A Sufficiently Complex Simulation
I’ve been circling a bigger question lately — what kind of memory a future organisation needs before an AI can genuinely advise it, rather than just autocomplete for it. I wanted a bench to test ideas on: some small, self-contained world I could run over and over, graft different machinery onto, and watch what happened. I was reading about Petri nets and wondering whether to build a business simulator from scratch.
Then I remembered Great Britain Ltd.
It was a game from 1982, the kind of thing you loaded off a cassette into a machine with a few kilobytes of memory. You play the government. Set income tax, VAT, duties on tobacco and petrol, spending on health and roads and schools — then a few years later there’s an election over 650 seats, and you find out whether the country still wants you. A whole national economy, boiled down to a handful of levers and squeezed into almost no memory at all.
I got curious about how they’d done it. Not nostalgic, exactly — I wanted to see the trick. So I had Claude reconstruct it clean-room from an archived listing: no original source, just the observable behaviour, rebuilt as a small Go engine. Taking something apart to see how it works is one of the oldest pleasures there is, and it turns out a 1982 designer modelling inflation and unemployment in a cassette game had made a lot of quietly clever choices.
Then, for fun, I put an API on it.
Two calls. Read the current state of the country; submit a budget for the year. That’s it. And here’s the part that started to interest me: I gave the player nothing else. No rules, no formulas, no explanation of cause and effect. You raise petrol duty and something happens to inflation and to your popularity, and you have to work out the relationship yourself, by living in it — the way the rest of us learn how the world works.
Which is when it clicked. That blindness is exactly the thing worth testing in an AI agent.
Most benchmarks hand a model a problem it can pattern-match against a million similar ones it has already seen. But drop an agent into a world it has never seen, withhold the manual, and give it a clear score — did you win the election, did you wreck the economy — and you’re testing something harder and more honest. Can it form a theory from observation? Update when the world pushes back? Get better over a run, or just thrash? A sufficiently complex simulation is a rulebook the agent has to reconstruct for itself, and there’s nowhere to hide.
So I ran it. I pointed a handful of language models at the two-call API and gave each a match of five games — no rules, just their own history to learn from between rounds. Most of them muddled along. One, GLM-5.2, did something I didn’t expect: it lost the opening game, then found its feet and won four in a row, crushing inflation to 4% and holding it there while its majority climbed. It hadn’t memorised a map; it had worked out the physics.
That was enough to convince me the shape was right — and that it deserved a proper test rig rather than a weekend’s tinkering. How the rest of the models fared, and how you build a fair test out of a forty-year-old game, is its own piece.
Great Britain Ltd is too small to be the real bench — five levers and a coarse economy only take you so far. But the idea it planted has since split in two.
One half is worth chasing for its own sake: building the benchmark properly. That’s a side project called Lorenzia — a world of sixty-odd thousand simulated citizens on a grid, grown from founders rather than sprinkled in, where the macro-economics — the revolts, the voting blocs, the Laffer curves — emerge from the population instead of being written in by hand. A world with no rulebook, and an agent measured on how well it learns to govern it.
The other half is the one I actually set out to build, and it’s still ahead of me. I want to simulate an organisation — probably a hospital — somewhere complicated enough that the original question has teeth: what does a place need to remember before an AI sitting inside it can offer genuinely good advice, and not just a fluent guess? That’s the bench that matters to me, and it needs more research before I start.
It started as a curiosity about a forty-year-old cassette. Most of the best things I build do.