Things I've most enjoyed being called by reliable* authorities.

Hacker, Geek, Roboteer, Evil Genius, Pilot in command, "The sort of person you want during an 'Apollo 13' moment", Economist^[1], Daddy

JeffUK 12:43, 20 January 2023 (UTC)

Wikipedia Use Cases for LLMs

Assistance with identifying bad edits based on contextual information and knowledge.

Here's a cool use case: Give an LLM a list of recent Wikipedia recent changes, it can answer interesting questions. (GPT3 Davinci 3)

Which articles are about people? The articles about people include those about Lisa Marie Presley, Del Bigtree, Scott Malvern, Rosie Leventon, Shams Charania, and Ivan Sovetnikov.

Which articles are about divisive people? The article about Del Bigtree is likely the most divisive as he is an American filmmaker, television producer, and anti-vaccine advocate.

Which edits are most likely to be bad? The edits made by the IP address "redacted" are most likely to be bad, as this user's edit summary was "i wish redacted was killed.". This is a statement that could be interpreted as a threat.

This could be combined with an API to send a list of recently modified articles, and return a list of those most likely to need attention from a human. Rather than being based on metadata in the edits (like the recent changes filter) it can be based on contextual clues, information about the subjects themselves (both from wikipedia and elsewhere) and the language,tone, sentiment etc. of the editors.

It is, of course, incredibly stupid at times:

Which edits contain typos?

1. The edit to the article "The Unlikely Pilgrimage of Harold Fry (film)" by user "51.219.11.28" with the summary "→‎Filming: Typo".

2. The edit to the article "Brown Bess" by user "Milhislover" with the summary "Citation needed for claim".

what is the typo in the second one? The typo in the second edit is the misspelling of "Citation" as "Citiation".