Anushrut Jignasu

@aphysicist Yes! Happy to help with it as well

should i drop this lol

@vincesitzmann nice blog post! we're already on the path to this with slice-100k (slice-100k.github.io). 100k+ paired G-code, meshes & renderings for finetuning on the printer interface.

In my recent blog post, I argue that "vision" is only well-defined as part of perception-action loops, and that the conventional view of computer vision - mapping imagery to intermediate representations (3D, flow, segmentation...) is about to go away. vincentsitzmann.com/blog/bitter_le…

@devonkcopeland @amypretzel @d4ferris Our viewer is limited to ~80-90 parts

@ajignasu @amypretzel @d4ferris Cool! Slice 100k is super fast. I like how the search updates in real time

We built a CAD search engine! @d4ferris and I indexed 1 million parametric CAD models using AI-generated captions and embeddings as a side project. Full write-up + open dataset: finalrev.com/blog/embedding…

@gkopanas They could be seen as a DSL, but its really just an API wrapper. Not built for gradient descent

@ajignasu Aren't blender scripts a DSL?

This work is amazing but makes me think that the blender scripting language as a proxy for communicating and generating in 3D is ad-hoc. These "languages" were never created for that purpose and I would be surprised if there is no better alternative.

@yacineMTB We’ve had good results with 3-axis aerosol jet printing using a conductive ink: sciencedirect.com/science/articl… Still lots of room to push this forward on in-the-wild geometries

this might be a stupid question, but couldn't you 3d print circuits using magnet wire?

@zachdive @adamdotnew normal map maxxing

the palace of fine arts, sponsored by @adamdotnew

Assembly simulation and auto-fix loops still seem like open territory for text-to-CAD

@AlexanderLong @ChaminHewa Let’s go @ChaminHewa!

Since I started getting interested in ML I got it in my head that all I wanted to do was one smart thing that I could look back on and be satisfied that I did. Most papers are kinda bad even if they get accepted - the idea is very incremental, or it's just not that good an idea, or it doesn't really matter. I never was able to do this all through PhD or my time at Amazon. All the papers I did there got into various places, but I never really thought they were actually that good. And I'd pretty much given up on this because Pluralis meant I couldn't really devote enough time to research myself. But in February I decided I didn't care and spend two months focused on a specific problem that had been going round in my head for about a year that I felt we needed to solve, and the solution came to me, and @ChaminHewa picked it up and generalised the approach and ran a bunch of novel experiments I hadn't thought of, and pulled everything together into an actual paper. And yesterday we presented this work at NeurIPS. This is the first and probably only work I will ever do that for me feels like "ok that was GOOD". I don't care if it racks up a bunch of citations and disperses into the field or not, I don't care if someone repackages the ideas and takes all the credit for it, I don't care. For me there is an internal checkbox that just got ticked after more than ten years of trying. Anyone in ML will understand what I'm trying to say. Special day I'm going to remember for a long time.

@yongyuanxi The adapter layers trick to project CAD tokens to the LLMs embedding space is kinda similar to how LLaVA uses projection layers to align visual features with text embeddings

CAD-Tokenizer 🪙: a tokenizer specific to CAD command sequences, results in 10% better generation performance

Domain specific outputs are too general right now (number of parts, types of parts, their functional nature). For part decomposition, prompt engineering is surprisingly useful. Part costs aren’t realtime/market costs. Overall, the model provides a good starting point for parts and instructions.

@ajignasu Super cool! How accurate was the BOM extraction / the instructions?

Predicting manufacturing costs directly from 2D CAD drawings using XGBoost on geometric features

@willccbb Reminds me of W&B’s naming scheme for multiple runs, like lone-aardvark or frosty-wolf

one of my favorite parts of working at prime intellect is getting to pick the silly names whenever someone launches a new instance

@Suhail Would be cool to have a cubesat stream images of earth/space to Playgrounds app in real time

3/ It has finally arrived! Pretend cubesat assembly begins.

1/ Oops new hobby 🫡

Excited to share that I’ll be presenting our work, “Slice-100K: A Multimodal Dataset for Extrusion-based 3D Printing,” at #NeurIPS2024 Session details: West Ballroom A-D, #5301 Wed, Dec 11 | 4.30 - 7.30 pm PT Feel free to reach out!

Delighted that Slice-100K was accepted to #NeurIPS2024 D&B track! 🎉 Website: slice-100k.github.io Github: github.com/idealab-isu/SL… Read here: arxiv.org/abs/2407.04180 Grateful for my co-authors and the @NeurIPSConf reviewers!

@zoodotdev Hi, will you have internship positions open anytime soon? Text-to-CAD is amazing and would love to contribute!

We believe SLICE-100K will be the first step in building multi-modal foundational models for digital manufacturing and beyond. This is a joint effort with @kellym11111 , @M17409753Ankush, Lucas Nerone Rillo, @baskinscience, Aditya Balu, @chegday, and @adarshk 5/n

To demonstrate its usefulness, we finetune GPT-2 on the task of G-code translation from a legacy format (Sailfish) to a modern and widely used format (Marlin). 4/n

🚨📢 Excited to announce SLICE-100K, a first-of-its-kind dataset of over 100,000 G-code files, along with their tessellated CAD model, LVIS (Large Vocabulary Instance Segmentation) categories, geometric properties, and renderings. 1/n