Newest at the top
| 2025-01-23 21:43:49 +0100 | <dminuoso> | Not enough to consider it a pattern, but its a possibility |
| 2025-01-23 21:43:42 +0100 | <dminuoso> | But Ive seen reports from a few maintainers that have to fight floods of PRs of poorly generated code already. |
| 2025-01-23 21:43:32 +0100 | <homo> | also when updates happen rarely, it's easier to make independent review and decide whether or not they are desirable |
| 2025-01-23 21:43:24 +0100 | <dminuoso> | LLMs certainly seem like a promising solution to a lack of open source contributions - whether its for good or worse remains to be seen. |
| 2025-01-23 21:42:54 +0100 | EvanR | (~EvanR@user/evanr) EvanR |
| 2025-01-23 21:42:51 +0100 | <dminuoso> | sm: Especially in projects where there is 1-2 develoeprs, but dozens of issues a day from people that really want things fixed. |
| 2025-01-23 21:42:27 +0100 | <Rembane> | homo: Some ecosystems are so much more chill than others when it comes to updates. |
| 2025-01-23 21:42:15 +0100 | <Rembane> | homo: I think that depends on the ecosystem |
| 2025-01-23 21:42:12 +0100 | <dminuoso> | sm: Well most certainly do not donate 40 hours a week. When I say "little time" I mean it in the sense of "what little time they can donate" while splitting time with family, other hobbies and sleep |
| 2025-01-23 21:41:17 +0100 | <geekosaur> | (or modules in their private `lib` directories) |
| 2025-01-23 21:41:06 +0100 | <geekosaur> | some of us ignore the update culture. xmonad's core barely changes; it's solid and doesn't need to change, and most things can be done via hooks (which mostly come from xmonad-contrib but I've seen a fair amount of personal extensions in people's `xmonad.hs` files) |
| 2025-01-23 21:40:06 +0100 | <homo> | because if you don't push more code, people think your project is dead, irrelevant |
| 2025-01-23 21:40:03 +0100 | <haskellbridge> | <sm> me too. Not all projects can be done; either way, I love the ones that just keep working |
| 2025-01-23 21:39:16 +0100 | <Rembane> | Why? |
| 2025-01-23 21:38:26 +0100 | <homo> | I bet this kind of culture of frequent updates helps exploits like in xz happen |
| 2025-01-23 21:38:08 +0100 | <Rembane> | I love projects that are done. And that just keeps working. |
| 2025-01-23 21:37:09 +0100 | simplystuart | (~simplystu@c-75-75-152-164.hsd1.pa.comcast.net) (Ping timeout: 248 seconds) |
| 2025-01-23 21:36:01 +0100 | <homo> | "oh no, last time it was updated was 2 years ago, it's so bad it wasn't updated yesterday" |
| 2025-01-23 21:33:20 +0100 | <homo> | it also depends on how much sense it makes to focus on project, what if it's complete and thus requires updates very rarely? |
| 2025-01-23 21:32:43 +0100 | simplystuart | (~simplystu@c-75-75-152-164.hsd1.pa.comcast.net) |
| 2025-01-23 21:31:51 +0100 | <haskellbridge> | <sm> eh.. for some value of "very little time" |
| 2025-01-23 21:30:52 +0100 | <dminuoso> | Already now, people spend very little time on their open source projects. |
| 2025-01-23 21:30:29 +0100 | <dminuoso> | Why would free software be unimpacted? |
| 2025-01-23 21:30:11 +0100 | <homo> | I welcome it too, because then proprietary software will be so shitty that more and more people will switch to free software |
| 2025-01-23 21:30:00 +0100 | <dminuoso> | I believe to be above the job security threshold. |
| 2025-01-23 21:29:38 +0100 | <dminuoso> | Someone needs to fix the hallucinated garbage in the future, someone who actually knows how to build programs without the tool that broke it in the first place. |
| 2025-01-23 21:29:16 +0100 | <dminuoso> | I welcome it. |
| 2025-01-23 21:29:14 +0100 | <dminuoso> | homo: Personally I dont mind if companies adopt LLMs large scale for programming. |
| 2025-01-23 21:28:59 +0100 | <haskellbridge> | <sm> nobody has trained one focussed on rust/haskell porting yet, but something like that is quite possible in the near future. Much sooner than anyone will manually port darcs or pijul I predict |
| 2025-01-23 21:28:58 +0100 | <homo> | dminuoso don't tell that to companies that want to cut their budget on workers, let them go bankrupt after they replace programmers with llms |
| 2025-01-23 21:28:17 +0100 | <haskellbridge> | <sm> I used an LLM very successfully yesterday for some PHP and JS work. (Sorry to say it, because I don't like their costs and don't want to promote them. But I'm hoping to mostly use a local one soon.) |
| 2025-01-23 21:26:54 +0100 | <dminuoso> | You and I must be reading very different HN articles. :) |
| 2025-01-23 21:26:37 +0100 | Midjak | (~MarciZ@82.66.147.146) Midjak |
| 2025-01-23 21:26:21 +0100 | <haskellbridge> | <sm> homO: I meant that the more successful and more efficient new LLMs seem to be benefitting from a more carefully structured training process (from my HN reading). I'm not saying anything about companies |
| 2025-01-23 21:26:19 +0100 | <dminuoso> | Besides, everything I have seen suggests that LLMs in programming perform rather poor on any non-trivial problems, and yet they get used much in boiler plate code. To me that suggests a negative quality feedback loop, as the generated low quality code increases rapidly |
| 2025-01-23 21:26:19 +0100 | acidjnk | (~acidjnk@p200300d6e7283f96b914b0b581af9338.dip0.t-ipconnect.de) acidjnk |
| 2025-01-23 21:25:01 +0100 | <dminuoso> | For recursive training to add value, we would need LLMs to generate better output than their input, especially in programming since auditing large piles of codes for quality and correctness is just not feasible. |
| 2025-01-23 21:24:18 +0100 | Feuermagier | (~Feuermagi@user/feuermagier) (Ping timeout: 276 seconds) |
| 2025-01-23 21:24:07 +0100 | <haskellbridge> | <sm> eh ? |
| 2025-01-23 21:23:43 +0100 | <homo> | sm you are very optimistic about how companies are structured internally when in reality they suffer a lot of disorganization |
| 2025-01-23 21:23:41 +0100 | <dminuoso> | The publications I have seen and skimmed suggest that LLMs break down on recursive training. |
| 2025-01-23 21:20:45 +0100 | <haskellbridge> | <sm> I think training often involves synthetic data and is carefully structured these days |
| 2025-01-23 21:20:21 +0100 | son0p | (~ff@2800:e6:4001:6cc3:2e2c:4b4e:bc2a:6f17) (Remote host closed the connection) |
| 2025-01-23 21:19:57 +0100 | <dminuoso> | It is not unreasonable that the big companies have taken big snapshots from the pristine state to avoid training from poisoned sets. |
| 2025-01-23 21:18:51 +0100 | <haskellbridge> | <sm> the poisoning thing doesn't seem to be happening, or can be avoided |
| 2025-01-23 21:18:34 +0100 | <haskellbridge> | <sm> they have been so far |
| 2025-01-23 21:07:29 +0100 | <dminuoso> | You need to scale up training sets, already the well is poisoned because the internet is getting flooded by LLM generated code. |
| 2025-01-23 21:06:56 +0100 | <dminuoso> | Especially in programming tasks. |
| 2025-01-23 21:06:33 +0100 | <dminuoso> | I am not convinced LLMs will continuously improve over time. |
| 2025-01-23 21:04:21 +0100 | <haskellbridge> | <sm> maybe LLMs will get good at converting rust to haskell sooner |