018 min
An unequal starting shelf
Allocate counters using previous clicks and calculate expected new clicks.
Ask: “Can an unseen book collect clicks?”
Listen for: “No, it first has to be shown.”
Brightlab
brightlab.Made for curious mindsbrightlab.Made for curious mindsBrightlab
Year 5 · Feedback loops
Can a recommendation create the popularity it measures?
Go to the investigation ↓
Investigate with Kiki the kookaburraA recommendation policy note with exposure histories and a feedback diagram.
Can recommendations change the evidence used for the next recommendations?
A recommender chooses what receives exposure. In this model, expected clicks equal exposure multiplied by fixed synthetic click rates. A popularity-only policy uses the resulting clicks to allocate the next round, so an early advantage can grow. Exploration reserves some exposure for less-seen items. These are deterministic expected-value calculations, not measured behaviour from real readers.
Before this lesson: Describe how collected examples affect a later choice. Useful earlier investigations: A reef with two stories.
A recommender chooses what receives exposure. In this model, expected clicks equal exposure multiplied by fixed synthetic click rates. A popularity-only policy uses the resulting clicks to allocate the next round, so an early advantage can grow. Exploration reserves some exposure for less-seen items. These are deterministic expected-value calculations, not measured behaviour from real readers.
Describe how collected examples affect a later choice. Useful earlier investigations: A reef with two stories.
Prepare four fictional reading topics with fixed click rates and unequal initial counts. Rehearse reset and single-round stepping so comparisons begin identically.
ACARA V9 AI curriculum connection ↗ · Technologies ↗
A complete teaching sequence · 65 minutes
Allocate counters using previous clicks and calculate expected new clicks.
Ask: “Can an unseen book collect clicks?”
Listen for: “No, it first has to be shown.”
Predict the dominant topic after five rounds with zero exploration.
Ask: “What part of the starting advantage can feed itself?”
Listen for: “More clicks cause more future exposure.”
Advance one round at a time, tracing exposure to clicks to next exposure. Record shares and concentration.
Ask: “Did we change what readers prefer?”
Listen for: “No, the synthetic click rates stayed fixed.”
Inspect the high-click-rate topic with almost no initial exposure.
Ask: “Why did popularity fail to discover this item?”
Listen for: “It barely showed it.”
Reset, reserve an exposure fraction for all topics and replay five rounds. Compare discovery and concentration, acknowledging any short-term click cost.
Ask: “What did exploration make possible?”
Listen for: “The less-seen topic could gather evidence.”
Explain the feedback loop and policy choice with two saved runs.
Ask: “Would clicks alone tell us whether readers benefited?”
Listen for: “No, clicks are only one behaviour.”
Popularity only reflects preference.
A high-interest topic stays hidden because it starts with almost no exposure, so the system gathers little evidence about it.
Choose an exploration fraction and compare five identical-start rounds with a popularity-only baseline.
Check learners reset both counts and round number. Ask them to point to the feedback arrow and explain its causal role.
| Criterion | Beginning | Secure | Extending |
|---|---|---|---|
| Loop explanation | Describes popularity as fixed | Connects exposure and later counts | Explains an initial-condition effect |
| Controlled comparison | Compares different starting rounds | Resets and compares equal-length runs | Reports discovery and short-term trade-offs |
Use two topics and physical counters before running four-topic rounds. Provide multiplication results if needed.
Add a decaying memory of old clicks and compare how quickly a newly useful topic becomes visible.
No real student browsing, clicks or reading histories are tracked. Expected clicks are synthetic model outputs and must not be described as observations.
Shared device? Turn remembering off. A project file lets you continue on another device.
Your browser is the laboratory
Predict which topic gains exposure over five popularity-only rounds with all click rates fixed.
Initial clicks: Adventure 20, Nature 12, Hidden histories 0.1, Poetry 5. Fixed click rates: 0.35,0.30,0.55,0.25. Each round allocates 100 exposures from previous popularity plus an exploration share.
The experiment opens after your prediction.
There is no penalty for being surprised.
Apple Silicon · PyTorch MPS
Simulate many initial exposure allocations with fixed fictional click rates. Change exploration and compare hidden-topic exposure and concentration across twenty rounds.
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-loop-mac-pathway.zip -d y5-loop-mac cd y5-loop-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
Simulate many initial exposure allocations with fixed fictional click rates. Change exploration and compare hidden-topic exposure and concentration across twenty rounds.
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-loop-dgx-pathway.zip -d y5-loop-dgx cd y5-loop-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 recommendation policy note with exposure histories and a feedback diagram.
Download the editable handout →