Understanding the tools we depend on
Jason Kridner · Northville educator workshop
The opening five-minute story below reproduces revision 3 without changing its spoken narrative. The full story and scene notes remain available. The remaining notes are a proposed facilitation script.
Run of show
0–7: stakes and story. 7–10: mechanism and agency. 10–23: card activity and proof. 23–27: universality. 27–31: Mechanical Turk and human work. 31–43: next-word hat and discussion. 43–49: learned parameters and InstructGPT. 49–54: training economics. 54–60: classroom reflection.
Each table needs nine distinct cards, a cup, small slips, pencils and a response sheet. No student login, phone, microphone or BeagleY-AI is required. If running late, shorten the second hat run and historical discussion; retain the proof, the training/generation distinction and the final reflection.
Ideas are equal.
Opportunities are not.
Opening, 0:00–2:00. Introduce the line as a saying RJ uses, with the wording Jason recalls. Do not claim a verified original author. Ask the audience to hold two children in mind: one discovering what she can make, another surrounded by extraordinary things she cannot see how to make. We owe children more than exposure to powerful products. The ability to read a program’s logic and reason about what it will do is part of their opportunity to direct their own lives. Today we will inspect two mechanisms on paper, then ask who can understand, change and retain the tools. This hour will not solve the economic or institutional barriers. It should make those barriers harder to overlook.
The opening illustration is an existing AI-generated image. The untouched gun is symbolic background inspired by a shirt image, not a binary choice offered to a child or a causal claim about violence. Do not point to it or build a threat narrative.
“Birds need somewhere to sit.”
Chloe draws houses with crooked chimneys.
“Birds need somewhere to sit,” she says.
She presses so hard that the purple crayon snaps. She examines the two pieces, peels the paper off one, and rolls its side across the page. Now the house has an enormous purple evening behind it.
She grins. That wasn't the plan.
“Could you make it darker?”
On a school trip to the crayon factory, Chloe stands on tiptoe at the rail. A worker shows them wax, color, a mold. She points to a tray of purple crayons.
“Could you make it darker?”
The worker shows her where the color goes in.
Back in class, they make their own. The teacher handles the hot wax. Chloe chooses the scraps. When hers has cooled, she rubs it across a corner of paper, frowns, and puts more blue scraps beside the mold for the next batch.
She takes both crayons home. She has plans for the second one.
“Can we stay?”
At a museum, a pencil drawing stops Chloe. A tree fills the wall. She feels she could walk into its shade.
She leans closer. The shadows break into little strokes, dark where they overlap. She traces a familiar motion against her palm.
“Can we stay?”
At home, she sharpens her pencil and tries one small branch. Then another.
She still has the two purple crayons.
Years pass. Her bedroom fills with houses, including a dreadful clay one she won't let anybody throw away. On a borrowed tablet, she learns to make their windows light up. Somebody is still reading.
One afternoon, on her way home, she sees the factory windows covered with brown paper. She slows down. The little entrance where her class lined up is chained shut.
She still has the two purple crayons.
The house answers, softly.
Now imagine Chloe years further on, working with AI. She sketches, asks, rejects, redraws. Together with the system, she makes a house you can walk through, though its rooms could never fit inside its walls. Rain on the roof becomes music. A visitor whistles, and a staircase curls toward the sound.
Chloe spends three evenings getting the rain right.
At the opening, a woman hums a tune her father used to whistle. The house answers, softly. She asks if she can stay a little longer. Chloe finds her a chair.
“Mine doesn’t do anything.”
Her daughter Bridget sees it too. Later, at the kitchen table, Bridget slides her own drawing under a placemat.
“Mine doesn't do anything.”
Chloe pulls up a chair. “You should see my clay house.”
Bridget almost smiles.
“I want mine to follow me home.”
Chloe taps the drawing. “Let's try.”
“But I want to take it.”
She opens the little animation tool she started with. Its old address leads to the company that now runs her studio tools. The beginner version is gone. She searches for another favorite. The same sign-in page appears.
She opens her own workspace. Adding Bridget means another paid seat. Chloe reaches for her card, then reads the small print: projects on this plan can run only inside the provider's exhibition service.
Bridget points at her drawing. “But I want to take it.”
Chloe opens the help window. She tries an export. A beautiful, motionless house lands on the desktop.
“Hang on,” she says.
Bridget waits. Chloe reads the instructions again.
“I don’t know yet.”
Then she opens her own house, the one with the rain. A notice says the version she used will stop running next month. The replacement offers to rebuild it. She looks for a way to keep the old one. The only download is a recording.
“Can you fix it?” Bridget asks.
Chloe puts her hand on the mouse, then takes it off.
“I don't know yet.”
The rain is still playing. Bridget's drawing is between them.
What can the child carry away?
Before our next classroom activity, hold that picture for a moment. When the activity ends, what can the child still understand, change, and carry away?
Understanding the mechanism
7:00–10:00. Return from the story to the educational responsibility. Students can use agents to help write code and still learn to inspect the logic themselves. The important difference is whether they can explain and test the mechanism, rather than depending entirely on another system to make it work. This is practical, conditional agency, not total independence. Reusing code is valuable; the question is whether the path to understanding and changing it is available.
Do not reduce this to “computers execute reliably, but AI makes mistakes.” The point is the accessible reasoning route: for an explicit algorithm and stated inputs, we can sometimes derive an outcome before running it, and teach why. Repeatability alone does not give us that explanation. Neither a successful demo nor access to a source file guarantees comprehension.
Invite one example of a classroom activity that ends at operating a product, then one that exposes an underlying mechanism. Keep this brief. The upcoming activity is deliberately small enough to understand completely.
Nine shuffled cards
10:00–12:00. Proposed adaptation of NRICH Spot the Card, not a final activity decision. Work in pairs, with any nine distinct playing cards per pair. Remove jokers; no full-deck sort or prepared arrangement is necessary. One person deals, one remembers a card. Let them shuffle openly. Before anyone selects a card, predict that after the two prescribed rounds it will be fifth from the top. Do not claim you can name it before receiving any information. The named column is an input to this algorithm.
Deal one card at a time face up, left to right, into three columns, building three rows top to bottom. The spectator silently picks one visible card and names only its column. Use the following slide and printed activity sheet for the collection order. An observer can check the operations; nobody needs to act as a hidden assistant.
The same short procedure
12:00–18:00. Exact ordering matters. Treat each column as a list from top row to bottom row. Pick up the other columns and the selected column so that the new packet, read from TOP to BOTTOM, is: all three cards of one unselected column in row order; all three cards of the selected column in row order; all three cards of the other unselected column in row order. Do not casually scoop the grid into a pile. One easy method is to lay those three lists end to end on the table, then assemble the face-down packet starting with the LAST card and finishing with the FIRST on top.
Deal that packet again, starting with its top card, left to right, three rows. Spectator names the column again. Collect exactly as before. Count from the top: the fifth card is the remembered card. Turn it over. If it fails, inspect the trace and packet orientation before repeating; human execution can make mistakes without changing the algorithm’s guarantee.
Allow roughly six minutes for pairs to perform and check. This is an estimate requiring rehearsal. Do not spend the entire hour recovering a confusing collection instruction. Use the numbered printed trace or demonstrate a single pair if needed.
Why position five?
18:00–23:00. Lay out the intermediate packet and count positions. After the first collection, the chosen column occupies positions 4, 5 and 6. In the second deal those three positions all land in the middle row, one per column. After putting the chosen column between the other two, its middle card occupies position 5. This works for every initial ordering, provided the procedure and the reported columns are correct.
Ask partners to explain this before running it a second time. The second run can start from a fresh shuffle. They now predict from the mechanism rather than merely believing a repeated performance. Extend only if time permits: what if we put the selected column first? Trace positions instead of guessing.
The test supplied with this deck enumerates all nine starting positions and both orders of unselected columns at both collections. It verifies 36 traces; it does not measure classroom timing or prove a general-purpose computer.
A small instruction set
23:00–27:00. The card procedure is one finite algorithm. A more general machine can read and write memory, choose a next step based on a value, and repeat. A Turing-complete model can simulate any Turing machine, given unbounded storage and enough time for a computation that halts. Finite physical machines have resource limits. Some problems are undecidable and some programs never halt; “any algorithm runs to completion” would be false. Not every simple machine is universal. The card trick does not establish universality.
The Turing test is a different question about behavior in an imitation game, not the range of computable functions. Keep the terms distinct. A compiler is a program that translates a language into another representation a machine can execute. Students need not become compiler specialists to benefit from seeing such a translation and understanding that it is a mechanism, not magic. No compiler activity is required in this hour.
Where is the work happening?
27:00–29:00. The chess automaton became a famous spectacle, with a concealed human player operating the mechanism. The museum reconstruction helps make the arrangement concrete. Use the illustration to ask what the audience could see and what they could not. It is an AI-generated artistic scene, not documentary evidence of a specific encounter. Do not assert that the pictured occasion with Franklin happened in 1770.
The bridge is the changing mechanism behind an impressive result. Modern language models are not people secretly typing each reply. Human labor matters in a different place, including preparing and judging training examples.
People behind the examples
29:00–31:00. Amazon Mechanical Turk is a marketplace for human work. Requesters can obtain labels or judgments that support machine-learning datasets. The name provides a useful connection with the historical machine, while the mechanism is different. Ask whose experience a label captures and who decides what good work means. Compensation and task design deserve attention.
We will use InstructGPT as a separate documented example of human demonstrations and rankings in training. Do not say Amazon Mechanical Turk supplied that project’s workers; this deck makes no such claim.
A word from the hat
31:00–33:00. Tables of four to six. Each table needs a cup or hat, identical small slips, pencils, one shared response sheet and a reader. Show this exact prompt. Begin with an empty response. Everyone independently writes ONE proposed next word, or STOP. Put all slips in the hat, mix, then draw one without looking. Keep duplicate words as separate slips: if three people suggested “We,” it has three chances. Append the drawn word to the response. If STOP is drawn, finish that run.
Set a cap of eight selected words per run to fit this rehearsal schedule. The cap is a facilitator time limit, not a model’s learned stopping behavior. Punctuation may attach to a word. Anyone may dictate a slip or take the reader role for accessibility. No phone, account or microphone required.
Read the whole response again
33:00–39:00. After each selection, read the prompt AND the entire response so far. Empty the hat. Everyone now writes a fresh next-word or STOP suggestion based on the new context. Repeat until STOP is drawn or the eight-word cap is reached. Never cut up a prewritten sentence or keep old slips for a changed context. Save the selected words in order.
Reset to the identical prompt and empty response. Repeat the same rules for a second run. Do not deliberately alter the first word, split the group, or change the rules to manufacture a difference. Random selection supplies the randomness. Different outputs are possible, not guaranteed. If two stories match, that is an honest result. A third run is optional only if ahead of time.
Six minutes for two short runs is a hypothesis, not tested classroom timing. Have slips ready. If groups are slow, finish the second run with fewer words and use the resulting partial texts in discussion. Do not silently turn the activity into a single person writing the story.
Same prompt. Same process.
39:00–43:00. Invite two tables to read their outputs. At a particular step count the candidate slips. If four of six slips say “door,” its draw probability was four sixths in our setup. That is an empirical frequency of these people’s suggestions, not a measurement of a language model. A selected word changes the context, so the next set of proposals can change. STOP is a possible draw, not a majority vote.
This models sampling during generation. It does not train a neural network. People supply the candidate distribution from their own learning and judgment. A language model calculates a distribution over tokens using learned parameters and the input context; tokens may be parts of words. Generation can sample from that distribution. Some decoding choices are deterministic, so randomness is not the defining difference between a learned model and explicit program logic. Temperature is optional extra detail and is omitted here.
Where do the probabilities come from?
43:00–46:00. Offer a bounded intuition: during training, examples supply a prediction task; a learning procedure changes many numerical parameters to reduce its training error. In a neural language model those parameters help transform the context into scores for possible next tokens. They are not a readable list of instructions written by an author for every situation. Learned associations can be useful and also reflect accidental regularities or gaps in the examples.
Compare the card proof: we derived an exact destination from the stated rules. The hat made the influence of context and selection visible, but it did not expose how a neural network learns representations. A model can repeat an answer consistently without giving us a comparably simple justification that it is correct on a new input. Explicit programs can also be buggy, complex or undecidable to analyze in general. We choose mechanisms and verification methods appropriate to what we depend on.
People teach and judge
46:00–49:00. InstructGPT began with pretrained GPT-3. Human example responses supported supervised fine-tuning. People ranked outputs; those comparisons trained a reward model. Further reinforcement learning used that model’s scores. This is a specific historical workflow, not a complete description of every current model. Ask: whose preferences became the target, and what happens when people disagree?
Human feedback improved the reported behavior, but the researchers still described factual mistakes and bias. The hat exercise sampled suggestions; this training process changes model behavior before later generation. The hidden human in the historical Turk, distributed paid work, and training feedback are three different arrangements of human work. Keep the distinctions visible.
Who can build the tool?
49:00–54:00. Compare what the classroom just needed with producing a large pretrained model: gathered and prepared data, substantial computation, technical work, evaluation and continuing infrastructure. GPT-3’s paper describes a 175-billion-parameter model. That is a historical scale example, not a current price quote or a minimum size for useful machine learning. Training a small model, adapting an existing model and training a frontier-scale foundation model from scratch are very different undertakings. Individuals can do valuable work with smaller and openly available systems, while competing at the largest scale is highly unequal.
Relate this to Chloe. Access to an impressive service is real opportunity, but it can carry conditions that the user cannot change. A downloadable result may not preserve the running system, the ability to edit it, or the right to take it elsewhere. Avoid claiming all subscriptions are bad or all open-source systems solve access. Ask which dependencies can be replaced, inspected, repaired or exited and which cannot. A crayon-making activity has no classroom-scale equivalent for reproducing an entire modern AI stack. That gap is part of the subject.
After the activity ends…
54:00–60:00. Return to Bridget’s drawing and Chloe’s uncertainty. Give pairs two minutes: choose an activity they teach, name the mechanism children actually inspect, and name a dependency that remains beyond their control. Then invite two responses. Leave at least one minute for the last question or an unresolved concern.
Re-state the central commitment: give students opportunities to read logic, predict behavior under stated conditions and make a machine carry out their intentions, even if AI helped produce the code. Successful exposure to a product does not by itself provide that understanding. The mechanisms we examined are small and comprehensible. The larger barriers include access, cost, ownership, complexity and intimidation, and an hour cannot remove them. Teachers can nevertheless become much more attentive to what opportunity an activity leaves with a child.
The phone tool and microphone-triggered image reveal remain optional future investigations. This deck contains no listener, recording, live model call or mandatory hardware demonstration.
Art and implementation
The existing opening image and Mechanical Turk illustration are AI-generated. The story, factory and later services are fictional; no real artist or museum has been selected for the museum scene. This draft uses local Reveal assets already included in the repository. See third-party notices. No analytics or live AI call is part of the deck.