LLMs for Building Design: Game-changer or Old Wine in New Bottles?

In this blog, we will dissect both perspectives to give a balanced primer on the application of LLMs, setting a more grounded perspective on computational design within the building and construction industry.

What might have sounded unthinkable only a year ago is now a subject of heated debate within the Architecture, Engineering, and Construction (AEC) industry. Using a type of AI known as a ‘Large Language Model’ (LLM for short), computing nerds and experts alike are exploring how this recent technology could be used to conceptualize, design, and execute projects in more efficient and creative ways. At the same time, skeptics argue the current AI moment is just ‘old wine in new bottles’; that it is the latest repackaging of digital trends which have unfolded over the past decade.

What is Prompt Engineering?

Before digging deeper, let’s first take a moment to discuss ‘prompt engineering’; an important concept for understanding how we can use an LLM, like GPT-4, to transform human language into a detailed 3D model of a building. In this use case, a user starts by describing their vision for the project to a computer, which is then refined into a ‘prompt’ to highlight key elements and conform to the AI’s operational logic

In essence, prompt engineering works by acting as a filter to transform natural language into a structured ‘prompt’ that is used as input for the LLM. These prompts act as a guide for the LLM, providing specific instructions and constraints on how to interpret user input and provide output in a specific format that can be read by other computer programs. In the case of generating 3D building models, the output of the LLM serves as input for another computer program called a generative algorithm, which produces a 3D model based on parameters and constraints.

In this use case, the LLM serves as an intermediary, operating behind the scenes to communicate complex design ideas to generative design software. One of the key benefits of prompt engineering versus using the raw output of the LLM is that it minimizes the occurrence of ‘hallucinations’ (made-up results) by constraining the inputs and outputs of an LLM to a very specific application.

We can think of LLMs as essentially large lookup tables. Using prompt engineering, we can modify their behavior to produce outputs in a specific format, which could be extremely helpful. But it does not save us from hallucinations, since sometimes even guided LLMs can make things up and disrupt the whole generation pipeline. — Aleksei, AI Design Engineer

Perspective: Old wine in new Bottles?

From the outset, the promise of using LLMs to create 3D models using natural language and no modeling skills represents a significant departure from the current paradigm; today, architects, planners, and engineers spend years learning CAD, BIM, and GIS software. Despite this promise, there are a few reasons why LLMs may not provide the significant leap over the current paradigm:

Argument 1 — Recycling Established Design Principles

While LLMs introduce faster and more natural interaction, the foundational principles guiding the outputs are tethered to the generative design software it is connected to. Since the output of the LLM is only text, it needs to be translated into geometry based on tried and tested computational design rules and methods that have been developed over the last half-century, raising the question: Do LLMs drive true innovation, or reinforce existing knowledge and design principles?

LLMs are not able to compose or reason through an elaborate solution or system of solutions to address any of the fairly complex or even less complex set of problems in the building industry. This is not surprising, since LLMs were not designed to reason but only to recompose text from a limited and discrete dictionary of words.’ — Aleksei, AI Design Engineer

Argument 2 — Limited Novelty in Problem-Solving

It is well documented that LLMs today are fundamentally poor at planning and reasoning tasks. As a result, the output proposed by LLMs, while varied, seeks to mimic traditional architectural problem-solving approaches and often make up results. This is why, in our own testing of the technology, we found it took several follow-up prompts and human brainpower to guide an LLM to produce a satisfactory building design.

Perspective: LLMs are a Game-changer

We have pointed out some limitations of LLMs. However, there is also reason to believe that LLMs can be a game-changer for building design. Here are 2 arguments why LLMs can be a game-changer for building design:

Argument 1 — Democratization of Design

LLMs can shape new ways of working by democratizing the design process. If they live up to their promise, it means individuals without specialized software knowledge can more actively contribute, making sophisticated architectural design more widely accessible. This shift towards inclusivity will invite more perspectives, potentially sparking unprecedented innovation in the field. Furthermore, since LLMs are trained on the corpus of all human knowledge, they can also make it easy to bring together insights from various fields, sparking a more holistic approach.

‘LLMs are impressive and remarkably capable of understanding and formulating coherent sequences of natural language (e.g., English). This has useful implications in terms of transforming the interaction between humans and digital solution systems. Rather than having humans learn to adapt to each tool and solution, they can collaborate with and command them simply using a language-based interface.’ — Houssame, Applied Data Scientist

Argument 2 — Faster Iteration leading to Better Design Concepts

It is hard to argue against the speed with which LLMs can be used to generate, iterate, and refine ideas. This is less relevant for the later stages of a project where more precise decisions are required; however, this is perfect for the conceptual phase. This means not only accelerating the design phase but making it easier to explore a broad range of possibilities, leading to more innovative designs. Furthermore, in our work, we have explored how we can fine-tune an LLM on a specific domain like architecture or urban planning, leading to more novel and practical designs.

Conclusion

Ultimately, the debate over LLMs in generative building design as a game-changer or merely old wine in a new bottle is nuanced; While it’s evident that LLMs carry forward some traditional constraints, especially in their application where precision or risk is a major concern, their potential to democratize design, enhance efficiency, and foster interdisciplinary innovation cannot be understated.

Looking ahead, the focus should perhaps not be on whether LLMs are revolutionary in themselves but on how we can leverage these tools with other approaches to address unprecedented design challenges, such as climate adaptability and material scarcity, where conventional wisdom often falls short.