elie

50 posts

elie

elie

@eliec_

compiler nerd working on scalable shape rotators @mecadoinc

Cambridge, MA Katılım Nisan 2026
55 Takip Edilen12 Takipçiler
Sabitlenmiş Tweet
elie
elie@eliec_·
What cannot be measured cannot be improved! As @mecadoinc shows in our work with @huggingface, there's a whole lot of work to do to create AI systems that really work for mechanical design. And these are just single parts - no requirements or assembly context!
Michael Rabinovich@MikushRab

Introducing CADGenBench: measure how well AI systems produce engineering-grade 3D parts! While current models can generate 3D parts, they are far from precise enough to build functional parts. We built a benchmark to systematically measure their capabilities on two tasks: 1. Generation from an engineering drawing of a part 2. Editing: given an existing STEP file and a requested change The benchmark is tool-agnostic. It makes no assumptions about how you build the model. You can vary the LLM, and you can vary the environment. Use build123d, Onshape, Autodesk, or a model without an LLM entirely. We open sourced the scoring engine and a reference baseline on top of build123d. A collaboration between Hugging Face and @mecadoinc! Submission space: huggingface.co/spaces/Hugging… Code repository: github.com/huggingface/ca…

English
4
6
11
2.2K
elie
elie@eliec_·
Value of CATIA, Creo, ESPECIALLY Mechanical/STAR is that they are not vibe coded and rather heavily validated. even if u could vibe code a geometry kernel (lol, lmfao even) dyk how hard it is convincing engineers to UPGRADE ts? let alone convince them to use a vibe coded alt
Thomas Ahle@thomasahle

Most expensive software you can actually install and run yourself. What's stopping you from vibe-coding a cheaper alternative? 1. Oracle Enterprise Identity Services Suite Plus — $250,000 per CPU 2. Oracle Business Intelligence Suite EE — $221,250 per CPU 3. Oracle Database Enterprise Edition — $47,500 per CPU (before add-ons) 4. Siemens Simcenter STAR-CCM+ Power Session Plus — ~$61,800 per license 5. Ansys Autodyn — ~$59,000 per license 6. Ansys Mechanical Enterprise — ~$43,000 per license 7. Autodesk Moldflow Insight Premium — ~$30,400/year 8. PTC Creo Design Engineering Professional — ~$29,400/year 9. Dassault CATIA Mechanical Designer — ~$7,560/year 10. MATLAB + Simulink + industry toolboxes — can exceed $10,000–$50,000 per seat depending on toolboxes

English
0
1
1
73
elie retweetledi
Connor Kapoor
Connor Kapoor@connorkapoor·
We just dropped the most robust DFM (design for manufacturability) guide for Investment Cast parts. Super useful for our awesome customers to have all this information right upfront. Really shows the amazing design freedom we unlock here @digitalmetalinc
English
12
22
250
10.4K
elie
elie@eliec_·
@BrierRat I think some of your criticisms are fair. We'll patch with this. But I still think the original drawing, while flawed, is reasonable and readable.
elie tweet media
English
1
0
0
23
Brian Ratliff
Brian Ratliff@BrierRat·
Note, it's not about "completeness". I'm talking about basic stuff like aligning views and extraneous lines. Overlapping dimensions and dimension lines merging with object lines. Extraneous datum callouts which aren't used in a GDT block. That sort of thing. You are allowed and encouraged (you can probably say, required) to fully dimension parts in both the ASME and ISO standards. Profile tolerances sans dimensions is really only a thing with MBD. Has nothing to do with a drawing's readability. And I don't think you are generally talking about tolerancing either, right? I didn't see tolerances called out on the drawings I saw, and in CAD, tolerances are generally not geometrically modeled, unless there is good reason to do so (tolerances are typically just called out on the drawing or in the MBD callout). We are simply talking about defining geometry fully and in a readable format. In tolerancing, there may be some ambiguity, but typically, there is no ambiguity for geometric definition. Even very complex geometry is going to be defined completely in a drawing. I've done literally thousands of CAD drawings in my career. All of them have been fully defined in the layout drawing.
English
1
0
0
40
Brian Ratliff
Brian Ratliff@BrierRat·
New contest part. Claude CAD still struggles, but with experience recorded in skills, it is getting better. The improvement is most noticeable in improved steer-ability Got the correct result of 1.88 lbs These contests make for good testing grounds for drawing interpretation
Brian Ratliff tweet mediaBrian Ratliff tweet media
English
2
1
2
315
elie
elie@eliec_·
@JasBilyeu i hear claude is really good at family mold optimization nowadays
English
1
0
1
99
elie
elie@eliec_·
@Dylan_Morri yea i’m building the next gen of hardware and by “building the next gen of hardware” what i really mean is: debugging my “Teamcenter Admin” Windows User 😎
English
1
0
6
211
Dylan Morris
Dylan Morris@Dylan_Morri·
HELP WANTED If you know how to setup teamcenter pls help my team. Just need initialization. Pls
English
14
3
39
17.3K
elie
elie@eliec_·
Half of the benchmark does use BREP for input ground truth, and all of them use BREP for output ground truth. And the latter asks for a very specific edit on existing geometry- maybe you can just evaluate your system on that. Also, drawings to ASME standards imply a lot. Your description that a drawing to standard would always produce the same model might be true in theory, but in practice not true at all, especially for more complex parts. Most of the time, for complex parts, engineers will use profile tolerances, which don’t work for this test case when there is no reference CAD. ASME 14.5 also implies a lot, for example RFS. But the whole world doesn’t use 14.5, and also these benchmark tasks aren’t meant for manufacturing! Because we needed the objective to have zero “implication,” because they aren’t going to be inspected after manufacturing, and because we wanted to avoid only American standards, we explicitly controlled every single feature, which made the drawing more complete at the cost of readability.
English
1
0
0
29
Brian Ratliff
Brian Ratliff@BrierRat·
Why not use a BREP for the ground truth? Drawings drawn to standards are not ambiguous in any way. Give a drawing to 100 different people to model and you'll get 100 identical models. That's the point of this exercise for Claude CAD. I'm training it to read drawings. But the drawings you are showing in your eval set are unreadable, because they are not drawn to standard. I would not expect my Claude CAD system to do well in your evaluation.
English
2
0
0
41
elie
elie@eliec_·
@BrierRat ground-truth, ie the "perfect geometry" that's why everything, including inner and outer fillet radii have to be dimensions, or else things are up to interpretation
English
1
0
0
57
elie
elie@eliec_·
@BrierRat Ha! Sorry, the space didn't make it through. Yes, huggingface.co/spaces/Hugging… Those drawings are made specifically to give a CAD generation model total gt to build the parts with, because the scoring relies on complete topological accuracy to a gt.
English
1
0
0
59
Brian Ratliff
Brian Ratliff@BrierRat·
@eliec_ I assume you are talking about CADGenBench, right? (Your link is 404). Why are the drawings almost unreadable? What's the intended test? Like this one here, I don't think I could even model this accurately.
Brian Ratliff tweet media
English
1
0
0
27
elie
elie@eliec_·
@camrrowe percentage of your industrial machine that is extrusions = skill level
English
0
0
2
30
Kyle Cothern
Kyle Cothern@risknc·
@mirandahw_ no approvals in the title block, missing a cage code, no date in the revision history, not fully defined and no minimally dimensioned drawing note, so there's ambiguity in the spacing of the angled curves on the left, no material specified wouldn't make it past check.
English
3
0
10
545
Miranda 🚀
Miranda 🚀@mirandahw_·
blue merch team cooked
Miranda 🚀 tweet media
English
10
22
211
5.5K
Jake Fitzgerald
Jake Fitzgerald@earthtojake·
you can design hardware in codex / claude code this is a 7dof robot arm with functional kinematics that was 100% prompted (no cad software) using custom skills, agents can generate STEP files, robot description files, and even run FEA checks from text / image prompts
English
8
19
213
41.2K
elie
elie@eliec_·
@emm0sh made steve so mad they started using windows emulators to get NX instead
English
0
0
12
698
Photon Commander
Photon Commander@photoncmndr·
Just a friendly reminder that AI still sucks at hardware.
English
11
1
42
4.1K
István Csanády
István Csanády@istvan_csanady·
MiT jUsT oPeNsOuRcEd A mOdEl ThAt WiLl EnD tHe CaD iNdUsTrY
English
7
3
108
12.8K

Keşfet

@digitalmetalinc @BrierRat @JasBilyeu @Dylan_Morri @camrrowe @elonmusk @BarackObama @taylorswift13