I was at a conference in San Francisco last week. The topic of the presentations at the conference was the state of AI and ML in AEC. After each presentation there was a round table discussion. I'd summarize the sentiment of the conversations in this way. AI is a fast horse. The AEC industry has fallen off that horse and is now being dragged at high speed towards an uncertain doom.
Mike Cummings wrote an article in the Yale News recently featuring Phil Bernstein. Phil took his students through an exercise using several AI models to create images of buildings. He points out that these are only pictures of buildings. And he correctly identifies one shortcoming that will keep these models from being truly useful in AEC. They don't know anything about space.
Phil has advised Hypar on an off over the years, and I respect him. But I think Phil goes way off track when he imagines what comes next. A period that he calls the "data interstice," where we figure out how to train AEC-specific models on data coming from the tools that we use today. It's a reasonable assumption for a next step given what we know about how today's models are trained on copious amounts of data. But I think it's fundamentally the wrong way to start the conversation of what comes next.
What's frustrating to me, both in the conversations at the conference, and in the article, is that I don't hear people imagining how we want our future to work. Where's the science fiction version of building design? A version that sounds unattainable?! It's like we've lost the ability, or at least the will, to even imagine what's beyond chat prompts.
In Hypar's earliest presentations, we showed an image from Gerald Delon's "Methodology for Total Hospital Design" of a hospital floor plan generated by a computer program. When this work was done in the 1960s, Delon was imagining computers not as a tool to make drawings, but as a tool to design buildings. Instead, in the 1980s we got AutoCAD to make drawings. Twenty years later we got Revit to make models, to make drawings. Here we are more than twenty years further on and we're no closer to Delon's futuristic vision.
What Delon was pointing at, and what I've come to understand deeply in the last decade of working on the problem of computing buildings, is that we know the rules. When a slab edge moves, a building's structure needs to support it. The facade needs to align with it. The mechanical systems need to adjust to a change in the volume of the building. And so on. Granted, there are hundreds of systems in a building. The problem of reconciling these systems when the design changes is complex. Today we parcel out this task to humans to do manual work, updating models, drawings, and analyses. We could create a bunch of automated design systems to do this today. Why don't we?
It has to do with trust. Architects and engineers take responsibility for life safety. We trust our engineer to specify the weight and depth of a beam when its length or load changes. Not our software. Our certainty in the safety of buildings sits on a bedrock of trust in human experts. However, this might be about to change.
Two years ago, if I asked an architect, "Would you use AI to generate an image of your building?" the answer would have been an emphatic "Never!". Last week, I was at a conference with architects swapping stories of how they're using AI. The trend is picking up speed. This is concerning because, as Phil points out, these AI are stupid -- they can't reason about space. But it's worse than that. They're also confident. You can ask ChatGPT to design you a mechanical system for a building, and it will do it. And it will be wrong.
We're placing an increasing amount of trust and hope in these systems. The hype is just too strong, driven by breathless pronouncements of artificial general intelligence (AGI) just over the horizon, fueled by an orgy of venture capital. As we invite these systems into our lives in ways that we wouldn't have imagined just a couple of years ago, our bedrock of trust in human experts will begin to shift. The line between human intelligence and machine intelligence will start to blur.
This will be a powerful force for change. The next generation of designers coming into the profession raised on AI, won't suffer complex, slow, tools that require them to do manual work when a design changes. Their expectation will be that intelligence is ambient and ever-expanding. The only apps they'll use are those that encourage exploration, collaboration, learning, and play. Perhaps we start with imagining what building design will look like for them?
This is what we've been doing at Hypar since 2019. When we started, Anthony said something that stuck with me. "Nobody wants drawings. They want a building." It's a concise way of reminding us that the goal is not "generative design" as it was called at the time, or "AI" as everything has been rebranded now. It's a building. Everything in between is an implementation detail.
We began with a few key insights. The first was that most of the systems in a building can be generated. That is, we know the rules. The second was that we have an abundance of cheap computing power distributed across the planet. The third was a suspicion that by moving from elements to building systems we could vastly simplify building design.
Even at the time, we understood the importance of trust. There was always someone in the front row whose hand would shoot up in the Q&A and they'd ask "How do you guarantee that these systems generate the right thing?" We had to get this part right. Trust is usually expressed in three layers. I built it. My team built it. Someone else built it, but I can verify how it works. You can handle the first two with permissions structures that allow you create something privately, then gradually share it with your team and the world. For the third, we took inspiration from the software world, creating open source software libraries and "building blocks" systems that could be inspected and verified, forked, and remixed.
With the engine in place, we began to experiment with radical new interfaces for building design. First through simple "workflows" where you could snap together blocks representing building systems. Then through an interface where you could generate a detailed BIM from text. It's funny to rewatch that video and understand that it was a bit of a goof. Andrew was playing with the possibilities. But every day I get someone asking me whether Hypar can still do "Text to BIM". The reason it can't goes straight to the heart of this discussion.
Text to BIM is a powerful demo because it goes from nothing to a building in seconds. A simple input creates a detailed output with a fraction of the effort required by today's tools. That jump twists a designer's understanding of their value in the same way that watching ChatGPT write an essay does for a writer. It's red meat for those with visions of a world where we go straight from a spreadsheet to construction without that pesky design process.
But it's a terrible interface for design. What if we built a world reminiscent of activities builders love -- playing with blocks and sketching -- powered by the same engine that can conjure building systems? A world where detailed building systems flow into existence through a user's direct interaction with things they understand like beams, walls, and doors. Not through secondary scripting interfaces, panels of sliders, or chat prompts.
A world where building designers feel powerful.
In September of last year, we released a new version of Hypar. It's as simple as playing with blocks, and it's powered by the same engine we've been building for years. Our focus on keeping designers in the flow led us to reimagine again how generative systems will augment design, seamlessly suggesting solutions as a design evolves. And it gets smarter as you design, seamlessly incorporating your designs into an ever-expanding sub-stratum of organizational expertise, while suggesting solutions from the same. When the time is right, we'll reintroduce the generation of connected building systems, instantly transforming spatial layouts into richly detailed models.
Building design will change, because we will change. We can choose to wait and see what OpenAI or Autodesk gives us next. We should experiment with everything new. But we should never settle. At Hypar, we never settled. We've been building a new foundation for building design. It's time to build.