I did profiling. But, there were many other more likely candidates, and, its a bit of a weird edgecase in our case (since we don't control the entire environment.. It's middleware)..

A profiler ought to show you where it's spending the most time.

Just because an issue isn't reported though, doesn't mean it hasn't been noticed by users.

We often see issues which are discussed in private forums, but were never actually reported to us (which is why we check them all)

I think we need a redefinition of "bug" though. If something hasn't been noticed for 10+ years, is it really a "bug"? I think the definition should depend on the probability of negative effects, not just on the possibility.

In fact, my impression is that any piece of code can be shown to have some bug, if taken from a particular angle or in particular contexts or scenarios.

(I think there are also connections with mathematical-logical foundations. For example, it is impossible to build an axiomatic system that is enough powerful and does not have undecidable propositions. Whenever you fix the axioms to fix one of the undecidable propositions, another one automatically pops up.)

Dependencies, dependencies everywhere! https://xkcd.com/2347/

(well, maybe not exactly right, but similar problem.)

The funny thing is that I wrote a Websocket server library years ago, and recently (as we scaled up), there were reports of bad performance in much larger systems

AI actually basically called my code slop (and it was), and made changes that improved performance, and fixed issues with incomplete frames. The solution it came up with was far better, and far cleaner than mine. I never would have come up with it..

And, I never would have guessed that's where the performance issues were either

I feel like there are people who are willing to point the finger instantly when AI generates bad code, but, it doesn't get credit when it improves it

your ai will test it out

Yeah but
humanSlop = good.
aiSlop = bad.

LLMs have no business near the kernel code or being used at all.

Cool. Let us know when your fork is packaged and I'll test it out.

Was actually going to mention that.

There are also ambiguities in specifications too.

Like one of our libraries is a websocket/http server I developed, that has been working fine for years. However recently, a system broke, and we actually discovered it's because there are even small ambiguities in http areas too that we need to accomodate, and I couldn't find anything in the specs that explicitly prevented it either.

Exactly this.

I've been doing code review on some of our code from 10+ years ago and I'm surprised by how many major bugs I'm STILL finding (different ones between Claude and ChatGPT). And this is code that's been in prod for years, and in heavy usage.

Yeah lol also reading through the actually interaction around the patch

> The uptodate check in filemap_read_rcu_can_copy() has to stay, though.
> The flag only says the folio was uptodate when it was marked. Some
> filesystems clear PG_uptodate on a folio that stays in the page cache,
> and that can happen after the folio got the flag.

Those filesystems are buggy.  We have assertions that the uptodate flag
is set on folios mapped into userspace, and filesystems going arund
clearing the uptodate bit will cause problems.  It's something I'm
working on removing, but since it's all connected to writeback errors,
it takes a lot of arguing with each filesystem maintainer.

So one of the reasons there’s so many changes in the patch series is to account for… human bugs

People really be thinking every fully human-rolled PR is gonna be flawless, yet the influx of fresh CVEs is higher than it's ever been right now, because LLMs are busy poking holes in existing human code.

From my understanding, that was a patch series that was still under development, and NOT in the kernel. And, it was caught in Code reviews.

Code reviews are normal, and stuff like this is NORMAL, and has been well before AI. Are you seriously claiming the kernel never received buggy patch sets (in staging / non-master trees) in the past?

Also, it is interesting that the same 3 links seem to be posted everywhere blindly. Did you even read what I wrote in my post?

Do you mind explaining the technical reasons why the patch is "slop"? Or, did you just see the LLM part and think "this is gonna be great pitchfork material and I'll pretend like I actually understand whats going on"

So... your evidence for slop is that he made a patch, they caught an issue in the review process, and then he had to go on vacation so he asked another maintainer to take over to get the change done correctly in the merge window?

How is that not the process working exactly as intended?

Do it.

Just wait until Anthropic makes this against the TOS because it's impacting their friends bottom line.

I don't want to have to spend 5 grand and monthly fees to claud just so I can vibe code my own apps because no one makes them anymore.

I think that's the scary part. When all human maintainers of essential libraries burn out, how will software even be developed? Everything will end up being made by agents, and possibly companies like OpenAI and Anthropic will need to be the new library providers so that agents don't have to reinvent the wheel all the time.

Yeah, kinda sucks that our only option is paying the families of vampires. It would be awesome if the code trained from opensource must be mandatory opensource, i can see local model of LLM being offer by site that host opensource project that trained their model from said opensource, i have quite a few very small personal project i really wanted to do with it but for now i'm holding it off until this AI craze die down a bit. I don't mind it being slow, i don't have the time to do it anyway.

At least this way all those chip won't be horde by big tech.

Don't even bother. Between me getting a bigger phone and changing keyboard apps I'm only going to progressively get worse as the predictive text remembers more and more of my mistakes. That and I'm illiterate

usecase

use case

No. Phew.

I am tempted to use AI to try and create firmware for RokuTV devices and televisions.