Wikipedia talk:Wikipedia Signpost/2023-10-23/In the media

"Slate" celebrates encyclopedic selfies

 * I think the wiki-link for Slate is wrong... Still, shout-out to and the legend Annie Rauwerda! Oltrepier (talk) 08:05, 24 October 2023 (UTC)
 * Thanks for the head-up; it looks like Ca already fixed it shortly before your comment. Regards, HaeB (talk) 03:53, 25 October 2023 (UTC)

Wikipedia is the second-fastest website in the US

 * so... Wikipedia is spending money elsewhere, which some believe is unnecessary, and disbanding teams that do essential work? Why am I not surprised? —usernamekiran (talk) 11:14, 23 October 2023 (UTC)
 * I'm sorry, but I can't figure out which section in this article you are referring to. Smallbones( smalltalk ) 13:55, 23 October 2023 (UTC)
 * @Smallbones Counterfeit Purses (talk) 16:12, 23 October 2023 (UTC)
 * OK, I got it. With no disrespect to the performance team (which appears to have been fairly small and effective), I took the main reason for the speed of Wikipedia to be the simple design, as emphasized in the "TechNewsWorld" article. Smallbones( smalltalk ) 17:44, 23 October 2023 (UTC)

AI finding references
This item could have benefited from a bit more context, e.g. the fact that the paper was already published last year in preprint form and received media attention back then. We covered it in both "In the media" ("Facebook experiments with Wikipedia fact-checking") and "Recent research" ( "Facebook/Meta research on "'Improving Wikipedia Verifiability with AI'") at the time, and the current story doesn't really offer any new information about this research project. That said, we might still run a fuller review in "Recent research" now that the published version of the paper is out. Regards, HaeB (talk) 19:36, 23 October 2023 (UTC)
 * Yes - technically this is well above my pay grade ($0). And a short paragraph couldn't possibly cover it as well as an article in Recent research, it would be great for this. Smallbones( smalltalk ) 20:10, 23 October 2023 (UTC)

To follow up on a remark by Piotrus (moving here, as a more suitable location):
 * unlike most coverage (and research), this seems actually useful. Underlying research is here: https://www.nature.com/articles/s42256-023-00726-1 The research say the code to reproduce the study is somewhere here: https://github.com/facebookresearch/side . Can anyone convert it it into a usable tool, assuming this has not been done already?

I agree this could be super interesting. Two things to be aware of though:
 * The models alone appear to require two terabytes of disk space, so the server requirements are not quite trivial (e.g. it's not clear whether one could make this a Toolforge tool). That said, perhaps the Foundation's new "Lift Wing" machine learning infrastructure could be open to hosting such projects?
 * In July 2022, the lead author stated that Note, Side is a POC [proof of concept] that shows the technology is there. To build a production system there is still lots to do. :) and the code repository hasn't seen new commits since then.

Regards, HaeB (talk) 21:03, 23 October 2023 (UTC)


 * Empirically we found that only using the first sentence in front of the claim and also adding the Wikipedia article’s title to the query did yield the best BM25 results. I know from my experiments that their approach, undoubtedly the best of what they tried, doesn't isolate the entire claim being sourced more than half the time. Associating a reference with the article text it is intended to support is very difficult even after examining the source in full. Sandizer  (talk) 16:32, 25 October 2023 (UTC)
 * Interesting! Yes, I have been wondering what the state of the art is regarding this kind of entailment problem. This paper highlighted in the July issue of "Recent research" (see also talk page there) seemed to have encountered more difficulties than the authors of the Facebook/Meta paper. (By the way, in case you are interested, we would still like to run a fuller review of the paper in "Recent research", which could touch on the various issues mentioned above.) Regards, HaeB (talk) 03:42, 27 October 2023 (UTC)

Breaking news
See https://twitter.com/EdKrassen/status/1716103049863663962 and The Hill. Elon Mush offers $1 Billion if Wikipedia will change it's name to "Dickypedia" for 1 year. I got dibs on this story for the next issue! Smallbones( smalltalk ) 20:12, 23 October 2023 (UTC)
 * Weirder things have actually happened. ☆ Bri (talk) 20:39, 23 October 2023 (UTC)
 * And don't forget Joe, Montana Or the fact that somebody actually bought Twitter for $XX billion. Smallbones( smalltalk ) 20:56, 23 October 2023 (UTC)
 * Don't we have WP:POST/TIPS and/or the Newsroom talk page for this kind of remark?
 * Anyway, thanks in advance for your sacrifice ;) Some background on how what triggered this one-side schoolyard row: . And the Guardian is making some hay of it too . Regards, HaeB (talk) 20:50, 23 October 2023 (UTC)