“Let's simulate everything” experiments: what worked & what didn't

§1The claim, made testable

Post 1 ended with a promise. I had argued, from first principles, that every correction to my mental model pushed one way. A city twin must be scoped, or it won't work. But an argument from first principles is cheap until the record agrees with it. So this post brings the record.

Governments have funded city-scale everything-twins for more than a decade. Companies have sold scoped twins into the same market for just as long. Both histories are public. I explored both to see what worked and what didn't.

First, terms. I call a twin an everything-twin when its stated scope is the whole city — geometry, movement, economy, environment — in one system. A scoped twin answers a named question about one subsystem, at one cadence.

§2What happened to the everything-twins

Singapore built the most famous one. Virtual Singapore launched in December 2014, with S$73 million and a national ambition. It was the first digital twin of an entire country. Eleven years on, its durable core is a detailed 3D model of the island. Agencies use it for visualisation and for scoped test cases, such as crowd evacuation at the sports hub. That core is genuinely good. But look at what it is: geometry, plus scoped experiments run on the geometry. The living simulation of everything — people, money, behaviour — did not arrive.

The United Kingdom ran the experiment at national scale. HM Treasury launched the National Digital Twin programme in 2018. The prize was sized at up to £7 billion per year across the UK's infrastructure sectors. The delivery centre closed in September 2022. No national twin exists. What survived is instructive: a set of framework documents, a practitioner community, and one demonstrator. The demonstrator, CReDo, connects three utility networks to study one hazard — flooding. It is the most scoped artifact the programme produced, and it is the part still running. The programme itself was reborn under the Department for Business and Trade. Eight years after launch, it is appointing ontology suppliers and building prototype demonstrators.

Queensland is the local instance. The state prototyped a South East Queensland twin with CSIRO's Data61 in 2021, on the open-source TerriaJS platform. KPMG delivered a business case that December. The program page now describes the investigation in the past tense. The department has moved to a "statewide infrastructure insights capability" for its capital pipeline. Two of the twin's use cases survive, and both are about pipeline coordination. Six years out from the 2032 Games, no public twin has shipped.

The research literature reached the same verdict by another route. A 2025 critique in Computational Urban Science separates simple, complicated, and complex systems. Twins work well on the first two. Cities are the third, and there the limits are structural, not budgetary. A 2023 paper in Nature Computational Science, by Caldarelli, Batty, Helbing and colleagues, makes the matching argument from complexity science.

§3What the scoped twins did

The twins that shipped, and stayed shipped, were born scoped. OnePlan's VenueTwin was the official digital-twin supplier for Paris 2024. It modelled 49 venues, claimed large planning-time savings, and was retained for Milano Cortina 2026. It scopes hard: venues, overlays, and operations — not the city around them.

Aimsun Live runs as a real-time traffic twin inside road-operations centres. One subsystem, one cadence, one class of question. New South Wales and Victoria both shipped spatial twins on TerriaJS. They are honest about their scope. They visualise, and they do not claim to simulate. My own university runs the same pattern at campus scale. The UQ St Lucia campus twin models floods, movement, building occupancy, and energy use for one campus. It has flagged buildings that sit near-empty with the air-conditioning still running. Even CReDo, the one durable output of the UK program, obeys the same rule: one hazard, three networks, one region.

§4The pattern, live: a US$2 billion test case

You could object that §2 is a story about government mapping programs. Old platforms, slow procurement, no machine learning. So take the strongest counterexample on offer, funded five months ago.

Simile is the most credentialed attempt yet at an everything-twin of human behaviour. Its founders include the inventor of the generative-agents paradigm and the scientist who coined the term foundation model. Its stated mission is to "simulate all eight billion people on earth, accurately and honestly." It launched in February 2026 and reached a US$2 billion valuation by July.

The validated part is real. In the founders' study, 1,052 interview-based agents reproduced their subjects' survey answers at 85% of the subjects' own test-retest consistency. Enterprises including CVS and Gallup pay for the product. This is not vapour.

Now look at the envelope of that validation. Gallup is testing the method independently, and it has published the boundaries. The agents perform best on topics close to the original interviews. Accuracy decays with topic distance and with time. Gallup will not use simulated responses for published population estimates. That is a fidelity envelope, stated by a neutral party with no stake in my thesis. It is the scoped core, drawn inside the universal mission.

One more caution, in plain words. Simile's chief scientist says the product can answer "if we do X, what happens?" That is a hard claim to earn. The agents learned from people who live under today's rules. Change the rules, and people change what they do. Data from the old rules cannot say how. Economists call this trap the Lucas critique. I do not predict Simile's failure. I only observe where the evidence sits. Customers pay for the scoped part. The independent validators drew the boundary around the scoped part. The universal part remains a mission statement. The pattern held for a S$73 million government program. It is holding, so far, for a US$2 billion foundation-model company.

§5Capability maturity map & opportunities

Put the track record on a capability map. Venue twins are maturing. They shipped, and they are on their second Olympics. Traffic twins are maturing. They run live inside operations centres. Spatial visualisation is maturing, with two state-level examples in Australia. Land-use modelling is only part-way there. The open tooling is dormant, and the commercial platforms serve North America. Behaviour simulation is new, contested, and heavily funded. One capability is missing. Nobody credibly couples a spatial twin to a municipality's money.

The fiscal mechanism is simple. A council's rates revenue is close to a sum over parcels: rating category × statutory land valuation × a rate in the dollar. In Queensland, every input has a public counterpart. The cadastre is open. Valuations are open at trend level. The Brisbane City Council revenue statement is public, though as a PDF. A scoped twin can be built in this space. It uses the rates ledger, released every year. It focuses on comparing simulated outcomes of policy choices — this rating policy against that one.

The local timing is unusual. Queensland is the only eastern state without a shipped spatial twin. The 2032 Games give it a decade-defining reason to want one, and a public capital pipeline to model against. Industry sees the same opening. Consultants have called for digital twins to be mandated across Games infrastructure. Vendors argue Brisbane is falling behind on exactly this. The gap is open, the data is open, and the deadline is real.