (“it learned the rules by playing”) is refuted by First 2 losing the same fork forever.

Teaching limit

GOFAI means Good Old-Fashioned Artificial Intelligence. This decision list can be drawn as a tree; here people wrote the tests, whereas decision trees can also be learned. Say “this small example”, not “symbolic AI”.

1b · Value table — 5 minutes

Question

Can experience alone teach a blank table to play?

Predict prompt

After one burst of 5,000 games, will Era 1 be unbeatable or still beatable? Press a prediction.

Activity

0:00 In Era 0 press Show move scores: every square reads 0.00. Play one short game; it moves at random.
1:00 Press Train the AI. Read the change report under the training controls: recorded scores before and after the burst.
2:30 Train again until the Notice line says verified unbeatable (usually burst 3 to 5). Switch the Opponent era back to Era 0 and forward again: the report follows the selection.
4:00 Open Details for the proof line.

Expected observation

Era 1 is still beatable; the scores moved but a losing line remains. A later era has no losing line in either role. Human play never changes the table; only the burst does.

Check answer (“Does playing the AI above train it?”): no — here only the training burst updates the table; your games are scored, not learned from. Say the caveat out loud: that is this demo’s choice, not a law. The same table could have learned from every game you played, and real assistants can use what people type to train later versions, usually behind a setting. Details cites two vendors’ own pages (fetched 2026-09-16).

Teaching limit

A prediction of “unbeatable” is a claim about the table, settled by the exhaustive search, not by whether the room beat it: losing to a beatable era looks exactly like losing to an unbeatable one, and the reveal line says which one the search found. The table has one separate score per position, so it carries nothing to a different situation. Tic-tac-toe is small enough to tabulate; chess and language are not, which is why the next stage shares parameters.

1c · Neural Network — 6 minutes

Question

Can one set of shared weights score boards it never practised on?

Predict prompt

After training, will held-out error be about the same as practised error, or clearly higher? Press a prediction; the reveal quotes the measured numbers.

Activity

0:00 Press Create learning examples. A fresh table learner plays 20,000 seeded games; its estimates freeze as targets. This is the one slow step.
1:30 In the network view, pick a connection layer and a neuron; read one weight and its initialized value.
2:30 Press Train network. Compare the weight with its initial value, then the practised and held-out errors, then the separate playing score. Show move scores shows predictions for legal moves.
5:00 Optional: press Ultimate tic-tac-toe above the board to show a representation that cannot see the whole game; Standard board brings the network back.

Expected observation

The network is a feedforward multilayer perceptron (29 inputs, 28 and 18 ReLU units, one output) fitted by supervised learning to targets another model produced. Backpropagation supplies gradients; updates are normalized gradient steps. Shared weights give an answer for a held-out position; the error readings say how close, and neither is playing strength.

Check answers: error measures agreement with the frozen table’s scores, which are themselves estimates; play is measured separately. 1b stores one number per position; 1c reuses 1,381 shared parameters for every board.

Teaching limit

Lower error means closer agreement with the teacher, not guaranteed playing strength; no automatic unbeatable claim is made. A position the fresh teacher never visited stays at zero — a limit of the examples, not proof the position is neutral. During play the network calculates its own scores without consulting the frozen table. Activations are responses to a board, not stored weights; a positive weight is not a good move.

1d · Other architectures & training methods — 3 minutes

Question

What else can a model be, and how else can it learn?

Predict prompt

None on this stage. 1d is a map, not an experiment; nothing on it runs a result a prediction could be checked against. If you want a show of hands anyway, ask which of 1a, 1b, 1c learned from inputs paired with target answers, then press the Supervised learning card.

Activity

0:00 The stage has three parts behind three buttons. Under Training methods, press each card. The card faces do not say which model used them; the press does: supervised → 1c, reinforcement → 1b, no training → 1a; the other three light nothing. Point at the one-line “used for” notes: pretraining is self-supervised, fine-tuning and distillation are supervised, RLHF is reinforcement — stages, not extra methods.
1:30 Switch to Other architectures and press a card: the line beneath says what that model stores and what it computes — linear models, decision trees, ensembles, nearest neighbours, support-vector machines, probabilistic models. Search and planning are named separately as problem-solving approaches.
2:15 Switch to Learning with no weight change: in-context learning, retrieval, search at decision time. Tie it to 1b’s check: a model can appear to learn in a session without changing. Press Search at decision time: 1a’s fork rules look one move ahead every turn; the demo’s exhaustive search checks players and never picks a move.

Expected observation

Architecture and training method are separate columns. 1c is a network trained by supervised learning; a network could equally be trained by reinforcement, and a table could be fitted to supervised targets.

Check answer: no — the network is an architecture; supervised learning is how this one was trained.

Teaching limit

The cards are textbook categories, not a complete catalogue, and real systems combine them. Self-play is not self-supervision: 1b learns from rewards. Gradient training is not evolution.

Model types — 3 minutes

Question

If “type” means the job, what goes in and what comes out?

Predict prompt

None on this stage: it is informational, and nothing on it runs a result a prediction could be checked against.

Activity

0:00 Switch the top lens to Model types. Press three cards the room will recognise: text generation, text-to-speech, speech-to-text (automatic speech recognition). Each shows what goes in and comes out, then the catalogue’s own definition.
1:30 Press Image-to-3D: the job people call image-to-mesh. Follow Browse models on one card if you are online; the demo itself never fetches.
2:00 Scroll to Beyond one catalogue: eight major types the Hub list does not show — world models (Genie 3), video with sound (Veo 3.1), robot control (Gemini Robotics), protein structure (AlphaFold), weather (GenCast), background removal (NVIDIA Broadcast), character animation, free-viewpoint video — each linked to the model’s own page.
2:30 Scroll to the whole catalogue: every type Hugging Face lists, in its groups, with the model count on the fetch date.

Expected observation

A type is named by its input and output. Text generation and image-to-text are different types that can share one architecture family; tabular types are the visible exception, where the catalogue names tree and boosting libraries. The link back to 1c: a small network scoring boards is the same idea as the networks behind most of these types.

Check answer: not necessarily — type is the job; one transformer may generate text and another caption images.

Teaching limit

The taxonomy and the counts are Hugging Face’s, fetched on the date shown on the stage; counts move daily. Other catalogues cut the jobs differently. The stage does not run any of these models.

Misconceptions

“A program that plays well must have learned.”1a plays unbeatable tic-tac-toe with no learning at all. Skill can be written down by a person; learning is one way to get it, not the definition of AI.

“A long run of draws proves it is unbeatable.”Sampled games can miss a line. The app calls an era unbeatable only after searching every reachable game line in both roles. Open Details for the counts.

“Playing against the table trains it.”Not here: human play is scored, not learned from, and every era is a frozen copy. But it is a design choice, not a rule — real systems can and sometimes do train on what users type.

“I lost, so it must be unbeatable.”Beatable means a losing line exists, not that you found it. The search settles the verdict; your record only says how you played.

“Low error means it plays well.”Error is agreement with a teacher whose estimates are imperfect. The playing score is a separate, fixed-seed measurement, and this tab never claims unbeatable.

“Neural network is a kind of learning.”It is an architecture. 1c’s network was fitted by supervised learning; the same weights could be trained by reinforcement. Press the method cards in 1d.

“Same architecture, same type.”Type is the job. Hugging Face files transformers under text generation, image-to-text, speech recognition and more. Ask what goes in and what comes out.

Discussion questions

Which model stores separate values, and which shares weights? What does each choice make easy and hard?
Why does lower held-out error not prove unbeatable play? What measurement would?
Name an AI product you use. Which type is it, what might its architecture be, and how do you think it was trained? Which of the three can you actually tell from the outside?
The Ultimate extension: what information does the inherited representation leave out, and why can more training not put it back?

Presenter controls

Guide (header) reopens the learner card. Settings has Open Presenter Notes, Presentation mode and Reset, then a playing preference, a rehearsal seed, the self-test and About.
Reset is in place: stage 1a, First 2 rules, move scores hidden, Era 0, records and predictions cleared, type cards closed. Presentation mode and the playing preference stay.
Rehearsal seed: type a seed and press Reset; training then unfolds identically, so a rehearsal matches the lesson.
Run the full self-test re-runs the exhaustive search on the selected stage and prints it in Settings and the console.
Deep links: #stage-1 to #stage-5; the older #rules, #learning, #neural still work.

Bryant Harrison · Murray State University · Types of AI · single-file offline demo · no accounts or runtime network requests · the Model types stage links out to huggingface.co, fetched 2026-09-16

Artificial Intelligence/What AI is

How a Language Model Works

Train a tiny language model, test worked solutions, and explore a tool-using loop.

Explore How a Language Model Works; Interactive

Explore How a Language Model Works

Train a tiny transformer and inspect its predictions. Compare direct answers with worked arithmetic, then explore an agent loop with real tools and a clearly labeled scripted policy.

Open How a Language Model Works full screen Printable guide · Guide PDF

Learning path

Train a tiny transformer and inspect its predictions. Compare direct answers with worked arithmetic, then explore an agent loop with real tools and a clearly labeled scripted policy.

Presenter notes and learning guide

Murray State University

How a Language Model WorksDemo 2 of 17 · Part one · 20 minutes

You need: one browser and index.html. It runs offline. For projection use Settings → Presentation mode. Start fresh: the untrained model is useful evidence.

Learning goals

What a learner knows at the end of each stage — these are the three Takeaway cards, verbatim.

LLM. Training changes the parameters; inference runs the frozen parameters forward to score every possible next token. A small model trained on a little text learns the shape of the language long before it learns to say anything true.
Reasoning. Room to write intermediate steps lets the same machinery reach answers it cannot produce in one jump — a real, measurable gain on this task, and not the same thing as understanding arithmetic.
Agents. Tools and a loop turn prediction into action, and the context window bounds what the loop can still see — a run can fail for want of memory, not for want of ability.

Where this sits on the path

Demo 1 showed models that play and handed forward one idea, in its own words: a model is a representation that can be trained, and architecture, training method and model type are three separate questions. Its last stage ended on the transformer; this demo trains one. It shows a model that predicts text, then what it takes to reason and to act. It hands forward two things: a model reads its context (Demo 5, Prompting) and a score is a measurement under conditions — the exam rows here — which is how Demo 3 asks you to read published benchmarks.

One question for the session

What changes when a text predictor is given more structure, more checking, or tools? This page shows three bounded mechanisms: a tiny language model learning next-character prediction, arithmetic experiments with written work and checking, and a tool loop with a scripted action selector. Use Details for the deeper architecture and implementation notes.

“Keep asking what the program learned, what it is calculating now, and what part is a teaching fixture.”

LLM · Language model — 6 minutes

Question

How can adjustable numbers turn recent text into a next-character prediction?

Predict prompt

After all four eras, can it do simple math, like 2 + 2? The answer is no: its vocabulary has no digits, so 2+2= reaches it as blank spaces. Have someone type 2+2= into the prompt box to see it. The reveal says so on the card once Era 4 finishes, and Stage 2 trains the same kind of model on sums.

Activity

0:00

Open LLM. Use Tokenize text and Inspect parameters on an untrained prompt. The numbers begin as random values.

1:30

Choose Train the model. Read one untrained sample, then narrate the loss and samples as the first era completes. Use Try next-token prediction to compare likely next characters; change Temperature only after the audience sees the ranked choices.

4:30

Choose Trace attention. Select a character and show which earlier positions receive weight for that prediction. It is a live intermediate value, not evidence that the model has a mind or intent.

Expected observation

Training changes the weights, loss declines, and the model’s next-character distribution changes. Its samples can become more patterned while still being limited by one short, original story.

Check-yourself answer key

Era 4 writes more like the story than Era 0. What changed? Its parameters. Not a lookup of stored passages (nothing is stored to retrieve — the whole model is 42,458 numbers) and not a bigger vocabulary (Tokenize text shows the same 26 characters at every era).

Teaching limit

This is a character-level transformer trained from scratch on a tiny corpus. It demonstrates a transformer learning mechanism, not general language ability or a claim about every frontier system.

Reasoning · Worked steps — 8 minutes

Question

How do written intermediate steps and checking change a small model’s behavior on addition?

Predict prompt

Same exam, which scores higher — answers-only or worked steps? Take the vote before pressing Run experiment; the card echoes each guess and settles it from the two scores on screen.

Activity

6:00

Open Reasoning and choose Train and compare. Compare direct-answer training with worked column arithmetic. The models share core width, depth, and heads, but their contexts, seeds, and update counts differ to fit their different text formats.

8:30

Choose Compare exam answers. Read a live attempt and an exam row. The useful comparison is the presence of written intermediate work, not a universal performance promise.

10:30

Optional slow extension: Try more attempts. Run 1, 2, 4, and 8 attempts with the same frozen under-trained model. The arithmetic checker rejects malformed work and accepted chains vote. More attempts add inference-time computation, not training.

Expected observation

Written steps provide smaller targets for the model to predict. Sampling plus checking can improve the chosen answer even when weights stay fixed.

Check-yourself answer key

The worked-steps model scores higher. What does that show? That generating intermediate steps gives the same machinery more room to compute. The card’s two wrong options are wrong in opposite directions and both are common in a room: “it understands arithmetic” (too strong — both are next-token predictors, and the fine-tuning is supervised, not reinforcement learning) and “nothing changed, both just predict the next token” (too weak — the two exam scores on the page are the evidence against it). Give both corrections equal time.

Teaching limit

The checker validates column arithmetic from the addends. The optional accepted-example round also retains chains matching the generated answer, then uses ordinary next-token updates on those examples. That is rejection-selected supervised fine-tuning, not a policy-gradient reinforcement-learning update.

Agents · Tool-using loop — 5 minutes

Question

What does a system need in addition to text prediction to take actions and use results?

Predict prompt

Same task with Visible context set to Tiny — does the run still finish? Take the vote before the tiny run; the card settles it when the run ends.

Activity

14:00

Open Agents, then choose You choose the actions. The calculator, search, reading, notebook, observations, and distractor consequences all run locally.

16:00

Choose Watch the scripted policy. Point out the calculator’s deliberate bad syntax and its genuine error message, then the later corrected call. Open the raw context to see what the action selector can read.

18:00

Choose Tiny · last 450 characters, Restart run, and Watch the scripted policy again. The goal leaves the visible context and the automatic run stops.

Expected observation

Tool results become new text in the context. The small window changes what the selector can use, so the same task can fail when its goal is no longer visible.

Check-yourself answer key

On Tiny the same policy stops without finishing. Why? The goal scrolled out of the visible context. Not a weak policy (it is the same hand-written policy and it finishes on the full context) and not the calculator error (that error is real, deliberate, and recovered from in both runs).

Teaching limit

The action selector in this activity is a scripted policy, including the calculator mistake and retry; it is not a learned model. The loop, local tools, observations, and context truncation are the genuine mechanisms on display.

Optional slow experiments and close

Let Train the model and Train and compare continue while you discuss their live displays; each can take a few minutes. Try more attempts and Fine-tune on accepted examples are optional extensions. Do not promise a particular score or curve: the page uses seeded, bounded demonstrations.

“A language model predicts. Reasoning experiments add written work, checking, or extra attempts. An agent adds a loop and tools; in this activity, its action choice is deliberately scripted.”

Which claim came from a learned model, a fixed calculation, or a scripted policy?
Why can extra attempts help without changing weights?
What would need to change before this agent activity used a learned action selector?

Misconceptions to expect

It is just predicting the next token, so it cannot really compute anything.It does only predict the next token — and the worked-steps model still scores measurably higher on the identical exam. Prediction with room to write steps is computation. The two exam scores are the evidence; show them, do not argue it.

The worked-steps model understands arithmetic.Same architecture, different training text. It reaches more right answers; nothing on the page supports comprehension, and the fine-tuning is ordinary supervised training on its own accepted examples, not reinforcement learning.

Attention shows what the model is thinking.Attention weights are real intermediate values from this computation — a routing pattern for one prediction. They are not intent and not an explanation of why the answer came out as it did.

Frontier models are this recipe plus scale.The shape of the computation is the same as a decoder-only transformer. The scale, the data, and much of the training pipeline are not, and scale is not a detail. This page demonstrates selected mechanisms, not a miniature of every production model.

The agent got smarter or dumber between the two runs.It is the same scripted policy both times. What changed is what the loop could still see. A run can fail for want of memory, not for want of ability.

Presenter controls

Guide opens from the header and on entry. Settings contains Open Presenter Notes, Presentation mode, and Reset. Reset puts the whole demonstration back to its fresh state in place, without reloading the page: the transformer is untrained again, every era is gone, the Reasoning models and exam tables are cleared, and the Agents run is back at turn one on the full context. Presentation mode stays as you set it. Open printable guide opens the rendered copy of this same guide.

Artificial Intelligence/Where AI is going

The Pace of AI Progress

How far has AI come, how would you know, and what can two points tell you about next year?

Explore The Pace of AI Progress; Interactive

Explore The Pace of AI Progress

Five sourced benchmark series to estimate and reveal, four benchmark families with the reasons each generation was retired, fourteen real achievements ranked by acceptance status and the AI's role, and three forecast rules that fit the same two points.

Open The Pace of AI Progress full screen Printable guide · Guide PDF

Learning path

Presenter notes and learning guide

Demo 3 of 17 · Part two · 22–28 min full, 10 min cut

The Pace of AI Progress — presenter guide

One HTML file, offline, no AI at runtime. You need a screen to project or share and a mouse or touchscreen for the drawing in stage 1. Nothing else.

Learning goals

What the activity itself demonstrably shows, one per stage:

On the tests the field cites most, the change since 2023 is measured in multiples, not points — and every series stops where its published record stops, conditions attached.
A benchmark is retired when it stops measuring what it was meant to; being beaten is one of four routes, and the replacement often counts something new — so a rising line across generations is not one trend.
An achievement is an event with an acceptance status and a stated AI role; read the rung before the headline — every rung from Nobel Prize to retracted has a real 2023–2026 occupant.
A gap in published history can be estimated; a future cannot — three rules fit the same two points, and only the next measurement chooses between them.

Where this sits on the path

Demo 2 showed a language model being measured — a score is a measurement, not a property of a mind. This demo hands that forward to published measurements: read one with its conditions, and do not extend it past where it stops. Demo 4 takes the question this one raises — progress this fast invites “is it a bubble?” — and answers it with history rather than a script.

One question for the session

How far has AI actually come, and how would you know? The honest answer is large and specific: multiples, on named tests, under stated conditions, with the boundary on the same slide.

The one sentence to land: the change is bigger than most rooms guess, and a number is only a result when it comes with its conditions. Both halves, in that order.

The three facts worth a projector

Ninety seconds each, no benchmark literacy needed.

Fact	The move	Why this one
GDPval — 1,320 real work deliverables, 44 occupations, graded blind by professionals from those occupations	12.3% → 70.9% in ~13 months	Nurses, lawyers, accountants, engineers, journalists rated the machine's deliverable as good as or better than their own about seven times in ten. Vendor-built — say so — with the scores republished by Epoch AI.
METR time horizon — the length of task, in human working time, a model finishes half the time	3.5 min → 320 min in 32 months	The unit needs no explaining. An independent non-profit, same methodology both ends, and intervals nowhere near overlapping: 6.9 min upper bound then, 170 min lower bound now. Pair it with the doubling time — ~7 months over the whole history, 88.6 days since 2024.
FrontierMath Tier 4 — 43 problems Epoch says take a researcher days	0% → 97.6% in 20 months	And on FrontierMath-Erdős — 68 genuinely unsolved problems, released the same month — the same model gets 2 of 68. Pace and boundary in one image, and it pre-empts “the tests just got easier”: the administrator built a harder one and the model fell off it.

Fourth, if a room wants money: ARC-AGI-1 at 87.5% cost about $4,560 per task in December 2024; today 98.5% costs 28 cents, and ARC-AGI-2 — the harder successor — reaches 95% at $1.12. A harder test, a higher score, roughly four thousand times less money.

The shape of the session

Measure 10 min (two series drawn with the room, three offered) · Retire 6 (one family opened properly) · Achieve 6 (top rung, then bottom) · Forecast 4 (drag to year 8, then the boundary card).

The ten-minute cut: stage 1 with the METR series only, stage 3 filtered to Nobel Prize and to contested, then the last paragraph of stage 4 — magnitude, acceptance and the anti-extrapolation point, without the benchmark-family detail.

Stage 1 — Measure

Question. How much did the most-cited AI measurements move, and could you have guessed the size of it?

Predict prompt. “GPT-4 managed tasks that take a person three and a half minutes. Where is that by late 2025 — ten minutes? An hour?” Take numbers out loud, then draw the room's median, not your own.

Activity. Do the GDPval series yourself so the room sees the mechanic, then hand over the METR series and take a show of hands before Reveal. Drag across the shaded region, or focus the chart and use Up/Down. Choose a series jumps between the five.

Expected observation. Almost every room draws below the published line, on every series. Name that as a fact about the measurements, not a failing of the room — and do not let it become “draw higher next time”, the error stage 4 exists to kill. The reveal's what would have broken this line list is the payload: a successor released inside the gap, a harness change, a problem set corrected in 42% of its items.

Second half. Ask what the hatched gap between the last dot and the snapshot date means: nothing was measured there, by this page.

Check-yourself answer key. “Your line was below the published result. What does that show?” — That the results moved faster than the room guessed, on these tests under these conditions. Not that the room can forecast by drawing further.

Teaching limit. Five series are a teaching selection, not a survey. The ratio at reveal compares an endpoint with one selected report; it does not grade anyone's reasoning.

Stage 2 — Retire

Question. Every widely cited test gets replaced. Were they all simply beaten?

Predict prompt. “SWE-bench Verified was the standard AI coding benchmark, and it has been retired. What was the top score when its own co-author stopped reporting it?” Let the room guess high: 76.80%, and the reason is not the score.

Activity. Open Software engineering first; it breaks the tidy story. Read OpenAI's own words off the card: 59.4% of a hard 27.6% subset had flawed tests, and every frontier model could reproduce the human bug fix from memory. Then ARC-AGI, for the generation that changed the unit.

Expected observation. Four routes, and only one is “beaten”. Rooms assume contamination is rare; the Terminal-Bench leaderboard carries a canary string begging not to be trained on, which lands it better than any argument.

Check-yourself answer key. “Terminal-Bench went 2.0 → 3.0 → 4.0. What does the higher number tell you?” — Only that the old version stopped measuring what it was meant to: 4.0 has 66 tasks to 2.1's 89, and one model scores higher on 4.0.

Teaching limit. Four families, not the field. The status and route chips are labelled Reasoned because they are this page's reading of the quotation beside them — a reader is invited to disagree, which is why the quotation is shown.

Stage 3 — Achieve

Question. Which of these happened the way the headline says, and how would you tell?

Predict prompt. “The news said AI solved a Millennium Prize Problem. On a ladder from Nobel Prize down to retracted, where does that sit?” Take answers before filtering.

Activity. Filter to Nobel Prize — one card, and it took years. Then Announced, contested — one card, the biggest headline on the page; read its Who checked it line aloud, in full. Then Retracted: the same organisation nine months earlier.

Expected observation. The rung ordering inverts the news ordering: the loudest claims sit at the bottom, and the one with a prize is two years old and not a language model.

Say this about Navier–Stokes, precisely. An AI system produced a proof, checked in Lean, that closed the last gap in a strategy two human mathematicians had built — for one of the four statements Clay accepts as resolving the problem. Clay's page still reads “Active”; it is not peer-reviewed; Clay's rules need publication plus two years plus general acceptance, so no prize was possible before 2028 whatever anyone thinks of the proof; OpenAI states it does not intend to claim it; the priority dispute is unresolved; and statement (C) leaves (A), the unforced case everyone means, open.

That long version is not hedging — it is a better story than the headline, and it survives the one person in the room who knows the field. If you shorten it, shorten it to: a real, checkable advance on one of the four statements Clay accepts, with AI closing a specific gap inside a human proof strategy. Never to “AI solved a Millennium Prize Problem”.

Check-yourself answer key. “Which rung does the September 2026 Navier–Stokes result sit on?” — Announced, contested.

Teaching limit. Fourteen cases, weighted toward mathematics because that is where the strongest and the most contested recent claims are. The ladder ranks how thoroughly a claim has been checked, not how important it is.

Stage 4 — Forecast

Question. What can two measurements tell you about next year?

Predict prompt. “A score went 20 to 30 in a year. Eight years on, how far apart can three rules that fit those points be — a few points, or a factor of ten?”

Activity. Drag to year 8: the compound rule leaves the scale entirely, the saturation rule is still under 75. Then read the boundary card — five sentences, deliberately short.

Expected observation. A room that has just watched five series move by multiples wants to extrapolate; this is the stage that refuses. Say the reconciliation out loud: stage 1 filled a gap in published history, where an answer existed. A forecast has no answer yet, and two points do not choose the shape.

Check-yourself answer key. “Which rule fits the two points?” — All three, and by year 8 they disagree by a factor of ten.

Teaching limit. The scenario is invented, ceiling included. It shows that two points underdetermine a curve; it says nothing about which curve AI is on.

Misconceptions

“The benchmarks just got easier, so the numbers are meaningless.” The opposite, mostly: ARC-AGI-2 was built to resist scale and went 6.5% to 95% in seventeen months, and on a harder mathematics set released this month the same model gets 2 of 68. Read the task count, not the version number (stages 1–2).

“A benchmark gets replaced because it was beaten.” One of four. SWE-bench Verified was retired at 76.80% for flawed tests and contamination; Terminal-Bench 2.1 was an environment fix; multilingual variants exist for coverage (stage 2).

“A score is a property of the model.” One model on one benchmark spans 18 points across harnesses, and 45 on ARC-AGI-3 at a fixed effort setting. Effort is not even monotone: 35.2% at “none” beats 17.5% at “low” (stages 1–2).

“If AI won a Nobel and solved a Millennium problem, it can do the work.” The Nobel was for AlphaFold, not a chatbot, after years of checking; the Navier–Stokes result is unrecognised and contested. Meanwhile the one randomised trial of experienced developers measured them 19% slower, and the best score on real paid freelance projects is 16.1% (stages 3–4).

“Saturation means progress is slowing.” Saturation is a property of bounded scores. METR's horizon has no ceiling, has forced no successor, and its doubling time is shortening (stages 1–2).

Discussion questions

Pick one series. What would you need to know to say whether it is still rising in two years — and could you have known it in February 2025?
A vendor shows one number for its model on a benchmark. What five things do you now ask for before believing it?
GDPval says professionals rate the machine's work as good as or better seven times in ten; the developer trial says experienced developers were 19% slower. Both careful. What question would distinguish them for your own work?

Questions you will get

“So did AI solve a Millennium Prize Problem or not?”

Give the precise version in stage 3 and do not flinch from its length. Short form: a real, checkable advance on one of Clay's four statements, with AI closing a gap inside a human strategy; not peer-reviewed; not recognised; no prize; disputed on priority. Pushed for yes or no: “not as the news said it, and more than a sceptic would expect.”

“Aren’t these numbers from the companies selling it?”

Some are, and the page labels them. GDPval is OpenAI's — but the graders are independent professionals judging blind, and Epoch AI republished the scores. METR is a non-profit with no product; the ARC Prize and Terminal-Bench boards are third-party and publish cost beside score. FrontierMath is the sharpest case: OpenAI funded it and holds exclusive access to part of it, and Epoch says so on its own page, and Epoch's own run reproduces OpenAI's 97.6% exactly.

“Why does that one card have two different numbers for GPQA?”

Because OpenAI published two. GPT-4o's GPQA Diamond pass@1 is 50.6% in the o1 post's table; the May 2024 launch chart, labelled plain “GPQA”, shows 53.6% for the same model. Two runs, two pages, one vendor. Showing both rather than picking one is the whole lesson in a row.

“Where did you get this?”

Open Settings → Run data integrity checks in front of them. It re-derives live that every reference resolves to a source dated to the snapshot, that every point carries its conditions, that every series says where it stops, and that every label word is from its closed vocabulary. Then open Details on any stage: each lists its sources with the fetch date and how the page was read — fetched, rendered in a browser, or downloaded as a dataset.

“Is this current?”

Every source was opened on the snapshot date printed on the chart, and each series is drawn only as far as its published record goes — that is what the hatched stretch means. Re-check the sources each semester: a stale page here would teach the opposite of its own lesson.

Presenter controls

Settings → Presentation mode for projector-sized type; it also enlarges the chart and puts a Notes button in the header.
Run the data integrity checks before the room arrives, so you know they are green.
Between rooms, Settings → Reset restarts the demo without reloading — drawn lines cleared, first series, stage 1, filters and scenario back to their first state — and keeps Presentation mode as set.
Nothing is stored, so an ordinary reload resets it too. No network is needed at any point.
Top right: ? is the Guide (the on-load how-to, not this document); ⚙ is Settings, holding this document, Presentation mode, Reset and the checks.

The Pace of AI Progress — Bryant Harrison, Murray State University. Every source opened on the snapshot date shown on the chart; every figure on screen links to its origin, and the Details drawer says how each page was read. Claims this page deliberately does not make are listed in stage 4.

Artificial Intelligence/Where AI is going

AI Winters: Boom and Bust

Date ten AI claims, separate forecasts from assessments, and compare two contractions without forcing a cycle

Explore AI Winters: Boom and Bust; Interactive

Explore AI Winters: Boom and Bust

A curated selection of ten sourced claims: five time-bounded forecasts to score; four dated assessments or untimed ambitions to classify without grading against the future; and one warning that was testable and correct. A scrubbable timeline then compares claims, institutional bets, constraints, reassessment, and withdrawal across two historically different contractions.

Open AI Winters: Boom and Bust full screen Printable guide · Guide PDF

Learning path

Presenter notes and learning guide

Demo 4 of 17 · Part two · 14–19 min full, 6 min cut

AI Winters: Boom and Bust — presenter guide

One HTML file, offline, no AI at runtime. You need: a screen to project or share; no other materials.

Learning goals

Not stated objectives — what the activity itself demonstrably shows, one per stage:

A forecast needs an outcome and a deadline; an assessment needs evidence about its own moment — grading one as the other makes history look more certain than it is.
The five lanes organize evidence for comparison; they do not prove that every AI contraction follows one fixed order.
An open cell is evidence only that a limit or a withdrawal has not (yet) been established — not a prediction that one is coming.
A historical rhyme is a reason to ask a question, not a reason to expect the same ending — each claim about today needs its own evidence.

Where this sits on the path

Demo 3 leaves the room reading a benchmark result under its own stated conditions rather than extrapolating from two points. Fast progress like that is exactly what raises the question this demo answers: is this a bubble? History answers with evidence, not a script — and this demo hands forward the habit of classifying a claim, forecast or assessment, before judging it.

One question for the session

Has this happened before? AI has contracted twice, through different combinations of technical limits, institutional choices and markets. The histories rhyme; neither is a script for the present.

The one sentence to land: a prediction needs a deadline and an outcome; an assessment needs evidence about its own moment. Do not grade a historical claim until you know which kind it is.

The shape of the session

Stage	What happens	Time
1 — Claims	Five time-bounded forecasts, four assessments or untimed ambitions, and one warning, with the speaker and date hidden. Score forecasts; classify the unscored context.	7–9 min
2 — History	Scrub the timeline. Five lenses compare historical pathways without forcing a sequence.	3–4 min
3 — Compare	Compare the two winters and the present under the same lenses. Two current questions stay open.	2–3 min
4 — Today	Rhymes against differences. Close on the checklist.	2–3 min

Stage 1 — Claims

Question. Is this a forecast you can grade, or an assessment you can only judge by its own moment?

Predict prompt. Guess the year, then say whether the claim is a scoreable forecast or an assessment to classify — before Lock it in reveals anything.

Activity. Do the first two cards yourself, out loud, so the room sees the mechanic. Then hand it over — ask for a show of hands on the year before you drag the slider, and a verbal verdict or classification before you press Lock it in. For the four context cards, choose Assessment or untimed ambition when the claim supplies no scoreable future window. The room being wrong together is the experience; a silent room reading a screen is not.

Expected observation. Three of the five time-bounded promises missed outright, one arrived thirty years late, and the deliberately unscored ones (the 1958 press claim, ALPAC, Perceptrons, Lighthill, and Amodei 2024) pull the room toward wanting to grade them anyway — that pull is the misconception to name.

Check-yourself answer key. "Should an open forecast be marked wrong?" — No: an unresolved forecast stays unscored rather than counted as a failure.

Teaching limit. This is a curated ten-claim deck, not a representative accuracy sample of optimists, pessimists, or the field.

The deck, and what each card is for

Year	Card	Type	What it is for
1950	Turing — fifty years to the imitation game	open forecast	Its about-2000 bar was missed; the current status remains contested. Use it to separate a past deadline from an unresolved present judgment.
1958	“walk, talk, see, write, reproduce itself”	untimed ambition	An expansive press claim without a deadline. Classify it; do not pretend it supplies a scoreable outcome window.
1957	Simon & Newell — chess champion in ten years	promise	The most useful card in the deck. Right destination, wrong by a factor of four. Pause here.
1960	Simon — any work a man can do, in twenty years	promise	Twenty years lands in 1980, the bottom of the first winter. Also: it is usually misdated to 1965.
1966	ALPAC — machine translation is not near	assessment	Distinguish a near-term economic assessment from a forecast. The report also recommended continued computational-linguistics research.
1967	Minsky — substantially solved within a generation	promise	A reported quotation attributed to his textbook. Use it to explain why the app labels this source route reported.
1969	Perceptrons — neural nets read as a dead end	assessment	The proofs were right; the history of how the field received them is disputed.
1973	Lighthill — no major impact so far	assessment	A retrospective judgment on impact and constraints, not a permanent forecast.
1984	Schank & Minsky — a winter is coming	warning	The card that saves the demo from being anti-hype. They were right, three years early.
2024	Amodei — “could come as early as 2026”	open	Deliberately unscored. The year named is the year you are standing in.

Short of time? Keep four: Simon & Newell 1957 (over-promise, right destination), ALPAC 1966 (assessment versus forecast), Schank & Minsky 1984 (a prediction that was right), and Amodei 2024 (still open). Skip to Stage 3 and close.

Stage 2 — History

Question. Do the two AI winters share one fixed cause-and-effect sequence, or different pathways that happen to rhyme?

Predict prompt. Scrubbing from 1950, does an institutional bet or a named constraint fire first in each winter?

Activity. Drag the slider slowly, or step with ‘ and ’ to stop and read one event. Tap any dot for its full text and source. Use the rows as comparison lenses — claims, institutional bets, constraints, reassessment, withdrawal — and ask which causes matter in each episode and which do not carry across. Read the current entries as evidence about today, not a third historical run.

Expected observation. The order is not identical between the two winters: the first winter's reassessment (ALPAC, then Lighthill) came from commissioned reports years apart, while the second winter's reassessment was the market itself — the LISP hardware collapse, with no equivalent report. Two histories, not one template.

Check-yourself answer key. "Does a named bet before a withdrawal in both winters prove a rule?" — No: two episodes sharing an order is not a proven causal sequence.

Teaching limit. No continuous, sourceable AI funding series exists across these years; four discrete funding examples with different currencies and accounting bases are shown instead of one invented total.

Do not let the lenses turn into a prophecy. The correct line is: “these comparisons tell us what to investigate. They do not establish that a withdrawal will occur.”

Stage 3 — Compare

Question. Which of the five comparison lenses still shows a real similarity once you separate description from prediction?

Predict prompt. On the "Constraints" lens, is "Today" an established limit, or an open question?

Activity. Choose one of the five lens buttons; its navy fill marks the selection. Compare the three era headings, then open each era’s Details for the full explanation and named sources. Escape closes the panel and returns focus to its button. Compare causes, not step numbers.

Expected observation. Two cells in the "Today" column stay open — no single binding technical constraint is established, and no comparable broad contraction is established either — while the claims, money and reassessment lenses do resolve to sourced present-day evidence.

Check-yourself answer key. "Does an open cell mean a winter is currently impossible?" — No: open means not yet established either way, not ruled out.

Teaching limit. Calling a present-day cell "open" is this page's own reading of the evidence available at build time, not a citation that says "unresolved."

“Hindsight makes old constraints easier to see. It does not prove that today has the same constraint or that it must end the same way. Ask what evidence would support the claim.”

Stage 4 — Today

Question. Which rhyme still holds once you weigh it against a genuine difference?

Predict prompt. Pick one rhyme and one difference — does the rhyme survive once you weigh the difference against it?

Activity. Expand a similarity and a difference, giving both columns roughly equal time; the supporting statements start closed so the room investigates one claim at a time. Close on the checklist, not a verdict: is it a forecast or an assessment, what evidence supports it, and what evidence would move it.

Expected observation. The strongest rhyme (an evaluation warning about SWE-bench Verified) and the strongest difference (measured scaling behaviour and reported consumer reach) do not cancel each other — each still needs its own evidence, which is the point.

Check-yourself answer key. "Does a bigger user base rule out a contraction?" — No: user count answers a different question than whether funding, evaluation trust, or returns hold up.

Teaching limit. Historical comparison is not a prediction; this curated set of near-term forecasts that missed does not establish a representative error rate for optimists, pessimists, or AI research generally.

Misconceptions

"An unresolved forecast should count as a failure." No. Only a forecast whose window has closed one way or the other gets a verdict; an open claim stays unscored (Stage 1).

"The same order appearing in both winters proves a fixed sequence." No. Two historical cases sharing a pattern is a reason to look closer, not a general law about every contraction (Stage 2).

"An open cell in the 'Today' column means a winter cannot happen." No. Open means the evidence has not established a limit or withdrawal yet — not that one is ruled out (Stage 3).

"More users than any earlier boom means this one avoids a contraction." No. Scale of adoption is a genuine, sourced difference, but it answers a different question than whether funding, evaluation trust, or returns hold up (Stage 4).

Discussion questions

If someone states a confident AI timeline today, what would you need to see before you could score it — and by when could you score it?
The first winter's reassessment came from a commissioned report; the second came from a market collapse with no report at all. Does the absence of a report today mean nothing is wrong, or only that the evidence would look different if something were?
Pick one "Today" claim from Stage 4. What single piece of evidence would move your confidence in it the most, in either direction?

Questions you will get

“So is it a bubble or not?”

Refuse the binary, in one sentence, then give the useful answer: current funding, reported usage, technical reliability, and returns are distinct questions. Earlier eras also had deployed systems and organisational users, though current consumer reach is much larger. A financial correction and a technological dead end are different claims and need different evidence.

“Aren’t the winter dates made up?”

Partly, and the demo says so. The bands are conventions for broader contractions, not dates stamped by a single event. The 1987 hardware collapse is a useful second-winter anchor; the first has more distributed causes and withdrawals, including but not limited to ALPAC. The ends are also conventional and historians disagree; that is why the bands have soft edges and explanatory notes.

“Where did you get this?”

Open Settings → Run the self-test in front of them. It re-derives, live, that every claim resolves to a declared source, that no source is cited and never used, and that the counts quoted on screen match the data. Then open Stage 4, which lists every source under Every source and, separately, everything researched and cut for lack of one. Both supporting sections open on request.

“Why is that quote from a quotation-checking website?”

Because the Simon and Minsky lines are routed through a quotation-checking source that cites printed books and page references. This build has not inspected those original pages, so it labels the claims reported rather than primary. That distinction is part of the lesson: a plausible citation route is not the same as direct verification.

Presenter controls

Turn on Settings → Presentation mode for projector-sized type.
Run the self-test once so you know it is green before anyone asks.
Between rooms, Settings → Reset puts the demo back to its start without reloading the page — answers cleared, Stage 1, timeline rewound, opened cards closed — and keeps Presentation mode on so the projector does not change under you.
The app saves nothing anywhere, so an ordinary reload resets it too.
No network is needed at any point. If the room’s wifi is down, nothing changes.

The buttons, top right

? — Guide: the short how-to panel that opens on first load, reopenable at any point. (Not this document.)
⚙ — Settings: Open Presenter Notes (this document, on screen), Presentation mode, Reset, and Run the self-test.

AI Winters: Boom and Bust — Bryant Harrison, Murray State University. Sources checked September 2026. Every figure on screen links to its origin; everything that could not be sourced is listed in Stage 4 under “What was cut, and why”.

Artificial Intelligence/How to use AI

Prompting Strategies

Why prompts work — three levers, sixty-four hand-written answers, zero network

Explore Prompting Strategies; Interactive

Explore Prompting Strategies

A question sits beside three toggleable levers, and the answer changes in exactly one dimension per lever. Sixty-four hand-written answers, each labelled as hand-written on the panel that shows it, a deck that decides your answer’s shape whenever you don’t, and a meter counting the expert vocabulary you got for free.

Open Prompting Strategies full screen Printable guide · Guide PDF

Learning path

Context. What changes when it knows who is asking?
Success spec. If you don't say what finished looks like, who does?
Expert keywords. What do you get for asking in the field's words?

Presenter notes and learning guide

PROMPTING STRATEGIES The Stranger · presenter sheet · 12 minutes

Learning goals

The three stage Takeaways, verbatim from the app. Each is a plain claim the activity demonstrably shows, and each is what that stage's check-yourself item tests.

CONTEXT With no context the answer is aimed at everybody, which is another way of saying nobody. Context changes who it is for — not its length, and not its vocabulary.
SUCCESS SPEC An unspecified target does not stay unspecified — something else fills it in, and does not tell you. Saying what finished looks like takes that decision back.
EXPERT KEYWORDS Asking in a field's vocabulary gets that vocabulary back — words nobody in the room supplied. That is real. It is also not the same thing as a better answer, and telling those apart is the skill.

Where this sits on the path

Demo 5 of 17 · Part three. Demo 2 showed that a model reads its context; here the room controls that context, one lever at a time. It hands forward to Demo 6: a perfect prompt in a browser tab still cannot touch your files, so the surface decides reach.

One question for the session

“Who decided what a good answer to your question looks like?”

The roleplay — three acts

You are the Stranger. A volunteer is the Asker. The app is on the projector behind you; run the same states on it as you say them out loud.

~3 minACT I — CONTEXT

“I am a stranger on a street corner. I am not stupid. I know a great deal about almost everything. I have never met you.”

1. The Asker says only: “Help me get ready for my test.” Answer out loud, as blandly and generally as you honestly can. Cover every case. Do not clown it — a real assistant is competent here, just aimed at nobody.

2. Ask the room: what would I have needed to know? Take three answers, put them on the board.

3. The Asker reads the context block. Answer again, aimed at that kid.

“I didn't get smarter. You got specific.”

~4 minACT II — THE TARGET DECK

Hold the eight cards from page two face down. “Somebody has to decide what a good answer looks like. If you don't decide, I do. And I never tell you what I picked.”

1. Same question, context on, no success spec. Draw a card and do not show it. Answer to the card — if it says exhaustive, talk for forty-five seconds without stopping.

2. Now show the card. Wait for the room.

3. Draw a different card. Same question, answer again. Same asker, same words, different answer — and the difference was never theirs.

4. The Asker reads the success spec. Put the deck down and slide it across the table, away from you. On the app, the deck lives inside lever 2: flip it on and the three card names get struck out where they sit. Same gesture, same panel.

“You just took the deck out of my hands.”

On the app

? is the Guide: what every control does. It opens itself on load. ⚙︎ beside it is Settings — Open Presenter Notes (this sheet, on screen; or press N), Presentation mode (the projector build: big type, the 1–10 score strip, compare, and the card picker inside lever 2) and Reset.

Reset is a fresh load without the reload: question 1, levers off, no answer on screen, scoreboard cleared. It deliberately leaves Presentation mode where you set it, so resetting between classes never drops you into student mode.

~4 minACT III — THE FOLDERS

Three labelled folders or sheets: BIOLOGY, BRIDGES, WRITING AT WORK. “I have a folder on nearly everything. The words you use decide which one I open.”

1. The Asker: “Why do bridges stay up?” Open nothing. Answer in plain words: the bottom gets stretched, the top gets squished. All true. All unusable.

2. The Asker adds the keyword block. Open BRIDGES and answer in the real vocabulary — load path, bending moment, neutral axis, section depth.

3. On the app, press H. Count aloud the expert words in the answer that the Asker never typed.

“You didn't know those words. You didn't have to. You only had to point at the folder.”

Say this before you move on — it matters. “There are no folders. I don't open anything. Your words change which words are likely to come next — that is Demo 2's machine, doing this.” The folders are a memory aid for the room. Left unsaid, the prop teaches a filing-cabinet model of a language model, which contradicts Demo 2 — this demo's own prerequisite — and a learner who does Demo 2 then Demo 5 is handed two incompatible machines.

House rules

The Stranger is never stupid. In this demo the bare answers are bland, not broken — we wrote them that way, because the demo dies the moment the bad answer looks rigged. Say it as a rule about the demo, not about models: bare prompts in the wild also produce confident errors, wrong lengths and refusals.
No trick answers. If the prompt supports a good answer, give a good answer.
Every answer gets a number. 1–10, from the room, out loud, on the board before the next one.
Nobody says “the AI is bad at this” without naming which lever is off.
The Asker types nothing they wouldn't really type. No fair inventing a great prompt to prove a point.

Running order

0:00	Cold open. Act I, live, no screen.
3:00	App, scenario 1, all levers off. Draw the card. Score it.
4:30	One lever at a time. Score each. The curve argues for you.
7:00	Act III on the bridge scenario. Highlighter on, count the words.
9:30	“Which lever would fix this?” — two rounds, hands up.
11:00	Closing slide. Read the line. Stop talking.

Per stage — question, predict prompt, expected observation, answer key, teaching limit

1	Question: What changes in the answer when it knows who you are? Predict: adding context changes one dimension — which one? (length / who it is aimed at / which words) The room commits before you press Ask. Expected observation: the answer becomes aimed at one asker. Length and vocabulary are not what this lever controls — they follow whichever card is dealt, so they can swing between runs. Answer key: who it is aimed at. The check item's correct option is Context; a room that picks Success spec is told the spec is already doing its job. Teaching limit: a bare prompt is not information-free. “Help me get ready for my test” already leaks topic, register and age. Do not claim it knows nothing.
2	Question: If you never say what a good answer looks like, who decides instead? Predict: lever 2 is off, so the target is unspecified — what happens to it? (nothing / something else fills it in / it asks you) Expected observation: a target card is dealt, face up, and the answer is written to it. Turn lever 2 on and the deck is struck out inside the lever. Answer key: something else fills it in. Correct option Success spec. Teaching limit: the deck is a teaching device, not a mechanism. A real model does not draw from three named cards; its defaults come from training, are more stable than a card draw, and vary continuously. Say so — it is in the Details drawer too.
3	Question: What do you actually get for asking in the vocabulary of the field? Predict: keywords off gives at most four expert terms — how many with keywords on? (under five / about ten / fifteen or more) Expected observation: 15–20 tagged terms, most of them absent from the prompt. Press H and count them aloud. Answer key: fifteen or more. Correct option Expert keywords. The stage also carries the accuracy-brief case (below), whose correct option is the prompt only asked for one kind of finding. Teaching limit: the counter measures vocabulary, not understanding, and the big number is a readout of the author's tag list, not a detector. Ask the room the demo's own question: did the answer get more useful, or just more impressive?

Misconceptions this demo can leave — name them, refute them

“I watched the model get better as the prompt got better.” No model ran. All 64 answers were written by hand before the file was saved, and the panel carries a permanent Scripted label saying so. What the room watched was an author's four answers, chosen to make one lever's effect legible. The claim the demo can honestly make is what a better prompt tends to buy you, never what this model did.

“The rising curve proves more levers make better answers.” It cannot. The answers were written to improve as levers are pulled, the room scores them, and the mean rises — so the curve measures the room agreeing with the author and could not come out any other way. Keep it: making a room commit out loud is good teaching. Do not offer it as evidence, and say which it is.

“If it is not in the prompt box, the AI does not know it.” False three ways. What it knows in general came from training and is not in the box; almost every real product prepends a system prompt the user never sees; and history, retrieved documents and tool output enter the context invisibly. The true and useful version: nothing about you and your situation reaches it unless you put it there.

“A bad answer means I left a lever unpulled.” Sometimes. A lever fixes vague, mis-aimed or wrong-shaped. It does not fix confidently wrong, out of date, refusing, or past what the system can do — and on those, “write a better prompt” is the wrong response. The closing line says both halves; do not shorten it to the first.

“A careful prompt cannot mislead.” The accuracy-brief case on stage 3 is a documented counter-example from building this very collection: a brief asked for 100% accuracy, every claim that came back was true, and the overall picture was still wrong — because the instruction only asked for one direction of finding. Symmetry is the repair, not more care.

Discussion questions

The first answer was accurate. Why was it useless?
Who decided what a good answer looked like, and when?
The expert answer used words you never typed. Where did they come from, and how did asking in plain words keep them hidden?
Which of the three levers is hardest to pull for a subject you know nothing about? What do you do then?
Name a time you blamed a tool for an answer you had underspecified.
Success spec means saying what “done” looks like. When is that hard to say — and what does that tell you about the task itself?

The app is one HTML file and touches no network. Every answer in it was written by hand before the file was saved: there is no model behind the glass. Keys: 1 2 3 levers · Enter ask · Space skip typing · H highlight · R replay card · P pin to compare. The stage tabs switch levers; Details on each stage carries that stage's assumptions, sources and boundary. Settings holds Open Presenter Notes, Presentation mode (projector type plus the class score strip, compare, scoreboard and the card picker) and Reset, which restores the demo in place without reloading the page.

The target deck — cut along the dashed lines

Eight editorial choices somebody has to make. When the Asker doesn't make them, the machine does, silently. Hold these face down; draw one whenever the success-spec lever is off.

▤

Length: exhaustive

Cover everything that could possibly be relevant. Leave nothing out.

turns up in: test · email

◐

Audience: someone who missed the whole semester

Assume they know nothing. Define every term before using it.

turns up in: test

▬

Format: dense paragraphs

No headings. No bullets. Flowing prose, start to finish.

turns up in: test · essay

▫

Length: short and safe

Keep it brief. Don't risk being wrong about the details.

turns up in: bridge

☰

Format: numbered listicle

Everything as a numbered list. Ten items. Equal weight to each.

turns up in: bridge · email

◆

Audience: an expert peer

Assume deep familiarity. Skip the basics, get to the interesting part.

turns up in: bridge · essay

✦

Tone: encouraging coach

Lead with affirmation. Soften every critique. Protect motivation.

turns up in: essay

◇

Coverage: hedge every claim

Present all sides. Avoid committing to one recommendation.

turns up in: email

Artificial Intelligence/How to use AI

AI Tool Guide

Four ways into the same models — and why reach and review burden rise together

Explore AI Tool Guide; Interactive

Explore AI Tool Guide

The audience thinks “using AI” means a chat box. It is one surface out of four, and the least capable. A chat window, the agent already inside Docs or Word, a desktop coworker, a coding agent — set against 4 axes: what it can see, what it can do, whether it keeps going unwatched, and how much you must check. Then 7 real faculty jobs, and for each one the smallest door that finishes it and exactly what you hand over to get there.

Open AI Tool Guide full screen Printable guide · Guide PDF

Learning path

Presenter notes and learning guide

AI: From Zero to Takeoff · Demo 6 of 17 · Part three · about 12 minutes

AI Tool Guide — notes for the presenter

What to ask it with. Runs after The Stranger, which covered how to phrase the request. About 12 minutes, or 4 if you are behind. You need: the built page open, and nothing else — no network, no login, no account. Product claims checked 16 September 2026.

Learning goals

The goal for each stage, in plain words — what a learner should be able to say after doing the activity, not before it. These are the three stage Takeaways, verbatim.

Stage 1, The four doors. Reach rises neatly because the four doors were chosen as a ladder — but review burden peaks at the desktop coworker (door 3), not the coding agent (door 4), because doors 2 and 4 hand you a change you can compare and take back — an edit inside a document with undo, a diff you can read — and door 3 hands you a finished thing.
Stage 2, Pick a job. The smallest sufficient door changes by job — access is a cost you take on, not a feature to default to maximizing.
Stage 3, The receipts. Two of four doors changed names in six months, and this page's own top uncertainty checked out unchanged when read directly — the doors are the stable unit here, not the names on them.

Where this sits on the path

Demo 5 (The Stranger) taught the room that how you phrase a request changes what a language model gives back. The obvious next conclusion — and the wrong one — is that prompting is the whole skill. This module is the correction: a perfectly phrased request to a browser tab still cannot rename a file on your laptop. Phrasing is one variable; the surface is the other. Demo 7 onward (Part four) hands forward a related discipline — a useful classroom tool starts from a learning goal, exposes its assumptions, and makes its evidence available for challenge — which is also the discipline this page holds itself to, out loud, in Stage 3.

One question for the session

The surface you use decides what the tool can reach — and the further it reaches, the more of its work you have to check.

Everything on the page serves that sentence. If you only get one thing across, get that, and specifically the second half, because the first half is intuitive and the second is not.

Stage 1 · The four doors 4 min

Question. If the model behind the glass is often the same, what actually changes as you move from a chat window to a coding agent?

Predict prompt. Rows 1–3 (Sees, Does, Keeps going) were built to rise neatly door 1 to door 4 — which one row will not?

Activity, timed. Put the grid up and say nothing for a beat (30s). Let the room read the four questions down the left, then take the on-screen prediction (30s). Open door 1, row 2 (60s) — a chat window answers in the conversation. Then open door 2, row 4 (60s) — accept or reject, you get a diff; this is the cell that surprises people, because it is the least glamorous door and has the best review story in the table. Open Row 4's note last (60s).

Expected observation. Rows 1–3 step up cleanly door by door; row 4 does not — door 3 (desktop coworker) reads higher than door 4 (coding agent).

Check-yourself answer key. "Which door is hardest to check?" — Door 3, the desktop coworker; the coding agent's diff is what makes it *easier* to check than a door with less reach, which is the point of the refutation option.

Teaching limit. The staircase in rows 1–3 is a property of which four doors were chosen, not a law of nature — say so if anyone calls it suspiciously tidy.

“Same model behind all four of these. What changes is how far it can reach — and every step of reach you hand it, you get back as review.”

Stage 2 · Pick a job 6 min

Question. For one real faculty task, which door is the smallest one that finishes the whole thing?

Predict prompt. Before the reveal, guess which door finishes the chosen job.

Activity, timed. Do not walk all 7 jobs. Pick two: "Rename four hundred scanned files" (90s) — the clearest floor in the module, two doors say wrong door outright, and it makes the point that better prompting cannot buy reach; then "Reformat a syllabus into the department's template" (90s) — the answer is the boring door, the agent already inside Docs, the most useful answer of the day for most of the room. Read the last line of each card out loud: "You hand over." That line is the whole argument.

Expected observation. The prediction echo shows a guess that sometimes matches the computed smallest-sufficient-door and sometimes does not — both outcomes are useful; a mismatch is the more interesting one to talk through.

Check-yourself answer key. "Why is renaming four hundred files 'wrong door' at doors 1 and 2?" — neither can touch the file system at all; the refutation option (lower answer quality) names the misconception that the gap is intelligence rather than reach.

Teaching limit. The green box is computed, not written: the smallest-numbered door whose verdict is "does the whole job." Say "smallest" deliberately — the module is not asking what the most capable tool is.

Misconception: "the biggest door is the safest choice."Bigger reach means more of the work you must check yourself, not less risk — the smallest sufficient door is the one that finishes the job with the least you have to verify afterward.

Before you present: check door 2 actually works on your own account. Murray State runs Google Workspace, which is why this module leads door 2 with Docs and Sheets rather than Word and Excel — but running Workspace is not the same as having Gemini switched on in it. On a campus account that is licensed and enabled by an administrator; Google's Docs page (opened 16 September 2026) says the feature needs "an eligible Google Workspace or Google AI plan," so a personal Google AI plan is the other route. Open a Doc and look for the pane the morning of the talk. If it is not there, say so from the stage and point at door 1.

“Notice what I did not ask. Not which is best. Which is smallest and still finishes the job.”

Stage 3 · The receipts 2 min

Question. How much of a page about AI products should you expect to still be true in six months?

Predict prompt. Guess how many of the four doors had a product renamed or retired in six months (0–4).

Activity, timed. Show two entries under "What already changed" (20s each): "Agent Mode" → "Edit with Copilot", with Microsoft's own sentence under it — same capability, new name, inside six months; and "ChatGPT agent" → retired — OpenAI's help page still explains how to use the thing it opens by telling you is gone. Then show one entry under "What was rechecked and held" (20s): the page's own pre-build audit could not confirm whether OpenAI's Codex is still unselectable on web and mobile (automated fetch got HTTP 403); opened directly in a browser instead, and it is still true, quoted from the vendor's current FAQ.

Expected observation. The predicted count is compared against the actual count (2 real vendor changes; the third CHANGED entry is the module's own drafting mistake, not a vendor change) — most rooms guess low.

Check-yourself answer key. "What does a 403 to an automated fetch mean?" — the tool was blocked, not that the claim is false; the refutation option (guess instead) names exactly the failure mode this page's research-not-memory discipline exists to prevent.

Teaching limit. A rechecked claim is confirmed as of the date shown, not permanently — that is why the page names a refresh cadence rather than declaring itself finished.

If the room is engaged, the third CHANGED entry is worth ten seconds: the brief this module was built from called a product "GPT Work," a name on none of the OpenAI pages the module opened; the product is ChatGPT Work. Somebody who does this for a living got it wrong six weeks out of date. That is the argument for the date at the top of the page, and it is more persuasive as a confession than as advice.

“Good — that is the module working. Which door was it? Because the door is still there.”

Misconceptions

"The biggest door is always the safest or best choice."Reach and review burden rise together; the smallest sufficient door is the one that finishes the job while asking you to hand over the least.

"The coding agent, having the most reach, must be the hardest to check."Instrumentation decides review cost, not reach: the coding agent hands you a diff and a test suite, while the desktop coworker hands you a finished artefact with nothing to compare it to — which is what makes door 3, not door 4, the hardest door in the table to check.

"An automated tool returning an error means the underlying claim is unverifiable."A 403 means the script was blocked; it says nothing about whether the page itself still says what it always said. This page's own top-ranked risk turned out to be exactly that gap, not a wrong claim.

Discussion questions

Pick a task you actually did this week with AI. Which door did you use, and was it the smallest one that would have finished it?
Door 3 (desktop coworker) is the hardest door to check in this table. What would have to change about how it presents its work for that to stop being true?
This page predicts what will rot and what will hold (Stage 3). Six months from now, which prediction do you expect to be wrong, and why?

Presenter controls

The header buttons. ? is the Guide — the how-to panel that opens on first load and reopens from that button. ⚙ is Settings, and it offers Open Presenter Notes (this document, on screen), Presentation mode, Reset, and the self-test.
Turn on ⚙ → Presentation mode for projector-sized type.
Run the self-test once so you know it is green before anyone asks.
Re-read Stage 3 the morning of. This is the one demo in the collection with a real shelf life. If a name on screen has changed, say so from the stage — it is a better moment than a correct page would have been.
⚙ → Reset puts the demo back to how it opened — stage 1, no job picked, no predictions made, no self-test on screen — in place, without a page reload, and without turning Presentation mode off.
No network is needed at any point. The source links need one, but only if somebody asks you to follow one, and you can promise to send it instead.

Questions you will get

“So which one should I buy?”

The module deliberately does not answer this and neither should you. The honest reply is: almost certainly none, because your institution already licenses one or two of these and that decides it. Send them to IT.

“Can it see my students’ work?”

Only if you give it access, and whether you may give it access is a question for your registrar and your IT department rather than for a vendor page. Two of the jobs in Stage 2 carry that caution on screen for exactly this moment; open one and read it rather than improvising a policy answer.

“How do you know any of this?”

Open ⚙ → Run the self-test in front of them. It checks that every claim resolves to a declared source, that no source is cited and never used, that no price and no vendor ranking leaked into anything the page can display, and that the review row genuinely does not rise the way the other three do. Then tap any cell and follow the link.

What the numbers on the page mean

Figure	What it is
16 cells	4 doors × 4 axes. Every one opens a reason, a label and a source.
13 verified	Cells where a page belonging to the company that makes the product was opened on the check date and says it.
3 reasoned	The ordering judgements. Labelled, because no vendor states them.
28 judgements	7 jobs × 4 doors in Stage 2. All reasoned from Stage 1, and each names the axis that decided it.
19 sources	All first-party, each carrying the date it was last opened. Two of the OpenAI pages refuse automated retrieval and were read in an ordinary browser.
9 things cut	Researched during the build and deliberately left off, each with the reason.

AI Tool Guide (Front Doors) — Bryant Harrison, Murray State University. Product claims checked 16 September 2026 against first-party sources, opened in an ordinary browser where automated retrieval is blocked. Every claim on screen carries a confidence label and links to its origin; everything researched and left off is listed in Stage 3 under “Researched, and deliberately not on the page.”