| 3 Jan 2024 |
 rawwerks | An extreme example classifying 2M recipes for $19 (https://openpipe.ai/blog/from-prompts-to-models) | 05:36:38 |
 aflofo | Right now I'm working on an auto-differentiable CAD kernel in JAX for a different objective https://x.com/OpenOrionOrg/status/1721994755498786903?s=20, however, this could provide the missing link between the CAD and text. I think generation would be a lot more accurate when you directly interface with the kernel and train on those gradients rather than just running text based code generations. | 05:46:12 |
| aflofo | * Right now I'm working on an auto-differentiable CAD kernel in JAX for a different objective https://x.com/OpenOrionOrg/status/1721994755498786903?s=20, however, this could potentially provide the missing link between the CAD and text. I think generation would be a lot more accurate when you directly interface with the kernel and train on those gradients rather than just running text based code generations. | 05:46:24 |
| aflofo | Getting pretty far | 05:47:20 |
| aflofo | https://x.com/afshawnl/status/1742241487570584028?s=20 | 05:47:21 |
| aflofo | https://x.com/afshawnl/status/1733004955571610054?s=20 | 05:47:30 |
| rawwerks | gogogo! | 21:51:59 |
| rawwerks | just added basic CadQuery to my text-to-3D AI tool this morning (after discovering CadQuery last night)
https://nething.xyz/ | 21:53:14 |
| rawwerks | * just added basic CadQuery to my text-to-3D AI tool this morning (after discovering CadQuery last night) => https://nething.xyz/ | 21:53:27 |
| rawwerks | i didn't import all of the extras and the AI is not very bright yet (just GPT3.5 with some basic instructions). so be gentle | 21:54:18 |
 Roger Maitland#7070 | Iād be interested in how cq compares to https://build123d.readthedocs.io/en/latest/index.html# | 21:54:32 |
| rawwerks | oh jesus, now i need to add another one? š | 21:55:03 |
| rawwerks | already have SDF and CadQuery - how many freaking python CAD tools are there? | 21:55:29 |
| Roger Maitland#7070 | Please š | 21:55:48 |
| rawwerks | definietly don't have time this week, but will take a look. | 21:56:04 |
| rawwerks | * definitely don't have time this week, but will take a look. | 21:56:11 |
 shehabattia96 | From my experience with ChatGPT 4 and CodeToCAD, given the docs, the LLM is able to call methods, but sometimes the order of operations would be wrong or missing steps. So for example, if I asked it to make a green cube that is 2x2x2, it would correctly call Part("cube").create_cube(2,2,2).set_material("green") without creating a material first and setting its color to green, so you have to tell it rules like "If you use a material, you have to create it first". Ergo, someone has to sit down and create those rules..
I think it's possible to get a general outline of code-CAD that a human can then tweak. It's basically the same problem with for anything you ask ChatGPT for that's code related. It won't be fully self-driving any time soon.. or ever. | 22:02:20 |
| shehabattia96 | Not that many, actually haha | 22:02:50 |
| shehabattia96 | Not that many good, feature-complete ones | 22:03:19 |
| rawwerks | i agree with a lot of this. for https://nething.xyz/, the color/material is set manually, and i added an "editor" so that the human can fix the AI mistakes. | 22:03:44 |
| shehabattia96 | Curious, and you don't have to give a specific answer, how much dough are you throwing at optimizing this? š | 22:35:31 |
| shehabattia96 | * Curious, and you don't have to give a specific answer, how much dough are you throwing at tuning this? š | 22:36:18 |
| shehabattia96 | * Curious, and you don't have to give a specific answer, how much dough are you throwing at tuning params for this? š | 22:36:25 |
| rawwerks | fine-tuning? none right now. just gpt-3.5 and a lot of examples | 22:46:10 |
| shehabattia96 | you're just feeding the examples via a plugin? | 22:46:58 |
| rawwerks | i am interested in fine-tuning though, we were chatting a bit about this last night. if there are enough example cases i think it could work. | 22:47:00 |
| rawwerks | no plugin, i'm just calling the api | 22:47:17 |
| shehabattia96 | Sounds interesting š | 22:47:35 |
| rawwerks | the challenge i see is that everything will need a human eval (other than does the code parse) - because in CAD there will be many "equally correct" or "equally wrong" answers that look fine in Code but then when you examine the 3d model it's not quite right. | 22:48:37 |
| rawwerks | (whereas most fine tuning i'm aware of uses automated or semi-automated evals). | 22:49:36 |
