There appears to be a few bad actors in the Lemmy community spreading ragebait for personal gain
In fact, the same mechanisms added at the beginning to catch bad human written code are the exact same that catch bad AI code
- No, the Linux kernel DOES NOT use AI to automatically approve code for inclusion. As far as I can tell, there are no plans to do so
- The Kernel actually has a huge number of maintainers. As part of the process, all patches are reviewed by a human who provide feedback as part of the process, and patches get bounced back and forth multiple time.
- Patches don’t just magically go straight to master. The maintainers generally have their own branches where it goes to, then it goes to staging trees like Linux-Next. It goes through many eyes, and many tests by the end. The idea that AI code is overrunning the kernel is rediculous
- Maintainers also check that the code meets coding guidelines and is written properly, and solves the problem they are trying to solve. Yes, they also ensure the code isn’t slop
- LLM’s though might be used to identify obvious problems with the code (which they are VERY good at), or even identify code that doesn’t follow coding guidelines.
- The only thing that matters is the human who submitted the code, as they should be checking the code submitted anyway (including LLM generated code)
Writing bad code isn’t something new, and isn’t unique to LLM’s. Coders (and it seems some “journalists”) have been writing slop since computers were invented.
The process hasn’t gotten less stringent (probably even MORE stringent as there is an additional set of robotic eyes that can also check), and the same mechanisms that were put in place to catch badly written code 10 years ago, are still there now… With no changes.
I agree with you that the Kernel isn’t doing slop. But there are a few points I want to address, maybe even disagree. First let me say that people have different definition of what slop means. That is why we don’t always agree on statements involving slop.
True, but its still not enough to handle all slop. One problem is, its not always apparent at first glance. And more code is generated than ever before, surpassing the capabilities, time and nerves of all maintainers. That means they start using Ai to combat the amount. And that’s an issue on itself.
This might be true, but that is not a reason to accept or normalize slop. It does not matter how much slop was before, we always try to minimize it. The difference to before is, that nowadays its easier and faster to write (or generate) slop in mass that looks legit. Finding out its bad code or wrong can suck up huge amount of time of maintainers, which would have worked on something more productive instead.
For you maybe, but there are issues using Ai regardless of how the code looks. And I don’t even talk about the amount of code it can generate and outpace every maintainer capacity. I mean stuff like how the LLM was trained on in example, and if the licensing will become an issue one day.
While the Linux Kernel operates very securely and in a controlled manner, that we cannot call it slop (yet). But there are inherent problems using Ai that are unsolved.
There is no policy that will stop that. It’s happening whether you want it to or not.
What every project needs is a way to handle submissions that ensures a level of quality. The kernel has this. Also needing to police whether code was llm generated would be an increased burden.
Yeah, this was the dumbest point OP made IMO, there are non-experienced people but they don’t become Kernel maintainers or contributors. He’s admitting LLMs slops this up and LLMs are yet Kernel maintainers and contribitors.
Well put, thank you :)
There’s always been code that has been bad, like Linus never liked the Con Kalivas performance patches, or Reiser4.
It is also finding a lot of security issues and bugs faster too though
I’ve actually written optimisations for code that I thought would be faster, and turned out to be a waste of time in production
What I would like to see is a LLM that’s developed exclusively using Open source compatibility licences with FULL traceability to sources uses, and a way to specify licencing allowed (or something similar).
There is one issue with AI that you didn’t mention however. The biggest problem with AI is that it adversely benefits closed source companies via Open source code. Because closed source code isn’t easily accessed by the AI for training, but Open source is. The other issue is that hackers won’t avoid LLM’s based on the ethical issues, so, unfortunately, it forces developers to use them to keep up with the cat and mouse game.
As a developer, am I happy about AI? For code review yes, but, it makes my job suck and not fun. And no, I’m not happy about the ethics… There are a lot of things that sh** me, like AI companies SHOULD be forced to pay royalties to everyone they stole info from.
However, what sucks, is that we can’t pretend it’s not a thing, and if Linux wants to remain competitive, it does need to use it.
The best thing we can do though, is petition to tax AI companies and pay producers of the information. That would make AI unprofitable immediately and kill the industry
Nobody ever remembers the environmental issues of using AI, like climate change and displacement of poor people. And that’s happening inside the Evil Empire itself, imagine what will do to the rainforest and other ecosystems.
Everyone knows about that… But, that doesn’t have anything to do with “AI Slop”. that comes down to ethical factors and is unrelated to the quality of the code
Funny that someone thinks it’s acceptable to compartimentalize the main issue with AI usage while being trigger happy to label outcry as bad actors.
That’s not really possible. I mean, you can list the training data, but you can’t attribute a specific output to a specific piece of training data, because all of the training data influences all of the weights. It’s not possible to know which training data influenced which weights and how, because that data is not retained (and it would be massive and difficult to interpret if it was -on the order of weights × number of training context windows, and the weights alone are already terabytes)
Besides all those Ai users, where code is uploaded to their servers.
On commercial accounts, you can turn off training, and they have policies generally to avoid code leaks (another bad thing for open source, the free accounts you generally can’t, because you are the product).
Greg KH (the second in command of the Linux kernel, below Linus) in a recent talk said anything uploaded to a corporate AI is going to be used for training, and to assume anything uploaded or pasted into it is insecure and certainly not private.
If you believe that they’re not using corporate code to train their models… These guys are all total lying sociopaths who think they’re on a mission from the machine god, everything is permissible in their minds.
I work for a HUGE organization that uses all the big AI companies products. We have a contract for each that specifies that they MUST NOT train their models with our code.
If they do it anyway, my employer’s vast teams of lawyers will make them take out brand new billion dollar loans to pay off the settlements. They might not even survive such a lawsuit because if they were caught using our code you can bet your ass they used everyone else’s that has a similar contract.
I think in a sane world that would be the outcome. But this world (and particularly the AI cultists) is not sane.
I also worked at a couple of huge corporations (which are both major hyperscalers) and they tried to steal ip from each other pretty much continuously, that was one of the reasons that you had to keep switching companies to advance your career.
Look at how open AI straight expected candidates to bring development models of upcoming Apple devices to interviews. These dudes are more degenerate than you think, I promise.