Oct 27, 2026, 3:43 PMEvery OSS project maintainer I hear from is like "we can't ban AI because it's an unenforceable policy. There's literally no way of knowing whether something is written by AI" and every OSS project contributor I hear from is like "yeah, contributing is really miserable because I'm so busy reviewing this PRs and every single one is so obviously AI"
Oct 27, 2026, 3:47 PM@mcc I maintain HardenedBSD. I also maintain that HardenedBSD will not accept AI/LLM generated patches.
The only problem we have is that we maintain our patches on top of our upstream (FreeBSD). It's possible that an AI/LLM generated patch would get accepted there and be brought into HardenedBSD that way.
I'm under the (currently unfounded) suspicion that OpenZFS might have already accepted at least one AI/LLM generated patch. Those patches could be brought into FreeBSD, too, which then obviously dominoes into HardenedBSD.
It really is a tough problem. I myself don't want code that is potentially tainted with regards to licensing and IP rights.
Oct 27, 2026, 3:52 PM@lattera Yeah, this is why i'm looking at Redox. There is a long list of reasons why Redox is not an adequate replacement for FreeBSD, but for my purposes maybe I can get away with it.
Oct 27, 2026, 4:05 PM
Oct 27, 2026, 8:01 PM@sackofbones @reflex @lattera Is there a reason you are telling us this? It's obvious that projects use libraries and a project's policies may not re.
The way I see it, if you have a project that's committed to no-AI, and the language or the libraries become tainted, languages and libraries are "interfaces" which can be swapped out at a later time for clean ones. The important thing is the project commit to an intent; that gives you a "safe" zone you can stand in and clear out slop from there.
Oct 27, 2026, 8:02 PMAlso uutils (1) is command line, so i don't have to use it; I can make a functioning system without a command line prompt, and
(2) they forked uutils, I don't think they're still downstream; I was involved in writing docs for them and when I asked if I should describe the core command line tools as "from the uutils project" they were like "no not anymore"
@mcc @sackofbones @reflex @lattera We're using upstream uutils aside from a single minor patch: https://gitlab.redox-os.org/redox-os/redox/-/blob/master/recipes/core/uutils/recipe.toml?ref_type=heads#L4 In any case I don't like AI usage spreading everywhere either and I'm glad Redox has banned AI contributions.
Oct 27, 2026, 8:10 PM@sackofbones @mcc @lattera Never let the hypothetically perfect be the enemy of the better.
With projects deciding *not* to flag LLM commits we aren't going to be able to avoid all the slop. We can, however, state our principles and do our best to stand by them.
Oct 27, 2026, 8:19 PM@sackofbones I suspect that, for example, the maintainer of HardenedBSD has enough information to figure that out themselves
Oct 27, 2026, 8:28 PMThe important thing is the project commit to an intent; that gives you a "safe" zone you can stand in and clear out slop from there.
I agree, but only to the level the project's maintainer(s) can and want to support. For example, I have absolutely zero time (or desire) to audit each and every commit from each and every third-party vendored code in FreeBSD. It's already difficult tracking FreeBSD src on a commit-by-commit basis. (Thanks to the FreeBSD folks for providing per-commit emails!)
(Sorry to go on a tangent parenthetical, but that last paragraph touches on something I've learned as the long-term maintainer of a FOSS project: we, as a community, need to be careful about placing expectations on how random other people on the internet use their time. Only my family, my employer(s), and I can set my daily schedule or priorities.)
So, coming from the aspect of the futility of the lack of time: the intent (for me and the HardenedBSD project specifically) to be an "AI safe zone" is there--but the ability is not 100%. My time is much better spent implementing exploit mitigations and doing other R&D than reading git commit logs.
If someone in the community wanted to take that effort on, by all means--and that's part of what makes open source great! I myself can't.
Oct 27, 2026, 8:31 PM@mcc @sackofbones @reflex (I apologize for the long post--nuance is warranted here)
Oct 27, 2026, 10:32 PM@gooba42 @mcc Folks are finding success in running many Steam games under FreeBSD's linuxulator, which HardenedBSD inherits from. Though, intentionally HardenedBSD makes it rather painful to enable the linuxulator (it's possible--you just have to run a few extra commands to say "I know the risk I'm taking, please let me proceed").
Oct 27, 2026, 5:56 PM@lanodan @lattera @mcc @linear Yeah. If they want to go back and strip out the slop taint in tests, they can just rm all the affected tests. Zero impact on actual functionality. And until then, the potentially-dangerous code doesn't end up running on users' machines. Just git forge CI infrastructure. 🙃
@dalias hardcore blackbox testing.
Hmm. I had not considered moving them all the way out of the repo. That's really interesting
Oct 27, 2026, 6:48 PM
Oct 27, 2026, 6:58 PM
Oct 27, 2026, 7:07 PMOct 27, 2026, 7:13 PMAlso many of the gcc tests are only compiled and not run tests (this is even more true of the ones created by LLM it seems).
Plus ones reported via bug reports are small self contained testcases which have been audited when writing fixes for it, I am not sure if those are even less cared about.
Yes a compiler might be a special case for testcases unlike say a library.
Oct 27, 2026, 4:29 PM
Oct 27, 2026, 6:02 PM
Oct 28, 2026, 12:09 AMOct 28, 2026, 12:17 AM@jeremiah_ @mcc I've never heard of Contributors Covenant. HardenedBSD-specific code uses either the 2BSDL or the 3BSDL depending on the specific code.
Oct 28, 2026, 12:32 AM@lattera @jeremiah_ "Contributors Covenant" is a code of conduct, and I'm not sure if it's relevant. Do you have a DCO?
Oct 28, 2026, 12:33 AM@mcc @jeremiah_ I'm unsure what a DCO is, either. Focus has been on hacking on code.
Oct 28, 2026, 12:34 AM@lattera @jeremiah_ A DCO is an agreement some projects make contributors agree to where they simply affirm that they are the author of the code and have permission to use it. As opposed to it being stolen from IBM on an usb thumbdrive or something.
Oct 28, 2026, 12:39 AM@mcc @jeremiah_ Ah. We're still such small fry that we haven't had to worry about any of that.
Oct 28, 2026, 3:16 AM@lattera @mcc You mentioned you a concern about copyright and potentially other legal issues, especially as related to LLMs used in coding. This is why I thought that things like the CC and DCO might help, they can help secure provenance and reduce your risk from bad actors. They don’t help much if you’re referring to the copyright theft that may occur when training a model, but many FOSS licenses explicitly permit that kind of use.
@lattera what about code that's co-authored by AI?
You're AMD64 only, correct?
Oct 29, 2026, 1:25 AM@fuzzy any disclosed involvement of AI/LLM code directly to HardenedBSD would result in rejection of the patch, no matter how small the involvement--perhaps save only finding the bug in the first place. Patches should be 100% human generated.
Re architecture support: "what we have functional build infrastructure (physical hardware)" which is amd64 only right now.
I would love to bring back arm64 support, but we lack sufficient hardware.
OpenZFS and AI
Co-authored, 2026-10-22:
vdev: atomically account checkpoint space during parallel load · openzfs/zfs@1ad9b78 — https://github.com/openzfs/zfs/commit/1ad9b782c3c1513204ad32dab628611708bc08b9
Assisted, 2026-10-29:
ZTS: isolate the legacy mount option test dataset · openzfs/zfs@cec1aba — https://github.com/openzfs/zfs/commit/cec1aba086f9ea6fc0dfdc867800aa351481fcc5
Generally:
https://tinyurl.com/openzfs-master-avapet
@lattera ignore my question about AMD64, in an earlier scan I was looking at the wrong branch via <https://tinyurl.com/openzfs-avapet>. The URL that it targets has somehow changed.
I prefer to use short URLs for the ava.pet service because its use of the word "slop" is unnecessarily derogatory.
https://mastodon.social/@mcc/116110922163062791 by opening poster mcc was an instablock.
Oct 27, 2026, 3:47 PM@mcc part of me feels like this is just the extremely perfect-as-the-enemy-of-the-good Software Engineer way of looking at problems. like the VLC guy saying "seeking by one frame backwards is difficult or even intractable on some worst-case video formats, so until we have a solution that always works all of the time in all situations, we will never support the feature ever"
Oct 27, 2026, 3:48 PM@jk i would be ok with a back one frame that took a second or so to work. or that printed an error if it detected it was going to take 30 seconds
Oct 27, 2026, 4:26 PMOct 27, 2026, 4:28 PMOct 27, 2026, 4:32 PM
@jk @mcc or like, "actually this is NP hard so there's no reason to even try". Look, that's not what that means, you almost never need an ideal solution!
"Oh no that's basically travelling salesman"
Hey dumbass it's like traveling salesman but the salesman is clark kent and it doesn't matter if they retrace their steps, greedy solutions are fine, just try it.
@kevingranade @jk @mcc an optimizing compiler is literally the traveling purchaser problem, code can never be good
Oct 27, 2026, 5:46 PMAlso solving "*basically* traveling salesman" is actually O(n^3) and often good enough.
Oct 27, 2026, 3:50 PM@mcc ah yes, the two genders
Oct 27, 2026, 3:51 PM@lumi possibly literally accurate
@mcc paraphrasing:
"we can't have a policy against trolling, derogatory comments, public or private harassment, and other inappropriate conduct[0], because it's an unenforceable policy. there's literally no way of knowing whether something is inappropriate"
also
"we can't require attribution in our license, because it's unenforceable. there's literally no way to know if someone used our code without attribution, and no way we can stop them from doing it"
[0] contributor covenant code of conduct a.k.a. the "we have a coc" checkbox
Oct 27, 2026, 4:07 PM
Oct 27, 2026, 5:01 PM@reflex
The panacea fallacy is incredibly annoying because we don't hold *anything else* to a literally perfect standard. Nobody that pushes this notion has ever expected 100% efficacy of any other rule or norm they want to implement in their lives. It's equivalent to "one or two bacteria may live so I just won't brush my teeth or shower at all going forward".
@outfrost @mcc
Oct 27, 2026, 5:13 PM
Oct 27, 2026, 5:27 PM@kiri @reflex @outfrost @mcc I think "even with a no-LLMs rule your project will be tainted by someone who lies" is an extension of "linux/LLVM/GCC/Python are tained, lots of things will have been tainted before they wrote a no-LLMs rule, everything's tainted so you should just give up". This is exactly the "you can't avoid it so you shouldn't even try, get with the times" argument the so-called "AI" grifters push.
"No ethical consumption under capitalism" doesn't mean we shouldn't try at all!
(It would be nice if there was better tooling for removing things from VCS history, but that's a different rant.)
Oct 27, 2026, 5:45 PM
Oct 27, 2026, 4:16 PM@mcc this is changing quickly, but you can still "sense" whether something was created by an LLM or not: they completely lack a sense of "elegance". In general terms, their output feels brutish. That's the reason why enforcing such a policy is hard: "the contribution must be elegant" does not translate well.
There are some papers on using deep learning models to output a probability that a piece of code comes from an LLM. I don't know if people who do not want those kinds of contributions are willing to use models to detect them.
Oct 27, 2026, 4:25 PM@mem a good thing about your criteria is you don't have to worry about false positives. you can just reject stuff like "sorry, this patch is big and the code is not good and you can't clearly answer questions about it"
@mcc@mastodon.social @mem@mastodon.cr
…I am now remembering Torvalds rejecting Kent Overstreet's kernel filesystem patches for clearly flounting the rules and getting annoyed all over again.
Oct 28, 2026, 10:07 PM@mcc @mem
This is the big thing.
You don't have to "prove" that it's made with AI or whatever.
You can just reject with "sorry, this doesn't look right to me, I don't feel comfortable engaging with this". You *may* choose to engage in some education, but you cannot be required to do so no matter what.
Anyone who insists you have an obligation towards them, other than being decent and suitably polite, automatically disqualifies themselves.