Loading video...

Video Failed to Load

Go Home

don't let them fool ya with video game assets! Working feature tree, working assembly, 100% AI generated! Text-to-CAD is not the hard part and it is not the interesting part!

18,445 views • 2 months ago •via X (Twitter)

20 Comments

Matt Ferraro's profile picture
Matt Ferraro2 months ago

Early access signup here:

Jern's profile picture
Jern2 months ago

I would argue that text-to-CAD is actually very hard. It is a fundamentally bad idea because a two sentence prompt is completely inadequate to describe a CAD design of almost any non-trivial complexity. Text-to-CAD creates instant garbage

Matt Ferraro's profile picture
Matt Ferraro2 months ago

I think we are saying the same thing but differently! Text to cad is garbage unless you have powerful ways to validate that the parts are actually manufacturable and performant, which we believe is mostly an simulation and FEA problem!

Brian Ratliff's profile picture
Brian Ratliff2 months ago

It's not validation, in my experience. It's communication which is the problem. Mechanical design is a much higher dimensional space than most coding applications. For instance, take the napkin sketch and the associated model below. How many dimensions are required to fully define this shape, even given drafting conventions such as "lines appearing straight are straight"? And this is just simple throwaway test model geometry with ~8 features.

Matt Ferraro's profile picture
Matt Ferraro2 months ago

@Jern123d exactly! that's why we built our system as a really good CAD program first. If you need a specific dimension set, it's faster to just go in and set it than to ask an AI to do it. But if you need to change all your countersunk screw holes to counterbored, that's a job for AI

Andres's profile picture
Andres2 months ago

dark mode + some real bike rpms

Matt Ferraro's profile picture
Matt Ferraro2 months ago

Engine you gotta stop. You smoke too tough. You spin too fast. They'll kill you

MilkDromeda's profile picture
MilkDromeda2 months ago

I have investigated AI to CAD for a period of time, I think model tree is not needed for AI parametric design, and it is also the most significant difference between AI CAD modeling and human design, if we think deeply through it. but still excited to see this browser platform.

Space Colonize's profile picture
Space Colonize2 months ago

Text to additive manufacturer is coming too

Matt Ferraro's profile picture
Matt Ferraro2 months ago

can't wait!

Jake Dahn's profile picture
Jake Dahn2 months ago

Oh em gee, I can almost hear it

Connor Kapoor's profile picture
Connor Kapoor2 months ago

Looking good! Did you guys reset accounts?

Matt Ferraro's profile picture
Matt Ferraro2 months ago

Yes we reset accounts, I’ll get you sorted here in a few mins!

SC Maker Space's profile picture
SC Maker Space2 months ago

Makes me think I should sell my CAD library as training data. 32k files at last count.

Carl Lumma's profile picture
Carl Lumma2 months ago

What geometric modeling kernel are you using? And does it run on a sever or in the browser?

Matt Ferraro's profile picture
Matt Ferraro2 months ago

It’s parasolid under the hood, so it runs on a server. The client gets a brep and mesh, so the browser is doing the actual render work!

Chizzydigital's profile picture
Chizzydigital2 months ago

This one is more impressive

Alex's profile picture
Alex2 months ago

what's the hard part then? tolerance stack-up validation? or getting the assembly to actually survive a physics sim?

Maicon Lourenço's profile picture
Maicon Lourenço2 months ago

Awesome motion!

Daniel Owsianski's profile picture
Daniel Owsianski2 months ago

I wrote my own browser viewer for OpenSCAD because reviewing the result was the bottleneck, never producing it. Whether the part is actually right I learn 3 hours later at the printer. What do you give people to check an AI assembly before that point?

Related Videos