Mastodawn

David Chisnall (*Now with 50% more sarcasm!*)3d ago

@davidgerard

As I understand it, the state of copyright for LLM-generated output is:

The person who promoted the LLM cannot assert copyright on the output (US copyright office ruling, confirmed in court).
People whose copyrighted materials were used as the training data may be able to assert copyright on the output as a derived work (inconclusive case law).

So accepting LLM-generated code seems very high risk. Especially for something like ext4, where it’s likely that there is exactly one implementation in the training set.

EDIT: I don’t expect the second to be settled conclusively any time soon, because there’s a lot of subjectivity in existing human-only derived-work cases.

John-Mark Gurney 1d ago

@david_chisnall

I agree that the second won't have a court ruling anytime soon, but the fact that some AI companies are settling instead of going to court is a strong signal. If they truly thought they'd win in court, they wouldn't settle for $1.5b. (yes there is risk reward to consider when analyzing this, but that just, imo, shows that the legal standing is even weaker, that if they lost they'd likely be on the hook for even larger penalties.)

@davidgerard

@davidgerard Okay so if I understood correctly, that means NetBSD and OpenBSD remain slop-free while FreeBSD and Linux won't. Didn't have "Run BSD" on my bucket list, but here we go I guess 🙃

波鉄 (Hatetsu)3d ago

@lu_leipzig @davidgerard Linux, I think I've heard enough about, but FreeBSD? Is the guy who submitted the slop a contributor or something? Has something been moving there? Or is this simply an inference from current discourse over there?

@HaTetsu @davidgerard That was based on FreeBSD's GitHub repo showing the infamous "Claude was here" banner:

@lu_leipzig @HaTetsu oh noooo

@lu_leipzig @HaTetsu @davidgerard that banner appears even if claude contributed to one PR that was accidentally accepted and then promptly reverted, as happened with Godot for example. Manual audit is still necessary, unfortunately.

@lu_leipzig @HaTetsu @davidgerard In this case, claude does not (yet?) have contributions on the main branch.

@ratsnakegames @HaTetsu @davidgerard Found this commit on main from a couple of weeks back: https://github.com/freebsd/freebsd-src/commit/6495dafd58b94a44fc9bc966ef47d6bc6916f5b9

It's only a trivial change and it seems to originally come from this PR at openzfs: https://github.com/openzfs/zfs/pull/18255

So yes, technically there's a Claude commit in FreeBSD's main branch, but this particular one is probably too trivial to directly affect FreeBSD's quality or legal situation. It would have been great if they still rejected it as a matter of principle though.

range_tree: use zfs_panic_recover() for partial-overlap remove · freebsd/freebsd-src@6495daf

zfs_range_tree_remove_impl() used a bare panic() when a segment to be removed was not completely overlapped by an existing tree entry. Every other consistency check in range_tree.c uses zfs_panic_...

GitHub

https://github.com/openbsd/src/commit/9c2b8e445a0bdfafdd6148b1760f00aa5429627b

The_Decryptor 3d ago

@lu_leipzig @ratsnakegames @HaTetsu @davidgerard And here's a similar situation with OpenBSD

It's more involved, but also from a 3rd party project.

Add support for applications to use synchronized output mode (DECSET · openbsd/src@9c2b8e4

2026) to prevent screen tearing during rapid updates. When an application sends SM ?2026, tmux buffers output until RM ?2026 is received or a 1-second timeout expires. From Chris Lloyd with the as...

GitHub

Rafael Martins 3d ago

@ratsnakegames @lu_leipzig @HaTetsu @davidgerard not so fast... the "claude user" only appears if the person was lazy enough to ask claude to even commit for them. it is perfectly possible (and even more common) that people just commit their changes made with claude code manually, with no traces of anything you could track, other than the low quality overall. this is a mistake I see frequently from people that had no exposure to claude (and AI in general) as an user.

@rafaelmartins @ratsnakegames @lu_leipzig @HaTetsu yes, people can lie, it's true

Rafael Martins 3d ago

@davidgerard @ratsnakegames @lu_leipzig @HaTetsu unfortunately it is not so simple. Only the official Claude client instructs the models to do that via system prompt, purely as a marketing stunt. Anyone using the Claude models via other clients won't have that, even if they ask the model to commit, and it is not their fault, they are not lying to anyone just because of that. My point is that you should not assume that something is clean purely based on a header missing on every commit message

@ratsnakegames @HaTetsu @davidgerard That's good to know, thanks for pointing that out. Now I'm curious what exactly happened in FreeBSD's case.

@lu_leipzig @HaTetsu @davidgerard their CONTRIBUTING.md does allow AI contributions:

"Low quality submimssions will be rejected. This includes submissions generated by AI tools that lack sufficient validation, testing, or understanding by the submitter."

@lu_leipzig @HaTetsu @davidgerard @ratsnakegames

God, that's so depressing…

@lu_leipzig @HaTetsu @davidgerard @ratsnakegames

Godot does actually accept AI slop code though, right?

@cloudskater @lu_leipzig @HaTetsu @davidgerard Not to my knowledge.

@lu_leipzig @davidgerard

I knew Linux was doing a stupid and considering AI "reviews" and stuff, but are they actually "generating" code now? Jesus Christ…

@cloudskater @lu_leipzig openly vibed code has made it into the kernel

@davidgerard @cloudskater Yikes, but yeah I'm not surprised. When did that happen?

@lu_leipzig @davidgerard

Seriously? Fucking hell. Yeah when did that start happening? Pls source so I can reference it later.

@cloudskater @davidgerard Not that I know of, but since they seem open to the idea in principle (and also since they take a lot of corporate code contributions), my hopes of having no slop code in the Kernel in the long run are rather low.

I'm also a bit suspicious of Rust in the Kernel in this regard, but that's just an uneducated gut feeling.

https://lwn.net/Articles/1026558/

@lu_leipzig @cloudskater nah it's in there

performative dickhead put it in without telling the maintainer

see comments for the maintainer unhappy at being sandbagged

Supporting kernel development with large language models

Kernel development and machine learning seem like vastly different areas of endeavor; there are [...]

LWN.net

Kevin Granade 3d ago

@davidgerard I see in the comments, "why are technical issues taking a back seat to license issues" and just wanted to give a maintainers perspective on this.
Fundamentally, if you don't have license compliance you don't have a project at all. A lot of us might dislike pretty much everything about copyright, but we don't have the luxury of pretending it doesn't exist.
You can discuss technical issues in the name of knowledge transfer or whatever even then, but that submission is already dead.

@kevingranade and a main BSD OS accepting a bot crib from GPL code? yeah, *not* happening