AI Lab

Explore transformational models

I have been fascinated by models all my life: how a diagram, a notation, a simulation can make an invisible process suddenly visible. Powerful AI lets me act on that fascination. An idea now becomes a working model in hours rather than months: real variables, interacting over time, running live in your browser.

That changes how we can learn. Instead of reading about evolution, inflation or inequality, you move the levers yourself, watch the consequences unfold, and come to your own conclusions. This is a fascinating shift: it will transform understanding and explanation. Have a go yourself.

The Economics Lab: interactive economic models

Every economic theory is a model: a deliberate simplification you can argue with. These let you argue with your hands: move the levers, watch twenty years unfold, and see which stories survive contact with their own arithmetic. All models are wrong; these are useful.

Gini 0.94 wealth tax inheritance RUN ▶

The Monopoly Machine

economic model

Why does the game always end with one winner? An inequality sandbox with taxes, inheritance, basic income and jubilees. Change the rules, watch the destiny change.

Run the model →

GOLDEN AGE can’t find a sitter → nobody wants to sit GBP 187 hrs

The Babysitting Economy

economic model

The famous Capitol Hill babysitting co-op, live: too little scrip and nobody dares go out, too much and nobody will sit. Recessions, made small enough to touch.

Run the model →

12345678 parents’ deficit 59 = kids’ savings 59

The Business-Card Economy

economic model

Warren Mosler’s kitchen-table parable: why taxes give money its value, and why the parents’ deficit IS the children’s savings. Four levers, one unbreakable identity.

Run the model →

social security healthcare education defence EU 60% €10 mld toggle: ON

The Dutch Economy

economic model

The Netherlands in honest round numbers, through three lenses: the EU rulebook, the investor’s ledger, and the what-if of a currency of our own. Includes the €10 billion mortgage-deduction switch.

Run the model →

restriction 35%

The Artificial Scarcity Machine

economic model

One lever: how restricted is access? Five stories, from a blockbuster drug to a designer handbag, and one number to remember: for every euro the owner captures, how much value for everyone is destroyed?

Run the model →

×6.6 well-being

The Transfer Flywheel

economic model

The same billion, given to different hands. The poor spend it onward, the rich file it away: watch spending ripple down the parade of incomes, and see the trickle-down test fail in person.

Run the model →

at home at work ×4,000,000 CO2e leverage

Climate Positivity

leverage calculator

What you can save as a private person, next to what you can shift through your work. Pick an action, pick a profession, and watch the private choice vanish against the professional one: the gap runs to a thousand times and far beyond.

Run the calculator →

Download the research paper (PDF) →

Playing with AI: turning an idea into something that exists

The surprise of these tools is not that they answer questions. It is that a stray notion, an old song, a small daily habit you wish existed, can become a real thing in an afternoon. Two of mine.

No Worries Mate

music video

An AI reproduction of a song I wrote and played with my high school band, Midnight Mojo. Video made with Sora, December 2024.

Watch it on Vimeo →

All I ask is the chance to prove that money can’t make me happy Otherwise

Otherwise

daily nudge

One quote a day to shift your thinking: a small daily crack in the frame. Built with my friend Chris. Free, on iPhone and Android.

Get the app at otherwisequote.com →

Learning machines: live demonstrations of how AI learns

One thesis, three engines: competence condenses out of randomness. What separates them is a single question: where does the signal that changes the machine come from? A verdict, a score, or a direction. Everything below runs on one of those three, or is a way of feeding one of them better material. Each demonstration starts from nothing and learns in front of you, with no pre-trained shortcuts.

Selection: the signal is a verdict

Vary at random, then keep or discard. It needs almost nothing to work, which is why it was the first idea anyone had, and why it is slow.

WDLTMNLT DTJBKWIRZREZLMQCO P METHINKS IT ISWLIKE B WEASEL METHINKS IT IS LIKE A WEASEL generation 43

Weasel

cumulative selection

Dawkins’ 1986 program, live: blind mutation plus selection assembles a line of Shakespeare in sixty generations, and the Monkeys toggle shows why randomness alone never would.

Watch it learn →

function risk(t) { if (t.amount > 1000) if (t.amount > 6750 && t.amount < 11450) } random: 5 38 of 60 40 generations, live

The auditor that rewrites itself

cumulative selection (a miniature AlphaEvolve)

2,000 transactions, 60 of them fraudulent, and time to check only 100. Random sampling finds five. Nobody tells the machine what fraud looks like: it writes its own rule, rewrites it, and finds 38. Weasel’s engine, with a language model proposing the mutations instead of a dice roll.

Watch it rewrite itself →

Download the explainer (PDF) →

Reinforcement: the signal is a score, and it arrives late

Nobody tells you the right move. You find out at the end whether it went well, and the hard part is working out which move deserves the credit. MENACE does it with beads in matchboxes and no arithmetic at all, which is why this counts as an engine in its own right rather than a kind of gradient descent.

XXOO beads = confidence win +3 · draw +1 · loss −1

MENACE

reinforcement learning

A tic-tac-toe machine made of matchboxes and beads that learns to win: 1960s machine learning, rebuilt in your browser. It learns from you, too.

Watch it learn →

€ 3.512,– week 27 · btw-aangifte factuur: € 840 − 2% korting? PAY HOLD

The Cash-Flow Game

reinforcement learning

An AI bookkeeper learns which invoices to pay early and which to stretch, and beats three benchmark strategies at a Dutch company’s books.

Watch it learn →

best bookkeeper 60 seconds, live 4.9M decisions

The Cash-Flow Game — live

reinforcement learning

The same network, untrained. Press start and watch it learn in sixty seconds, from paying every invoice the moment it arrives, to beating the best human rule of thumb. Nothing is replayed: it is training in your browser as you watch.

Watch it learn live →

Download the explainer (PDF) →

−4.8% vs 2-opt

The Routing Apprentice

reinforcement learning

A neural network teaches itself delivery-van routing by trial and reward, and beats the classic textbook method by five percent.

Watch it learn →

Gradient descent: the signal is a direction, for every knob at once

Not “you were wrong” but “this weight down a little, that one up”. Far faster than the other two, and it only works when you already have an answer to measure the distance from.

loss 4.9 → 1.0 jq3Òwp…fz kx∅v the probleem of scarcty is **3.1 Reframing the problem** 67 min on M4

nano-Karim

gradient descent

A tiny language model trained from scratch on my own writing. Watch it go from symbol soup to my style, attempt by attempt. Style without thought, made visible.

Watch it learn →

same net, other soul tWφe ol…nq srr∅h KING HERNY: what sayst thou FIRST CITIZEN: Speak, speak. iteration 50k

nano-Shakespeare

gradient descent

The same tiny model, raised on Shakespeare instead. Same architecture, different soul: the data is the destiny.

Watch it learn →

And one thing that is not an engine: self-play

Self-play does not teach anything. It solves a different problem: where do the games come from? Play a copy of yourself and the opponent is always exactly your own strength, and you never run out.

self-play only — no human examples generation 7,038

Connect Four

all three engines, fed by self-play

The full DeepMind recipe in miniature: no human examples, just a copy of itself to play against. Self-play makes the games, reinforcement supplies the win or lose, gradient descent moves the weights. It made itself effectively unbeatable overnight.

Watch it learn →

Ask of any learning machine you meet: where does the signal come from? Verdict, score, or direction. That one question sorts the whole field, and it is why “is it reinforcement learning or gradient descent?” is a badly formed question: reinforcement learning is the framing, gradient descent is the machinery, and AlphaZero is both at once.

Scroll to Top