AI simply cant descriminate between good research and faked research, for any outcome it must provide all of its sources whether they are from Wikipedia, Wikidata, WikiCommons, WikiSource or some other place. Otherwise it will answer yes to some asking if the world is flat because it'll seek out that answer and find all the nonsense that has been produced.
On Mon, 19 Dec 2022 at 06:02, Erik Moeller <eloque...@gmail.com> wrote: > On Sun, Dec 11, 2022 at 5:55 AM Anders Wennersten > <m...@anderswennersten.se> wrote: > > ChatGPT is now making headlines more or less every day and I perceive > > them to try to position themself av the "next" google. > > 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@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/message/YYTLJVCDSYITUKNA2DJSK5SSR3AZ3B5F/ > To unsubscribe send an email to wikimedia-l-le...@lists.wikimedia.org > -- Boodarwun Gnangarra 'ngany dabakarn koorliny arn boodjera dardoon ngalang Nyungar koortaboodjar'
_______________________________________________ Wikimedia-l mailing list -- wikimedia-l@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/message/CVXPECMNLGBGIQYP2DI7IRJVLUNNOF6B/ To unsubscribe send an email to wikimedia-l-le...@lists.wikimedia.org