Hacker Newsnew | past | comments | ask | show | jobs | submit | cdr6934's commentslogin


Also added a little more information here: https://x.com/generatecoll/status/2018202734734741589?s=20


Hey HN,

I’ve been working on a project called Paramancer: think Claude Code, but for iterative 3D modeling.

Instead of opening Blender, Fusion, or OnShape and building a bracket or pegboard mount from scratch, the idea is to describe what you want in plain language and have an LLM generate a parametric 3D model you can iterate on.

You can then refine the design conversationally: adjust dimensions, add mounting holes, tweak tolerances, and regenerate until it’s right.

I’ve had some early success producing useful and interesting models, but the real challenge is scale. To get there, I’m building a feedback loop and dataset that captures how people describe parts, how models fail, and how they get corrected—so the system can steadily improve.

So far I’ve run ~100 different iterations by hand and built internal tooling to help refine and fine-tune the model, but I’d love to take this further with real users and real use cases.

I’m very early, but excited about where this could go. Would love feedback, criticism, and ideas from folks who design parts, print them, or just hate CAD UIs as much as I do.


My guess is we are either at the top or rising to the top of cyclical curve of the trend.


The ease of quantization in the DAW is pretty easy to do as well. So I am not sure that would be unique to music / live coding sessions.


It is unique because everything is quantized. I've never used these tools but I am assuming you could give it some level of randomness but as someone who has performed and recorded a non-quantized performance is not random. So sure, it's super easy to quantize in your daw but it is a tool to be applied when needed, not something that is on all the time by default.


yes exactly, and when I say "quantization of every dimension of composition" I mean an application of quantization to every aspect of composition not just pitch and rhythm.


Quantization and repetition are what some genres depend on. It won't be the right instrument for a Rock ballad, but for a Techno track you need this kind of "everything being quantized". That said, in loopmaster you can add swing and noise to the note offsets to humanize a sequence, a lot is left to the imagination and ability of the creator.


No one in this thread is saying quantization is never appropriate.


Currently working toward this at 38, but my goal is to start building a team of individuals to create a research / design firm that studies symbiotic relationships in nature in order to discover and pair natural additive processes (think spider producing webs as one of these additive process) starting with bespoke pieces such as a spider woven glove.

This would help create buzz and intrigue with the objective to attract top talent and essentially the seed money to self funded a hybrid medusa that is studying "organic 3D printers" with the objective of being the "Manhattan project" size of integrating nature into the manufacturing process.


The speed at which this is happening could be a masterful execution of getting out of under the non-profit status.


The corporate structure is so convoluted, OpenAI is only part non profit.



I wonder if those structures were compared to what AlphaGo output and found significant differences.


These aren't protein crystal structures, they are metal-organic frameworks (MOFs), so AlphaFold probably wouldn't work well on these ones.

It would be really interesting to see an equivalent model trained to predict these structures. The physical chemistry of transition metal complexes, especially when multiple metals are in close proximity to each other and connected by shared ligands, is much more complicated than proteins. The reason is because of multireference effects - essentially the quantum entanglement of multiple possible electron configurations. These are exceedingly difficult calculations to perform - common approximations are O(n^8) or worse and require highly specialized knowledge to apply correctly - so an ML model that can efficiently make predictions in this space would be a major transformative breakthrough.


A collection of links, tutorials and philosophical thoughts on generative art


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: