For the past six years, I have been intensely interested in cities because of their complexity. They are living, breathing organisms that respond in complex ways to interventions, both planned and unplanned, and the process of making them smarter, more advanced, more efficient, and better for their citizens is intricate. Cities are a confluence of wicked problems, and many urbanists have referred to cities simply as 'the urban problem'.
Despite this complexity, I've always felt that there could be a principled, formulaic way, not too different from a formal mathematical standard, of understanding cities and ultimately improving them. With the rapid growth of AI over the past few years, my view has been simultaneously supported and challenged. At times, it seems that AI could be a fundamental tool in decoding the way cities work. On the other hand, models which simply predict the next word could hardly offer the solution to many urban problems, let alone the urban problem.
Before diving deeper into whether LLMs and foundation models could offer solutions to the urban question, let's first define it. An understanding of the problem and the opportunity is crucial because it is from here that we define the parameters of execution and the criteria for success. I am a trained mathematician so I will approach this as such. What is the optimization problem? Succinctly, it is to maximize urban welfare for residents. Certainly, this description should raise many questions. What determines the optimal allocation? What is the optimal allocation? Those are certainly important considerations, among many others, but I will defer them for now.
If the objective above is unsatisfactory, we can provide a slightly more expansive objective: to maximize the gain in utility that the average immigrant to a city gains by moving in and minimize the subsequent average loss in utility to existing residents. Admittedly, this still is rather lacking, not least because we need a principled way to measure both previous levels of utility
A natural solution to improving the objective is to enhance our understanding of how cities work. By having a better grasp on the constraints in an urban system, we can structure our objective function in a better-informed way. It follows that any attempt to fix the city or improve it without a granular understanding of it is, at best, misguided and ineffective, and at worst, harmful. This means that the optimization problem is no longer merely about maximizing utility, but about maximizing our understanding of the city. We've gone up a layer of abstraction and arrived at a meta problem: our ability to improve utility for urban residents is directly tied to the ability to sufficiently model the city to a satisfactory degree of error. To be able to solve urban problems, we need to be able to learn and predict their occurrence based on historical patterns and current trends.
The optimization problem is now: what is the most accurate representation of the city that we can form? This is, in a sense, finding the optimum in the field of formulations of the problem rather than the field of the problem itself. From a granular representation, the optimal way to improve a city is now reachable. The result is a double-edged sword. By being agnostic with respect to different views of how cities should work, the framework is general enough to be applied in different urban contexts. At the same time, the liberality with which different interpretations of urban improvement can be applied has the potential to exacerbate urban inequality.
Smart and efficient cities are those that invest time, effort, talent, and technology into developing this principled understanding of cities. This is why I'm bullish on world models; models with a physical understanding of the world will be key in transitively improving urban spaces.