Module 8 Questions

Module 8 Questions:

Part 1 -

The problem was that AutoCAD had inconsistent layer names where architects spent hours manually classifying layers before converting drawings into Revit. As such, the process was repetitive and error-prone. So KLH came in and used 38,000 previous layer-ma[[ing decisions as training data to build a machine learning model in Python to automatically predict the correct layer category. This was then integrated into a model that could be directly implemented into their AutoCAD to Revit workflow.

I was surprised that the LLM was simple and only took a few hundred lines of code to make. Additionally, the hardest part was collecting and organizing good training data, not building an algorithm.

I would like to try to pursue a similar workflow for a facade modeling software where engineers repeatedly categorize or organize data. The requirements to make the model would be lots of historical project data, straightforward inputs and outputs, as well as a repetitive task with recognizable patterns. A successful LLM in AEC does not need to be complex - it works best when solving a repetitive problem using existing project data.

Part 2 -

Something that I have struggled with is debugging Grasshopper logic/Dynamo logic as it takes me a while to find the problem and sometimes even if I have worked on it for a long time, I still cannot find the perfect fix. This is especially true when I am trying to fix geometry that wasn’t behaving correctly when I am setting up generative design studies. I also struggle to understand dynamo logic and much prefer grasshopper logic. The best fix I have found is using repetitive parameter testing and adjusting as I go.

If I were to use AI, I would take screenshots of my work and feed them into an LLM to identity any broken nodes and fix any geometry errors. AI can also suggest ideas on how to set up a generative AI model for something a user might not know how to do. Ai can then set up the optimization study and then afterwards, summarize design tradeoffs and results.

I think I would like AI for debugging and for doing repetitive tasks (however, I feel like there are not that many repetitive tasks when setting up a grasshopper model). I also like it when I have an idea and it can suggest how to do it. I, finally, also think that it can explore design iteration quickly, which can be useful in a time crunch. I would not use it for finalizing design decisions, because it does not know exactly what I like. I think some friction is useful for learning how the grasshopper worked and understanding parameter relationships.

Part 3 -