An Intuition About Simulation
How the Use of A.I. Agents Could Transform the Temporality of Design
In this Post:
Architect: Sociologist, Soothsayer, Therapist, Urbanist, etc.
Could Future You Help Present Me, Help Future You, Today?
A.I.’s, Now, with Personality
Designing for the Future, with the Future
An Intuition About Simulation
When asked for advice on A.I., one of the things I always tell architects is to stop worrying about how to use AI to do what you already do, just faster and better. That part will take care of itself, whether you want it to or not. Instead, try to imagine how AI might lead to totally different roles & processes, that get us to a result that's either similar or superior. Uber found success because it realized that people had no intrinsic attachment to taxis, they just needed a way from Point A to Point B that didn't involve public transportation, or driving themselves. Fundamentally, it was technology that enabled a completely different model, and along the way, dissolved all the processes (and the industries) on which the old model was based.
The lesson? Be committed to the result – not the process. That includes the design process, which some researchers at Stanford may have just fundamentally changed. Time will tell.
Architect: Sociologist, Soothsayer, Therapist, Urbanist, etc.
Good design is hard to do well, for many reasons. I would say my top three are:
Designing well asks you to see a future over an undefined time horizon.
It asks for superhuman empathy, asking you to inhabit the dreams, motives and behaviors of people you don't know. Â
It asks you to do those first two things without often giving you a chance to see whether you did them well. Â
An architect must understand how the building's intended occupants will behave within the space. Not just now, but far off into the future. Architects must therefore be master interviewers - getting to know and understand their clients. But what happens when the building is sold to someone else? Or used by unexpected visitors? Or the demographics of a neighborhood change?
Product designers get constant feedback on their work, in the form of sales data, reviews, etc. Lawyers get instant feedback on their work when the judge slams the gavel and renders judgment. But architects rarely get either. They often only get feedback when something goes terribly wrong.
The closest we come to a formal feedback loop is the post-occupancy evaluation. For you non-architects, that's when an architect is invited back to the building after some period of time to investigate certain things. The nature and scope of the survey is usually determined at contract inception, and is almost always tragically limited.
The optics and politics ensure that most parties, most of the time, have little incentive to pre-arrange such a thing. An architect doesn't have much incentive to revisit her old projects, find mistakes that the client hasn't noticed, and bring them to the client's attention. Clients typically scoff at the idea of paying a professional to validate their own work after the fact - the initial fee should have ensured that the project was done correctly the first time! When I pay a mechanic to fix my car, I don't invite him back over in six months to ask him to tell me whether he did the job correctly.
And for both parties,
there's little joy in discovering errors about which nothing can be practically done, which is often the case when concrete is involved.
But what if we could circumvent all of that? What if there were a way to peer into the future, and understand how the building is used, over time, and make adjustments to our designs in the present? Â
Could Future You Help Present Me, Help Future You, Today?
I read a paper early last year that got me dreaming of an entirely different way to do Post-Occupancy Evaluations, and would, in the process, lead to an entirely different approach to design.
In early 2023 Joon Sung Park, a PhD Candidate at Stanford University, published a research paper entitled Generative Agents: Interactive Simulacra of Human Behavior, wherein he gave separate personalities to 25 different A.I. Agents, and turned them loose in an artificial town.1 The technology is simple enough, and built on top of LLM technology like ChatGPT. If you tell ChatGPT 'imagine yourself as a baker, opening your bakery for the day' and tell another ChatGPT 'imagine yourself as a customer, entering a bakery, looking to buy some doughnuts' you can expect those two GPTs to have a lively chat. Imbued with LLM technology, the 25 A.I. Agents were able to interact, and displayed surprising, spontaneous behavior (at one point, they collectively got together and decided to plan a Valentine's Day party).Â
The idea of a bunch of virtual characters running around a virtual space and getting into hijinks is hardly new. In video games, they're called NPC's (Non-Playable Characters) and they can appear quite lifelike. They're automatons, programmed with a series of if/then reactions. If the protagonist (the player, you) asks 'where can I get some food' they're programmed to respond 'there's a diner down the road.' Video game programming has gotten sufficiently complex that when the NPC is confronted with something outside of their programming, like 'where can I find a play-doh and cheese sandwich', they'll respond with some stock, evasive response, like 'you ain't from around here, are yah?' But fundamentally, they never do anything we don't explicitly program them to do. That makes them great for video games, but problematic for investigative simulation, because the only actions they can take are ones that you've contemplated in advance.
But Park's virtual people weren't NPC's. They weren't preprogrammed, and each AI Agent was autonomous. Each AI Agent merely reacted to things said and done by the other AI Agents. And that got me thinking about the design possibilities within such an experiment. Â
After reading Park's initial paper, it occurred to me then that this would be a way of doing the Post-Occupancy Evaluations while still in design. You could model a design in BIM, and somehow populate it with artificial people, and then observe how they react to the space, which spaces they favor, etc. Then, if things were going in a really wild direction, you could amend the design appropriately. When asked by the client why you put the gym here and the cafeteria there, you could say that you had modeled all the future occupants of the building for the next thirty years, and knew that the two spaces were exactly where they should be.
As excited as I was, there were some hiccups.
The obvious and immediate limitation would be that you would have to craft identities for all of those artificial people to resemble the clients and/or occupants of whatever building you were designing. For a high school, you would want to make sure that a certain percentage of the A.I. Agents were teachers, some were students, and you'd also have to account for the fact that different generations of students would rotate through the school.Â
To design that school well, you'd also have to have A.I. Agents that reflected that particular school. High school students in rural Montana might likely behave very differently than those in inner-city Los Angeles.
After talking with some tech friends, the idea of designing a custom, A.I. Agents for every occupant (or groups of occupants) in every project seemed out of reach. Any architect reaching for such technology would find themselves designing bespoke A.I. Agents for each new project. The only alternative would be to build out generic, autonomous virtual people and use them as one would use silhouettes in a photoshop rendering - stand-ins for the real thing.
That's the opposite of what architects do. Architects’ superpower is to understand people, at an individual or collective level, and design a space around their needs, not the needs of staffage.
A.I.’s, Now, with Personality
Park, et al, seemed to have recently solved this problem, and inched us closer to the idea of designing in a humanized, simulated reality, with a new paper ‘Generative Agent Simulations of 1,000 People.’ They created a protocol for imbuing A.I. Agents with the attitudes and behaviors of real-life, specific individuals, giving the A.I. Agents the ability to 'act' like a particular person.
The researchers interviewed 1,052 demographically diverse Americans for two hours each, then used these interviews to create 1,052 A.I. Agents that could predict how each participant would respond to various scenarios. They tested the accuracy of their model by subsequently polling, in parallel, the individual, and their new A.I. Agent Avatar, with General Social Survey Questions.2
The A.I. Agents achieved an 85% accuracy compared to the respective humans on which they had been designed. And they performed similarly well on personality assessments, and other assorted tests.
I don't know if 85% sounds high, or low, to you, but the A.I. Agents were better able to predict a human's behavior/reactions to various tests and scenarios than the methods social scientists typically use. If you want evidence of this, consider the last presidential election. Or any presidential election. We always hear highly-paid pollsters make broad claims like 'we think suburban white women in the Midwest are going to swing left.' Well, that's probably between 15 and 20 million individuals, and the pollsters certainly didn't interview them all. They interview a sample, and then extrapolate outwards based on race, gender, education level, income level, shoe size, etc. I don't know enough about polling methods to tell you whether that's theoretically sound, but the results speak for themselves - these guys are wrong more often than Jim Cramer. So maybe generalizing about people based on their race, gender, education level, income level, shoe size, etc. isn't the best method, but short of interviewing everyone, what can you do?!
Well, using A.I. Agents, Park and his team were able to significantly reduce those kinds of demographic biases in prediction accuracy across political ideology, race, and gender groups. They weren't completely accurate (only 85%) and nor would we expect them to be. It seems fantastical to try and capture a whole person in a two-hour interview. But the results were meaningfully more accurate than what we do now.
Park, et al, then did the next logical thing, and created a 'bank' of these A.I. Agents, which they think could be used by other social scientists. I think that's just the beginning. We can create these A.I. Agents with a 2 hour interview of an actual person – creating the person takes at least 9 months. Another couple of years before they can even talk. It seems inevitable that this will be the way that all polling and marketing – all social science, really – is conducted. If you have a new product to sell, and want to know how it's going to land with college-educated urbanites under 30, you can rent/recruit a sample population of 1000, or 10000 such A.I. Agents, and ask them what they think.
Designing for the Future, with the Future
I'm not sure if anyone considers architecture a social science, but you have to be something of a social scientist to do it well. So how can architects jump on this train? A few ideas:
A.I. Agents, based on real people, have obvious utility in emergency planning, allowing the designer a window into how people might respond to different evacuation routes or disasters.
You could use them to simulate how different types of people would interact with and move through proposed spaces, observing circulation, flow patterns, and general use of the space. I see it as especially useful where many crowds of diverse types come together and intersect, as in airports, universities or public stadiums.
They could also help architects understand longer-term patterns of building use - how spaces might be adapted over time by their occupants, which spaces get overused, and which get underused, etc. We could model how the use of proposed spaces changes under different potential changes in demographics. If we could understand those futures, we can design buildings to flex with them.
Architecture has made various efforts at simulation before. After seeing The Matrix for the first time, I developed a serious infatuation with Space Syntax Theory. The idea of it seemed so beautiful - the idea that you could mathematically predict how people would react to space! Whatever successes it had, one of the enduring criticisms of Space Syntax theory is that it is overly deterministic - it treats building occupants as units in a maze, that when given the stair/elevator choice, will make a predictable one. It smooths over cultural, social, and economic factors that determine a lot about what people do, including how they move through buildings. More abstractly, it focuses on space as a determiner about how people use space. Acoustics, materiality, program, even smells will affect how someone uses a space.
But I'm not sure that's a fair criticism.Â
Architects have always faced the same challenge in simulation: namely, that the subject of our work is too complex to be simulated.Â
Sure, we can model the environmental systems, because those obey known laws of thermodynamics and are easily programmed into an algorithm. But all the human stuff? How does one model all the complex, human interactions that people have with their built environment? How does one model the emotional effect that a particular space has on a particular person?
And if we ever got to a point where we could model all that complexity, would there not be something intrinsically dystopic about having all the wonder and mystery of human behavior reduced to bits and bytes?
Well, we're not there yet. Only 85% of the way there. Which sounds like too little to create a dystopian hellscape, but plenty enough to revolutionize how we design buildings.
Park's work takes us a long way towards such a revolution. His 1st paper suggested the ability to populate a virtual space (like a BIM Model) with un-programmed virtual beings, thereby resolving the 'NPC' problem.
His 2nd paper resolves the genericity problem. The A.I. Agents whom you select to inhabit your virtual building needn't be generic, soulless or vanilla. They can actually be based on the clients and occupants that you're designing for on that particular project.
Because A.I. Agents are built on top of LLM technology, they have a pseudo-consciousness. They're not conscious, of course. But they can talk. They can respond to questions, just as ChatGPT does. The combination realizes the possibility of asking future occupants of a building what they think of a design now.
It raises the possibility of, instead of interviewing your client for 8 hours, and still not getting all the answers you need, interview them instead for 2 hours, and use that interview to build a simulacra of the client. The Client's A.I. Agent, could then be asked questions about the design while the design process is ongoing, meaning the architect no longer has to guess about whether the client is going to approve of this choice or that choice. The A.I. Agent may very well be more articulate than the actual Client.
An Intuition About Simulation
Will such simulations devalue the 'intuition' that architects bring forward about how people can, and will use a proposed space? Maybe, but that's a negative way to look at it, I think. Twenty years ago, an architect traded on his or her ability to intuit the interior climatic and comfort levels of a building they were designing. Even from a sketch, a good architect could tell you, 'this room will be too hot' or 'the air in this space will feel stagnant.' By having a broad knowledge of the physical world, the sciences, materials and human behavior, a good architect could summon judgment about such things. And now, we use a plethora of digital tools to do robust simulations of any building's environmental performance. I don't think that such environmental analysis devalues what an architect does. It makes it easier for them to prove to the client that their judgment is correct. And it gives them the ability to test out many different ideas with precision and speed.
I expect that simulations using A.I. Agents, in a virtual space, during the design phase will play out in a similar fashion. AI will be able to show us the future of our design ideas, but it won't be able to tell us which future to choose. That remains up to us, for now.