I recently went one-on-one with Gal Rozensweig, founder and CEO of FormX
Adam: What is generative design in architecture, and what should everyone understand about it?
Gal: Yeah, it’s a great question. I think if you ask most people, they may imagine that we now have this AI that is very smart and is basically becoming an architect-level kind of entity. You ask it to generate something, you describe it well enough, and then, like magic, you get a brand new design for a house or a building where everything is already there and already thought through. Maybe one day. I do think it’s going there. But today, I think the generative attributes of AI can help design happen faster. They can help designers and professionals solve problems faster and get options for a design they’re trying to do in a certain area of a building. For example, if you want multiple layouts of a bathroom, AI may be able to help you with some options, and then you can pick the one that fits best. So the way I think about generative AI right now is more as a helper for the professional and a way to speed up the process, rather than as the architect that is doing everything for you. I think that’s the more realistic view right now, and that’s also what we’re trying to do in FormX with this technology.
Adam: How is generative design changing the way architects are approaching design today?
Gal: So with AI, you now have the ability to interact with software using natural language and voice, and I think the user experience is becoming much more immediate. There are many companies now using the power of AI to generate certain portions of the design faster. Generative AI basically means that you generate some architecture or some attributes for a building, or certain options for a design, or even visualizations for a design, in an automated way. There is some engine in the background, mostly large language models based on a lot of information and some machine learning algorithms, and they can generate results.
The challenge is how to take these capabilities and make them useful. How do you make something that at the end of the day becomes a buildable design, something usable, something with the right floor plan and the right flow, rather than something that is just nice to look at? For example, many people have experience generating renders online. They use some of the new tools and create very beautiful images. But if you actually try to use them, you can see that some of them change the architecture to make it nicer or more beautiful. They don’t necessarily do something practical. They definitely help with inspiration and concept development, but one of the challenges of the industry is how to make it closer to the actual design and not let these AIs and LLMs float around and change things on their own.
That’s one of the challenges. If you can write software where the automatic generation is under control and makes useful things, and the architect doesn’t have to spend hours trying to control it, that becomes very valuable. That’s actually one of the things I’m hearing from colleagues. They say, I’m using this AI software and these generative capabilities, but it’s running away from me. I spend two or three hours trying to get a result that I already knew I needed, but the tool keeps going right and left and up and down, and I’m trying to control it. So software that gives more control to the user while still using these tools would be very useful, and that’s one of the things some of the companies are trying to solve right now. One of the problems with AI is that it’s hard to control. It has its own behavior. There are all kinds of tricks to get it under control, and the companies that can harness the power while also controlling it are the ones that will do well and actually deliver something useful for the professionals in the industry.
Adam: What is out there today that can and should be utilized?
Gal: There are many things out there, but they are just at the beginning, so time will tell which directions produce the winners. Of course, something that almost everybody has seen is the ability to do visualization faster. Renderings and images were among the first things to come to market once AI became more available. Some of the challenges I mentioned, especially around control, are still part of the problem, but many new software products are available, and even the biggest players like Google and other leading companies are providing tools as part of their AI offerings that already have rendering and generative capabilities built in. So that’s something becoming available to everybody, and I think the industry is already starting to make use of it.
The second category is probably automatic planning. There are many companies trying to automate plan creation. So let’s say you have a design and you want to create the plan set, or even if you already have plans, you want to validate that the plans are according to code and identify certain problems. That’s another category that is becoming more and more developed, and there are several startups working on it.
On the structural side, there are some companies trying to develop structural automation, or tools that help professionals create structural designs faster. Even leading companies like Autodesk with Revit are providing interfaces these days and plugins and different ways to manipulate the software using maybe some external AI tools. This is the beginning of what we’ve been seeing in the market. Some of the incumbents are adding interfaces and capabilities to interact with software differently.
And then there are startups doing automatic floor plans. For example, if you give them some kind of boundaries and dimensions and certain constraints, the software will automatically generate options for plans. If you say, I want something with three bedrooms, two bathrooms, and make it 40 feet by 80 feet, the software will try to come up with options. These are some of the directions, just as examples, of companies trying to use the new capabilities of AI and LLMs and the other tools that are becoming available.
Adam: You mentioned that a big part of generative design is creating options, and the conventional wisdom is that generative design is about creating options. But what really determines whether those options are useful?
Gal: Yeah, it’s a great question. Architecture is a combination of science and art. Some of it is science. Some of it is rule-based. There are certain rules that, if you meet them, you have a useful design. For example, you’re not going to create a hallway that is two feet wide. It’s too narrow. If you walk into that hallway, you’re probably not going to like the architect who created it. So you can actually set rules in the software that the generative AI or whatever tool you’re developing has to meet, and that is what makes it more useful.
But definitely some aspects are more about art, and those are difficult to capture. How do you really make sure this is a nice design? What actually constitutes a beautiful design, or an artistic design, or an innovative design? That is much harder to define. It will be very interesting to see how the industry and the technology solve that problem, because at the end of the day, you want architects and designers to stay creative and create something innovative, something artistic, something new.
But the first part, the rule-based part, the things you can define mathematically as constraints or goals, those are relatively easy. If you set those boundaries, then even if you have AI or LLMs or something working like an agent and doing some of the work automatically, if it abides by those rules, it will produce something much more useful.
Adam: When you’re starting a project, what are the key inputs or decisions that ultimately shape the outputs that you get?
Gal: There are many. That’s one of the challenges of great design. It has to adhere to a lot of different parameters and considerations. One of them is what the client really wants. What do they like? What is their inspiration? What is their vision? That is a guiding principle of any design. Then, of course, there is the site itself. Where are you building? What are the constraints of the site? What is the direction of the sun? What are the neighboring lots? What are the code constraints?
There are a lot of rules you have to follow for the city, the county, and the state. Then there is usability. What are you going to use the building for? How many rooms? How many bathrooms? What kind of spaces? What is the flow between them? What is the overall style? There are a lot of considerations that have to come together to get a great design.
The question is, what is the most efficient way to move from this wide variety of possibilities and all of these different considerations into a winning design? You can start from scratch and take everything into account from the beginning, or you can start from use cases that have already been solved and iterate from there. Those are kind of two different approaches. But yes, all of these considerations matter, and if you’re doing software that is supposed to help with architecture, it has to take all of them into consideration. It’s definitely not an easy problem to solve.
Adam: Once you have a large set of options, how do you decide what’s worth developing, instead of just relying on what the tool is telling you?
Gal: If, while you’re going through the design iteration, you can get signals about certain aspects, that becomes very useful. For example, cost. Or the time it would take to build the structure. Or whether some of the structural elements are becoming unreasonable, like needing a 40-foot steel beam to create the design, which may be too difficult or too expensive.
Software that can indicate the complexity of building it, the cost of building it, the time required, and whether it meets code requirements will definitely help speed up the process and help ensure that you’re not designing something that doesn’t make sense. There are also all the architectural rules, such as the minimum size of a bedroom, the minimum width of a corridor, ceiling heights, and many other considerations. If the software can take all of that into account and give you indications while you are iterating, that would be very useful.
I’ve heard so many stories, and maybe you have too, where someone goes to an architect or a design-build company, and they spend several months working on a design. Everybody likes it. Then they actually try to price it out by getting bids from general contractors, and it comes back at twice the budget. Then they have to start from scratch because it is so far beyond what they can afford that almost nothing from the last three or four months can really be used. That’s just one example, but I think I’ve heard that kind of story many times. So if software could indicate the rough cost of the building while you are designing, it could save months of going in the wrong direction.
Adam: That’s a great example. A lot of what we’ve been talking about really centers around where generative AI and the tools around AI can add real value, and where they can’t, and where the human being ultimately needs to step in. How do you think about that line? Where do the tools fall short? Where does the architect still have to step up and lead?
Gal: I think anything that relates to art and creativity is still in the domain of the professional. We still need the professional to pick the direction, control the process, and think about more complicated or more out-of-the-box ideas. Those things would still be driven by humans right now, much more than by AI.
Anything that is more mechanical, where you can generate options based on a lot of background data, or where there are a lot of examples to learn from, or where you are doing calculations, repetitive tasks, or local generation of options, all of that lends itself very well to AI and automation right now. I think that is where the industry is going to focus.
I can imagine software that gives many options and examples not only to the designer but even to a homeowner. If I think further ahead, I can imagine someone who wants to build a home using software that gives them a lot of inspiration, a lot of options, and a lot of possibilities to help them start thinking about what they really want and what kind of design would be right for them. I think the industry will move in that direction, too, where some things become more self-service and less reliant on professionals in the early stages. So I can imagine software that gives the homeowner or the developer a lot of information before they even talk to a professional, by presenting possibilities, inspiration, and preparation for the process. Then later, the professional comes in and does the fine-tuning. I can definitely see that as part of how the market evolves.
Adam: What are the most common mistakes that you see architects make when they’re setting up a generative design problem, and more broadly, what are the most common mistakes that you see architects make around generative AI and generative design?
Gal: Actually, I don’t think I’ve seen enough different architects using generative AI to date that I can speak about that intelligently, because I think these tools are only just arriving. I have seen a lot of architects using generative renders for visualization, and I’ve seen some architects using automation in software that provides help with certain tasks. But real generative design for architecture is still in its infancy, and I personally haven’t seen enough examples yet to point out consistent mistakes. But if you want, I can guess what some of these mistakes could be.
Adam: Another way of framing it would be, what are the pitfalls to look out for?
Gal: Yes, I can try to predict what those pitfalls could be. One is losing control to the generative software. You give the software too much power, and you basically lose control of the design process. That could definitely be a pitfall, both for the users and for the companies writing the software.
Another pitfall is not giving proper boundaries or constraints to the generative capabilities of the software, so it starts producing things that are not usable. We talked about dimensions, codes, buildability, and cost. If the software doesn’t take those into account, then you can get something that is not buildable, too expensive, or doesn’t meet minimum requirements. So I think setting the boundaries and making sure the software always checks the design or the options it generates against those constraints would make it much more usable and much less likely to lead to those pitfalls.
Adam: You bring up a really interesting topic, which is understanding when to use generative AI and when not to use generative AI. When should you use it? When shouldn’t you use it? What are the design problems that are great for it, and what are the design problems that just are not the right fit for it?
Gal: Yeah, it’s a good question. Generative AI, or generative software, is about creating new options and new designs very quickly. That’s basically the summary of it. So if the solution is very straightforward, or if it is extremely repetitive or mathematical, for example, if you’re designing a hospital where there are a ton of rules and you repeat the same room again and again, I don’t know how much generative value is really created there. I think generative AI is more useful when there is more variety and more room for the software to bring up different possibilities.
So maybe that’s one way to separate it. I wouldn’t use generative AI for an entire design right now. As I said earlier, I don’t think software right now or AI and LLMs are ready to become the architect and just create a whole building for you. At least not now. But actually, thinking about it in real time, I think generative AI is useful in maybe 90 percent of cases. The only two cases I can think of where it is less useful are first, when you give it too much responsibility and ask it to do everything, and second, when the problem is highly algorithmic and not really creative, like very repetitive buildings or projects where you just have to follow very specific rules. Other than that, I think it can be useful in most cases.
Adam: Are there specific parameters you would advise architects to use when they are thinking about how to utilize generative AI?
Gal: I think a good professional can often tell just by looking at a design whether it was done with quality or not. An experienced architect can look at alternatives and quickly tell whether this is a good design or whether it meets basic criteria. That’s just part of experience.
But you can also use AI to analyze. There are software tools today where you upload plans, and AI goes through them and highlights problems with the design. That’s one way to assess. It gives you an automatic review from multiple angles, maybe evaluating floor plans, elevations, or different sections of the building.
Personally, if I were using generative AI software, I would want to be able to set the criteria of what a good design is. Maybe that’s the right way to look at it. If the software gives you no control and has some secret sauce behind it that you can’t influence, then it may be less useful. But if I can set the criteria, and I know the software always follows them because it’s built that way, that would be the fastest way for me to know that the results I’m getting are useful. Still, I think the first level is the professional eye. Good professionals can usually see immediately if something is wrong.
Adam: Have you ever seen an instance where the best option according to the tool wasn’t really the best design? Is there a moment or a series of moments that stand out to you?
Gal: Absolutely. Many times I’ve seen two or three options where, if I were ranking them myself, they would maybe be my fifth, sixth, and seventh choices. I would already have four better ones ahead of them. And again, if I have no way to control the software, then it becomes less useful. I want to be able to tell the software, next time consider this more than that, or make sure you don’t do this. If it doesn’t have that capability, then after a few bad experiences, I’m probably just going to stop using it.
That’s exactly one of the challenges of building generative software. You don’t want to lose your audience. If you give people poor options very quickly, they will just abandon the tool because they can’t control it. It’s more forgivable when you are doing renders, because then you just want to look at nice things and get atmosphere and inspiration. But when you need to do a design that is actually going to be built, if you don’t have control, you’re probably going to give up.
Adam: You have a really interesting perspective because you’ve been on both sides of the table. You were working as an architect at a very senior level, and now you’re building a business for professionals trying to solve their problems. How does your experience impact your approach and perspective?
Gal: By the way, my experience in the first part of my career was actually technology, not architecture and construction. It was pure technology, coming from networking, security, voice over IP, optical networks, really pure tech. Later in my career, I went to study architecture and started doing design. And now I’m basically combining my technology background with this newer part of my career. I love it because I get the opportunity to bring together everything I enjoy in one shot.
Looking at it from both the technology side and the architecture side, I think these new tools, AI, LLMs, and the software that is coming on board, will speed up the entire process. Let’s take one example, and it’s exactly what we’re trying to do in FormX. You start by creating a design. I think software will speed up that process dramatically. Today design can take weeks, and there is a lot of back and forth with the homeowner or whoever the client is. Every time you stop, do something, meet again, go back and forth, and it’s just not fast enough, at least not for my taste. It takes a long time. Every change affects many things and takes time.
I think these new approaches and technologies will make iteration much faster. Some of the work can even be done collaboratively in real time, so when you meet with your client, you can make progress together right there, which saves a lot of back and forth because you already looked at it together and made decisions together. Of course, some cycles still need time. You don’t design your home in one hour. You want to talk to your wife, sleep on it, look at examples, and come back. But instead of weeks and months, maybe it becomes days, or a couple of weeks.
Then there is the phase after the design, where all the planning comes in. There are plans for the city, plans for the general contractor, engineering, all of it. I think that phase will also shift in a big way. It used to take weeks and months and involve a serial process with many people. I think that will speed up, too, with more automation and much more synchronization between the different layers.
And finally, there is the question of how this technology can speed up the actual construction in the field. That is also a challenge, and I think some of the software will help make that process more efficient and faster as well. This is exactly what excites me. It’s what I think about every morning when I get up. How do I speed up these three things? If we can speed up the process from design to planning to building, then this becomes a completely different experience from starting a project to handing a key to the final homeowner.
That’s the exciting part. Think about flying from San Francisco to New York. It takes six hours, so you think about it as basically taking the whole day. But what if it took two hours instead? Then flying to New York would feel more like flying to Arizona. That’s the analogy. If the process from beginning to end, which today may take a year and a half, could become a few months and be much easier to do, then it becomes a completely different experience. I think that would also increase the market quite a bit, because more people would consider going through the process if it didn’t feel so painful.
Today, we are focusing in FormX on residential homes, less on commercial or large multi-unit projects. We’re more focused on single-family homes. And I think most people, if you ask them to get into this kind of process, are honestly scared. It absorbs all your time, it can even cause divorce, it takes forever, and there are so many surprises in cost and time. My passion is to solve this and make it more predictable, faster, more fun, and more like a 21st-century experience for people. That is where I’m coming from.
Adam: It sounds like we’re slowly moving into the 21st century. And from everything that I’m hearing from you, we’re really in the infancy when it comes to generative design in the field of architecture. Is that fair to say?
Gal: I think it’s fair to say. It’s just the beginning. If you think about it, AI has maybe only been generally available for a couple of years. It existed before that, but in terms of general availability, it’s only been a couple of years. The newer and more powerful LLMs are really only about a year or so old. Agentic AI is something everybody has been talking about for maybe six to eight months. There are so many new tools coming every month, and so much progress, and everything is changing very rapidly.
So yes, I do think it is still in its infancy. But because the software and capabilities are changing so fast, it may happen faster than anybody can imagine. I don’t have a good crystal ball, so if you ask me when exactly we will be in the middle of it, I don’t know whether it’s one year, two years, or three years. I just know the tools are changing quickly, and the capabilities are changing quickly, so it really does look like it’s still the beginning, especially because many of these tools have only been around for a year or two.
Adam: You have such an interesting perspective, because as you mentioned, your background was originally in technology, then you pivoted into architecture, and now you are in the marriage of technology and architecture. Even though we’re in the early innings, that might actually make it harder, because there is so much coming at us from so many different directions. You have a deep tech background, so maybe for you it’s easy to try new technology. But most architects don’t have that background. What advice would you give them?
Gal: I think what is actually interesting about the new technology is that the learning curve is becoming easier and easier. So even though I have a background in technology and somebody else doesn’t, that is becoming less important, because the way you interact with technology is much more intuitive than it used to be.
Take Revit, for example, which everybody in architecture is familiar with and which is still the predominant software for architects in the US and in some other parts of the world. You used to have to take a huge course just to know how to operate it. There were so many menus, so many options. There were specialists for families, specialists for other things, and even people who learned to code around it to automate tasks. It was an entire suite of capabilities that could take years to master.
The new type of software coming to market is much more intuitive. The learning curve is much simpler. Maybe I have an advantage when talking to my R and D team, defining the software, deciding on the right path, understanding what clients want, and how the product should look. But from a usability perspective, I don’t know if I really have such a big advantage, because any company building software today needs to think about making a tool that people don’t even have to formally learn. You just start using it, and you learn as you go. It should be intuitive and right in front of you.
That’s how companies need to think about designing software today. Natural language interfaces, voice interfaces, very minimal menus, almost nothing. It should just be available to you. And if you don’t know something, you ask it, and it tells you what to do. You can even have help tools built into the environment, where if you don’t know something, you ask, and it tells you which option to use. That is the philosophy of new software. If companies are still taking the old school training course kind of approach, I think they will be left behind.
Adam: That’s a great mindset shift, because most people in this field don’t have a software background. Recognizing that you don’t actually need a software background to capitalize on all the tools that are out there in the marketplace. What you do need is the desire to roll up your sleeves.
Gal: Just be curious. Curiosity. Just try. Just have the energy to go and hustle with it.
Adam: Yeah. With that said, there can be a sense that there’s so much stuff out there. How can you figure out, among all these different options, how to focus your time and energy?
Gal: Sure. First, define your pain point. What exactly are you trying to solve? There is a wide variety of different tasks and challenges. Some people need to speed up the design. Some need to speed up planning. Some need help with visualization. Some need help with engineering. So first, define exactly what you are trying to achieve.
Then, actually use AI to check what is out there and what the most popular tools are, what the best tools are, and what tools are available for your specific task. I think the LLMs and AI will tell you a lot of that. They look at so much information very quickly, so they can usually identify the winners, the losers, what is available, and what the right approach might be. Then go try it.
You may find that the first two tools you try are not useful. In that case, refine your question. I think it’s all about asking the right question, especially with LLMs. Go to OpenAI, Claude, Gemini, whatever it is, and ask a better question or a more precise question. Instruct it exactly what you are trying to solve if it gave you the wrong advice the first time, and narrow it down by iterating until you really understand what you need. Then you will probably start getting the right tools and the right available options. And then just hustle, try, and experiment.
It could be that you conclude it’s not there yet, and then you have to wait a bit. But not for too long. Even if someone finds that it’s not there yet, I would try again in three months, because things are moving so fast. This world is really changing quickly, so you have to keep checking what’s available and what’s possible, because the answer may be different every month. Even if you find a good tool, check again in a couple of months to see whether there is something better or something complementary.
Adam: What will great architects do differently with these tools compared to those that are not great?
Gal: Great architects will use these tools for two main things. First, they will use them to get better inspiration. At least for me, when I start an artistic project, something completely new, I try to think about art, examples, inspiration, and objectives. What does this need to achieve, and what is the best artistic way to capture it? These tools can give you options, inspiration, and things to think about. They let you start from a place where you have already seen many things visually in front of you. That’s one thing I would definitely do with them.
Second, I would try to find tools that do not replace me as an architect. I want to stay in control and keep doing the more creative parts myself, but use the tools to speed up the process. The things that are more technical and more mechanical, I would give to AI. The things that are more creative and artistic, I would keep for myself. Then you are working almost as if you have a companion or a helper, maybe even several helpers, who help you achieve the result much faster.
So at least in today’s world, that is what I would look for. Tools that help me develop a better concept faster and get more inspired, and tools that help me speed up the repetitive, mechanical, less interesting work, while leaving the creativity to me.
Adam: What do you think will be out there that isn’t out there right now? What’s around the corner?
Gal: I think the broader capability is what will grow. At the beginning, I said that AI is not yet an architect or a great designer that is just going to do everything on the fly, but it can help speed up certain things, give options, and support you along the way. I think that part will become bigger and bigger. Over time, the capability will broaden. Instead of handling very narrow tasks, it will handle wider tasks that you can actually use in a more meaningful way.
We also talked about the process being multi-step, starting from ideas, to design, to planning, to execution, and everything that needs to be produced along the way. I think you will see more integration between different components. You will see software that either integrates with other software or has more functions built into it so that you can do more in one place. Right now, maybe you need seven different software tools to get a project done. In the future, maybe one tool can do two or three of those things instead of requiring seven. I think that is another theme we will see. Software will become more capable, and you won’t need to jump around as much between different tools to get the project moving. And at some point, real generative architecture will happen. I think it is inevitable. But I think it is still some time away.
Adam: How far away do you think we are from that?
Gal: I don’t know. I don’t have a good crystal ball. It’s really hard to predict. I think we are talking about years, not decades, but I don’t know whether that means a couple of years, five years, or seven years. It’s hard to predict.I think it’s the evolution of AI itself. It’s everything the huge companies are working on right now. Smarter AI, more general AI, all these acronyms that get thrown around. You heard the news recently from Anthropic about what their new models can do. Things are happening very fast. Multiple huge companies are working on this right now, and I think once they release more powerful models, more powerful LLMs, agents, whatever you want to call them, at some point, they may become good enough to create a complete generative architect. I’m sure it’s going to happen. I just don’t know how many years it will take or how fast it will move, but everything feels like it’s accelerating. It feels like things are being delivered faster and released faster, and that’s why it’s so hard to predict.
Adam: Looking ahead, what advice would you give to architects to prepare for that moment when that happens?
Gal: It’s a great question. I don’t think anybody is going to use software to completely remove the need for a professional architect. There is always going to be a need for a professional who helps the client make the right decisions. There will still be choices, and there will still be gaps that need to be filled.
So the advice I would give is to keep up with the new tools, continue to focus on great service and creativity, and the artistic side of the business, because that will be very difficult to replace with technology. And to get an advantage, make sure you are always on the lookout for what can help you do a better job for your clients, speed things up, and remove the things that are not unique to you, so you can give faster service, better service, and more value.
I don’t imagine somebody designing and building a new home entirely with software and without talking to a person. At least for the foreseeable future, the interface will still be a person. Behind that person, there may be a lot of software, automation, generation, and whatever else, but at the end of the day, I think people will still design and buy a house through a person for many years to come.
Adam: Is there anything else you would like to share?
Gal: Just that I’m really excited about the opportunity I’ve been given to be at the forefront of this revolution. Solving this problem and being part of this revolution is something that really excites me and gets me going every morning. I think the fact that FormX is touching this whole long pipeline between design and build is very exciting to me. And yeah, thank you for having me. It was great talking to you. I really enjoyed it, and you asked me some questions I hadn’t thought about before, so thank you for the great questions.
