rust-lang/rust is adopting an LLM policy
56 points by wezm
56 points by wezm
It's fine to use LLMs to answer questions, analyze, distill, refine, check, suggest, review. But not to create.
It's nice to have such a clear and reasonable policy.
I wish this single-line guidance included "[...] review, assuming you clearly mark LLM output and exercise your own judgement." The section below clarifies things excellently, but I find this excerpt a little too vague on its own (or perhaps unintentionally giving a false impression).
I think they mean that you're allowed to use LLMs to review your own work before sending in a PR, in which case there's nothing to mark.
But the "General Rules" section says this:
No one except the author is required to read LLM output unless they choose to: LLM output isn't allowed in public docs, PR descriptions, or Github comments unless it's clearly marked; reviewers aren't required to look at LLM PRs if they don't want to.
[...]
Disclosure is required for machine translation, "trivial" changes, discovering bugs, and reviewing other people's work using an LLM. We welcome messages posted in your native language; English translation is not required to contribute.
Don't get me wrong, I welcome this policy, I just wish the "one line zinger" was a little more detailed, because if you come away from the article without reading on (or, more realistically, if secondary sources pick up this quote without the surrounding context), one could easily get a wrong idea.
This appears to be only a summary. There are multiple carve-outs for generated code linked in the article. However the explanation is quite long and I am not sure I understand the details.
We welcome messages posted in your native language; English translation is not required to contribute.
I very much welcome this. I will post all my comments in Korean to rust-lang/rust from now on.
I’d always rather someone post in their native language than machine-translated output in a language they don’t speak at all. That way, I can:
If I just get machine-translated output, I can’t do any of those things - I just have to hope the translation was accurate.
Of course, I’d the author speaks some of my language, it becomes more of a balancing act, as most people probably will just machine translate it, and the author may be able to do better than the machine translation.
Commendable in spirit, but any kind of friction drives away your audience. In practice, I suspect that the overwhelming majority of readers would simply scroll past the comment rather than machine-translate it, even if a "translate only this message for me" were right there (and for a significant subset of users it currently won't be). After all, they likely don't know what they're missing.
As a poster, the incentive for engaging with an existing community is to use the language most of it clearly understands at least to some extent.
I agree with this, it does seem baffling to me why Rust and Zig are adament on no machine-translated comments. A PR containing 20 different languages would straight up be chaos.
I think it's at least partially to cut down on the "I didn't write this using AI, I just translated it using AI" excuses.
One feature that I'm really enjoying in vivaldi right now is the translation sidebar. I can highlight text and open the sidebar, and get a cromulent machine translation.
Firefox have also added this kinda. If you enable the local translation you can select text and right click and select "translate section" for a pop-up
I thought so but now that it’s becoming more common I changed my mind. It’s much preferable to get someone’s AI translation if it’s done well. My favorite comments now are coming from people who send me their AI translation but also leave the original text behind an expandable block. I have seen a few Chinese users do this now and I think it’s the optimal path at the moment.
I’d always rather someone post in their native language than machine-translated output in a language they don’t speak at all. That way, I can: read it myself, if I speak the language to any extent <...>
Sorry, but this does not compute.
Purely probabilistically, there's a much higher chance that the speaker of uncommon language $X also speaks English (the lingua franca of tech) to some extent, than the chance of an arbitrary English-speaking reader to also speak uncommon language $X.
From this it obviously follows that there's a much higher chance of the sender being able to (somewhat) vet a machine translation of a message he's sending, than of the reader being able to vet his own machine translation of a message he received.
And, of course, I don't even mention the fact that multiple readers of such a message will come up with their own machine translations, all of which will be unique, and some of which likely won't perfectly match each other, and the chaos will ensue if this rule actually gets exercised to any significant degree.
This (banning LLM translations on the sender side) is a completely illogical move, which makes no rational sense other than to plug the "I just translated it" loophole at a great expense in communication quality.
This seems all very sensible and reasonable and I’m happy to see such a well thought take in a project like that. I hope more projects may take similare stances.