Playground
Play with the ideas.
Six small toys about metrics, models and machines. Break them, game them, argue with them. Everything runs in your browser on synthetic or stylised numbers, so no data leaves the page.
The Goodhart game
You run a support team judged on tickets closed per hour. Split your team's time, press Run, and watch both lines.
Stylised toy model. The weights are invented to illustrate the essay When the score becomes the goal; they are not data.
Will this customer churn?
A fictional subscriber appears. Guess, then see what a simple logistic model thought and why. Can you beat it?
Synthetic customers drawn at random; outcomes sampled from the model's own probability. Feature names echo the churn case study; the coefficients here are illustrative.
Scorecard cutoff
Credit scorecards trade approvals against bad debt. Move the cutoff, change how well the model separates good from bad, and see the trade-off.
Two normal score distributions, so Gini follows directly from separation. Stylised; see credit-risk-scorecard for a WoE/IV build on synthetic loans.
Brakes simulator
1,000 cases. The AI acts on its own above a confidence bar; the rest go to a person. Slide autonomy up and decide where the brakes belong.
Stylised: human review 40 units per case, a routine error 150, a high-stakes error 5,000. Illustrates Faster cars need better brakes.
Attention half-life
Attention fades; skill compounds. How long does attention have to last before the quieter curve stops mattering?
Stylised curves. From the essay Attention is not the same as value.
Spot the prompt injection
An AI agent is about to read each snippet. Is it ordinary content, or an attempt to hijack the agent?
Made-up examples for practice. Real defences: treat retrieved text as data, least-privilege tools, human approval on consequential actions.