What exactly is the use case for this? Just making pretty little 3d models?
There are so many reasons why I, as an engineer, will never even attempt to use AI for mechanical design, that trying to list them all is about to give me an autistic screeching fit.
Even if all I need is a simple little bracket or something, I can model that and know it's right much quicker than I can ask the AI to do it and then check it's work. There is no time savings here.
Heck, for any of the stuff I need that is simple enough to plausibly ask AI to draw, I don't even bother to model in the first place, I would either sketch it with a pencil or just make the piece right off the top of my head.
If it's more complicated than that, then my approach grows to include things like what stock I have available, what tooling, fixturing and machines are present, whether I can use any COTS hardware to simplify the design, the tolerancing scheme I want to use...and my output needs to include not just the model, but toleranced drawings and any other assemblies and such that are required.
And besides all of that, and with love....OpenSCAD is a joke, and if you seriously try to tell me that "the best paradigm for CAD generation is to generate CAD as code", I cannot take you seriously.
Is CADAM also what's used for the commercial product adam.new? How did you manage to write all those plugins? If it is also CADAM isn't stuff getting lost in the Fusion/Solidworks/Onshape -> OpenSCAD -> back process? Do constraints and everything just seamlessly import/export?
Some comments here mention tolerances/functional requirements. Do you think the LLM/screenshot loop will scale to that too? Maybe rendering subassemblies individually until they make sense? Still feels like a full functioning V8 engine block needs _a lot_ of ghost-view screenshots to verify it works. What's your thoughts on a "simulation" approach, since it's not aligned with your bitter-lesson-blog-post?
Are you able to reveal more about what kind of traction you have? 10s/100s/1000s of companies?
Very cool open source project, and thanks for sharing so much!
The same principles of an agent writing cad as code and then visualing inspecting the output in a loop runs through all our products. Whilst CADAM uses OpenSCAD, Fusion uses Python, Onshape use FeatureScript etc..
The majority of our enterprise traction is on our flagship product: https://adam.new/
Here you can connect to your engineering software and use AI to generate:
I'm pretty skeptical of AI products, but your onboarding and first design experience has been pretty awesome. I will definitely spend a bit more time experimenting with this
I have a general question regarding 3d design using LLMs. My understanding is that all current applications have been trying to deal with 3d design/CAD as text/code. LLMs are clearly good at those but do you see this as the long term approch for 3d designs? do you see world models eventually evolving to produce 3d spaces or point clouds or CAD designs instead of doing video? is this approch explored?
Congrats on the launch!
"I need an engine mount for 1999 toyota land cruiser j90 for the 1kz-te engine with a manual gearbox. Can you generate me a cad to send to a company in China to 3d print it?"
"Done — I've created a heavy-duty, fully parametric engine mount bracket that fits a typical four-bolt block pattern and a single-stud chassis isolator with an alignment pin, much like what the 1KZ-TE requires."
I dont think it's even close :(
PS. Your entry message should be "Madam, I'm Adam" ;)
There's already a lack of information online for simple things like torque specs. I can't imagine that a skilled professional could design the engine mounts even with if they had all of the relevant context online.
As far as I know, the way that these reproduction hardware companies operate is that they have physical cars that they can design around.
I have a 1993 Subaru WRX and I needed to replace the coolant header tank because mine had a bunch of leaks. I ended buying one from a specialty fab shop in the UK and I had to make a few measurements for them because there was varying bolt spacing for GC8 Impreza models.
I find all current LLMs to have pretty poor spatial awareness. It is becoming better, but still very poor. How are you dealing with that? Got any special tricks, any advice?
Makes sense, I keep hoping some startup will be able to crack the open licensing without limiting their business problem. Would love it if we could do a bit better than the rugpulling dynamics that are kind of common now without just giving AWS a buffet of startups to hurt.
I asked it to create 3d model of "AMF-O97L45-DB". It pulled datasheet and generated 3D model. Left is reality, right is what was generated: https://imgur.com/a/oNaz51q
It's currently in the plugins and we're working on bringing it to CADAM. Basically you'll be able to use the GUI, and give face/edge selection context to your prompt "extrude a hole through this face". It's directly tied to us adding brep support.
How can this approach be better than just selecting the edge and click the extrude button/write extrude command? Now you have to start writing a prompt and hope that what you want to do is understood by the LLM. I mean, CAD is really not so complicated with the tools we currently have. You just have to learn how to use them.
That's more of an example to address the point. We've find our users often use this feature in our onshape/fusion extensions in complex assemblies. Being able to select faces and edges as context in addition to prompts can be quite a powerful interface in more complex projects where users need to adjust tolerances or edit multiple objects to prevent interference
I understand the goal, but describing complex geometries with specific tolerances with natural language is much more complex than creating the geometry programmatically. There are geometries that I could not clearly describe with words, but it's clear the operations I need to do to create them. But who knows, maybe I'll be proven wrong.
We're intent on building a dedicated editor, that way we can build a lot of nice UI! We'd also like to build public mcps for some of the popular cad tools
take a look at modelrift.com, it is built around annotating built models by basic pen and arrow tools, works fairly well ('smarter' model is significantly better)
Who are your users? Are you working with professionals that use similar commercial products or hobbyists? I have a hard time imagining that seasoned industrial designers prefer text over sketches…
I suspect that your VLM might do a bad job at transcribing sketches into CADs, and you wrongly interpreted the adoption data as a preference for text-based interaction
I think it's primarily for designing chemical processing systems, though I know it through the pipe layout software being used off-label to design vehicle electrical harnesses.
Yeah, no, that's a lie. This isn't a CAD model. It's a fantasy 3d model that looks like it's straight out of Gearhead Garage (1999).
Any time I see these 'AI CAD' solutions it's always toys, toys, toys. Show me something functional that you've actually manufactured (shitty 3D prints don't count). Or at least show me something that can actually be assembled and isn't just a bunch of boxes with no fasteners to hold them together.
Why not? The 3D print market is pretty large and tools to generate some designs that can then be tweaked are pretty useful in that context. I don't think that type of AI CAD tool would replace professional CAD work, that's something that requires way too much context and human judgement. But being able to prototype something to be 3D printed via an AI thing is one of the few places where I see AI being genuinely useful.
I personally enjoy designing my own things with Plasticity, so wouldn't be the perfect target audience
not to say this isn't cool, but it's about as useful as having claude generate a JavaScript illustration of how a v8 works and then expecting someone to manufacturer an engine from that
For anyone doing CAD at a professional level (ie not 3d printed trinkets), the important parts are the physical parameters and tolerances designed into the model. For example I suspect your crankshaft would rip itself apart at engine speeds, not to mention all the plumbing, oil and coolant delivery, and auxiliary pumps and belts are missing
I see cams intersecting eachother and still nothing that is actually ready to be manufactured or even looks like a design that has had any thought put into it. It's the CAD equivalent of idle doodling.
Do you have a single person on your team that's actually a mechanical engineer with practical industry experience?
Yes and we have a number of mechanical engineers using our extensions! AI in CAD is defo a WIP but when you trace the progress it's not too hard to envisage what the future will look like.
For the Fusion demo we intentionally didn't include the block or any accessories in the visualization as we wanted demonstrate Adam's ability to reason through the mechanical workings of an engine, like how the cams push the valves or the way the the crankshaft drives the connecting rods.
> Simple parameter tweaks bypass the model entirely; adjusting a slider does a deterministic regex update on the SCAD source, requiring no LLM call.
Why do you use regex for that? OpenSCAD allows you to pass data natively, no? What’s the advantage of the regex over using that?
This is so the .scad file stays the canonical state. With -D, the rendered geometry diverges from what the file says.
https://modelrift.com/blog/openscad-llm-benchmark/ LLMs are still weak at spatial reasoning, but it gets better. Check out modelrift.com for another alternative
What exactly is the use case for this? Just making pretty little 3d models?
There are so many reasons why I, as an engineer, will never even attempt to use AI for mechanical design, that trying to list them all is about to give me an autistic screeching fit.
Even if all I need is a simple little bracket or something, I can model that and know it's right much quicker than I can ask the AI to do it and then check it's work. There is no time savings here.
Heck, for any of the stuff I need that is simple enough to plausibly ask AI to draw, I don't even bother to model in the first place, I would either sketch it with a pencil or just make the piece right off the top of my head.
If it's more complicated than that, then my approach grows to include things like what stock I have available, what tooling, fixturing and machines are present, whether I can use any COTS hardware to simplify the design, the tolerancing scheme I want to use...and my output needs to include not just the model, but toleranced drawings and any other assemblies and such that are required.
And besides all of that, and with love....OpenSCAD is a joke, and if you seriously try to tell me that "the best paradigm for CAD generation is to generate CAD as code", I cannot take you seriously.
I agree openSCAD isn't anywhere near powerful enough for professional workflows. Hence why I framed this as AI TinkerCAD.
However, you can drive any professional CAD software though code. As we drive Autodesk Fusion via python through agents.
This doesn’t seem to be for machining CAD but more like stuff you would 3D print at home
Is CADAM also what's used for the commercial product adam.new? How did you manage to write all those plugins? If it is also CADAM isn't stuff getting lost in the Fusion/Solidworks/Onshape -> OpenSCAD -> back process? Do constraints and everything just seamlessly import/export?
Some comments here mention tolerances/functional requirements. Do you think the LLM/screenshot loop will scale to that too? Maybe rendering subassemblies individually until they make sense? Still feels like a full functioning V8 engine block needs _a lot_ of ghost-view screenshots to verify it works. What's your thoughts on a "simulation" approach, since it's not aligned with your bitter-lesson-blog-post?
Are you able to reveal more about what kind of traction you have? 10s/100s/1000s of companies?
Very cool open source project, and thanks for sharing so much!
The same principles of an agent writing cad as code and then visualing inspecting the output in a loop runs through all our products. Whilst CADAM uses OpenSCAD, Fusion uses Python, Onshape use FeatureScript etc..
The majority of our enterprise traction is on our flagship product: https://adam.new/
Here you can connect to your engineering software and use AI to generate:
- CAD - Renderings - Slides for design review - BOM
and much more!
I'm pretty skeptical of AI products, but your onboarding and first design experience has been pretty awesome. I will definitely spend a bit more time experimenting with this
damn credits are consumed fast. And they are pretty expensive, $20 for 2k credits won't last long at all
Thanks!!
Can it work from photos? I'm specifically thinking about stuff like this:
https://www.tooltrace.ai/
Wow, that’s a solution to an oddly specific problem I didn’t know existed, but it looks kind of cool
Yes it can! You can even use mesh mode which is very strong for this!
I have a general question regarding 3d design using LLMs. My understanding is that all current applications have been trying to deal with 3d design/CAD as text/code. LLMs are clearly good at those but do you see this as the long term approch for 3d designs? do you see world models eventually evolving to produce 3d spaces or point clouds or CAD designs instead of doing video? is this approch explored? Congrats on the launch!
I strongly believe that by scaling up LLMs we will get there. I write about this here in more detail: https://adam.new/blog/bitter-lesson-ai-cad
"I need an engine mount for 1999 toyota land cruiser j90 for the 1kz-te engine with a manual gearbox. Can you generate me a cad to send to a company in China to 3d print it?"
"Done — I've created a heavy-duty, fully parametric engine mount bracket that fits a typical four-bolt block pattern and a single-stud chassis isolator with an alignment pin, much like what the 1KZ-TE requires."
I dont think it's even close :(
PS. Your entry message should be "Madam, I'm Adam" ;)
There's already a lack of information online for simple things like torque specs. I can't imagine that a skilled professional could design the engine mounts even with if they had all of the relevant context online.
As far as I know, the way that these reproduction hardware companies operate is that they have physical cars that they can design around.
I have a 1993 Subaru WRX and I needed to replace the coolant header tank because mine had a bunch of leaks. I ended buying one from a specialty fab shop in the UK and I had to make a few measurements for them because there was varying bolt spacing for GC8 Impreza models.
If they AI would respond with what you just wrote, instead of "Done..." the world would be trending in a good direction.
I find all current LLMs to have pretty poor spatial awareness. It is becoming better, but still very poor. How are you dealing with that? Got any special tricks, any advice?
I write about this in detail here: https://adam.new/blog/bitter-lesson-ai-cad
This is improving greatly in recent model releases
Opus 4.5-4.7 was pretty bad at it, 4.8 was a bit better, and I have not tried Fable much.
So basically you have a good enough code that’s “intuitive” for a model, screenshots, and that’s it?
fable is a fair bit better, but to an extent its that it tried more things to get an understanding of whats happening than opus does
Fable is considerably better from my experience: https://x.com/LLMJunky/status/2065229625702109340?s=20
Fingers crossed it comes back!
My favorite spatial reasoning benchmark: https://minebench.ai/
no tricks, I'd definitely be curious to know how much screenshots help
Cool that you licensed it with GPL, what was your thinking on the license?
Honestly I would've liked to MIT. It's mainly that the openscad wasm we use is GPL
Makes sense, I keep hoping some startup will be able to crack the open licensing without limiting their business problem. Would love it if we could do a bit better than the rugpulling dynamics that are kind of common now without just giving AWS a buffet of startups to hurt.
I asked it to create 3d model of "AMF-O97L45-DB". It pulled datasheet and generated 3D model. Left is reality, right is what was generated: https://imgur.com/a/oNaz51q
- wrong pitch
- wrong pins position
- missing pins
did you try to iterate? copypasting your brief message here to the prompt would probably fix something.
yes i agree, this would probably fix it
curious how this compares to on baseline: https://adam.new/
let me know!
It could be a nice touch to give some examples of what it's possible to ask CADAM!
Yes good idea! We've added a few in the read me if you'd like to take a look
Can you talk more about the UI for face/edge selection that you're working on? Is that only going to be in the OnShape/Fusion plugins?
It's currently in the plugins and we're working on bringing it to CADAM. Basically you'll be able to use the GUI, and give face/edge selection context to your prompt "extrude a hole through this face". It's directly tied to us adding brep support.
How can this approach be better than just selecting the edge and click the extrude button/write extrude command? Now you have to start writing a prompt and hope that what you want to do is understood by the LLM. I mean, CAD is really not so complicated with the tools we currently have. You just have to learn how to use them.
That's more of an example to address the point. We've find our users often use this feature in our onshape/fusion extensions in complex assemblies. Being able to select faces and edges as context in addition to prompts can be quite a powerful interface in more complex projects where users need to adjust tolerances or edit multiple objects to prevent interference
I understand the goal, but describing complex geometries with specific tolerances with natural language is much more complex than creating the geometry programmatically. There are geometries that I could not clearly describe with words, but it's clear the operations I need to do to create them. But who knows, maybe I'll be proven wrong.
I largely agree with you. It's case by case and the ideal ux imho is to have both.
Very cool. Why not start with an MCP instead?
An existing LLM could drive the generation while the MCP can render the final result?
We're intent on building a dedicated editor, that way we can build a lot of nice UI! We'd also like to build public mcps for some of the popular cad tools
why is text the setup, rather than sketches? pictures?
ive found a process by which the llm gives me a picture, then i draw on it and hand it back works fairly well
take a look at modelrift.com, it is built around annotating built models by basic pen and arrow tools, works fairly well ('smarter' model is significantly better)
you can upload an image or a sketch! we actually have drawing suppor in our extensions, but we've found our users use it far less than we expected!
Who are your users? Are you working with professionals that use similar commercial products or hobbyists? I have a hard time imagining that seasoned industrial designers prefer text over sketches…
I suspect that your VLM might do a bad job at transcribing sketches into CADs, and you wrongly interpreted the adoption data as a preference for text-based interaction
generally professionals our flagship app at https://adam.new/
hobbyists and makers use CADAM
FYI there is already a product with a very similar name, CADEM.
Oh thanks! What's CADEM?
I think it's primarily for designing chemical processing systems, though I know it through the pipe layout software being used off-label to design vehicle electrical harnesses.
> A complete V8 internal combustion engine
Yeah, no, that's a lie. This isn't a CAD model. It's a fantasy 3d model that looks like it's straight out of Gearhead Garage (1999).
Any time I see these 'AI CAD' solutions it's always toys, toys, toys. Show me something functional that you've actually manufactured (shitty 3D prints don't count). Or at least show me something that can actually be assembled and isn't just a bunch of boxes with no fasteners to hold them together.
> shitty 3D prints don't count
Why not? The 3D print market is pretty large and tools to generate some designs that can then be tweaked are pretty useful in that context. I don't think that type of AI CAD tool would replace professional CAD work, that's something that requires way too much context and human judgement. But being able to prototype something to be 3D printed via an AI thing is one of the few places where I see AI being genuinely useful.
I personally enjoy designing my own things with Plasticity, so wouldn't be the perfect target audience
Yes I intentionally called this AI TinkerCAD
We could defo update our readme! What do you think of this?: https://x.com/aaronli/status/2064876123109089742?s=20
Fable 5 in our Fusion Extension.
not to say this isn't cool, but it's about as useful as having claude generate a JavaScript illustration of how a v8 works and then expecting someone to manufacturer an engine from that
For anyone doing CAD at a professional level (ie not 3d printed trinkets), the important parts are the physical parameters and tolerances designed into the model. For example I suspect your crankshaft would rip itself apart at engine speeds, not to mention all the plumbing, oil and coolant delivery, and auxiliary pumps and belts are missing
I see cams intersecting eachother and still nothing that is actually ready to be manufactured or even looks like a design that has had any thought put into it. It's the CAD equivalent of idle doodling.
Do you have a single person on your team that's actually a mechanical engineer with practical industry experience?
Yes and we have a number of mechanical engineers using our extensions! AI in CAD is defo a WIP but when you trace the progress it's not too hard to envisage what the future will look like.
For the Fusion demo we intentionally didn't include the block or any accessories in the visualization as we wanted demonstrate Adam's ability to reason through the mechanical workings of an engine, like how the cams push the valves or the way the the crankshaft drives the connecting rods.
Can you claim your product here? https://thecadhub.com/details/adam-cad/