I'm a PE who has been tinkering with LLMs for a few years now. Opus 5.5 does seem to be a game changer for drafting. I don't generate with python tho, I have it drive AutoCAD headlessly and use our firm's drafting standards/templates. Its a junior-level EIT at first, but I give it redlines via Bluebeam MCP and it updates the relevant CAD SOP for all future sessions.
Given how good the leading models got at 3D model creation over the past few months - Blender, Three.js, etc - it's not surprising to see that their 2D CAD-style abilities have improved to this extent as well.
(Amusingly the pelican-on-a-bicycle test doesn't capture this very effectively, despite testing 2D drawing with SVG - because pelicans can't write bicycles, so the results are inherently unrealistic. Having the model create an architectural diagram from a concept sketch is a much more useful 2D test.)
Civil engineering rules are mostly written in blood and costly and dangerous accidents. At least in my country. If somewhere else inspectors don't provide actual value, they should either get regulated or good riddance and AI might actually be better. I feel old, sigh...
That's not what we're talking about. By all means, follow the structural engineer's specification. But if you've ever submitted plans to a city to build a home, there is endless nonsense full of red tape: Drainage plan, soil analysis, tree survey, topographical survey, environmental site assessment, wind analysis, noise assessment, and more.
Sometimes these things make sense, but a lot of the time the work was already recently completed 3 years ago and yet they'll make you do it again. Last month I paid $3,900.00 to get 3 tiny trees and one medium sized oak surveyed here in the Bay Area. It's really about politics and control- the city wants to make a bunch of fake jobs and collect very high fees along the way. Welcome to insanity!
1) Permitting is slow and, at some point, if you fuck it up big enough, the inspectors will probably take further steps, but otherwise not care much. You simply won't get a permit.
2) Who will care a lot are the builders and creditors. Building is expensive and if there's a couple of months of set back because of your fantasy drawings, that's not going to be in your favor.
Which is not so say AI planning is not going to happen. It will. But at that point it's probably mostly going to work.
Might be nice to scope out where exactly it is good and where it is lacking ; residential vs commercial vs industrial , A vs S vs MEP, style of architecture? We need better benchmarking here to substantiate.
Agreed. Also the consistency between the different drawings would be nice to have an understanding of. This is where a higher level 3d domain model (such as IFC), do validation, interference checking etc., and generating drawings from those makes more sense (in my mind; but I'm biased). We did a quick demo once [0] but it ran out-of-date quickly due to the hardcoded model ids.
Washington State is trying out a 'Permit-Ready Plans' program in a few cities, were pre-approved plans are offered at no cost. According to their initial study, building plans account for 5-20% of total construction costs. Software that might be able to shave 10% off new construction costs is a big deal.
A big reason LLMs can write code productively is cause software dev tends to be pretty resilient to failure. And at the pessimistic end, lots of companies can make money off buggy products, so it’s easy to hand wave the failures away. Does that hold true for architecture?
That being said, I’m certain they’ll be extremely useful in other areas such as review, auditing, etc - if those even matter for architects.
I'm a PE who has been tinkering with LLMs for a few years now. Opus 5.5 does seem to be a game changer for drafting. I don't generate with python tho, I have it drive AutoCAD headlessly and use our firm's drafting standards/templates. Its a junior-level EIT at first, but I give it redlines via Bluebeam MCP and it updates the relevant CAD SOP for all future sessions.
It would be interesting to add Deepseek 4.1 flash 1017 into this lineup. Deepseek's vision and 3D modeling capabilities are very impressive.
Given how good the leading models got at 3D model creation over the past few months - Blender, Three.js, etc - it's not surprising to see that their 2D CAD-style abilities have improved to this extent as well.
(Amusingly the pelican-on-a-bicycle test doesn't capture this very effectively, despite testing 2D drawing with SVG - because pelicans can't write bicycles, so the results are inherently unrealistic. Having the model create an architectural diagram from a concept sketch is a much more useful 2D test.)
I feel bad for all the inspectors who are going to have to do inspections on vibe-architected buildings in the next 12-18 months.
I don't. A lot of what they do is "make work" that slows down building homes and spaces with dilbert-esque nit picks.
Civil engineering rules are mostly written in blood and costly and dangerous accidents. At least in my country. If somewhere else inspectors don't provide actual value, they should either get regulated or good riddance and AI might actually be better. I feel old, sigh...
That's not what we're talking about. By all means, follow the structural engineer's specification. But if you've ever submitted plans to a city to build a home, there is endless nonsense full of red tape: Drainage plan, soil analysis, tree survey, topographical survey, environmental site assessment, wind analysis, noise assessment, and more.
Sometimes these things make sense, but a lot of the time the work was already recently completed 3 years ago and yet they'll make you do it again. Last month I paid $3,900.00 to get 3 tiny trees and one medium sized oak surveyed here in the Bay Area. It's really about politics and control- the city wants to make a bunch of fake jobs and collect very high fees along the way. Welcome to insanity!
Most of that stuff is important. That is, unless you like sewage pouring into the aquifers.
I doubt that's going to happen.
1) Permitting is slow and, at some point, if you fuck it up big enough, the inspectors will probably take further steps, but otherwise not care much. You simply won't get a permit.
2) Who will care a lot are the builders and creditors. Building is expensive and if there's a couple of months of set back because of your fantasy drawings, that's not going to be in your favor.
Which is not so say AI planning is not going to happen. It will. But at that point it's probably mostly going to work.
Pretty big claim, attention grabbing for sure.
Might be nice to scope out where exactly it is good and where it is lacking ; residential vs commercial vs industrial , A vs S vs MEP, style of architecture? We need better benchmarking here to substantiate.
Agreed. Also the consistency between the different drawings would be nice to have an understanding of. This is where a higher level 3d domain model (such as IFC), do validation, interference checking etc., and generating drawings from those makes more sense (in my mind; but I'm biased). We did a quick demo once [0] but it ran out-of-date quickly due to the hardcoded model ids.
[0] https://ai-chat.ifcopenshell.org/
Washington State is trying out a 'Permit-Ready Plans' program in a few cities, were pre-approved plans are offered at no cost. According to their initial study, building plans account for 5-20% of total construction costs. Software that might be able to shave 10% off new construction costs is a big deal.
I don't know how much time is spent 'drafting' at architecture practices vs modelling.
The drawings can be resolved from a data-enriched 3D model, which is normally where the hard hours go.
A big reason LLMs can write code productively is cause software dev tends to be pretty resilient to failure. And at the pessimistic end, lots of companies can make money off buggy products, so it’s easy to hand wave the failures away. Does that hold true for architecture?
That being said, I’m certain they’ll be extremely useful in other areas such as review, auditing, etc - if those even matter for architects.