Proposing A Spatial Artificial Intelligence (V1, archived)

“Spatial AI” refers to all emerging applications of AI that interrogate, model, and speculate on spatial design problems and propositions. I propose grouping all AI techniques, existing and future, and particularly those regarding architectural and urban spaces, under this term.

“Spatial AI” refers to emerging applications of AI that interrogate, model, and speculate on spatial design problems and propositions.

AI already pervades the built environment. Governments monitor and track people in public spaces; semi-autonomous vehicles and dog-like robots continuously map their environments in real time; tech companies listen to and record our homes; satellite data and imagery feed the predictive models of urban growth and environmental destruction.

AI already pervades the built environment.

Despite such surveillance strategies and governance challenges, the same enabling AI technologies present an opportunity to develop sophisticated design methods. Yet still, today’s spatial AI methods are in a fragmented and early state, with a range of established techniques being used in real-world systems leading a longer tail of experimental ones.

There’s a huge opportunity to formalize and expand upon spatial AI, especially by researching and building open generative models inspired by today’s LLMs and text-to-image tools (like DALL-E, MidJourney, and stable diffusion).

There’s a huge opportunity to formalize and expand upon spatial AI, especially by researching and building open generative models inspired by today’s LLMs and text-to-image tools.

And today, computational design depends heavily on software tools. Coding does afford designers nearly infinite flexibility, but the path to developing sophisticated, one-off, project-specific algorithms (or robust shared frameworks) can be long and complex. Products fitting a range of markets, methods, and professions, like Grasshopper, Jupyter Notebook, and Processing, can make exploration easier, but they still carry their own affordances and constraints that don’t necessarily achieve the ideal.

These tools (including generative AI) are built by companies with profit motives or open communities who aren’t addressing the same design challenges.

And it remains that these tools (including generative AI) are built by companies with profit motives or open communities who aren’t targeting the same design challenges.

I posit AI can create new potential and flexibility for computational design. With the right models, we could explore new interactions, create new semantically rich design systems, incorporate and interact with real-world data, even program with natural language and thus describe design problems to our machines as we might to our colleagues, rather than through the formal language of computer code.

What if in the same ways we explore ideas with ChatGPT, we could interrogate our design ideas with the same fluidity and “understanding” but with real-world data and the visual and representational power currently at our disposal?

But for new spatial AI methods to germinate and grow into their creative potential, computational designers must guide the development of their own generative AI models and agents, those tuned to spatial design problems and the semantics of space.