Hacker News new | ask | show | jobs
LLM Policy for Rust Compiler (github.com)
100 points by liyanage 38 days ago
14 comments

> ## Other organizations

> These are organized along a spectrum of AI friendliness, where top is least friendly, and bottom is most friendly.

This section is an extremely useful reference

This policy is straightforward and shouldn't be particularly controversial (I'm sure it will be bikeshedded to death though). It basically bans the obvious stuff ("don't just drop LLM generated comments onto PRs") and allows the important stuff like LLMs writing code so long as you disclose.

edit: Wow people did not read the policy. It's literally just "if you use an LLM you are responsible for it, we will reject low quality PRs, please disclose that you have used an LLM". This is bog standard.

So...big caveat that this is still under review, so what we're talking about is a moving target, but based on what I can see, it seems considerably more nuanced than that. They basically ban LLM-authored code, with a careful carve-out to run an experiment to try to get only high-quality LLM PRs:

> It's fine to use LLMs to answer questions, analyze, distill, refine, check, suggest, review. But not to create.

> We carve out a space for "experimentation" to inform future revisions to this policy.

Importantly, the LLM contributions must be solicited, i.e., the people responsible for reviewing the final implementation have to opt in explicitly beforehand.

I think that the only significant caveat here is the need for reviewers to opt in, otherwise it's effectively "you can do it if you are open about it and are responsible for the output". The only notable ask here that's different from other policies is "if it's an LLM, tell reviewers beforehand".

TBH I think that makes no sense ("I have an LLM written PR ready, can I open it?") but yeah the policy is also in draft and has actually already changed since my first comment.

"allows the important stuff like LLMs writing code so long as you disclose."

Are you sure? It says:

"It's fine to use LLMs to answer questions, analyze, distill, refine, check, suggest, review. But not to *create*."

Yes. The policy is pretty clear on what the rules are for LLM generated code. You need a reviewer to agree to review LLM generated code, you need to read the code yourself, etc.
It's a pretty strict ban, with an exception.

That exception is experimental and somewhat limited; Only allows "well tested, high-quality" PRs on parts of the codebase that have a low probability of causing soundness issues, and it has a seperate review process with much higher standards.

And it requires the reviewer to agree to the use of LLMs ahead of time, before the PR is opened.

IMO, it has a high likelihood of degrading to a closed system, where some programmers with a good track record have little issue merging LLM generated PRs, while anyone without a reputation will struggle to even open an ai-assisted PR.

The discussion thread in the PR is also interesting to got through, lots of people concern in the HN discussion are already well discussed there
I like the 'better, but not faster' phrasing.

I'm not gonna say no to some system that validates my code before I present it as a PR, and if said system (it being static checking, dynamic checking, or LLM), gives me a 'comment', I will interpret the output, validate, and decide whether to take action, and how.

And maybe there I disagree a little with the proposal. It's me, it's my code. I stand behind it. But I get where the authors come from, and I believe it's a fine compromise.

I think more importantly in the world of Software Engineering we're seeing a split.

On one hand, people who go all-in on AI and take the output as 100% correct, copy-paste it as theirs without reviewing, prompt it to create PRs and submit them as-is, and worse, as their own.

Then there is the other side, people who use AI as a tool to validate and go deeper.

The problem with the first group that everywhere they feel entitled to shift the validation onus from themselves to the receipient of what they're sending, it being PR or review comment or message or whatever.

In PRs now the repo maintainer has to do a lot more work, as they cannot rely on the social construct of "OptionOfT wrote this, they have 10+ years of experience in these and these systems, so we can look at the PR through that lens".

Equally, I've been on the receiving end of AI PR comments (the PR was human-authored), but copy-pasted by humans, presenting the comment as their own, without properly (if anything) validating the correctness of the comment, or whether it actually makes sense for the PR. Lots of derailment there. This now increases the workload of the PR author, as above. We need to validate and cannot rely on the social construct. Is the comment even correct? What's the context? Why? Is it a hallucination?

And the downside is that it looks like the first group is now going faster, but the second group is actually slowing down due to the increased burden.

This is highly interesting. It seems clear to me that a lot of thought and work went into this. If I ever were to write a similar document, I'm sure I could learn a lot from this one. Props to the authors and all involved.
Here are the actual policies, not a comment:

https://github.com/jyn514/rust-forge/blob/llm-policy/src/pol...

It's in-line with the 'nanny' stereotype of the Rust community that they give you permission to act in a way they would never be able to verify anyways:

> The following are allowed. > Asking an LLM questions about an existing codebase. > Asking an LLM to summarize comments on an issue, PR, or RFC...

Like seriously, what's the point of explicitly allowing this? Imagine the opposite were true, you weren't allowed to do this - what would they do? Revert an update because the person later claimed they checked it with an LLM?

The Linux policy on this is much superior and more sensible.

> Like seriously, what's the point of explicitly allowing this?

Explicit permission can be useful to preemptively cut off some questions from well meaning people who, acting in good faith, might otherwise pester for clarification (no matter how silly / "obvious" it might otherwise be), or get agitated by misconstruing an all-banned list as being an overly verbose "no LLMs ever" overreach.

> It's in-line with the 'nanny' stereotype of the Rust community that they give you permission to act in a way they would never be able to verify anyways: [...]

Many of us work or have worked in corporate settings where IT takes great pains to help detect and prevent data exfiltration, and have absolutely installed the corporate spyware to detect those kinds of actions when performed on their own closed source codebases. Others rely on the honor system - at least as far as you know - but still ban such actions out of copyright/trade secret concerns. If you're steeped deeply enough in that NDA-preserving culture, a reminder that you've switched contexts might help when common sense proves uncommon.

While nannying can be obnoxious, I'm not sure that having a document one can point to/link/cite, to allay any raised concerns, counts.

> If you're steeped deeply enough in that NDA-preserving culture, a reminder that you've switched contexts might help when common sense proves uncommon.

What?

> If you're steeped deeply enough in that NDA-preserving culture

If you've throroughly absorbed a culture of honoring non disclosure agreements (NDAs), which are legal contracts demanding you keep secrets and avoid sharing sensitive data or code...

> a reminder that you've switched contexts might help

A reminder that rust-lang is a transparent, open source project, with no non-disclosure agreements or trade secrets to keep private unto itself might help [1].

> when common sense proves uncommon.

Because everyone misses the "obvious" sometimes. And because "obvious" is a subjective value judgement, meaning people will disagree what is or is not obvious.

-------

1. That said, if you've got a private, corporate-internal, closed source fork, you might still be bound by such concerns. For example, various people have ported rust's stdlib to work on various consoles (xbox, playstation, etc.) - and one of the reasons you don't see that upstreamed is because doing so would require violating console vendor NDAs, as well as possibly their company's NDAs - possibly for such banal reasons as not wanting to leak a hint of a console port or new title before their marketing plans are ready to go to capitilize on any hype.

> Like seriously, what's the point of explicitly allowing this?

I would have LOVED if the university course I took last winter had this. I had to take a very paranoid attitude to what was allowed.

What they're trying to avoid is a lot of unnecessary conflict with zealous anti-AI people calling for your exclusion for admitting to doing these things. There are people who would ban this too.

So then the Rust maintainers are going to give you an F on your report card?
Try using allman braces and see how far you get on a basic issue like that
No they’ll just drop() you
> Like seriously, what's the point of explicitly allowing this? Imagine the opposite were true, you weren't allowed to do this - what would they do?

Imagine if they just say "LLMs are banned" then there's a lot of ambiguity. So they specifically outlined that generative uses of LLMs are banned, and that non-generative ones are not banned (i.e. "allowed").

I think it's a poor choice of words on their part, but it makes sense (considering what their policy is). It's more of a "we're not disallowing use in these particular scenarios, so you can still use LLMs for these if you want". Remember: it's a big project, and if they don't explicitly state something then people will ask and waste everyone's time.

If anything, it reads to me as a proactive rebuttal of complaints that they don't allow LLMs; they're definitively stating that they do allow using them for very specific purposes.
Needs to be "solicited" from a senior dev. How many requests for ai code do you think they will be making?
I can't find any reference to using LLMs to ask questions in the ways that are cited from the parent comment needing to be solicited.
Y tho? It's already bad enough a programming language wants to play politics (doesn't matter what my politics are if I want to code in the c "community"), now they're taking purely emotional stances like "AI evil"
> now they're taking purely emotional stances like "AI evil"

But they aren't? Nowhere in the document it says this; in fact, it says the opposite - that they don't want to make a moral judgement.

> It's already bad enough a programming language wants to play politics (doesn't matter what my politics are if I want to code in the c "community")

It also doesn't matter what your politics are in the Rust community. My personal politics don't agree with the majority of prominent Rust contributors either, and that's fine. It doesn't (and hasn't) stopped me from being able to use Rust for over a decade now. Ignore politics and just engage on a purely technical level, and you'll be fine.

> It also doesn't matter what your politics are in the Rust community

Largely true, apart from the trans issue. Refuse to agree a man is a woman because he says it's true, and you're a bigot

You don't have to agree to treat people with respect. Using someone's preferred pronouns doesn't hurt you.
They're just giving examples of what you can do and explicitly saying so. Saying "you couldn't stop me" is completely missing the point.

This is not very different from the Linux kernel's policy so it's an odd comparison. It's actually almost identical in practical terms.

edit: lol proof that this doc needs to be stupidly explicit is in the pudding with the HN comments going out of their way to radically misread it

It feels telling that it reads like university course guidelines.
What do you mean?
Does the policy fix the issue of many low quality PRs being submitted? Unlikely.

Will it fix a related but different problem? Likely.

The people who submit low quality LLM-generated PRs often don't bother to read the policies first, but at least it will be easier to reject those.
The point here if you read contributor comments is mainly to allow people to shut a PR down without having claims of “unfairness” because some other PR wasn’t shut down. These are “moderation policies” in the style of old internet forums, their primary purpose is to clear up ambiguity and make maintainer’s (moderators) lives easer.

The birth of vibe coding has seen interactions on public FOSS projects increasingly reminiscent of the flame wars and moderator hammers of the old forum days. A lot of projects have been behind the curve on preparing and codifying the hammers, probably because no maintainer really wants to be a moderator, but thats where its naturally landed unfortunately.

> probably because no maintainer really wants to be a moderator, but thats where its naturally landed unfortunately.

Yeah this is autistic bunk. If you run an open source project, dealing with people is part and parcel of it, disagreements as well.

Not sure what is particularly autistic about the observation that "couldn't we just all get along" is unfortunately not how the world works even if we would like it to?
While we surely hope that at least some people will read and honor the policy, of course we know not everyone will. But creating a policy gives us teeth. Currently sending such PR is not disallowed, provided it doesn't fall in the thin area of some previous policies about slop PRs. With this policy, doing it will be escalated to the moderation team. First time you'll get a warning, second time you'll be banned from the project.
Ok but what if their OpenClaw reads it for them
Kudos to the team for this. I think it’s brave of them to stand up for their own experiences and push back against the hype train.

Before you knee jerk hate on the team for being luddites, consider:

1. For a language like rust there’s too few eyes and too many mouths. Reviewing is a job, and is extremely taxing. 2. The code base needs to be highly hermetic because it’s load bearing across the global economy 3. Most changes are only relevant if they’ve followed extensive process, including community feedback.

Some of these are just straight up unhinged.

> Using an LLM to discover bugs, as long as you personally verify the bug, write it up yourself, and disclose that an LLM was used.

What are they going to do go back and reject a bug if someone later admits they found it with an LLM? Honestly they and most other project would probably be better off just ignoring the situation until norms start developing.

They're trying to avoid a Boy Who Cried Wolf situation.

If they get swamped with 100 bugs that turned out, after they investigate them to be hallucinations then it's likely they will ignore or lose in the noise a real bug.

A llm generated bug that pretends it was a human created bug would be trying to abuse that presumption of validity, and therefore considered a dick move.

> If they get swamped with 100 bugs that turned out, after they investigate them to be hallucinations then it's likely they will ignore or lose in the noise a real bug.

But theyre saying if they're 100 correct bug reports it's still banned.

That's hysterical

That's the baby out with the bathwater.

No, you read incorrectly. Using LLMs to discover bugs is allowed given that you personally verified them.
The assumption here is that people act in good faith. If you break the rules, this indicates that you are not acting in good faith, and perhaps should no longer be welcome.
Sounds very welcoming. I came here for the code, not to join some social club
Then don't go interact with the social club?

If you want to interact with their project, follow their rules on their terms, otherwise just grab the binaries and don't stir things up.

What are you even talking about lol the policy doesn't imply that at all.

That's in the "allowed with caveats" section. It's just saying to not open bug reports without first reading them yourself or your bug may be closed. No one is saying "by policy we will have to add the bug back in" jesus christ

The policy is insanely straightforward, idk how you can be misinterpreting it this badly. It's just "Disclose that you use a model, you are on the hook for reviewing model output as a human" and then some clear cut examples.

Oh no where is Bun gonna be ported to next?
Nothing. You can always vibe-code in Rust even when the rust-lang/rust repository itself largely forbids vibe coding.
But one of the reasons they switched was because the compiler upstream for the original language they used, Zig, wouldn't accept slop contributions they wanted to make for Bun perf. What will they do when they need to try to push a slop contribution upstream to rust?

At this point they will probably just fork yet again and maintain some vibe compiler.

Huh. I wonder if the original intent was to merge an AI generated PR to a high-profile project like Zig. It makes the headlines and generates hype. But that went embarassingly bad for them so they had "port Bun to Rust" as a backup.
They should make FullstackLang. It compiles English in .md to machine code that can directly run on the specialized hardware it designs for it that you have to 3d print at runtime. Every program gets its own custom hardware. Composability and reuse be damned. Pay the token masters for every thought you have
Is there a citation for this? This was my suspicion but it's quite amazing if this was the actual reason for the bun spectacle
No, they've explicitly denied it.[0] However, they do regularly dig at how much faster their fork is[1][2] that they can't merge because of Zig's AI policy.

[0]: https://x.com/jarredsumner/status/2051600118886138262

[1]: https://x.com/bunjavascript/status/2048427636414923250

[2]: https://x.com/jarredsumner/status/2053050239423312035

> even when the rust-lang/rust repository itself largely forbids vibe coding.

This policy does not seem to forbid vibe coding?

It does in the narrower sense of vibe coding (as opposed to more general agentic coding, which is also called vibe coding from time to time...).

> Solicited, non-critical, high-quality, well-tested, and well-reviewed code changes that are originally authored by an LLM are allowed, with disclosure.

Vibe coding (in its original meaning) would have hard time arguing it's of high quality.

I guess that's the problem with the term. It should likely be left entirely out of a document like this since it's just confusing.
I read it as a hypothetical
Nim
Note that there are currently several proposed policies (plus hundreds of discussions mostly in private channels), and frankly I'm not sure we'll ever reach a consensus (I'm a Rust project member).
I wonder what will happen once these guidelines end up in the LLM training datasets
if an LLM says "I can't open a PR automatically until you solicit a review from a maintainer", i think that's good actually. likewise for proactively following the rest of the rules.
It's not the submitter who solicits, but the reviewer. They can't give code, AND THEN get approval, they need to be asked specifically for an llm created PR.
Content not viewable in I and I's region
it says “that’s right” in bold red cursive font. i know what the policy says, i wrote it, you nincompoop.
Github just won't respond at all.
On a general note, I like vouch by mitchellh.

> People must be vouched for before interacting with certain parts of a project (the exact parts are configurable to the project to enforce).

https://github.com/mitchellh/vouch

I think many projects will adopt this instead of allowing everyone / blocking everyone

Many projects have "ai slop" check in place to directly close and ban user if it is "ai slop". Else, it will be hard to handle the velocity of PRs

Maybe a network of ppl who can vouch they meet in real life?

I don't know if having your name/ face a secret is still acceptable? Maybe tiers of devs (anon vs other) on that one?

Saying "LLM" now sounds dumb. Just say "model". Some are no longer "large" and that is arbitrary.