Before building an AI code editor, the Cursor founders were working on a copilot for CAD. The idea was to help mechanical engineers in tools like SolidWorks and Fusion 360 by predicting the next geometry change while designing a part. They explored two directions: A pure 3D approach. A text-based approach, where CAD actions were converted into sequences of method calls. That second path sounded clever, but it was very hard in practice. The model had to do more than predict the next action. It also had to mentally reconstruct the geometry from a sequence of operations, which is difficult because CAD kernels and 3D geometry are complex. The bigger issue was data. There was far less CAD data on the open internet than code, so training useful models was much harder. The science also was not ready yet. Pretrained models were still weak for 3D tasks, and the team had to do a lot of scraping and data work just to improve performance.…