Co-Lab: The Designer and the Tool
Research / Thesis · SCI-Arc, UAI · 2023
How much of a design does the software really make? Co-Lab reflects on the agency designers hand over, from parametric rules to generative models.
Co-Lab is the thesis under most of the work on this site. Design software spent decades as a prosthesis, extending what a designer could reach. Clark and Chalmers called that an extension of mind. Then the software started proposing, and the thesis argues it became a second agent with an agency of its own.


Amplified creativity
In 2015 neural style transfer repainted a photograph as a Van Gogh. Three years later a GAN portrait sold at Christie's for 432,500 dollars. In 2019 Kartell presented the first mass-produced chair whose form was dictated by AI. Art absorbed the shock first because it could chase beauty with no constraints. Design has to work.


Form-giving
For two decades the frontier was form-finding. Rules go in, the software searches, and a shape comes out that nobody drew. Co-Lab argues the successor is form-giving. What comes back from a generative model is an approximation of the request, and the designer accepts it. Furniture was the test.
That acceptance is a small transfer of authorship, repeated hundreds of times per project.





The illusion of good design
A photoreal render that once took a studio days now arrives from a single prompt. It is easy to read that beauty as quality. The image is a reflection of the dataset it came from. It looks resolved without being resolved. Design stays incomplete without the input a model can only simulate. Emotion, memory, context, the reason the thing should exist at all.
Watch: Co-Lab: The Designer and the Tool
Designing the collaborator
At SCI-Arc the study moved inside the models. Prompts, CFG values, datasets, the pipelines connecting one program to the next. Building a tool raises a harder question than using one. The designer sets the agency that the tool will later exercise on the project. [Sketch](/work/sketch-gen-arch-platform/) came out of that, a platform that reads a site and generates massing before anyone draws.

As generation gets cheap, judgment gets expensive. Case Miller calls the result a digital selector, a designer choosing which of a hundred machine-made candidates answers the brief and able to say why. The practice continues at Judson Studios, folding AI workflows into stained glass, a craft that has survived every tool change since the middle ages.
