I was just watching a video with a chart that showed how long it took different technologies to get to 100 million users: the telephone, Facebook, and DeepSeek. It was 100 years for the phone, four years for Facebook, two months for Deepseek. Things are moving quickly, almost too quickly, which means you don’t have a lot of time to plan and test pedagogy. 

We are trying to do three things here. One, make sure that students understand the philosophical, legal, and disciplinary implications of using these kinds of technologies. I’ll be giving a talk to our incoming students as part of their orientation about the relationship between generative technology, architectural intellectual property, precedent, and academic integrity. And why you’re here to learn: not how to teach algorithms to do things, but to do them yourself. That’s one dimension. 

The second dimension is, we’re big believers in making as much technology as we can support and afford available to the students. So we’ve been working with the central campus to provide access to larger platforms, and to make things as available and understandable as we possibly can.

Thirdly, in the classroom, individual studio instructors are taking their own stance on how they want to see the tools used. We taught a studio last year where the students tried to delegate a lot of their design responsibility to algorithms, just to see how it went, right? 

Control. You lose a lot of design autonomy when you delegate to an algorithm. We’ve also been teaching a class called “Scales of Intelligence,” which tries to look at this problem from a theory, history, and technological evolution perspective, delving into the implications for practice and design. So it’s a mixed bag of stuff, very much a moving target, because the technology evolves literally during the course of a semester. 

It’s getting more interesting, minute to minute, very shifting ground. I was on the Yale Provost’s AI Task Force, which was the faculty working group formed a year ago that tried to figure out what we’re doing as a university. Everybody was in the same boat, it’s just some of the boats were tiny, paper boats floating in the bathtub, and some of them were battleships—like the medical school, with more than 50 AI pilots. We’re trying to keep up with that. I don’t know how good a job we’re doing now. 

It’s difficult to generalize, because there are a lot of variables: your willingness to experiment, a firm’s internal capabilities, the availability of data, and degree of sophistication. I’ve been arguing that because this technology is expensive and requires a lot of data and investment to figure it out, the real innovation will happen in the big firms. 

Everybody’s creating marketing collateral, generating renderings, all that stuff. The diffusion models and large language models, the two things that are widely available—everybody is screwing around with that. The question is, where’s the innovation? And it’s a little early to tell.

The other thing you’ve got to remember is the basic principle of technology adoption in the architectural world, which is: When you figure out a technological advantage, you don’t broadcast it; you keep your advantage to yourself for as long as you can, until somebody else catches up. A recent example: It’s not like there were firms out there helping each other adopt building information modeling.

I don’t know. The reigning thesis—I’m simplifying this—is that you can build knowledge from which you can reason inferentially by memorizing all the data in the world and breaking it into a giant probability matrix. I don’t happen to think that thesis is correct. It’s the Connectionists vs. the Symbolic Logic people. I believe that you’re going to need both of these things. But all the money right now is down on the Connectionists, the Sam Altman theory of the world. Some of these things are very useful, but they’re not 100% reliable. And in our world, as architects, reliability is kind of important.

Why do you say that? There’s a conclusion in your statement. 

That’s true, but you better check it.

So, therefore … what? 

I don’t know that it’s inevitable that machines could design entire buildings well …

The essence of my argument is: there are many places where AI is very useful. Where it begins to collapse is when it’s operating in a multivalent environment, trying to integrate multiple streams of both data and logic.

Exactly. Certain streams may become more optimized. For instance: If I were a structural engineer right now, I’d be worried, because structural engineering has very clear, robust means of representation, clear rules of measurement. The bulk of the work can be routinized. So they’re massively exposed. But these diffusion models right now can’t draw a damn floor plan with any degree of coherence. A floor plan is an abstraction of a much more complicated phenomenon. It’s going to be a while before these systems are able to do the most important things that architects do, which is make judgments, exercise experience, make tradeoffs, and take responsibility for what they do.

For purposes of this discussion, let’s stipulate that artificial general intelligence that can do anything isn’t in the foreseeable future, because once that happens, the whole economic proposition of the world collapses. When that happens, we’re in a completely different world. And that won’t just be a problem for architects. So, if that’s not going to happen any time soon, then you have two sets of questions. Question one: In the near term, does AI provide productivity gains in a way that reduces the need for staff in an architect’s office?

OK, in the near term, maybe we won’t need as many marketing people. You won’t need any rendering people, although you probably didn’t have those in the first place. But let me give you an example from an adjacent discipline that’s come up recently. It turns out that one thing that these AIs are supposed to be really good at is writing computer code. Because computer code is highly rational. You can test it and see if it works. There’s boatloads of it on the internet as training data in well organized locations, very consistently accessible—which is not true of architectural data, by the way. 

It turns out that many software engineering companies who had decided to replace their programmers with AIs are now hiring them back because the code-generating AIs are not reliable enough to write good code. And then you intersect that with the problem that was described in a presentation I saw a couple of months ago by our director of undergraduate studies in computer science, [Theodore Kim,] who said that so many students are using AI to generate code that they don’t understand how to debug the code once it’s written. He got a call from the head of software engineering for EA, who said, “I can’t hire your graduates because they don’t know how to debug.” And if it’s true here, I guarantee you, it’s true everywhere across the country. So you have a skill loss. 

Then there’s what I would call the issue of the luddites. The [original] Luddites didn’t object to the weaving machines, per se, but they objected to the fact that while they were waiting for a job in the loom factory, they didn’t have any work. Because there was this gap between when humans get replaced by technology and when there are new jobs for them doing other things: you lost your job plowing that cornfield with a horse because there’s a tractor now, but you didn’t get a job in the tractor factory, somebody else did. These are all issues that have to be thought about.

But transcriptions are so easy. I do not disagree that, over time, these algorithms will get more capable doing some of the things that architects do. But if we get to the point where they’re good enough to literally replace architects, we’re going to be facing a much larger social problem. 

There’s also a market problem here that you need to be aware of. These things are fantastically expensive to build, and architects are not good technology customers. We’re cheap and steal a lot of software—not good customers for multibillion-dollar investments. Maybe, over time, someone builds something that’s sophisticated enough, multimodal enough, that can operate with language, video, three-dimensional reasoning, analytical models, cost estimates, all those things that architects need. But I’m not concerned that that’s going to happen in the foreseeable future. It’s too hard a problem, unless somebody comes up with a way to train these things on much skinnier data sets. 

That’s the other problem: all of our data is disaggregated, spread all over the place. Nobody wants to share it, because it involves risk. When the med school has 33,000 patients enrolled in a trial, they’re getting lots of highly curated, accurate data that they can use to train their AIs. Where’s our accurate data? I can take every Revit model that Skidmore, Owings & Merrill has ever produced in the history of their firm, and it’s not nearly enough data to train an AI. Not nearly enough.

That’s always been under pressure. It depends on what we as a profession decide. I’ve written extensively about this. We have two options. The first option is a race to the bottom: Who can use AI to cut their fees as much as possible? Option number two, value: How do we use AI to do a better job and charge more money? That’s not a technology question, it’s a business strategy question. So if I’ve built an AI that is so good that I can promise a client that x is going to happen or y is going to happen, I should charge for that: “I’m absolutely positive that this building is going to produce 23% less carbon than it would have had I not designed it. Here’s a third party that can validate this. Write me a check.” 

Featured image courtesy of Easy-Peasy.AI.