The original note, as written at the time
I think the way forward to achieve a general AI for the architecture field would be to hit the base at the most mundane duties. Draft writing and report writing. Report writing is a bit trickier but drafting is quite straightforward. We could start by documenting the daily routine of a draftsman or anyone doing drafting. We document both the instructions that the Architect issues the draftsman, the work that he does in response to that instruction, the sketches that correct the preliminary output, and the kind of response that the draftsman carries out to effect.
Ongoing development
The way forward to a general AI for the architecture field is not the drawing. It is the errand.
I have watched draftsmen work for years and the thing that strikes you, once you stop looking at the screen, is how little of the day is design and how much is response. An architect issues an instruction — often verbal, often half-formed, usually meaning something slightly different from what it says. The draftsman interprets it, executes a version of it, brings it back, and is corrected. The correction is not a new instruction. It is a gesture at the thing that was meant. And the draftsman, who has been doing this for a decade, understands the gesture in a way that no brief has ever captured.
That loop is the asset. Not Revit. Not the plugins. The loop: instruction, interpretation, execution, correction, re-execution, until the drawing stops being argued with.
So the way to build this is to document the loop before automating any part of it. Record what the architect actually said. Record what the draftsman did in response. Record the sketch that arrived afterwards to correct it, and what the draftsman changed to satisfy the sketch. Do that for a hundred ordinary days and you have something no dataset has: the tacit negotiation between two people who are making a building together.
Then the mundane duties come first, and they should come first. Draft writing and report writing are unglamorous and enormous, and they are where the profession loses its hours. A machine that can produce a competent drawing register, a schedule of finishes, a set of as-built notes, is already worth its fee, and it is learning the language of the office while it does it.
The dishonest version of this project starts at the top — a machine that generates the concept — and it fails because the concept is the smallest part of the work and the least traceable. The honest version starts at the errand, and it has a much longer road, and the road is made of ordinary afternoons.
Written before any of it existed, and the drafting half has since arrived: generative tools now produce drawing sets, schedules and the paperwork nobody wanted to do, and practices are quietly reorganising around them. What has not arrived is the part I actually argued for — nobody documented the loop. The instruction, the misinterpretation, the sketch that corrects it, the draftsman’s reply. The tools were built from drawings, and drawings are the residue of the conversation rather than the conversation itself.