Add rules on AI generated code

[Ratkiller446]

I wonder if it's okay to submit AI-generated code.

[vaguinerg]

if it works who cares

[neuralpain]

if it works who cares

"I got a hole in my pants and you can kinda see my underwear, but it looks good so who cares"

@Ratkiller446 Better to thoroughly review the code before submitting anything rather than outright trusting AI generated output. The AI will never fully understand what you want; and that "if it works" mentality should not be encouraged here.

[ToneyFoxxy]

if it works who cares

"I got a hole in my pants and you can kinda see my underwear, but it looks good so who cares"

@Ratkiller446 Better to thoroughly review the code before submitting anything rather than outright trusting AI generated output. The AI will never fully understand what you want; and that "if it works" mentality should not be encouraged here.

Indeed, I have a feeling "If it works" is how X11 got so mangled up to begin with. Medux has been trying to un-fuck the codebase so that he can actually get work done

[dec05eba]

You can use AI tools to find bugs, but verify if the code is indeed incorrect yourself

[metux]

Using AI as an assistant, eg. finding bugs, giving hints what you should look out for, etc, is a really good thing. Doing this myself.
Of course, the AI can't write actual (good) Xserver code.

[Soccera1]

Using AI as an assistant, eg. finding bugs, giving hints what you should look out for, etc, is a really good thing. Doing this myself.
Of course, the AI can't write actual (good) Xserver code.

I suspect Gemini 3 Pro will have a large enough context window (llama 4 would be great if it was a good model) to be able to implement some things. Gemini 2.5 already has a 1M context window which is not enough for XLibre, but I'm fairly sure 3 will have more.

Disclaimer: this is purely speculation from the roughly two years I've been an AI enthusiast but I do not work in the sector, nor do I have any inside info.

[svin24]

IBM 1979:

I don't think it would hurt in repetitive,effectively boilerplate code, but do make sure to actually read and understand what it does.
It's a tool like many others and it so far has kept improving but it can fail.

I think a policy where it is allowed, but one needs to understand what it is doing, and clarifies that it was used and how it was used, makes most sense

[fredvs]

Yes but IBM said also that Wayland is the future.

[svin24]

@fredvs IBM much like all the other major old and new companies quite literally do not care about desktop Linux.
That is: Unless IBM got in the business of making embedded devices like car media systems(which sometimes use wayland)
In addition to that I believe IBM still serves their own enterprise Unix offering which I don't think ships wayland at all.(I don't think it ships any graphics these days tbh)
If there are people in IBM who care enough about the Linux desktop to push for wayland, then it's people like you and me.

[reaperx7]

If you use AI for any code, make sure you double check it yourself, and if necessary, have someone pull your code to re-check it.

I'm not against A.I. being used to assist with anything, but A.I. has fallacies and it's best to always maintain a human standard for everything.

Also IF you do any generated code, there is a problem we might face with the license. Code has to be owned to be licensed in the first place, and written by someone. Grok, for example, is not a person nor a sentient entity. Grok can not claim primary ownership to license the code under GPL, BSDL, MIT, CDDL, etc.

[smj-cc]

Very graphic in the end

I stole that phrase from someone who was describing gas operated repeating arms. Made an impression on me.

[metux]

Sorry, I did not mean to imply "legal" rights, but rather the principle that spawned those rights. Who knows what the vestigial "courts" will actually do. The only plan there is a "war chest" and council on retainer.

Average judges btw should be replaced by LLMs, they're already better. And when we're already saving billions by throwing out the robe perpetrators (in costs as well as the damage they can't cause anymore), we can easily put each trial that actually counts something (eg. all criminal trials, and everything where either the govt or a large corporation going against SMEs or individuals) on grand jury.

Large Language Models, are able to do what they do by being trained on data largely scraped from the internet at large. This data is presently almost entirely generated by human beings, who in principle still retain ownership of the fruits of their labor.

Sure. And where's the fundamental difference to humans who're just reading a lot of other people's work (books, code, whatever you get in the internet) - except that LLMs just scale better ?

This fact has caused many to speculate that LLMs may be in their heyday, and about to collapse, because:

They take from the works of those they train on, to produce product which competes with the data they train on, thereby reducing any incentive for that data to be produced by humans in future, thereby reducing both the quantity and quality of production of it's own training data.

They're actually competing with the bio robots. Just seeing it live (in my family) in a different profession: medicine.
Most MDs, even the "specialists" are just are just following dumb procedures, w/o actually thinking through the individual cases any average plumber shows more creativity. Diagnostic geniuses like Dr.House are more or less fictional (there are some rare exceptions).

They then provide their product to the public via the internet, thereby polluting their own training data. In other words, it shits where it eats.

Welcome to the "scientific community" - they're doing exactly the same (eg. with their quotation cartels).

The actual disruptive change created by LLMs is making the bio robots obsolete and visible to all. Those now have the choice between finally starting to actually use their brain (and heart - that's what LLMs don't have) or falling into the abyss of meaninglessness.

[smj-cc]

Average judges btw should be replaced by LLMs, they're already better. And when we're already saving billions by throwing out the robe perpetrators (in costs as well as the damage they can't cause anymore), we can easily put each trial that actually counts something (eg. all criminal trials, and everything where either the govt or a large corporation going against SMEs or individuals) on grand jury.

Could work... but:

Legal systems only exist as an alternative to violence, which is the historic default for human interaction. Courts only work when otherwise potentially violent people trust and respect court decisions and abide by them in lieu of violence, or when such people feel overwhelmed by the greater violence of the state backing it's courts (when such people are a clear minority of the violence capable)

Trust is earned. LLMs still have some work to do there. Time will tell.

It's hard (for me) not to notice the differences between El Salvador and Brazil, as a result of how differently they managed their judges. As always, everything depends on what dangerous men are willing to do... Time will tell.

Sure. And where's the fundamental difference to humans who're just reading a lot of other people's work (books, code, whatever you get in the internet) - except that LLMs just scale better ?

The "NPC" meme would suggest that, at least for some, there is no difference. I suspect that there is a difference, at least for some. I could be wrong. I think this issue is traditionally called "free will" or "agency". I believe that "agency" can be abdicated, but that we start with "free will". True or not, such a belief is prerequisite to (the illusion of?) creativity.

They're actually competing with the bio robots. Just seeing it live (in my family) in a different profession: medicine. Most MDs, even the "specialists" are just are just following dumb procedures, w/o actually thinking through the individual cases any average plumber shows more creativity. Diagnostic geniuses like Dr.House are more or less fictional (there are some rare exceptions).

Agreed. Again, "most", not all. Ok, in the MD example, perhaps all, as the exceptions have been excised. When people meet punishment for being creative, those who cannot choose to stop being creative, go "galt"... which makes their numbers difficult to assess. This is the same problem the soviets had: punish those who produce, and those who can't stop producing will either run away, or drink until they learn how to stop producing.

We have some here who have chosen to impale themselves on the issue, and are in court for faking medical records. One such appears to be winning... I'm just jaundiced enough to suspect it's just another of "Lucy's footballs". Time will tell.

thereby polluting their own training data

Welcome to the "scientific community" - they're doing exactly the same (eg. with their quotation cartels).

Trust is now so thoroughly violated in all "fields", that there is no longer such a thing as "authority". It's an overt attempt to destroy truth itself, and they've about sold the public on the idea that truth is subjective... that reality is a simulation (as if that changed anything), and that emotion is a valid measure of fact. ("All data is fake" -Adams) They are at war with reality. Trust takes a long time to earn, and has been lost in a heartbeat.

The actual disruptive change created by LLMs is making the bio robots obsolete and visible to all. Those now have the choice between finally starting to actually use their brain (and heart - that's what LLMs don't have) or falling into the abyss of meaninglessness.

Very Interesting point... It's hard to get "agency" back after abdication, once it's become a comfortable habit, but even LLMs (I expect other methods of "AI") are an easy replacement for the NPC meat robots. They will force change.

I see I have not had enough attention on exactly how this is going to play out. Many of the bio-robots are actually violence capable. Without agency, they can't initiate, but they're not hard to organize and manipulate, and they're going to be hungry if they can't compete.

[pugly-tech]

Average judges btw should be replaced by LLMs, they're already better. And when we're already saving billions by throwing out the robe perpetrators (in costs as well as the damage they can't cause anymore), we can easily put each trial that actually counts something (eg. all criminal trials, and everything where either the govt or a large corporation going against SMEs or individuals) on grand jury.

Could work... but:

Legal systems only exist as an alternative to violence, which is the historic default for human interaction. Courts only work when otherwise potentially violent people trust and respect court decisions and abide by them in lieu of violence, or when such people feel overwhelmed by the greater violence of the state backing it's courts (when such people are a clear minority of the violence capable)

Trust is earned. LLMs still have some work to do there. Time will tell.

It's hard (for me) not to notice the differences between El Salvador and Brazil, as a result of how differently they managed their judges. As always, everything depends on what dangerous men are willing to do... Time will tell.

Sure. And where's the fundamental difference to humans who're just reading a lot of other people's work (books, code, whatever you get in the internet) - except that LLMs just scale better ?

The "NPC" meme would suggest that, at least for some, there is no difference. I suspect that there is a difference, at least for some. I could be wrong. I think this issue is traditionally called "free will" or "agency". I believe that "agency" can be abdicated, but that we start with "free will". True or not, such a belief is prerequisite to (the illusion of?) creativity.

They're actually competing with the bio robots. Just seeing it live (in my family) in a different profession: medicine. Most MDs, even the "specialists" are just are just following dumb procedures, w/o actually thinking through the individual cases any average plumber shows more creativity. Diagnostic geniuses like Dr.House are more or less fictional (there are some rare exceptions).

Agreed. Again, "most", not all. Ok, in the MD example, perhaps all, as the exceptions have been excised. When people meet punishment for being creative, those who cannot choose to stop being creative, go "galt"... which makes their numbers difficult to assess. This is the same problem the soviets had: punish those who produce, and those who can't stop producing will either run away, or drink until they learn how to stop producing.

We have some here who have chosen to impale themselves on the issue, and are in court for faking medical records. One such appears to be winning... I'm just jaundiced enough to suspect it's just another of "Lucy's footballs". Time will tell.

thereby polluting their own training data

Welcome to the "scientific community" - they're doing exactly the same (eg. with their quotation cartels).

Trust is now so thoroughly violated in all "fields", that there is no longer such a thing as "authority". It's an overt attempt to destroy truth itself, and they've about sold the public on the idea that truth is subjective... that reality is a simulation (as if that changed anything), and that emotion is a valid measure of fact. ("All data is fake" -Adams) They are at war with reality. Trust takes a long time to earn, and has been lost in a heartbeat.

The actual disruptive change created by LLMs is making the bio robots obsolete and visible to all. Those now have the choice between finally starting to actually use their brain (and heart - that's what LLMs don't have) or falling into the abyss of meaninglessness.

Very Interesting point... It's hard to get "agency" back after abdication, once it's become a comfortable habit, but even LLMs (I expect other methods of "AI") are an easy replacement for the NPC meat robots. They will force change.

I see I have not had enough attention on exactly how this is going to play out. Many of the bio-robots are actually violence capable. Without agency, they can't initiate, but they're not hard to organize and manipulate, and they're going to be hungry if they can't compete.

Replacing judges with AI and rebuilding the legal system unfortunately fall outside of the scope of the Xlibre project

[callmetango]

fall outside of the scope of the Xlibre project

Best answer! I would also like to suggest to focus on the XLibre project.

[callmetango]

Thank you for your contribution! We currently restructured the "Ideas" discussions and accordingly this discussion will be moved to the X11Libre 4 Good Ideas For Later category.

[metux]

Just had a few experiments with chatgpt:
It helped me creating a little script for preparing/submitting PRs from outside an incubator branch:
https://github.com/X11Libre/misc/blob/master/scripts/x-make-pr.sh

Lessons learned:
a) it can be a nice assistant for writing little helpers (that don't matter so much in the greater context)
b) it does create lots of bugs (you really have to carefully read the generated code!)
c) it can be a tedious process to tell it what you exactly want - often it forgetting things again that it already did before
d) similar situation like with classic rapid prototyping tools

So before merging AI generated code, we need to make sure:
a) the code is really good (usually needs manual fixing) --> careful review
b) the used AI service needs to grant you full copyright, so you really own the output (chatgpt does that, no idea about others)

[callmetango]

It would be good to state this in https://github.com/X11Libre/xserver/blob/25c002c54b58a81518f7cac6ab6dffa52e53ee56/CONTRIBUTING.md . And require contributors to also state whenever parts of their contribution was generated by some machines so care can be taken.

[Soccera1]

For XLibre it should be fine since it's far enough away from any nonfree code out there as afaik all modern nonfree display servers are no where close architecturally to X11, but sometimes AIs will yoink nonfree code from leaked or source available repos and not tell you. You therefore can't be certain that you actually own the copyright, even if the AI company tells you that you do, as they might not legally have the rights to do that.

[pugly-tech]

AI generated code is dangerous because it looks fine if you do a quick overview but can contain dangerous bugs that require someoen going through exactly how the code operates to catch - and if there is not human actually typing the code then you don't get that oversight. It's fine for doing rough layouts and basic shit like a to-do web app, but it should never touch fundamental system code

https://xcancel.com/spendergrsec/status/1958265452561613236#m

You too can crash today's 6.12.43 LTS kernel thanks to a stable maintainer's AI slop. All you need is CAP_SYS_RESOURCE, modern systemd, and this:

40 b7 40 c1 e7 18 83 ef 08 57 57 31 ff 40 b7 07 31 c0 b0 a0 48 89 e6 0f 05 5e ff ce 31 ff b0 21 0f 05

Look at all this extra space!

the nr_open it just set. That will fail immediately regardless of privilege level in setrlimit at: git.kernel.org/pub/scm/linux…

Next, it tries to dup2 with a fd higher than the resource limit it just set, that too will fail immediately early in dup2 regardless of privilege level without calling alloc_fdtable or doing any allocations here: git.kernel.org/pub/scm/linux…

It really looks like an LLM was just pointed at: git.kernel.org/pub/scm/linux… and it failed to understand the system limitations and the relationship between the different values, but spit something out that "looked" similar, complete with the dup'd fd being offset by 64 from nr_open

But the important point is that this was supposedly code to reproduce a specific bug, a real one I had already tweeted about in January. But the bogus test tests nothing at all, it will simply fail 100% of the time without ever testing what it claimed. Worse than that because the "developer" didn't go through a real process producing the code and commit, slapping a stable@ on it, they never looked at what would happen to that commit on older kernels. It patched a function (alloc_fdtable()) that was changed in 6.13 to return ERR_PTR values...

On earlier kernels, alloc_fdtable() returns NULL on all errors, and callers will dereference any non-NULL return value. So when it got cherry-picked clean despite the obvious difference visible in diff context, you end up with a vuln in a core syscall.

There's no point in reporting this stuff because the vulnerability is simple:

stable maintainers using AI who have no idea what they're doing and are using it as a substitute for the work everyone else assumes they've performed

It'll keep happening until that bug's fixed.

[smj-cc]

Heh, responses fault the use of "plagiarized", but it is a fair argument that that is exactly what LLMs do: Plagiarize their training data, without credit or attribution.

[callmetango]

The big question is: Is this legal after all? Maybe in the US given "fair use" rights, but even then... What if you use GPL-licensed code for training???

[smj-cc]

They use everything for training, and in such a way as to ensure nothing can be traced back to it's source. If you look at all modern LLM projects, there's been very little attempt to hide the software, most of it is open source. The training data, however, is a closely guarded and protected resource. There will be no transparency there.

Whether a thing is "legal" or not is completely moot, if either there is no meaningful penalty, or if there is no meaningful enforcement, or worse: meaningful enforcement is provably not possible. ("And there's a legal limit to the snow here" - Camelot) A more useful question might be "what can be done about it?".

Presently, AI is the "golden child" of both government and large corporations. Nothing will be permitted to hobble it. Therefore, "intellectual property" will only be enforceable for the properties of governments and large corporations.

I suppose I expect there to be some legal challenges, but they will get no traction, and will be allowed only to set the precedence that training data cannot be challenged... much the way internet service providers were once declared "common carriers", so that they could not be held responsible for "content", so long as they did not attempt to control content. And of course, later, when government and corporations decide they want to control content, "common carrier" goes "out the window" in courts.

If you are doing something that's completely illegal, but nobody cares... nothing will happen to you. If you are doing something that's completely legal, but upsets everyone around you... something will happen to you. Laws are "really more like guidelines" (-Barbossa).

[callmetango]

A quick search on the internet regarding using copyrighted material in pattern matching training data resulted in:

• Copyright Implications of the Use of Code Repositories to Train a Machine Learning Model — Free Software Foundation — Working together for free software
• Legal Corner: The threshold of originality for copyrightable source code - FSFE
• non-ai-licenses/non-ai-licenses: This repository contains software licenses that restrict software from being used in AI training datasets or AI technologies.

[metux]

Well, I've seen lots of copy+paste programming in the industry. Especially in areas where one really should know what one's doing.
And often they copy+paste the wrong thing, and not reading that somebody, two posts later, already told that this specific code is wrong, and a good example how not to do it. Modern LLMs are at least clever enough to read the whole read, so already better than those bio robots.

Also happened to myself on certain concepts: anybody recalling the great disaster on the German lawyers digital post system (kind of "secure" email, but entirely homegrown, quite central like ancient Lotus Notes, typical bullshit made in Germany (TM)):
I once had a meeting with those guys about that stuff (back when we made a concept study for the same thing for the notaries). I've told them certain things they really must not do. And guess what, they did exactly that. Shortly after several hundred millions were burned and it came online, some people from CCC quickly hacked it. The right timing for nice xmas present.

Or another example in medtech. Some of the famous corporations in that field (with lots of high gloss paper). One of the tasks was a little bit of CAN communication from a Linux host to some MCUs. Pretty simple, just use CAN socket, takes about 10..20 LoC, mostly boring average BSD socket interface. It took them several 10KLoC to create a huge machinery for internally queuing and distributing some small structs (those knowing the fundamentals of Unix and sockets would just use several sockets on the same interface - a trivial idea that those jerks didn't believe to work, even after I've presented them the manpage). And guess what: the few lines doing the actual CAN communication were wrong. Copy-pasted from some forum (char by char, including surrounding comments). And two posts later explained where these were wrong.

So now tell me, how would make LLMs the situation even worse than it already is ?

[smj-cc]

AI is different -antirez

[Trklntry]

It'd be good to avoid anything which "thinks" for oneself, to become and remain proficient we must think and act independently.