 | CadQuery Showcase | 574 Members |
| 3 Servers |
| CadQuery Showcase | 574 Members |
| 3 Servers |
| Sender | Message | Time |
|---|---|---|
| 21 Dec 2023 | ||
 aflofo | My thinking is LLMs are terrible at ground up CAD but suprising good at putting together high level CAD API parameterization | 19:45:03 |
| aflofo | For now | 19:45:08 |
| aflofo | * My thinking is LLMs are terrible at ground up CAD but suprisingly good at putting together high level CAD API parameterization | 19:46:30 |
 Jojain | What would be amazing is drawing to actual part. I'm sure prototypes could already be made where you draw a hand sketch of what you want to do and it generates the code to make it. What a time to be alive 😄 | 19:47:56 |
| aflofo | You can potentially do that already with GPT's no multi modal API. Haven't used it my self but have seen things where people draw something and it renders it in matplotlib | 19:49:10 |
| aflofo | * You can potentially do that already with GPT's new multi modal API. Haven't used it my self but have seen things where people draw something and it renders it in matplotlib | 19:49:16 |
| aflofo | * You can potentially do that already with GPT's new multi modal API. Haven't used it myself but have seen things where people draw something and it renders it in matplotlib | 19:49:21 |
| aflofo | * You can potentially do that already with GPT's new multi modal API. Haven't used it myself but have seen things where people draw something on paper and it computes the geometry and draws it in matplotlib | 19:49:49 |
| Jojain | Yes I have seen it too, though there is A LOT of work required around the prototypes to make it actually useful. I feel a single person can build a working prototype alone quite fast but to have something actually useful requires a lot of non LLM logic around it ! Interesting stuff nonetheless | 19:52:17 |
| 22 Dec 2023 | ||
 deanzyne | Exciting but also Kinda scary.... AI moves fast Someone please tell me why the jobs I thought would be least resistant to ai take over are the jobs that are correctly the most invested into AI art AI video AI music AI Voice and voice acting AI CAD All the creative spheres.... Please tell me why Ai accounting isn't a thing , but Ai CAD is? | 12:45:14 |
| Jojain | Ai cad isn't at all. Cad is all about precision a complex iterative process which AI isn't close to just "replace" but it makes things much faster and that's great | 13:15:38 |
 regattaguru_15413 joined the room. | 14:48:07 | |
 dashlambda | You know, you just made me realize how unexpected that is. We always think of computers and AI as super precise and logical, so we would think it'd be best suited for rigorous and precise tasks, but the way we're going about it ended up with the exact opposite. The reason current "AI" is being applied to creative fields much more than technical fields is because it's extremely error prone, and creative fields are way more error tolerant. If a story generator has a plot hole or an image generator adds an extra finger it can be close enough, but if an accountant gets one character wrong it can cost someone thousands of dollars, and if a mechanical engineer gets a dimension wrong a part can fail. Its primary use in technical fields right now is to generate a starting point, not the final product but something that the engineer or whoever else can modify, fix, and use. In technical applications it's worthless without oversight --for the time being. | 15:19:33 |
| dashlambda | There's also the fact that LLMs still struggle with coherent long-term planning. You can generate a lot of code with an LLM but it can't architect good software. | 15:22:52 |
| dashlambda | Yet | 15:23:49 |
 roger_maitland | I don't remember the name of the company but a former co-worker left to join a forensic accounting company that was using AI to help detect fraud. The AI wasn't producing spreadsheets but looking for patterns in the money transfers that looked suspicious. | 15:26:18 |
| dashlambda | Yeah, that's what ML's primary applications were before generative stuff took off --identifying complex and unintuitive patterns in large datasets. I did a project using deep learning to detect biomarkers for Alzheimer's in saliva. The information was in the data but it was expressed in complicated relationships between variables rather than something directly measurable --I have no idea how long it would take a doctor to put it together but as a grad student with a GPU and a dataset I got it to 95-98% accuracy in a few weeks. Of course, I have no idea what's going on with the protiens, I just know how my model performed on validation data. | 15:34:14 |
| dashlambda | The whole field of machine learning is just curve fitting taken to its logical extremes. | 15:41:05 |
| deanzyne | Sounds phenomenal Amazing work In scenarios like that I believe that's the biggest impact AI can have, extremely useful! Making what might take months -years into something that takes weeks or days | 19:23:43 |
 poblazaid joined the room. | 19:55:59 | |
| 23 Dec 2023 | ||
 bernhard42 | I tried some more sophisticated open cascade question. It is amazing which API calls LLMs invent in order to answer. I always agreed with the LLM that it would be awesome if the API method would exist, but in reality it simply doesn't 😂 | 11:54:04 |
| bernhard42 | If you ask for some stuff where a ton of examples exists in the Internet, it works fine. The Websocket server in OCP CAD Viewer is based on an LLM response and also the Python client. Both worked immediately | 11:57:04 |
 medicationforall |  Download 2023_projects.png | 12:13:22 |
| medicationforall | All my 2023 stuff, written in cadquery | 12:13:22 |
 Karim joined the room. | 12:29:16 | |
| aflofo | I'm open to aggregating a large dataset of CQ codes and refining Llama if others want to collaborate on this | 20:46:33 |
| aflofo | As Jojain stated another cool approach is training from the drawing level to CAD and there is actually pretty cool literature for that already | 20:50:59 |
| aflofo | https://www.youtube.com/watch?v=Qm7ND25eMiw | 20:51:00 |
| 25 Dec 2023 | ||
 shehabattia96 joined the room. | 04:14:13 | |
| 26 Dec 2023 | ||
 halogenix joined the room. | 10:43:33 | |