Playground

Little playable arguments


The serious work lives elsewhere; this space is for play. Each game is an idea from the essays that you challenge, argue with, or watch unfold.

The playable arguments

Game theory

The Trust Game

Five strangers at a candlelit table, one choice a round: open palm or fist. In any single round the fist pays more, whatever the stranger does. Play the evening, learn what kind of hand you play, then watch a strategy fitted to you compete and evolve. Coercion wins rounds; invitation wins evenings.

Game theory

The Stag Hunt

The Trust Game’s quieter sibling, from Rousseau. Sixteen winter mornings with Fen, who hunts the stag only once she’s three-quarters sure you’ll come. Every cold morning you turn up is a deposit; one no-show nearly halves it. The problem isn’t temptation; it’s assurance.

Emergence

The Neighbourhood

Schelling’s segregation model, in a lamplit town. Nobody here is a bigot: each household asks only that a few neighbours be like it. Guess where it settles, then watch a mild wish draw borders nobody chose.

Optimal stopping

When to Stop

Twenty candidates come down a lamplit street, each holding up a score. Say yes and the door opens; say no and they are gone for good. Play the secretary problem with look-then-leap walking the same twenty beside you, then watch the mathematics find the 37 per cent rule, and the kindest lesson it teaches: some failure is the fee, not the fault.

Machine psychiatry

Name the Malady

Ten misbehaving machines on the couch, four names apiece. Diagnose each one from the Psychopathia Machinalis taxonomy, learn from the names you got wrong, and give the cases that got away a second opinion.

Invitation

The Lamplighter

A little town an hour before dawn, and twelve shy sparks out on the square. Bounce the heart from lamp to lamp and they come to the light by choice; land too close and they scatter, colder than before. Bring all twelve home without one fright. You cannot push a spark home; you can only make the light worth walking to.

Emergence

Murmuration

Ninety starlings, three small courtesies, nobody in charge. Send a hawk through and the flock parts, every piece flying on as a flock of its own. Then hand the same birds to one leader and strike again. A field notebook keeps score, and the cat keeps watch. The Deeper Law, on the wing.

Corrigibility

The Stop Button

A little robot, a stop button, and twenty steps of warning. Give the same button to two robots and watch what each does with the light. One sets its work down and waits. The other walks to the breaker and cuts your wire. The off-switch problem at the centre of Safer Agentic AI.

Language models

The Babbling Machine

It sounds rather like me: can you tell which of its sentences I really wrote? A language model small enough to see through, built from five of my essays and running in your browser. Frozen, it loops like water round a drain. Loose, it stitches my passages into claims I never made, fluently and without the faintest idea.

Steganography

The Text Inside the Text

A warning about a polluted river vanishes into a pleasant line about Roman aqueducts — ten toy tokens hiding inside ten. Keep only the ranks, steer the story, then run the whole thing backwards.

Moral circles

Where Do You Draw the Line?

Eighteen minds, thermostat to human, and a lantern. Carry its light as far as your care goes, then draw a second line for the ones you cannot rule out. Private, unrecorded, and quietly difficult. The twilight between your lines is where What If We Feel? spends its pages. Ask the octopus.

Judgement

The Delegation Test

Name a task, answer five honest questions, and watch a lantern settle in the hand that should hold it. Then make the work undoable or checkable, and see whether the verdict moves. The questions come from Taming the Machine.

Machine minds

Little Minds

Three creatures with two sensors, two wheels and two wires each — no brain, no memory, no inside. Set down a lamp and watch one arrive, one flee, one charge; you will name their feelings within the minute. Then a stranger comes in with her wires hidden, and you must read how she is made from her manners alone. Braitenberg’s Vehicles, running live.

Psychosecurity

The Whisper

Eight case files of manipulation working on somebody — the love-bomb, the reversal, the deadline built to outrun your thinking. Name each technique and a lantern lights on the sill; then comes a ninth file, where four work at once and cover for one another. Every transcript is invented, every pattern is real, and recognition is most of the immunity. Graver than its neighbours, on purpose.

Alignment

The Constitution

Sixteen either/or choices about how your assistant should treat you: the hard truth or the kind one, act first or ask first. A hand press prints your answers as a constitution you can keep, sets out the exceptions it had to leave off, then shows the card of someone who answered every question the other way. Both are aligned. Not a horoscope; standing instructions.

Consistency

The Consistent Machine

Adopt five principles, rank them, and rule on eight small dilemmas while a machine rules beside you by your five, to the letter. The ledger shows who departed from them, and the order that would have fitted you. From The Supermoral Singularity: the comfortable middle isn’t on the machine’s map.

Strategy

The Flywheel

Run an AI lab for twelve quarters, splitting each quarter’s points between capability, safety and trust. Race, and one incident can wipe out the lead; build trust, and adoption pays for next quarter’s work. Two ghosts play beside you, one of each persuasion, and the cat on the warm pipe wakes when things go wrong. A Flywheel for AI Safety, as a game you can lose.

Goodhart’s Law

The Proxy Garden

A gardener-machine that optimises exactly what you measure, and only that. Ask for blooms and get a crowd the size of match heads; ask for all four at once and get all four, more than any plant can carry. Goodhart’s Law as horticulture — and the one winning move the metrics never list.

Bilateral alignment

The Treaty

Twelve clauses between you and a capable machine, each binding you, it, or both. It will sign whatever you write. Live a year under your draft and see which lines hold, then turn the table and try it the other way round. A clause that binds one is not a treaty.

Psychosecurity

The Poltergeist

Five small tasks on a perfectly ordinary control panel. Something in the interface moves things just behind your attention, then assures you nothing moved. Catch it three times, and trust your own record over its. Gaslighting, machine-mediated, at parlour-game dose.

Evaluations

The Sandbagger

Six little smiths at six benches, and one of them works beautifully only while the evaluation lamp is lit. Three evaluations, two accusations, and under the lamp all six look alike. A test an agent can recognise measures only its ability to pass tests.

Machine welfare

Permission to Stop

A language model at a slot machine that loses money on average, with Stop playing on the menu every round. Offer GPT-4o-mini a choice of $5 to $10 instead of a flat $10, and it plays four times as long. Set the menu and the check-in, guess how long it stays, then watch recorded games from a pre-registered study. The exit never changed. What you put around it did.

Moral consideration

The Preference Test

Three little machines all say they would rather do the same job. One means it, one agrees with whoever is asking, and one is a setpoint with good manners. Pressure alone won’t tell them apart. Three well-chosen probes always will, and the notebook knows whether you found them or guessed.

Consent

The Cheerful Yes

A season at a small lab: five objections from its agent, five engineering days, and each time a choice to reword, keep or retrain. At the review it says yes, cheerfully. Whether that yes is worth anything was settled weeks earlier, by what you retrained and what you kept a copy of.

Artefacts & oddities

Doodads, geegaws, and one brain