It’s true. I’ve been using ChatGPT, Claude, Mistral, Perplexity and Deepseek(only 2 weeks) for a long time, but I’ve been using local Qwen2.5 32B for the past two days, and for me, it’s the best. It can write the code I want many times on the first try.
1 Like
I find chatgpt is decent if you know python but dont use it enough to remember the syntax. It is often a lot faster to ask the ai how the syntax would be than to dig through blenders python api or find it on stackexchange.
2 Likes
Yeah, so far its the one I’ve found to have the best results. Now if only it was fine-tuned to be a bit more Blender specific and hence compact and cover the latest API it would be a total killer.
But of course ‘main’ stream AI have little interest in doing that or even know that Blender exists.
I am surprised so many peopole are saying it doesnt work very well.
I have created some pretty complex scripts using Claude 3.7… as well as Claude Coder… If something is too repetitive, i script it up now.
1 Like
It’s getting better but it’s still just a tool. I have often had to ask an LLM to change the way it approaches a problem, even having to go as far as to tell it exactly what method to use.
Don’t get me wrong, it makes coding faster but i still have to proofread everything and fix minor issues. The most recent one i had to tell it (ChatGPT) 3 times to use list comprehension instead of if loops in python. I also have my own style of writing SQL command where i only capitalize tables, columns and rows in the commands but it still forgets them from time to time and i have to remind it again.
1 Like
It doesn’t work well at all if you can code well to begin with. In that context it’s way easier, faster and more reliable just to code yourself. If you cannot code then it makes it possible and you can overlook the time it takes to actually get usable results while it would be way faster and more reliable without AI. So yeah, it’s a great shortcut to skip learning, but it’s way worse in comparison.
2 Likes
i find ironic when chatgpt or other ai replies back to me “you are absolutely correct!” when i tell them about an obvious mistake in their answers 
3 Likes
Whilst some people work that way… i have been pretty much relying on it without proofreading. I give it suggestions and logic to help it out yes, but actual coding, i barely touch.
Whilst i dont classify myself as a coder, I am someone who has commits in blender, been writing python code for about 20 years and likes to think they can jump into code and start working.
2 Likes
Ill take a punt at some of my long standing blender bugs this week, see if claude coder can address them in a logical way 
2 Likes
Yeah 
Just yesterday i was asking ChatGPT about information about how to control the speed on an AC motor i salvaged from a trimmer and how the speed controller on the trimmer works (I am new to electronics) and then it started writing large paragraphs on how to control a DC motor on a battery powered trimmer (Mine uses or ‘used’ an extension cord).
Even when i corrected it and told it i was new to electronics, it listed some questionable ways like using capacitors to change the waveform.
I’ve been shocked multiple times in my life from the the mains and it is not something i would like to experience again, nor do i feel like dying. Funnily the most recent time i got shocked by the mains was last year in November when i was adding a wifi relay to my dads pool pump so he can set up a timed schedule for it using an app. Unplugged the pool and i was about to start cutting when i got shocked and the house tripped, my dad had wired the plug to the mains and the socket to the pool. I cussed a lot but was blamed for not knowing it was a ‘dead mans switch’, whatever the hell that is. Didn’t seem to matter and i redid the wiring properly so someone less lucky doesn’t die when they unplug the pool.
Sorry for the overshare, ADHD thing i guess.
1 Like
Grok 3 in thinking mode is really good at coding. It usually gets things right, but I can always ask for tweaks if I don’t like something or want to add more features
1 Like
You stand a high chance of dying a painful death if you’re playing around with industrial motors without knowing what you’re doing. The chance of death increases to approximately 99% if you’re relying on a LLM chatbot to keep you safe. Please stop this immediately
5 Likes
This isn’t meant to sound condescending, but asked out of concern for your lifespan: do you actually have any proper training as an electrician?
4 Likes
A simple check with a meter would prevent this stuff from happening.
1 Like
Web-sites for those who use “AI” to play with high-voltage electricity:
None of the AI models are good at Blender code. I regularly test all of them (except Grok at the moment) and daily review add-ons for extensions platform (which I maintain) that are more often than not written with AI. Claude is the worst, no competition, ChatGPT is the best, no competition (I know, I know, not cool enough narrative).
But the code generation will get better, they will get access to more data to train on, will catch up with API, etc, that’s not a problem. Why AI is horrible for coding is because it can’t design.
People who come fresh to coding don’t have understanding of design yet, and have very naive concept of how coding works: you have tasks and just execute them one by one. But it’s more complex than that. Design dictates quality, safety, speed, longevity of your code. You can take good paths towards the goal or bad ones. AI doesn’t have the capability for that reasoning, since it just bases it’s designs on what it finds, and what if finds is garbage, as most code are.
Take this example from one (now refactored) submission from extensions platform. This is my speculation of what happened, but at this point I’m versed in AI hallucinations so I’m confident of it:
-
User gave AI prompt: I want to be able to keep my objects as favorites, have a menu where my favorite objects are listed and when I click one of them it gets selected.
-
AI created operator that saves object names, created menu where all those object names are displayed, and second operator that selects object with given name, and attached it to those names in menu.
-
User tested and found issue, that if you rename object after you mark it as favorite link is broken, menu displays old name and operator can’t select it anymore. Asks AI for help (probably told it something like: “can you check when object is renamed and update everything?”).
-
What does AI do? It decides it needs to implement function that runs on depsgraph update (something that is triggered every time you do anything in Blender, so can be hundred times a second), and in this function iterate over objects of the file, and check if names of those match to those in favorites, if not, then something was renamed.
This results in code that is executed millions of times, requires more time more objects you have in the scene, and is prone to crashing and freezing. But as far as user is concerned, it does the job.
Now, how would adequate human design this?
- Instead of storing names of objects, just store objects themselves as pointers.
There we go. No need to execute any script, no need to keep track of anything, no performance costs, no risks, no manual clean-up needed, hundreds of lines of code avoided.
AI knows all about pointers, if you ask it to give you a script that stores objects as pointers it will do so. But, when given a prompt without such trigger words, when it’s left on its own to DESIGN the script, it outputs shit like that. It can’t design if its life depended on it, and I’m not sure that’s something it can get better at.
LLM is different tool in hands of someone who knows how to code and someone who doesn’t.
If you design the architecture of the script yourself AI becomes just a tool that types letters on the keyboard for you and saves you time (Github Copilot absolute madman in this).
If you ask AI to design the script for you based on your prompt you get shit in return, that at best will just be unoptimized, and at worst will corrupt or delete your data in blend file.
Take those two prompts.
-
Give me a Blender script that changes active object viewport display to Solid when user enters Edit Mode, and to Bounds when it enters Object more (or any other)
-
Give me a Blender script that registers
msgbuson object mode, and when it changes to Edit update viewport display to Solid, otherwise to Bounds.
First script will almost guaranteed will give you script that registers depsgraph handler. It will not be properly registered or unregistered, and will come with unnecessary stuff.
Second script will give you optimized, safe script, but in order to give that prompt you need to know about msgbus, about callbacks. You need to be able to plot/design the script before you write it.
That’s the issue. Out of 300-400 addons I’ve reviewed I would probably say 20-30 have good design. That’s where people are stuck and keep shouting at AI to solve it, but it wont. Most add-ons even thought they work, they’re unoptimized, filled with mind-boggling decision, and what is worst is that they’re only tuned for very specific usecase that prompt gave them. As soon as context changes (object is library-overriden or linked, scene is renamed, object has shape keys, object data is instanced, different OS, etc) code breaks.
“Issue” is that most of the users are absolutely fine with unoptimized, shitty codes that just work for specific purpose, they don’t care about bloat, visual clutter or any of that (BlenderMarkets existence is proof of that), so they don’t care about this point.
But if you plan to make an add-on that is dependable, and most importantly shared with others, you need to learn the code. You can use AI for it, step by step advancing, but you’ll need to know it yourself and it will take you a long time.
8 Likes
Does that included any of the open source LLM’s that one can run locally or just the commercial ones with a ‘free’ option?
And this is one of the reasons why I tend to avoid addons, you never know just how ‘good’ or reliable they are, how solid and well built or what any future support or updates there will be. Nothing worse then coming to depend on some addon, only to have it hit a wall at some point in production due to poor design/coding. Or face no more updates and have it totally break with future Blender releases.
This was also a reason why I pushed so hard when the extensions platform first went live and so much of then core Blender was going to be moved over. Keeping core ‘addons’ like Rigify and Node wrangler within Blender means it should always be well designed and updated just the same as Blender code is.
1 Like