<kintel>
Someone from https://usestash.com, who are familiar with OpenSCAD, reached out and offered us a free subscription. Looks benign enough. Any opinions on whether we should test it out?
<InPhase>
It's an interesting notion if it's accurate.
<InPhase>
Am I understanding that it like, adds a post to new issues or something?
<InPhase>
And if yes, how do we address errors? I assume even if good, errors won't be zero.
Smeef has joined #openscad
mmu_man has quit [Ping timeout: 252 seconds]
snaked has joined #openscad
J25K70 has joined #openscad
J25K has quit [Ping timeout: 250 seconds]
J25K70 has quit [Quit: Client closed]
J25K70 has joined #openscad
kintel has quit [Quit: My Mac has gone to sleep. ZZZzzz…]
krushia has quit [Quit: Konversation germinated!]
teepee has quit [Remote host closed the connection]
<nihil>
teepee: I'm sure both of the blind people who use OpenSCAD already have screen readers and if desired, an AI assistant that functions for their whole environment, not just one app
<nihil>
give it an API to make such integrations easy sure, but adding it into the application itself is feature-creep suicide
<nihil>
I've gone through the motions with helping make computing accessible for disabled people, and that's something that needs to be done at base OS level
<nihil>
I've also played with AI enough, in the context of having it try to generate openscad code, that it's going to be a fruitless time-suck that will yield questionable results even on a good day
<nihil>
I'm not against AI in and of itself
<nihil>
but I'm very against the widespread FOMO causing software devs to want to integrate it into their one application, when it's just going to fragment the ecosystem and not be nearly as useful as they're hoping
<nihil>
dealing with this a lot at work lately
<nihil>
users think it's going to magic their work to get done, and it's causing more problems than it even pretends to solve
<nihil>
and the ridiculous amount of companies competing to be the model these applications use is absurd
<nihil>
openai got it right, let the user pick their model and just provide the interface
kintel has joined #openscad
<kintel>
I think there is useful stuff we could do with AI an OpenSCAD, but I think it's outside of the domain of a chatbot sidebar where you can only copy-paste stuff
<kintel>
e.g. bootstrap some design idea, read the rendered output and have an AI iterate a few times if something looks off
<kintel>
refactor code
<kintel>
explain code
<kintel>
find common bugs, e.g. also by reading the rendered output
<kintel>
Perhaps building an internal MCP-like server hand have an AI call into it for common operations could be a good starting point.
<kintel>
-> lots of work of course, not a quick PR
<kintel>
smart auto-complete may be the lowest hanging fruit, especially if we could ingest a sizeable body of good code as context
<kintel>
We could also standardize some AI readable "design intent" as code comments, and let the AI compare that to the actual code and suggest changes
<kintel>
It feels a bit like all of that should ideally belong elsewhere, and OpenSCAD simply becomes yet another MCP service ;)
<kintel>
InPhase Yeah, I've seen interesting results for things like image-to-3D using gaussian splats. There is an opportunity for image to scad for sure. Not particularly trivial of course.
<kintel>
..but if solved, the opportunity for image-to-CAD in general is huge. But that's for someone else to take on :)
J25K70 has quit [Quit: Client closed]
<nihil>
kintel: agreed
J25K has joined #openscad
<nihil>
I'm not saying AI is horrible, just that it's overused and is best positioned in a more generic wide scope (and entirely optional) context
<kintel>
yup, so I think the best bang for the bucks would be to wrap OpenSCAD in an MCP server and let any existing AI automatically be able to use it
kintel has quit [Quit: My Mac has gone to sleep. ZZZzzz…]
axelrod has quit [Ping timeout: 256 seconds]
<J25K>
so the AI will use the png output like humans - or is there an llm understanding 3d files
howiemnt has quit [Read error: Connection reset by peer]
howiemnt has joined #openscad
sculptor has joined #openscad
J25K has quit [Quit: Client closed]
J25K has joined #openscad
teepee_ has joined #openscad
teepee has quit [Ping timeout: 272 seconds]
teepee_ is now known as teepee
L29Ah has quit [Ping timeout: 260 seconds]
<teepee>
nihil: fragmenting what? it's just a simple gui addition calling a rest endpoint, that's it, someone already wrote it
<teepee>
making it more open and better integrated would make sense, but even that should not be a huge amount of code
<teepee>
we have the 3d printing service to which is probably not used a lot, I still think it's fine to have it
t4nk_freenode is now known as t4nk_fn
<nihil>
fragmented as in you end up with a dozen different AI implementations trying to do a dozen different things, none of it ends up working very well so it becomes stale code that gets dragged along in the codebase for however long
<nihil>
often times that sort of feature creep just ends up as bloat that drains performance without offering anything truly useful or tangible
<teepee>
the core logic of this exactly one curl call, you can do it in one line shell script
<nihil>
it's the buzzword of 2025 so everybody puts AI in the apps, then it'll eventually get purged around 2029 or so
<teepee>
we can drop it in 2029 then, it's not a problem like taking out code from the language
<nihil>
sounds like it's just a chat pane and not inserting generated code or doing its own rendering?
<teepee>
at this point it's just that
<nihil>
may as well leave it to thr browser at that point
<nihil>
s/thr/the/
<teepee>
as said before a slightly better integration would make sense
<nihil>
as long as it can be disabled, that's my only plea
<teepee>
I would not mind having even a dedicated switch to make it show up in gui at all
<teepee>
my personal requirement is only that it must support self hosted stuff
<teepee>
just like I wanted to have octoprint as local print service so we don't have only remote paid services
<nihil>
yeah, I keep a local ollama instance for things like that
<teepee>
same here, I would not like an openai only way
<nihil>
mostly for the HA voice assistant
<nihil>
well that, and getting the kid prepared to live in a world where you'll need to know how to interact with it
<nihil>
speaking of, I have had him making simple models in openscad
<nihil>
he loves watching me do it, so I taught him some basic create object and bool stuff
<teepee>
cool, what age?
<nihil>
just turned 10
<teepee>
ah, nice, starting early
<nihil>
yeah trying to get him prepared to live in a fucky technological world the likes of which I never saw coming when I was his age
<teepee>
true, the world is not really ready for the technology we have now
<nihil>
he has already managed to circumvent the youtube block on his school chromebook
<nihil>
one of their learning programs lets you find and embed media
<nihil>
so he just embeds it there
<teepee>
good solution, simple, even following all rules :)
<nihil>
I dont think they had 10 hour loops of maxwell cat in mind when they implemented that feature
<nihil>
still not entirely sure how he's searching it
<nihil>
should probably get him to show me at some point