"What can AI bring into the field of architecture?"

In the wake of the groundbreaking releases of chatGPT and Midjourney in 2022, the conversation around AI's potential roles across various sectors surged with newfound intensity. The professional practice of architecture, traditionally marked by both curiosity and caution regarding technological innovation, found itself at a crossroads. During my tenure as a lecturer at Bangladesh University, I witnessed firsthand the mixed reactions to AI's encroachment into architectural creativity. A presentation on style transfer GANs left the audience—students poised at the threshold of their careers—simultaneously enchanted and unnerved, pondering the future of their profession in an AI-dominant landscape.

A year later, Andrew Kudless's workshop at CAPLA, University of Arizona, introduced diffusion image models and the art of prompt engineering, offering a glimpse into AI's burgeoning capabilities in enhancing architectural design processes. Yet, these explorations, while illuminating, seemed to merely graze the surface of AI's potential, akin to tracing the outline of an unseen elephant. Despite AI still being in its nascent stages, the promise it holds for revolutionizing architectural practices is undeniable, sparking both excitement and a tinge of apprehension about the future role of human architects in a discipline increasingly influenced by artificial intelligence.

Diffusion model image generators have sparked intriguing results for their ability to craft preliminary sketches from text prompts. These tools generate images by piecing together pixels to mirror training data, relying heavily on their associated keywords. But more often than not their way of "thinking" or "hallucinating" can be discarded by architects because of it's objective mimicry of their training dataset.

One of the core reason behind this thought should be, their understanding of context being superficial, rooted in a 2D perspective that lacks the depth and flexibility required for architectural design. While they can enhance image resolution and variety, they don't offer the replicability or the nuanced contextual awareness necessary for integrating into the architectural process. Essentially, these models serve more as digital sketching aids, similar to the role of parametric design tools in the past decade, rather than as comprehensive design solutions. Their grasp on context is limited to the two-dimensional images they've learned from, missing the quantitative insights critical to architecture.

This becomes clearly evident when these engines are tasked to produce 2D visualization of 3D elements such as architectural drawings as floorplans or construction documents.

Merging these insights with Stephen Coorlas's attempts to generates speculative axonometric construction drawings reveals that image models lack the ability to grasp the surrounding context, a critical element in architectural design. Despite expectations that parametric and AI tools developed within the architectural field should possess some level of understanding, a closer look at tools like "planfinder" and AI layout solutions such as Maket reveals no evidence of genuine "intelligence" beyond their algorithmic frameworks.

Maket generates unique plans in each iteration, showcasing AI's utility in producing varied solutions. However, it lacks the ability to discern which of its generated solutions is superior. Similarly, parametric programs like Pathfinder are confined to predetermined solutions, limiting their capacity to be recognized as a form of intelligence due to their inability to independently evaluate or adapt beyond their initial programming.

Interestingly, spatial relation between spaces can be algorithmized as shown by the developments of plugins like termite nest. So, we can produce functional and parametric floor plans even without using AI and that automation is less connected with context except for the spatial dimension.

In a simplified overview of a Generative Adversarial Network (GAN) AI system, it incorporates several key components:

a. Training is conducted using real-life samples to understand the patterns and data it aims to emulate.

b. It can generate new samples on demand through its generator component, creating outputs based on the learned data.

c. The system employs a discriminator to evaluate the generated solutions, distinguishing between the real training samples and the new, generated outputs to determine the success of the solutions.

The concept of a successful output in architecture is inherently multidimensional, defying straightforward definition. Although much effort has been directed towards crafting architectural visions and layouts through 2D images, architecture fundamentally operates in three dimensions. The layouts and conceptual visions represented in these images are, in essence, flattened versions of a more complex imaginative construct, underscoring the dimensional limitations when translating spatial ideas into two-dimensional representations.

My initial exploration was sparked by questioning the role of humans within the somewhat constrained framework provided by the current "discriminator" modules in Architectural AI systems and pondering the future of AI-human collaboration. Specifically, I'm intrigued by scenarios where humans no longer hold exclusive rights to "design intelligence." This exploration aims to uncover how AI and human intellect can merge, particularly in contexts where the distinction between creator and tool becomes blurred, signaling a shift towards a more symbiotic relationship between technology and human creativity in the architectural design process.

My xArch'23 paper Meta-morphing Architectural Domains: The Role of Humans and AI in Post-human Architecture is an attempt to shade light on that question.

Shape synthesis through 3D Generative Adversarial Networks (GANs) has significantly transformed object-oriented thinking in recent years. The pioneering project by MIT Media Lab in 2016, which generated 3D data from 2D images of a chair, marked the inception of this innovative merger of 3D modeling and AI technologies. In the years that followed, 3D GAN technology has advanced considerably. Nvidia Omniverse, for instance, is pushing the boundaries by using computer vision-driven objects to simulate expansive environments within software like Blender. Similarly, companies like SWAPP are leveraging these technologies to make Building Information Modeling (BIM) documentation more dynamic, showcasing the growing potential and application of 3D GANs in revolutionizing design and modeling practices.

Despite the advancements in integrating AI with architecture, a fundamental question remains unresolved:

"How well can AI comprehend the context and surroundings of an architectural design while making decisions?"

This inquiry highlights the ongoing challenge of developing AI systems that can fully grasp and consider the intricate details and environmental nuances crucial for informed architectural decision-making. 

The fundamental idea behind this enquiry is to understand that Architecture is not the craft of forming mere objects, as Kengo Kuma articulates in "Anti-Object," architecture should transcend its object-like attributes to embody an integrative approach that harmonizes with the natural environment and the local context. 

If we take a step back and try to reconstruct different aspects of a GAN system using existing parametric tools like grasshopper plugins, this question can be explored furnther. For example, there are generative codes for simple architectural masses like one from "Simple Sketch" .

Relying on single-factor optimizers like Galapagos for decision-making can lead to overly quantified, one-dimensional outcomes. This approach often simplifies complex decisions to numerical optimizations, potentially overlooking the multifaceted and qualitative aspects essential to holistic and nuanced architectural design resulting in a one room and and mimum sized plot for this.

The application of machine learning algorithms, such as LunchboxML, introduces additional dimensions to facade material responsiveness, hinting at the potential for developing multi-factor systems capable of self-prioritization in architecture. This progress underscores the ongoing research commitment to crafting "A contextually aware AI system that could serve as an autonomous design collaborator." 

The manner in which we address and navigate this challenge will critically shape the impact, transformation, and evolution of the design process within the architectural realm, marking a pivotal junction in the integration of AI with architectural practices.