AI & Architecture

Architectural Disruption and the Emergence of new Design Tools

8 provocations 

INTRO 

Thank you for your time Matias and for joining me on this little journey for the next 35 minutes or so. The questions I have prepared – mainly because I know that you are in the epi center of these investigations – are somewhat different from your recent publications and lectures. I formulated the questions as provocations to foster an engaged debate and discussion in regards to AI and its creative use in architecture, but also addressing the pink elephant in the room being prompt generated images, which currently oversaturates the internet. I am very happy that we can have this interview  in person here in Vienna. 

Q1: – the disruption

We often read, just like in your recent interview with archdaily that technology at large is “disrupting” the creative industries and architecture. You also say that neural Architecture is a true paradigm shift within the 21st century.  Do you think that recent phenomena and experimentation with these tools “disrupt”, meaning break with traditional notions of the architectural discipline or would you give here a different reading and reasoning?

MdC: Thanks a lot for having me and having an in-person interview. You of course also read my statement as a provocation to tease out the architectural community – in order to engage and debate. This is necessary and therefore I am happy about this interview. Let’s discuss first why I think that AI is the first true 21^st century novel design method, from my perspective or from the perspective of computational design. If you think about it, the majority of the tools we have been applying in the last 15 years were already present in the late 90 ies. They have been upgraded, optimized, polished, became faster. What started with an experiment of a dozen of people become the standard in the industry and profession today. But the tools and methods are not new. Its different with AI assisted tools in architecture. They are different – for a very simple reason, which is the technology. The technology to do these investigations did not even exist 5 years ago. Obviously, there is some trajectory of historic development, but the ability to apply them as a design tool, also with a specific speed and not total frustration – that has changed. The other thing is also the introduction of specific algorithms that have not existed before. GAN came up only in about 2014 -2015. So these are rather novel tools and the question is, how do we as architects respond and engage with these tools? I immediately aggressively went into this direction, having had previous interest with AI in Vienna with the OFA institute as a student. The reason why I think architects need to engage with these tools is because if we don’t somebody else will do it for us or instead of us – and that will not be very pretty. I am thinking about investors, developers, technology companies displacing architects – if we do not respond accordingly and start using these tools on our own terms and values.

Q2: – the design Process

I d like to dive in straight into the ins and out of prompt generated images and ask you the simple question how you read the novelty and meaning of these tools relevant for the architectural design process itself. I agree with the often cited notion that the images are, aesthetically pleasing and inspiring, but i disagree that they conceive spatial concepts –  and I would also claim that inspiration and imagination – even trough images and projects as through texts – has played an important role for architects for a long time. So what changes with prompt generated images within the design process or where do you see its meaningful application?

MdC: This question also goes back to a part of the first question – the disruption of the discipline, which we forgot. In these image based models, these diffusion models certain things happen there which are very interesting to how architects work. I d like to mention two of them. 

One, is the aspect of variation. Its how we have been trained and how we work, doing hundreds of sketches, models to iteratively approach design or a detail or a plan. This is nothing new and has been ongoing forever. The method of creating variation’s is very engraved in architecture. Prompt generated images and diffusion models, amplify this method. I have made a little calculation recently for a lecture and figured out that from May to September this year I have created 75000 images, just in mid journey. I am not even a power user. So the tools explodes the idea of variation, almost a bit overwhelming. How do you select the right image? Our intuition as architects though is challenged in a positive way by this overstimulation. Our sensibilities are trained with this tool. 

The other observation is that it turns around, or on its head the architectural production chain. Originally you do a sketch, go on to a 3d model, detail that – and also create renderings with a lot of production time – in order to present your design. Now it the opposite, you start with the image and work your way back into the production line of architecture.

I agree its not doing architecture per se, its not doing always good projects, but it definitely has the ability to challenge and push projects as a start. Its more like a sketching tool than anything else.

A sketch is not perfect because it needs room for interpretation, and that interpretation is what I think is important in diffusion models. So what are doing right now is finding in the latent space of the datasets we are given for this production things that we as architects have not seen before. This does not mean they are new.

Q3: – intentionality and adaptation

In one of your recent lectures you mentioned intentionality and adaptation of a system as its key characteristics to be declared as intelligent. I am wondering how these terminologies would apply to prompt generated images and the system behind it which heavily relies on machine learning, image recognition and evaluation of big data. I would claim to say that a simple google image search translates a noun into an image of that noun is sourced out of a large database. A response to an input. Where and how do you see intentionality and adaptation in these processes or also beyond MidJourney?

MdC: Well these 3 categories is obviously also the best proof that AI is not intelligent. Kate Crawford, in the Atlas of AI said: “ AI is neither artificial or intelligent” To perform intentionality you have to be conscious and AI is not conscious, we are pretty sure about that. But AI can mimic certain neurological processes that we know about. But also that’s about it. A researcher recently claimed that the best AI is half as intelligent as a cat.

Q4 – beyond the image

I think its essential now to get away from the blockbuster and look at AI in more general terms. You have recently visited here in Vienna several research institutions; you are leading your own laboratory “aril” interrogating the relationship between architecture and AI so i would be curious to hear where on a scientific level in more general terms the journey with AI speculatively is heading towards to.

MdC: Great question, but also a hard one to answer as there are so many different things going on at them moment. First we need to understand this journey as an interdisciplinary project not an architectural one only. We have to work together with computer scientist and roboticist, data scientists, and neuro scientists, philosophers – how AI can change human culture in general. Architecture is one achievement of human culture. What I see currently is already a shift of human culture because of AI. The anthropocentric idea, that within creativity the human is on top of the hierarchical pyramid and nothing else, is starting to erode and become a plateau where we have suddenly different players. This is also what I mean by “post human” design. Processes that are after the dominance of humans over creative processes. I am really interested in this shift. Accepting this shift will also enable us to push forward our own cultural understanding.

Q5: – shared agency

We can agree and its common consensus that within every design process with have a handful of authors and we have emancipated so far that we give credit to all shared and individual responsibilities within a design process. Now we have come to the point that we start to share authorship with machines as co-creators. Why is this suddenly different and what are the intellectual contributions of a machine in your opinion as opposed to using a personal computer 10y ago with modelling software.

Where is the line between a facilitator or enabler to a co-author?

MdC: This is a larger block of an investigation and conversation, but I have an example to answer the questions. But first id like to say that me personally I am profoundly suspicious of authorship these days in general. Authorship is a concept that got invented in the 18^th century, so its rather a new concept if you compare it in the lineage of human existence. It got also defined in a time where it was very clear what an author is. The guy who sits down and takes the ink and writes down something. He is the only one contributing, hence the author. This concept started to become vague with the rise of computers in general but much more now. 15 years ago if you were using word on your computer, every error that you make will be in that text. There was no auto-correct. We move now 15 year ahead and within the same software we have auto completion of sentences, alternatives of expression, auto correct in various languages, the mood of the text you are writing is recognized. This is the simplest tool we can imagine.

Now I have trained a neural network on all the texts I have written in my life in order to generate texts for me. So am I the author of these newly generated text. In a way I am because the network was trained on my own data, but still I have aid of a neural network. For me this is rather suspicious at the moment. We have seen this in the arts already, with the famous example of the portrait of Edmond Belamy sold at Christies. Who is the author here? The artist who came up with the idea to generate the image, or is it the programmer who wrote the algorithm, or is the author the artist in the dataset to generate the image? It’s a difficult question because the artwork is based on existing work but the result is something that did not exist before.

So maybe we have to get rid of the notion of authorship overall, because it does not work anymore.

Q6: – extended field of studies

How will architectural education need to change in the future after we have understood, interrogated and adapted these new tools into our design profession just like we have adapted digital tools and computational processes – and what are future competencies that an architect must have in order to stay relevant with these tools?

MdC: I see already changes happening in my own teaching in studio. I think the idea of a top down design studio doesn’t work so well anymore. One of the competencies that will need to be trained and mastered in my opinion is the ability to learn how to build a data set. It’s a simple rule of “garbage in – garbage out” as within any computational process. If your data set is not good you will not have good results. Architects will need how to do this – or let’s put this in another way – how to judge it. The other competency that will not only count for our profession but several ones is the ethical use of those toolsets. There are cultural, racial biases in the tools. How can you identify those and how do you work with them? There are also of course opportunities that are on the other hand fantastic. We were recently discussing int the lab, where we are building this large-scale plan data set – if the data set will be biased or culturally diverse as it is a participatory process of collecting. So the discussion is what are we achieving here? creating another international style, because we are using data from all over the world? Its like a plain level of architectural design. We figured that it is possible to do that, but also that this should not be the goal. The goals should be that because the data set is from all of the world and diverse it would be possible to respond to local contextual specificities in the design. It would have the ability to focus rather than generalize.

Q7: – datasets

Within the use of AI and prompt generated images the large amount of data set is set out as a positive criterion of the overall system. A necessity for the result. This makes sense again in a singular task with a given intention and goal, let’s say detect cancerogenic tissues within MRI images, but in design processes especially in the beginning often there is a multitude of directions and ambiguity. How does here the pooling from 5 billion images, or growing data sets in general – help in your opinion vs. the close study of 3 projects in detail. Would the use of AI within architecture not need to be isolated to singular entities, like f.ex structural optimization as practiced in some well-known engineering offices? 

MdC: The scale of the data set indeed is getting a lot of critique from the computer scientists’ community as they are aware that our own mind has the ability to do the same task with less data. They are trying to figure out why the brain still is better in let’s say object detection than machines? It should be the other way around. We still don’t know how this works. The other critique is the energy consumption to build and source from large data sets. Our brain comparatively uses a ridiculous amount of little energy to process enormous amount of data. I think we will see progress in how we build large datasets, and once we understand how the processes work we will also be able to reduce their quantity.

The other question on how to create a focused project with such large data in the beginning, is also interesting. I think it depends on the prompt in relation to what you want. So if you have  specific are you would like to investigate you can easily do that with diffusion models even with a large dataset. 

Everything is also of course dependent on the investigation or the question itself. Let s say you would like to research Adolf Looses “Loosbar” here in Vienna. Its probably better to go into the actual place, the library or read books because what can AI do here to help? Regenerate the “Loosbar” from datasets? Why would that be useful? It it a question of purpose often.

Q8: – multiple mediation

In my limited understanding of AI and its application i think that what AI is extremely good at is always a singular task. Recognize a human silhouette within an image. Quite useful for lets say self driving cars. Now within architecture i think that it is unavoidable to multitask and mediate several design intents and processes at the same time, not sequentially, not linear but literal as a field. I am thinking here about yes inspiration and intuition, organization, formal vocabulary, structures and materiality. To mediate and also understand how to adopt is a very human competency. How do neural networks mediate and adapt between tasks and goals?

MdC: You hit a very important point here, and I am going to tell you about an observation we made that tackles your question. Early on we were lucky to work with Justin Johnson who works with 3d models as datasets. It became during the research clear that these networks will be able to do either the interior or the exterior of a project but not both at the same time. So this is exactly the challenge your question is aiming at. We have not solved it yet, to be honest. We thought about possibilities to concatenate certain algorithms but then everything becomes so heavy computationally that it becomes frustrating to work with. Probably everybody is working on this pressing challenge that solves this dilemma on several levels. In cars they can do this already. We are lucky in Michigan to be close to these labs and be able to collaborate with Ford on concatenation of algorithms. One last things I would like to add here as to the aspect of intuition, creativity, organization – are very hard to capture and emulate with computational processes, let alone with AI simply because we don’t know how they inherently work. We don’t know what happens in the human mind when a creative thought is born. I put a lot of effort in researching if AI can be creative and I was convinced it is, but by now I know this is not the case. AI is not creative. The creative is the person reading the image and giving it meaning.

Dr. Matias del Campo is a registered architect, designer, and educator. He is an Associate Professor at Taubman College of Architecture and Urban Planning, University of Michigan, and director of the AR2IL – The Architecture and Artificial Intelligence Laboratory at UoM. He conducts research on advanced design methods in architecture, primarily through the application of Artificial Intelligence; collaborating with Michigan Robotics and the Computer Science department. Matias del Campo is the co-founder of the architecture practice SPAN. The award-winning architectural designs are informed by advanced geometry, computational methodologies, and philosophical inquiry. SPAN gained wide recognition for the design of the Austrian Pavilion at the 2010 Shanghai World Expo, and more recently for the Robot Garden at the Ford Robotics Building. SPAN’s work was featured at the Venice Architecture Biennale in 2012 and 2021; ArchiLab 2013, and the Architecture Biennale in Vienna and Buenos Aires in 2019. Solo shows include “Formations,” at the MAK in Vienna and the exhibition “Sublime Bodies” at the Fab Union Gallery in Shanghai, China. In 2013 the practice expanded its operations to Shanghai, China, where the practice is currently working on building projects of various scales. He earned his Master of Architecture from the University of Applied Arts Vienna and his Ph.D. from the Royal Melbourne Institute of Technology.