Godot Engine suffering from lots of "AI slop" code submissions https://www.gamingonlinux.com/2026/02/godot-engine-suffering-from-lots-of-ai-slop-code-submissions/
Timeline
Post
Remote status
Context
1@gamingonlinux@mastodon.social An experience programmer can see if the submission is AI slop within 30 seconds. A simple solution would be to have AI slop submissions be a zero tolerance permanent ban rule from the repository.
Replies
23
@SuperDicq @gamingonlinux @feld you submitted a change to a project using ai, didnt you? how did that go?
@mischievoustomato @gamingonlinux @SuperDicq some people will be able to quickly spot AI-slop
Most people will not be able to spot AI generated code that had the slop-patterns manually removed by the developer submitting it. Because it just looks like normal code.
Most people will not be able to spot AI generated code that had the slop-patterns manually removed by the developer submitting it. Because it just looks like normal code.
@mischievoustomato @SuperDicq @gamingonlinux also if you prompt the model to read other code and copy their code style -- that works great. They're absolutely clueless that anything was generated.
@feld@friedcheese.us @mischievoustomato@tsundere.love @gamingonlinux@mastodon.social If the developer edited the code that means they probably also caught the bugs which means that the code has at least already been checked by a human so it's probably fine to merge it.
I know it's trendy to have a full anti-LLM policy, but honestly it is impossible to prove if the LLM output is edited by a human.
@feld@friedcheese.us @mischievoustomato@tsundere.love @gamingonlinux@mastodon.social If you generate code with an LLM and then manually edit that also means a non-significant amount of work went into the submission and that the contributor has probably learned something and actually understands what it does.
The whole point of vibe coding is doing as little effort as possible and that's really what you want to prevent.
@feld@friedcheese.us @mischievoustomato@tsundere.love @gamingonlinux@mastodon.social Just to be clear I'm not full dogmatic anti-LLM. I can see that this technology has potential to be useful if used right.
I just really wish people stopped using proprietary LLMs. I personally see no problem with running LLMs locally using free software.
@SuperDicq @gamingonlinux @mischievoustomato
> I personally see no problem with running LLMs locally using free software.
sure that would be great if it was possible, but the size of the models required to get good results are too big to run on consumer hardware right now. We just aren't there yet.
> I personally see no problem with running LLMs locally using free software.
sure that would be great if it was possible, but the size of the models required to get good results are too big to run on consumer hardware right now. We just aren't there yet.
@feld@friedcheese.us @gamingonlinux@mastodon.social @mischievoustomato@tsundere.love Yes indeed. We aren't there yet, especially also in terms of free training datasets and stuff like that. But this day will come and we should strive for this. Only a matter of time.
@SuperDicq @gamingonlinux @mischievoustomato there are advancements coming that will crunch down the required hardware to run a large model too. I've seen one WIP inference engine for this. I have hope.
@feld@friedcheese.us @gamingonlinux@mastodon.social @mischievoustomato@tsundere.love Also I think a good solution is to just make smaller models. Why not just make a model that's good at one specific programming language for example? Why does it need all the knowledge in the world?
@SuperDicq @gamingonlinux @mischievoustomato this is true and I think that might be where we are heading. Just swap models. Working on frontend now? Swap in frontend model. etc etc
@feld@friedcheese.us @gamingonlinux@mastodon.social @mischievoustomato@tsundere.love Yeah I think that's also the biggest issue that these large proprietary LLM provider companies haven't really figured out yet.
In their blind chase towards AGI they really aim to make one single model that can do everything perfectly consuming so much power and data it has already gotten way past comical.
It would be much more productive for them as well to focus on making smaller models that have a very good domain specific dataset.
@SuperDicq @gamingonlinux @feld @mischievoustomato same reason a programmer who knows multiple languages (deeply) will be better at writing good code in each one. problem solving is general rather than particular, even if you're comparing vastly different toolsets
@HatkeshiatorTND@annihilation.social @gamingonlinux@mastodon.social @feld@friedcheese.us @mischievoustomato@tsundere.love Ever heard of the phrase "jack of all trades, master of none"? At some point there's gotta be diminishing returns on adding more data to the training data set.
@SuperDicq @gamingonlinux @feld @mischievoustomato there's good 2-3B open weight models that should run fairly well on most non-ancient machines. try one of those and tell me if they're good enough.
on a related note, i've been daydreaming for about a month of making a prose-only, en_US (1700-1900)-only dataset pruned from public domain datasets currently on huggingface. i've been trying and failing to figure out where to start but if i'm successful, that should create a very focused dataset for conversational and creative work. is that close to what you were asking?
on a related note, i've been daydreaming for about a month of making a prose-only, en_US (1700-1900)-only dataset pruned from public domain datasets currently on huggingface. i've been trying and failing to figure out where to start but if i'm successful, that should create a very focused dataset for conversational and creative work. is that close to what you were asking?
@HatkeshiatorTND @gamingonlinux @SuperDicq @mischievoustomato idk if this will help, but maybe???
https://github.com/stealthwater/model_tools
https://github.com/stealthwater/model_tools
@feld @gamingonlinux @SuperDicq @mischievoustomato there's also three open weight models now that are frontier level, kimi 2.5, glm-5 and qwen 3.5. they can be run by anyone who has the hardware.
@lain @gamingonlinux @SuperDicq @mischievoustomato can you define what these hardware requirements look like though?
@feld @SuperDicq @gamingonlinux @mischievoustomato (if you want to run it at home, that is. there's plenty of services that run it for you: https://openrouter.ai/z-ai/glm-5)
@lain @feld @gamingonlinux @SuperDicq @mischievoustomato no one has the hardware to run these models are 8-bit quants, and anyone who does is just gonna use Opus.
Local fags are so BTFO.
Local fags are so BTFO.