03:35:00 Experimental (bcrawl) branch on underhound.eu updated to: 0.23-a0-5261-gd9800d219b 08:45:28 bug7 (L7 DjFE) Crash caused by signal #11: Segmentation fault (D:6) 08:45:33 bug7 (L7 DjFE) Crash caused by signal #11: Segmentation fault (D:6) 08:45:43 bug7 (L7 DjFE) Crash caused by signal #11: Segmentation fault (D:6) 09:36:30 New branch created: pull/5255 (3 commits) 13https://github.com/crawl/crawl/pull/5255 09:36:31 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-365-ge41b80f063: feat(webtiles): add local PWA shell 10(14 hours ago, 11 files, 354+ 0-) 13https://github.com/crawl/crawl/commit/e41b80f063bd 09:36:31 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-366-g600cfed68a: feat(webtiles): make local PWA play ergonomic 10(26 minutes ago, 13 files, 675+ 95-) 13https://github.com/crawl/crawl/commit/600cfed68a8b 09:36:31 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-367-g8c6dd383c0: fix(webtiles): fit PWA menus and keys on phones 10(13 minutes ago, 1 file, 41+ 6-) 13https://github.com/crawl/crawl/commit/8c6dd383c07c 11:09:48 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-368-gb688a58333: fix(webtiles): shrink PWA message log 10(75 seconds ago, 1 file, 9+ 0-) 13https://github.com/crawl/crawl/commit/b688a58333bc 11:09:48 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-369-g15b582c39b: docs(pwa): map WebTiles PWA entry points 10(47 seconds ago, 1 file, 37+ 0-) 13https://github.com/crawl/crawl/commit/15b582c39b4e 12:28:44 <12g​e0ff> Even if you are a fan of ai-generated patches, there have to be some standards for the submitted code. Projects like linux kernel require contributors to take full responsibility for their submissions, which includes understanding of what their code does. If a PR message starts with "I'm not a JS/CSS expert", it really means "I don't understand the details of a change I'm submitting." But even a non-web dev could notice that AGENTS.md is 12:28:45 not supposed to be a part of the PR, or that they don't have to add copies of application icons and could just reuse the existing ones. 12:45:43 <11O​dds> (That led me down a rabbithole of whether and what people submit in their AGENTS.md - seems opinions vary quite a lot. But overall I totally agree, I think the standard of "take full responsibility for your submissions" is the correct one) 12:50:50 <09h​ellmonk> Do we need to write that ai submissions guideline thing 12:55:11 <12g​e0ff> it could be easier to adopt one from some other open source project 12:56:19 <12a​sciiphilia> I completely agree with this. Recently, one of my coworkers used an AI agent to modify around 100 files and opened a PR against a branch I maintain. I spent two hours leaving review comments, one by one, explaining why the code was wrong. But after reading my review, they simply asked the AI to fix the code again — and even used AI to write the comments responding to my review. It reminded me of a recent incident where I got 12:56:19 pretty angry about exactly this kind of thing. 12:56:46 <11O​dds> https://github.com/melissawm/open-source-ai-contribution-policies attempts to compile some policies 13:00:53 <11O​dds> TBH the very first one at https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#gen-ai-assisted-contributions seems pretty reasonable to me 13:06:51 (FWIW as a contributor I would support the use of LLMs being outright banned) 13:08:32 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-370-g8b8d055d9e: feat(pwa): improve mobile play controls 10(17 seconds ago, 2 files, 365+ 46-) 13https://github.com/crawl/crawl/commit/8b8d055d9e39 13:10:43 <11O​dds> (And FWIW, while I haven't used LLMs for crawlcode, I think they are very good for many things these days when used wisely and that banning them outright probably isn't the best approach) 13:13:03 <12g​e0ff> perryprog, some people use them as an auto-complete on steroids, some use llms to ask questions about the project, like where can i found a function that does x. There's a whole bunch of legit uses. Also, outright banning would just lead to people not disclosing the use of these tools. 13:23:14 ge0ff: 1) single-line locally driven code completion is definitely interesting and I do think it tends to avoid some of the more major problems of LLMs—when I used it (admittedly several years ago) I would just accept the completion if it was what I was already going to type. That being said, the line can get blurry as surely just accepting each completion blindly and hitting enter. 13:23:14 2) I will fully admit that prompting LLMs with questions about a codebase isn't something I've thought about that much, but in terms of maintaining codebase quality it doesn't intrinsically affect that... unless you start making decisions based on say some hallucinated response that ends up going against the codebase's standards. 13:23:14 3) Re lying: I suppose that can occur, but my entirely hypothetical feeling is that the people who lie about it aren't going to necessarily be good at "hiding" its usage. 13:28:26 (There's also some questions about whether LLM generated text can really be licensed under GPL, but I do think a lot of them are fairly fear-mongery. While the case law is still being developed, in most places besides the UK and China it's seemingly increasingly likely that the official stance will be machine-generated text isn't copyrightable. Obviously derivative works of that content which meet the threshold of originality would be 13:28:26 copyrightable.) 13:31:16 <11O​dds> Wait really? If machine-generated text isn't copyrightable isn't that a massive problem for all sorts of people using LLMs for coding? 13:32:21 <11O​dds> (I've got pretty different views on the utility of these tools, which I think is the key disagreement here but is not resolvable by argument) 13:33:09 Yeah no worries at all <3 13:38:36 Re copyright: It's still not entirely clear. With diffusion image generation there's some better examples of rulings, but as far as I know the only examples of copyright protection existing for machine-generated media is in the UK, (probably) China, and (definitely) Hong Kong. In most other jurisdictions only works created by a human are eligible, and the US copyright office has said that that applies for AI generated stuff too... except 13:38:37 strictly speaking what the US copyright office says isn't binding whatsoever. 13:39:07 (c..f, https://copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf) 13:44:33 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-371-gd84284b08d: feat(pwa): improve mobile touch controls 10(69 seconds ago, 3 files, 370+ 11-) 13https://github.com/crawl/crawl/commit/d84284b08d8e 13:46:36 <12g​e0ff> nice var dungeonTapMoveSlop = 14; 14:27:59 IDK if the above is related to the issue I opened saying "please _have_ a policy even if it's not 'burn it with fire and salt the earth'" :-) 14:29:21 <09h​ellmonk> world's first sloplike 14:36:31 <02M​onkooky> given that they're actively committing, might be good for someone to say 'hey we aren't going to accept a PR where you don't know what your code does' 14:36:39 <02M​onkooky> save them some time 14:38:08 <02M​onkooky> I imagine an AI contribution policy is going to take some serious discussion, but I don't think rejecting this specific PR will 14:42:24 It's not (alas) the first slop PR 16:06:52 <04d​racoomega> For the record, my stance on AI contributions can be more or less summarized as: no. Like, I realize (as Ge0ff mentions) that a blanket ban like that potentially encourages people to avoid disclosing the fact that their PR was generated using an LLM, or lie about it when asked, but several of these PRs are already being cagey about it. And, if anything, it makes me more inclined to reject complicated PRs from unfamiliar people out 16:06:53 of hand than it is to concede ground to this sort of workflow somehow being inevitable. As far as I'm concerned, if one can't understand and take ownership of every line of code they submit, they have no business submitting it. (To be clear, this is not that the code needs to be flawless - I can hardly claim mine is, either - but AI-generated code constantly dives way past that to "I don't even know how it's doing what it appears to be doing" which 16:06:53 feels like a inviting a minefield of unintended and opaque consequences.) 16:12:17 …which crawl already has enough of on its own 16:15:50 <04d​racoomega> I have certainly untangled my share of unintended and opaque consequences over the years, but at least most of those were written with intent 16:16:30 <04d​racoomega> (Code's hard; I'm not about to fault anyone for making mistakes, obviously.) 16:18:41 that's actually part of my point: is an LLM going to know those things? (granting that neither will a new contributor, vibe coding or not, but if they also don't understand the code they're submitting they're in _real_ trouble) 16:21:24 even if they test their code and it's behaving as they intended, they won't know about or what to do about knock-on effects elsewhere 16:21:53 (where have you heard that one before? 😈 ) 17:37:47 <08o​____0> When I submit PRs sometimes there's code I copied from another part of the code base or write something I think is safe but am not certain. In those cases I point out parts I am shakey on so others can pay special attention to those parts. With vibe coding it's a lot easier to submit code you know very little about 17:47:13 <04d​racoomega> That does seem to be a pattern I see, yes. (And frankly, even being like "I am not 100% sure why this code block is necessary, but I think that it's required for X" is still a certain level of understanding what one's code is, tbh) 20:20:17 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-372-g376bfc0094: feat(pwa): improve mobile control usability 10(10 seconds ago, 8 files, 1076+ 828-) 13https://github.com/crawl/crawl/commit/376bfc0094e9 20:36:12 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-373-gde71e876eb: fix(pwa): harden mobile control checks 10(31 seconds ago, 6 files, 50+ 33-) 13https://github.com/crawl/crawl/commit/de71e876eb9a 20:55:16 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-374-g511f33b793: refactor(pwa): single ESM entry, address review findings 10(37 seconds ago, 9 files, 1145+ 1207-) 13https://github.com/crawl/crawl/commit/511f33b7934a 20:59:30 03Etienne A02 07https://github.com/crawl/crawl/pull/5255 * 0.35-a0-375-gf4c1971c2d: revert(webtiles): keep $.toJSON in comm.js 10(63 seconds ago, 1 file, 2+ 2-) 13https://github.com/crawl/crawl/commit/f4c1971c2ddb 22:03:06 <06p​leasingfungus> these diffs lol 22:03:47 <06p​leasingfungus> i think it is completely reasonable to say "if we can tell this PR was AI generated, we will snap close it" 22:04:10 <06p​leasingfungus> insofar as the telltale signs of AI PRs are all bad things in and of themselves 22:08:01 <06p​leasingfungus> even before AI, it's been a struggle to keep up with submissions. there is negative value in having more, worse code to review 22:14:48 <09h​ellmonk> Ah, but what if I ask Claude to review the code 22:16:15 <06p​leasingfungus> my experience of claude code reviews is that they have about a 50% hit rate for actual issues vs nonsense 22:16:53 <06p​leasingfungus> that's reasonably useful if (a) you want to save your reviewer's time, (b) you're willing to slog through the nonsense, and (c) you don't mind paying $15-25 a shot lol (i.e. you have a company paying for you) 22:17:38 <06p​leasingfungus> that is, if you want to review your own code before PRing/merging 22:17:41 <06p​leasingfungus> but... that's neither here nor there 🙂 23:46:56 Monster database of master branch on crawl.develz.org updated to: 0.35-a0-364-g274e6c4f07