Projects

Relate, Relate, Relate: In the Age of Machine Learning

A large diagram with text connected by lines

by Henry Chung (MArch I AP ’24) — Recipient of the 2024 Digital Design Prize.

This thesis explores machine learning as an association machine, revealing its potential to historicize, contextualize, and generate architecture through relationships rather than categories. A book-manifesto anchors the work, archiving a semantic map of architecture: 28 GSD “five on five” lectures, 241 precedents, and 420,000 words analyzed to produce a cloud of relationships. Instead of hierarchies or lineages, this high-dimensional landscape blurs boundaries — pairing projects based on conceptual similarities and imagining entirely new architectures born from those pairings. It proposes a text-driven framework for understanding architecture beyond appearances — uncovering unexpected connections and offering an alternative way to produce and navigate architectural knowledge in the age of AI. The book presents 100 pairings and their imagined architectural descendants — an open-ended exploration of design through relationships.

Flip through the book at this link, or browse some excerpts in the image gallery below.

A large diagram with text connected by lines
More diagrams connecting text
Diagram connecting several words like "Kendall Square Garage," "The Projective Cast," etc.
Image of a book page saying "002 Meta-Reading of the Five on Five Lecure: The exercise works closely with the five on five lecture series at GSD; It is a series of panels organized by Mark Lee. Each lecture, Mark invites two architects to each share five projects that they find stimulating in some ways, for sparking new ideas, finding similarities, and identifying differences. The lecture is on the same line as the thesis, which is about contextualizing and relating. Yet, the lecture format is limited, and only allows for relating and differentiating projects in a limited scope of 10 projects per lecture. What if we can create a meta-reading of the 5 on 5 lecture, and re-relate the whole body of precedents together?"
A book with text "001 Collecting Text 28 Transcriptions/ 420,000 Words The lecture videos of the 28 lectures are available online, each paired with subtitles. They are downloaded and translated into text files through simple code. Example: 'Welcome everyone to our very first Five on Five session for the fall semester. My name is Andrew Holder. I'm an associate professor of Architecture here at the GSD and the program director for the MArch I programs. This is a special edition of Five on Five. Not only is it our first in the semester series, but it is our welcome event for the incoming class. I can think of no better participants for our inaugural episode in this series for the year than Teran Evans..' 241 Architectural Precedencts The transcription text files are unformatted. They are processed by a Large Language Model, which separates them into paragraphs, each corresponding to one of the 241 architectural precedents. Additionally, manual research is conducted to gather details such as the name, location, architect, images, and year of construction for each precedent. All this data is then compiled into a comprehensive Excel file"
A book with text "002 Analyzing Text BERT Model To discover semantic similarities between the precedents, Google's BERT, a Large Language Model, is trained to generate embeddings or coordinates for each precedent. Essentially, the model comprehends the meaning of all the texts and assigns a coordinate to each one. The more similar the texts or precedents are, the closer their coordinates will be. Dimension Reduction/ T-SNE The coordinates for each text contain multiple numbers because they represent a point in a high-dimensional space. A set of mathematical operations, called dimension reduction, is performed to simplify the coordinates into x and y coordinates"
A book with text "003 Mapping Precedents Orquideorama Plotting A python program plots the precedents onto a 2d map according to the coordinates generated by the LLM. The map is a 2d representation of an architectural landscape, in which similar projects are plotted close to each other. Columbus House Drawing Connections A python program plots the precedents onto a 2d map according to the coordinates generated by the LLM. The map is a 2d representation of an architectural landscape, in which similar projects are plotted close to each other. Each node is connected to the 3 closest codes, to pair the precedents. the 241 precedents generate 460 edges and 460 pairings"
A book with text saying "004 Finding Similarities Prompting LLM/ OpenAI API 460 prompts are sent to GPT, to find out the similarities between the 460 pairs. Each prompt includes the two texts that correlate to the pair of precedents. Here is an example: 005 Foregrounding the Edge Node & Edge The 2d map consists of the node and the edge, in which the node represents the existing precedents. On the other hand, the edge can be read in multiple ways. It can be read as the relationship/ similarities between the projects; it can also be read as an unexplored area in the architectural landscape that has a close relationship. Visualizing the Edge/ Mid-journey 460 prompts are sent to GPT, to generate 460 paragraphs that describe an non-existing building that reference to the pair of existing precedents. Each prompt includes the two texts that correlate to the pair of precedents. These 460 paragraphs are sent to Mid-journey, to visualize the relationship through" and some diagrms
A book with text and images of the Chicago Tribune Tower, Cover, Zabriskie Point, and others, with more diagrams
Page with text "Edges as an unexplored area in the architectural landscape that has a close relationship. Paragraph 01 Chicago Tribune Tower Chicago, 1922, Adolf Loos, Mentioned by Scott Cohen And then there were some other things that happened that I have analogized to the idea of the readymade. I mean, what could be more radical than just making a whole tower be one column? And this is not a critique of anything other than architecture itself. This is a very interior critical act within our domain as architects to say now, I will not continue making skyscrapers the way everyone Paragraph 02 BEST Notch building 1977, James Wines, Mentioned by Andrew Holder Number five, other things ruins can do, split the present into two parts that coexist with one another. So here, let's look at James wines and one and a half of his projects for BEST. And this was about a 10-year endeavor. I'm not sure how many total stores, maybe six or seven, that he did for BEST. This is one of my favorites. It's the Notch building. And the only feature is at this corner. I The Edge "The building features a lange hydraulic corner that can open and close, releasing balloons on special occasions, symbolizing a dynamic and interactive facade. This corner is designed like a gigantic, classical column, reminiscent of the singular, exaggerated column form used in the Chicago Tribune Tower project. The rest of the building integrates glass and ghostly mirror images, blurring the line between the real and the imaginary, enhancing the structure's surreal ap-pearance. This combination results in a visually striking skyscraper that merges whimsy with classical architectural references" and a photo of a building with shiny columns