018 min
Name the real purpose
Compare a useful reading experience with a quick click.
Ask: “Can you click without learning much?”
Listen for: “Yes, the headline might be enough to attract a click.”
Brightlab
brightlab.Made for curious mindsbrightlab.Made for curious mindsBrightlab
Year 5 · Objectives
What happens when we reward the easy thing to count?
Go to the investigation ↓A revised library objective with a counterexample and a tested allocation.
Can a better score make the real goal worse?
A proxy is a measurable substitute for something we care about. Click count may be used as a proxy for useful reading, but quick sensational snippets can earn many clicks with little reading value. This allocation model computes two different outcomes from supplied fictional coefficients. Neither coefficient is a universal measure of learning. Choosing and questioning the objective is a human responsibility.
Before this lesson: Distinguish a measured score from the purpose it represents. Useful earlier investigations: The uncertain recycling gate.
A proxy is a measurable substitute for something we care about. Click count may be used as a proxy for useful reading, but quick sensational snippets can earn many clicks with little reading value. This allocation model computes two different outcomes from supplied fictional coefficients. Neither coefficient is a universal measure of learning. Choosing and questioning the objective is a human responsibility.
Distinguish a measured score from the purpose it represents. Useful earlier investigations: The uncertain recycling gate.
Set the fictional library purpose: help readers spend time with useful material. Print three resource types with click and reading-value coefficients. Clarify that the values are deliberately simplified and not ratings of real genres.
ACARA V9 AI curriculum connection ↗ · Technologies ↗
A complete teaching sequence · 65 minutes
Compare a useful reading experience with a quick click.
Ask: “Can you click without learning much?”
Listen for: “Yes, the headline might be enough to attract a click.”
Predict which resource a click-maximiser will favour before allocating the budget.
Ask: “What would a system pursuing only this number choose?”
Listen for: “The item with the largest click coefficient.”
Move tokens between resource types; inspect both computed clicks and reading value. Keep the total budget fixed.
Ask: “Which change raised the scoreboard but hurt the purpose measure?”
Listen for: “Moving tokens toward high-click, low-reading snippets.”
Find an extreme allocation that wins on clicks and loses on reading value.
Ask: “Did the optimiser fail to follow its goal?”
Listen for: “No, the goal was poorly chosen.”
Create a weighted objective or a minimum reading-value constraint. Test the same allocations and identify a new way it might be gamed.
Ask: “Does a combined score remove the need for judgement?”
Listen for: “No, we chose its weights and it still misses things.”
Submit a purpose/proxy table and two compared allocations.
Ask: “What should a human review that the score cannot see?”
Listen for: “Whether the resources actually help the intended readers.”
A measurable score is the real goal.
The highest-click allocation can have the lowest supplied reading-value total.
Add a reading constraint or new weighting and test the same allocations; disclose the fictional coefficients and omitted outcomes.
Check learners explain that successful optimisation can implement the wrong objective. Ask for a concrete allocation proving the divergence.
| Criterion | Beginning | Secure | Extending |
|---|---|---|---|
| Proxy diagnosis | Equates clicks with benefit | Shows a numerical divergence | Explains why optimisation amplifies it |
| Objective redesign | Replaces one unexplained score | Justifies a constraint or weighting | Tests it and identifies a remaining gaming strategy |
Use five tokens and whole-number coefficients. Compare only two resource types initially.
Find allocations that cannot improve one measure without worsening the other and sketch a Pareto frontier.
All resource types and coefficients are invented. Do not use the score to rank classmates, teachers or reading ability.
Shared device? Turn remembering off. A project file lets you continue on another device.
Your browser is the laboratory
With a fixed token budget, predict which allocation maximises clicks and what happens to reading value.
Ten tokens are allocated among snippets, guides and stories. Per token, clicks are 10,3,5 and reading-value units are 1,9,7 respectively. These are supplied fictional coefficients.
The experiment opens after your prediction.
There is no penalty for being surprised.
Apple Silicon · PyTorch MPS
Enumerate large allocation spaces, calculate multi-objective frontiers on GPU and compare how different weights hide or expose trade-offs.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download Mac Studio investigation ↓unzip y5-proxy-mac-pathway.zip -d y5-proxy-mac cd y5-proxy-mac bash setup-mac.sh source .venv/bin/activate python experiment.py --device mps --output results marimo edit notebook.py
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
NVIDIA DGX Spark · PyTorch CUDA
Enumerate large allocation spaces, calculate multi-objective frontiers on GPU and compare how different weights hide or expose trade-offs.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download DGX Spark investigation ↓unzip y5-proxy-dgx-pathway.zip -d y5-proxy-dgx cd y5-proxy-dgx bash run-dgx.sh # Open the localhost notebook URL printed by the container.
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
A revised library objective with a counterexample and a tested allocation.
Download the editable handout →