L
> rather than relying on external (apparently non-free) services. Speaking as an emeritus Debian Developer, that's the part of this that saddens me the most, that Debian (which doesn't allow any non-DFSG-compliant software in the main section of its archive) relies on proprietary tools in such a central place.
L
> Everything useful that they found would have been caught by simply building (and then cleaning) the package in a deliberately minimal, non-network-connected VM, and, if the build is successful, noting whether any spurious artifacts remain. Quite possibly. But without this LLM review process being run, would you have thought of that as a possible improvement? And then comes the question: building such an automated test also takes time and effort, and has to be maintained. Are you going to…
L
> I read through some of the LLM generated reviews and they look absolutely reasonable Ok. But they seem to often be violating the actual debian policy they're supposed to help enforce. So being "absolutely reasonable" is not the goal here. The goal is to respect the debian policy.
L
> but the decision is still with a human being who presumably knows when to disregard the LLM, and/or who can improve the prompt. "presumably" is carrying a _lot_ of weight here. IME, the only "decision" being made is to outsource "decisions" to an LLM in the first place; once that happens the inevitable conclusion is "computer says no... <cough>". Taking humans out of the loop (usually so they can be laid off) is the entire *point* of LLM investors continuing to light money on fire.
L
Actually, I would argue it's (a) not uploading (it's a *response* to an upload), it's pushing back at the uploader, and (b) it's not "blindly accepting without human review" - it's actually *asking* for a human review. So while it might raise hackles, it's actually completely orthogonal to the GR wording, and it's the review team asking for help with their workload. Hopefully the uploader will come back and say "I've checked the points, and fixed the ones I think need fixing. Several of the…
L
The solution is simple. Require all files to have SPDX identifiers, then instead of an LLM you can run a purely automated scanner across the source code. If there are files without them, well, submit an upstream PR or carry a local patch. We did that kind of thing once already, working towards reproducible builds, so why not do it again? If that's not feasible … let me put it this way: as a reviewer I'd rather slog through a ten-point LLM-generated bullet list, even if six of these are…
L
If you have to review the prompt and update it, then you are training the LLM (poorly).
L
An AIs output always is the result of its training data and its context. While I don't know how Debian set this up, I am almost certain that they are not simply telling the AI "Review this package". I assume they provide the AI with proper context, e.g. requirements, rules, documentation, ... In this context you can easily address false findings. So no, if the AI context is properly managed, it will absolutely not continue flagging unwanted issues.
L
The Linux CNA is following the rules for the entire CVE system - anyone who's not doing what they're doing for projects they control is technically in breach of the rules of the system. This is the CVE system working exactly as designed, and exposing just how many security bugs Linux fixes in a typical release.
L
"Any sufficiently advanced incompetence is indistinguishable from malice." - Grey's Law. One might also allow the variation of "negligence" or "indifference" in place of "incompetence"...
L
The LLM does not reject the package. It finds issues which may or may not be relevant, but the decision is still with a human being who presumably knows when to disregard the LLM, and/or who can improve the prompt.
L
We've added a whitelist feature to kmod so that we can have all modules readily available but it needs an explicit administrative action to allow them to be loaded. There's an upstream PR, although I think I like our original out-of-tree variant more. https://github.com/kmod-project/kmod/pull/449
L
… is to revoke the CNA for the l’AInux people again. Or to declare bankruptcy for the entire CVE system.
L
… was an affront in the first place. It was apparently vibecoded, which was found out by accident only much later. The new DFSG team also was brilliant in all the communication it had with the previous team members, meaning about none…
L
This doesn't require training. It just requires review prompt instructions about how to properly check this.
L
I have made an attempt at a bug report, we will see. <https://salsa.debian.org/dfsg-team/dfsg-omp-review/-/work...>
L
> A human could easily get this wrong as well. But a human would learn. Then, the next time the issue came up, the reviewer would have the knowledge necessary not to reject the package. For an LLM, you can't have it both ways. Either they're training on the data we (debian developers uploading packages to NEW) are providing, or they're _not_ training on that data and the LLM will continue to incorrectly flag that as an issue.
L
> So that justifies adding even more noise? How about just ... not sending it to begin with? "The right not to see this" is a very dramatic way to ask for a minuscule improvement in email Signal/Noise Ratio. Nothing wrong with that option. Everything wrong with caring about it and trying to pretend LLMs don't exist when are turning the whole industry upside down.
L
Then the review can be adjusted - we've been going through this on my project, and the tool now does *better* than humans on a lot of stuff. For example, it now reviews all automated test failures and analyzes them - much more thoroughly than humans generally did. It's a godsend, and the persistent errors are correctable with adjustments in the instructions.
L
> LLMs, on the other hand, did not deserve the same consideration. ""We are right to complain that it is used to impose on our work and our attention."" Maintainers are volunteers. By submitting packages for review, you are imposing on their time and attention. And if they choose to prioritise packages where the submitters engage with the process, rather than scream "foul", that's their decision. If I want somebody to do something for me, I have to meet them on their terms. Whether it's a…
L
> Yes. One can complain about AI slop non-bug reports. But when there are real bugs, how is it AI's fault? The fault is insuffient quality of software produced, as AI has raised the bar for quality ny a significant amount. The promise (it seems) is/was faster development of new features, however due to the side effect of much better QA done by AI, the development has slowed down instead. This is not unique to Postgres, it's not unique to Open Source software, it's everywhere. Countless projects…
L
System containers require good supervision to be viable. When Docker came along, there was no good ubiquitous system-level supervisor, so Docker took on that role by association (which is only possible if you run application containers). Now that we have systemd in ubiquity, system-level containers in form of LXC/LXD/Incus/nspawn are absolutely sound, viable, and well-developed.
L
I remember using OpenVZ for more than 10 years. It was fantastic. Of course, they developed it entirely out-of-tree and it was a gigantic patch to primarily RHEL kernels... and there was no way that it would ever be accepted in the mainline kernel. Unfortunately Docker came along in 2013 and kinda spoiled the scene and everyone seemed to think "application" containers were the way to go... and we've been off on that tangent ever since. Virtuozzo has changed hands a few times I'm not sure who…
L
Yes. One can complain about AI slop non-bug reports. But when there are real bugs, how is it AI's fault?
L
> Firefox 159, released on October 27, listed more than 100 high-impact vulnerabilities including 72 use-after-free flaws and five privilege-escalation flaws. Chromium and Firefox, however, do have a much larger set of developers working on them and quite a few people scrutinizing the projects' code. This. A software project is (has been?) much more than its code: it's the people who can _understand_ that code and stand by it. And guess what: they can use LLM assistance too. Given the choice…
L
This Pixel 7 (updated to Android 17) is running kernel 6.1. That was released at the end of 2022 so almost 4 years old. It's not an OEM problem, it's every vendor including Google.
L
>Sure, it might technically have a test suite, but when you request a change that then breaks a test, >a LLM is just as likely to just alter the test in order for the test suite to go through than to see why the code stopped working Yes, that used to be true >= 10 months ago or so. Current deep reasoning models don't do that anymore. Instead they will find logic bugs in the test suite, missing corner cases in the test suite, then fix the code accordingly.
L
Come on - when I compare the code quality of a purely vibe-coded project (I'm not talking about someone using LLMs in general, I'm specifically talking about vibe coding, which I define as letting the LLM do all the interaction with the code and the user only writing prompts and never reviewing any of the code itself), it's way worse than most human-created projects I've seen. Sure, it might technically have a test suite, but when you request a change that then breaks a test, a LLM is just as…
L
Probably more significant than the triage resources is the actual bug fixing resources. If suggest ramping up the selectivity to them point where there's only one bug report, fix that bug. Then loosen the constraints a smidge and rinse and repeat. Obviously, there are some assumptions* baked into that simplified description, but you get my point, I hope. I think having an LLM adversarially challenge the reports generated seems like a sensible stage to include. It could probably to spot…
L
> It is not possible to send open file descriptors across the boundary (in either direction) using Unix-domain sockets This seems like a rather heavy restriction: fd-passing across the boundary isn't needed for the use-case where what you want is "a virtual machine, but cheaper", like a typical use of Docker or Incus; but in less-restrictive container use-cases like Flatpak, Toolbox or the sandboxed helpers used by e.g. web browsers, there are various protocols (Wayland, Pipewire, D-Bus…