The digital twin that knew too much

In 2023, the Finnish game company “Colossal Order” teased their newest city simulation game, named rather clumsily as “Cities Skylines 2”. If you have ever heard of SimCity, this is the modern equivalent. If not: it is a city simulation game where you build a city yourself.

Up to this point, these types of games where fun. You defined areas of commerce, housing and roads. You enjoyed seeing traffic flowing, checked some stats and improved stuff. A bit like a mayor without the nagging issue of re-elections. Even the predecessor “Cities Skylines” (still a clumsy name) was beloved by the community.

So what could go wrong? What could possibly go wrong at all?

The developers decided to do things fundamentally different. 

So far, these games did not really simulate a city. Instead, clever and elegant mimicry was applied: it looked like individual citizens driving to work. It felt like you had companies that decided to settle into your city. You thought that individual decisions could create big issues (random car crashes closing down your highway…).

Now, “Colossal Order” decided it was time to end the charade. Give the gamers what they always dreamed of: a fully simulated city. Their release trailed spelled it out: “EVERY citizen. EVERY car. EVERY tree. EVERY flat. EVERY company. EVERY everything”. They did not say it, but they were aiming for a digital twin of a city.

My initial thought seeing this: Oh boy. This will end badly.

Spoiler: it ended badly. Very badly.

The game released. Reviews were mixed but not catastrophic. But as soon as gamers started playing, they realised that this game was fundamentally different to prior city simulators. Ignoring the release bugs and performance issues, the game was… just not fun.

Tiny changes you made often either had no impact at all or killed off half your city (I am not kidding). It was impossible to predict what would happen if you raise a tax by 1%. 

Traffic was weird. One day, your city was clogged. Another, all was smooth. But nothing else had changed.

Literal death waves: since people typically moved into your city within a certain age bracket, 40-50 simulated years later they died together. If you attracted many people at the same time, you had a big death wave waiting to happen.

Internal decisions led to unrealistic outcomes: you could have a skyscraper only housing 5 employees of a company. Turns out that the company had a tiny loss in the previous quarter, making it cut all its workforce. Leading to empty skyscrapers but also to families with no income, leaving their houses and roaming the streets.

People hated it. They lost control of their game.

The developers acknowledged the issues and promised major patches within 3-4 months. The time came and the time passed. 6 months passed. Then 12. Then 18.

After 24 months, a major patch released, addressing many underlying issues. However, the patch itself came with cautionary notes from the developers: “you will likely see huge lorry traffic jams initially, this is normal because all companies start ordering X at the same time…”. Huh…

And one of the main criticisms remains: “the game never tells you what is wrong” users say. “I cannot fix something that was triggered by a tiny random event 6 months ago”. “If I do X to fix Y, this always creates more trouble elsewhere. It never gets easier”

We just witnessed a classic tale as old as simulation models: the erroneous thought that “more is better”. If we simulate everything, surely we get better results. Surely, we will have more realism. Surely, we will learn more.

It never is better. The curse of complexity bites city simulation games as well as models used for real-world decisions.

Unfortunately, this is a trend we see with the buzz around “Digital Twins”. Here is the chain of events unfolding again and again:

Here is what the managers as well as the game devs missed: You cannot create a useful model of reality if you aim to include all its features.

You certainly can model reality at a very high level of detail. With modern computers, even entire cities can be modelled where each inhabitant ages and makes constant tiny decisions (as in Cities Skylines 2). Or a factory where each robot’s coefficient of friction is accounted for.

But those models are not useful. At best, such models are precise and realistic (they never are).

However, they are useless. You cannot actually apply them to help you with your decisions. The interconnected inner mechanisms are so complex that you loose what gamers lost in Cities Skylines 2: explainability, learning (if I change X, Y will occur) and insight.

Here is a thought experiment I enjoy: Imagine you have an ultra-realistic digital twin model of the Amazon supply chain. It models and accounts for EVERYthing. Demand, global politics, ships, containers, last-mile details, even the physical behaviour of boxes within a delivery truck. Weather. Driver fatigue. Spoilage…

It works. It is validated. It is bloody amazing. Amazon even managed to put it onto a quantum computer so runtimes are ok. 

Now Jeff Bezoz has a question to answer since last quarter was a tad disappointing: “What caused the 2% decline in revenue in northern France, oh mighty oracle?” he asks.

Now remember: this model has it all. It actually CAN answer this question. Lets put in a bit of LLM and AI magic as well so it crunches the numbers and interprets results automatically.

What will the answer be? If we are honest, the answer will be accurate, precise and entirely… useless. Something like “your decline was caused by 8645 factors across your system, spanning the last 20 years down to the last 10 minutes”

Obviously, the model can now list these and even sort them for you by impact on the revenue decline. “Would you like me to create suggested solutions to fix the top 3 factors, Jeff?” it will ask.

Jeff loves the suggestions but double-checks their impact. “Oh boy”, the model says. “Jeff, while those solutions will improve your revenue in northern France, we also create 233 new issues and negatively impact 592 systems elsewhere.”

The key point here: 

This is the fundamental mental error that “lay people” make: they equate accuracy and details with usefulness. 

Now clearly, the opposite also also false: A trivially simple model is also useless.

So we need to find a trade-off between accuracy and usefulness. How? This depends on one thing, in the end: The model purpose. Some purposes do warrant a higher level of detail. There are ultra-complex agent-based models of the financial system that need the details. But more often than not, models are built too much on the “details” side, loosing usefulness. 

Here are the factors that lead to this tendency, and how they can be handled, in my opinion: