[Wikimedia-l] Re: ChatGPT and Wikipedia

On Sun, Dec 11, 2022 at 5:55 AM Anders Wennersten <mail(a)anderswennersten.se> wrote: I suspect OpenAI will continue to focus on generative applications (images, code, text for purposes such as copywriting, eventually music/video) and won't attempt to compete with Google directly, but we'll see. Currently GPT-3.5 (which ChatGPT is based on) is very prone to generating nonsensical answers, citations to works that don't exist, etc. But it is pretty cool if you keep its limitations in mind--for example, it's quite good at bootstrapping small scripts in various programming languages (with mistakes and idiosyncrasies). Google has one of the largest AI research programs on the planet, they just are extremely conservative about letting anyone try their models (due to reputational concerns, e.g., that generative AI will spit out racist output within about 30 seconds of people poking its guardrails). This blog post from September is instructive about the direction they're taking with what's called retrieval-augmented generation; see the paper linked from the post for details: https://www.deepmind.com/blog/building-safer-dialogue-agents (DeepMind is part of Google) That is likely to yield significantly more accurate answers than what ChatGPT is doing, and is difficult to replicate for folks like OpenAI without being dependent on the search APIs of big search companies. It's worth noting that Google has also started to incorporate language model tooling into how it's presenting search results (e.g., summarizing or highlighting different parts of a website to make the result snippet more useful). A retrieval-augmented approach that leverages Wikidata could IMO be quite powerful and could be a useful research program for Wikimedia to pursue, be it independently or in partnership with others. The resulting technology should of course be fully open source. Querying Wikidata via SPARQL is currently still a bit of wizardry (and the query builder is extremely limited). To pick a completely random example not at all inspired by current events, if I wanted to see a list of journalists with Mastodon accounts & a picture, I currently have to do this: SELECT DISTINCT ?personLabel ?mastodonName ?pic WHERE { ?person wdt:P4033 ?mastodonName ; wdt:P106 ?occupation . OPTIONAL { ?person wdt:P18 ?pic . } ?occupation wdt:P279* wd:Q1930187 . SERVICE wikibase:label { bd:serviceParam wikibase:language "en" } } Make a small mistake (a curly brace missing) and you'll get a red error message. Forgot the * after wdt:P279? A different response set in ways that are difficult to spot or reason about. Why can't I type "list of journalists with their picture and Mastodon account" as a natural language query? (You can try it in ChatGPT and it'll get you started, but it'll generate nonsense P/Q numbers.) If such queries could be produced reliably, it could be a very useful tool for readers as well. Warmly, Erik _______________________________________________ Wikimedia-l mailing list -- wikimedia-l(a)lists.wikimedia.org, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l Public archives at https://lists.wikimedia.org/hyperkitty/list/wikimedia-l@lists.wikimedia.org… To unsubscribe send an email to wikimedia-l-leave(a)lists.wikimedia.org