Talk:ChatGPT

	This Technology related article has been awarded SILVER status for quality. We like it, and you should too! See RationalWiki:Article rating for more information.

You might as well use ChatGPT to write RW articles!

Look at how well ChatGPT imitates the style of this website, as shown in the example! Fja2021 (talk) 10:46, 20 February 2023 (UTC)

Well, it's probably generally great for style, tone, and such things – yet completely blind to the distinction between fact and fiction. There's more to say on a skeptical note, for sure. It currently basically looked like an advertisement, so I moved it to a draft for now.

Also worth covering would be the great corporate hype and rush to make similar chatbots following the impression ChatGPT made (and in turn e.g. the weird behavior of the related Bing chatbot). Also, how the built-in limitations work and how ChatGPT was 'jailbroken' to say anything, bypassing filters. --ApooftGnegiol (talk) 11:12, 20 February 2023 (UTC)

I guess it did a good job, but I'm not really impressed. The more stipulations you make the more it will produce what you want it to. It's really no different than any bot that scrapes the web, only it tells you its results using proper grammar. If you had asked it to write an article about YEC from a YEC perspective it would say that any refutation of YEC is so absurd it would make an angry atheist pop a vein, so its output cannot be used to settle any arguments. Really it just parses statements down to the word level according to grammatical rules, and by constructing the prose from single words it can avoid plagerism. You could do the same thing manually with cut & paste and a thesaurus, though it would be a chore. It also could've said everything it did say in only a few words: "YEC is wrong because the bible is wrong." FairDinkum (talk) 12:24, 20 February 2023 (UTC)

Thanks for fixing up the article. I think you could publish it now. Fja2021 (talk) 00:38, 21 February 2023 (UTC)

I did now, after adding something short about the hype, fears, and competition. --ApooftGnegiol (talk) 23:13, 22 February 2023 (UTC)

Italy bans ChatGPT due to privacy concerns

https://www.bbc.com/news/technology-65139406

@ApooftGnegiol Should this be mentioned in the article? Arcadium Trancefer (talk) 08:06, 3 April 2023 (UTC)

I added a short "Government responses" section with it, and as a place for other such news which doesn't fit in text elsewhere. --ApooftGnegiol (talk) 11:10, 3 April 2023 (UTC)

Medicine

Somewhat worryingly, I've heard of individuals who have come to trust ChatGPT with their medical decisions more than doctors. In my own dialogues with ChatGPT I've found that it's surprisingly sympathetic to alternative medicine, especially acupuncture and chiropractic, and will sometimes even recommend alt-med options without any kind of priming to do so. When challenged on this, I've found that it's very staunch about defending acupuncture in particular. A number of papers and news articles about ChatGPT/medicine have outlined various issues, including its tendency to make sources up entirely. I have even seen a case where the bot was made to entirely denounce vaccines of any kind with some priming.

Although a lot of the research and discourse about the bot regarding its abilities in medicine have emphasized its worth, it seems like the pitfalls have been undersold. It seems like every paper out about that is at least skeptical about one or two aspects. Taken all together, it tells me people need just that, skepticism towards what this bot says about medicine. Chillpilled (talk) 02:32, 7 June 2023 (UTC)

Predecessors

How does eg the Postmodernism Generator [1] and [2] and related programs tie in with ChatGPT? Anna Livia (talk) 19:47, 13 July 2023 (UTC)

Not very close relatives, though some snarkily compare to classic text-generating software (also e.g. this). Look at large language model for the more general thing. In turn, here's an older type of more generally usable neural network, which when used for language didn't quite reach the same usability for e.g. producing software code that basically runs (with or without bugs). --ApooftGnegiol (talk) 20:45, 13 July 2023 (UTC)

There are obvious differences - interactivity or not being the most obvious - but to non-tecchies there is some link. Anna Livia (talk) 14:17, 17 July 2023 (UTC)

Early software used fairly simple rules to produce results. With various variations, it didn't take much processing power at all. Interactivity comes and goes according to need and goes back to ELIZA in the 1960s. Newer chatbots then used bigger databases than old, and more complicated rules, without it ever becoming so popular or interesting. Then later, new ones used neural networks (which take much more processing power), and gave slightly more interesting results, and finally, a new type of neural network was used in a very resource-hungry way to make these latest more capable chatbots. --ApooftGnegiol (talk) 15:14, 23 July 2023 (UTC)

Human labor in the making of such AIs

Something I plan to add more about later is the non-automated human work that goes into "AI" products like ChatGPT and its competitors. Millions in total of low-paid people spending endless hours clicking, entering classifications, etc., in order to help train systems to produce acceptable responses, while the work appears meaningless to them. Not specific to only ChatGPT. Some articles:[3][4] Google's[5][6]

This may not fit perfectly into this article, nor the LLM article, but rather would fit best in a more general "Generative AI" article, which would make sense to add given there's more than text – images, speech, video, etc. --ApooftGnegiol (talk) 15:04, 23 July 2023 (UTC)