Fun With AI

I noticed somethig different in WordPress yesterday. The Image source window now has Generate With AI as a new choice.

Clicking it gave the following screen inviting me to describe the image I’d like to create. The theme of this post was already about the difference between the image I’d like to create (on the basis of some preconceived notion called a design idea) and the image I’d like to have created for me (on the basis of some vaguely articulated wishes).

What follows is what I’d already written yesterday.

• • • 

Suppose for example, I didn’t have any idea what I wanted or how to express it so, for my text prompts, I went with a modern villa with three-dimensional organic curves despite not knowing what I even meant by ‘three-dimensional organic’ curves? I expected to be served some warmed-up Zaha Hadid.

Example 1a: a modern villa with three-dimensional organic curves

And yup, I wasn’t disappointed. I reset the random number and gave it another spin just for the fun of it.

Example 1b: a modern villa with three-dimensional organic curves

Okay-ish, but what’s with the pink – is pink the new white? But more to the point, is this closer to what I wanted? I don’t know because I never had any preconceptions to begin with. But having said that, it’s not far from what I was expecting. Q: In the field of design, when did having a preconception become a bad thing? Was it about the same time intern farming began? I think it was circa 1990 when I first noticed projects attributed to Herzog de Meuron or Jean Nouvelle having no one style.

Example 1c: A dynamic six-story building with three-dimensional organic curves

I’ve also had no idea what a ‘dynamic’ building is either but I was sure I’d find. Rightly or wrongly, curved surfaces have been mis/understood as an indicator of creativity for some time now but I’m not thrilled with this proposal. Either AI is lacking or I am. One peculiarity I’ve noticed is that AI doesn’t seem to be very good with numbers. I asked for a six-story building not a nine-story one. Perhaps it’s the number of stories that’s dynamic? In which case, fair play, but we’re entering strange territory if ‘dynamic six’ means seven or nine.

At one of these urgent client presentations AI is said to help us with, it’s not helpful if the client is wanting a six-story building and we give them a nine. Clients notice things like that.

Natural language suddenly doesn’t seem like how we should be talking to algorithms. We’ve nothing in common so it’s unrealistic to ask it to articulate our wishes and desires.

• • •

Before we go any further, I want to see if WordPress text-to-image AI can do any better. You judge. On the left below is modern villa with three-dimensional organic curves and on the right is dynamic six-story building with three-dimensional organic curves [even though I count seven. Ish.]

• • •

These images above are variations on what to expect if you don’t have any idea what you want.

But suppose you do.

Suppose you have one of those preconceived ideas called design ideas? How does it play out?

I recently learned about a phenomenon called the “Tom Cruise problem”. It seems that some people – let’s call them researchers (or “university researchers” as in “university researchers have proven that” as in old-time toothpaste commercials) – asked ChatGPT a question they knew the answer to, but when they asked the same question in a different yet directly related way, the response was less certain. It’s called the “Tom Cruise problem” because the first question was “Who is Tom Cruise’s mother?” – to which the answer is Mary Lee Pfeiffer. The answer to the second question “Who is Mary Lee Pfeiffer’s son?” should have been obvious and immediately forthcoming but wasn’t. This made me wonder what would happen with text-to-image architecture generation if we already knew what we were looking for? After all, somebody knowing how they wanted something to be once used to be the paradigm for architectural design. Only later did it come to be about selecting from multiple alternatives and only later still with the advent of parametric design did design come to be about selecting from infinite arrays of alternatives. With AI now it’s apparently about “navigating solution spaces.” But do we really want a design process based on efficiently and inexpensively satisfying some client appetite for multiple alternatives? We’re told we do. I don’t know if architects suddenly started saying design was all about layering as soon as tracing paper was invented, but we did hear a lot about design being about double curvatures the moment they could be generated. The history of the relationship between technology and architectural design is one of tail wagging dog and AI is just its latest manifestation. So let’s get some agency back and ask AI to generate proposals for design problems that aren’t problems because we already know the answer we’re looking for. Let’s start with an easy one.

Example 2: a white, modern glass and steel house in a grassy clearing

Architecture has always adored a grassy clearing and it’s not difficult to make a building look good in one, especially if two of the prompts are ‘modern’ and ‘whilte’. Of course, I intended the above prompts to guide me towards The Farnsworth House that I had in my mind as my design idea I wanted options for. I wanted to know if AI would be able to help me arrive at my image of The Farnsworth House for a world in which The Farnsworth House didn’t exist. Looking at this image though, I think I’d like to see it a bit more off the ground, and the corner columns less thick. I want a kitchen and fireplace feature, no full-height internal walls, and maybe some intermediate level or even a large step up to the door. I get the feeling that, after a few tweaks it could easily be made a lot closer to my ‘original’ idea but this wouldn’t be using AI in the way it was intended to be used. Bad me.

Example 3: a modern white two-story building on columns in a grassy clearing

I didn’t want to give the game away by using the word ‘pilotis’. I didn’t say anything about the type of tree making this clearing and so they seem odd. Now I think about it, the columns seem a bit on the thick side and this entire ‘modern’ house is a bit bottom-heavy overall. The grass needs cutting. More work needed. We all have an image of Villa Savoye in our minds, but it took LC a few iterations to get to the one we know now. [c.f. Architecture Myths #20: The Villa Savoye]

Example 4: a building shaped like a duck

I’m in no rush to see this one built but it’s quite okay in a stupid kind of way. I was surprised when it turned out not to be a white duck so I obviously need to work on this one some more if AI is going to help me realize my dream duck. It’s no good the idea staying in my head, after all.

Example 5: a sail-like opera house by water

AI “read my mind” on this one but let’s cut with the casual anthromorphization. This response was pretty-much determined the moment I input the word-string ‘opera house’. It’s a good approximation of my design idea but only because images of a sail-like opera house by water were data-scraped way back. As an educator in a world where a building called the Sydney Opera House actually exists, this is plagiarism and not creativity or even intelligence – all of which, btw are also casual anthromorphizations. The question remains: ”If Sydney Opera House didn’t already exist, would have I stopped at this and shown it to the client?” I’d like to think not because it’s not exactly what I had in mind. (To be fair, what was eventually built wan’t exactly what Utzon had in mind at the beginning either.)

Example 5: a villa above a rocky waterfall in a forest

I almost like this but it’s a bit grey and the waterfall my client has in mind isn’t a Chinese one. My idea for my house (that I’ve provisionally called “Fallingwater”) has a closer and more picturesque (a.k.a. compositional) relationship with the water and waterfall. This Fallingwater I was working towards was more spread out and almost hemmed in by the surrounding trees. I do like how the blockiness of this proposal house mimics the cliff edge but, all in all, I think the landscape needs to be dialled down a bit.

AI image generators generally produce villas in dramatic settings – at least going by the numbers being shown off on LinkedIn. I’d be more likely to think of AI image generators as a design tool if these landscape representations were based on actual contour data fed in somehow but, if there was an appetite for such reality then this functionality would’ve been demanded and duly satisfied by now. Similarly, there doesn’t seem to be any desire to know the estimated carbon cost or even the construction cost of the buildings in these images. We get the technology we deserve.

I chose Fallingwater for this experiment because we all know it and because many people believe it to be the epitome of architectural creativity and a masterwork of pure genius. We don’t know how many sketches Frank Lloyd Wright made before he sat down and drew it. I suspect none because, if there were any sketches or exploratory massing models in foam or clay or whatever, then books and papers would have been written about them by now. Whether or not you believe the story that Wright drew the entire house in one night, it seems like he had either had a complete or fairly good preconceived idea of what the finished house would look like. I doubt he was designing it as he was drawing it. Fallingwater is the conclusion of a little understood process we call the design process, and it is how it is because somebody imagined it to be and feel that way. It wasn’t selected from a range of design options or alternatives or plucked from a solution space or, even if it was, the process was unconscious and not necessarily by the same process we now have a representation of.

HOW TO PLAY

1) Type some prompts such as ‘villa’ ‘above’ ‘rocky’ ‘waterfall’ ‘in a forest’ into your AI text-to-image generator of choice.
2) See how close you can get to the Fallingwater you have in your mind. Will you settle for something less? Or will you get distracted by other potential outcomes?
3) Try to focus on the design and not on getting the landscape or the image right.
4) Don’t use the word Fallingwater (you’ll only get an image of a waterfall anyway), the giveaway Frank Lloyd Wright, or upload any ‘reference’ image.
5) Remember that the Fallingwater you have in your head is YOUR design idea. It’s what you want AI to help articulate for you. It’s the potential outcome you want. I don’t think anyone can claim AI can ‘help’ them design if it can’t help vizualize an architectural idea that pre-exists either in reality (as in this game) or in someone’s imagination (as designing used to be). If it can’t do that then it’s not an aid to imagination but a substitute for it, providing approximations of novelty with no benchmark by which to judge them.
6) Play responsibly. Remember that redefining design as a process of searching for potential outcomes may be a cheap way of “staffing” an office but its costs aren’t exactly hidden.

• • • 

Revisited this week

• Misfits’ Guide to HONG KONG
• Infinite Monkey Architecture
• AI & I