Talk:ChatGPT

Page contents not supported in other languages.
Page semi-protected
Source: Wikipedia, the free encyclopedia.

New logo?

When I go to https://chat.openai.com while logged out, I see on the top-left corner of the screen a new logo, that isn't OpenAI's signature flower(?) design. The logo is literally just ChatGPT●, but I can't find an SVG version of it, it's displayed as text on the website.

Should the logo be changed in the article? QuickQuokka [⁠talk • contribs] 23:21, 3 September 2023 (UTC)[reply]

However, it does retain its old logo in some other places. QuickQuokka [⁠talk • contribs] 23:22, 3 September 2023 (UTC)[reply]
That's not the logo. The logo is unchanged Pksois23 (talk) 08:53, 5 March 2024 (UTC)[reply]

"ChatGPT." listed at Redirects for discussion

The redirect ChatGPT. has been listed at redirects for discussion to determine whether its use and function meets the redirect guidelines. Readers of this page are welcome to comment on this redirect at Wikipedia:Redirects for discussion/Log/2023 November 6 § ChatGPT. until a consensus is reached. Gonnym (talk) 12:16, 6 November 2023 (UTC)[reply]

Nomination for deletion of Template:OpenAI

Template:OpenAI has been nominated for deletion. You are invited to comment on the discussion at the entry on the Templates for discussion page. InfiniteNexus (talk) 06:17, 14 December 2023 (UTC)[reply]

Editnotice

Several people now have also posted messages at Talk:Gemini (chatbot) believing — or pretending — to ask Gemini a question. As the number of chatbots continues to grow, I think it may be beneficial to have a generic editnotice template for talk pages of chatbot or generative AI software. InfiniteNexus (talk) 06:37, 25 February 2024 (UTC)[reply]

Agree, this has also been happening over at Sora on a near-daily basis. Since it doesn't seem likely to end this would be a good way to at least reduce the number of edits that need to be undone. Jamedeus (talk) 07:13, 25 February 2024 (UTC)[reply]
Done, see {{Generative AI editnotice}}. InfiniteNexus (talk) 08:48, 25 February 2024 (UTC)[reply]
Wow this looks fantastic! The automatic link to the AI tool is a nice touch, I hadn't thought of that. Thanks for your work on this.
If I could make one suggestion, it would be nice to add a bool parameter publicly_available that changes the text around the link. If set to False the text becomes something like "<tool> is not available to the public, but you can read more about it <here>". This would be handy for Sora, Codex, etc (and probably more in the future with the endless hype). Jamedeus (talk) 09:09, 25 February 2024 (UTC)[reply]
I like this a lot. I think it could help significantly with this issue. I also agree with the idea for a publicly_available parameter. I would additionally suggest maybe including some way (probably either another parameter, or just changing the wording of the notice in general) to adapt the notice to non-chatbot generative AI tools, such as the Sora and Codex models mentioned above, as the existing wording of the notice seems to assume that the tool will be a chatbot. –Gluonz talk contribs 14:25, 25 February 2024 (UTC)[reply]
I added a |public=no parameter to change the wording from "to do so, you may visit" to "learn more at". As for the non-chatbot concern, what wording change do you have in mind? Personally, it sounds pretty general to me, but I'm open to suggestions. InfiniteNexus (talk) 19:38, 25 February 2024 (UTC)[reply]
I agree that the wording seems mostly neutral. I think my concern would mainly be with the wording of “to ask [the AI system] a question”, since that does imply a chatbot. However, I am unsure whether there is an easy way to reword that without making it sound awkward, and the current wording should probably be understandable enough anyway, so I think that the template is essentially good to go for the time being. –Gluonz talk contribs 17:21, 27 February 2024 (UTC)[reply]
I have added {{namespace detect}} templates to detect if it's being used in mainspace and tweak it a little bit so it can be used on vandalized articles (not just talk pages). ~~2NumForIce (speak|edits) 03:34, 11 March 2024 (UTC)[reply]
I am sorry, but how to add it somewhere? I cannot find the template on Talk:Sora (text-to-video model) source code. RodRabelo7 (talk) 18:01, 27 February 2024 (UTC)[reply]
You must be an administrator, page move, or template editor. What page are you trying to add it to? InfiniteNexus (talk) 18:48, 27 February 2024 (UTC)[reply]
Great idea. This has been an annoying problem pretty much ever since GPT became a household name. popodameron ⁠talk 21:53, 25 February 2024 (UTC)[reply]

Financial markets

I think there is a need for a discussion with @

CommonKnowledgeCreator and @JPxG and anyone interested about this modification
.

Personally, my impression is that the content removed by

CommonKnowledgeCreator seems pretty accurate in regards to the sources (these studies indeed says that, although the phrasing should more clearly indicate that ChatGPT outperformed the average of the 10 investment funds, not necessarily all). And there is a desire to give the full picture of Patronus AI's study. But it's also confusing and I think the part on Patronus AI goes too deep into technical details that should be left in the references for further reading. I would probably prefer having just a few simple sentences with easy-to-interpret statistics concisely explaining what ChatGPT is good at (making quick and performant sentiment analysis and investment advice when fed a lot of data, integration with plugins) and what it's currently bad at (causal reasoning and mathematics, lack of knowledge of recent events, need for double-checking, reading facial expressions), e.g. a synthesis of the points in this article
that look still relevant today.

That's my opinion, but I confess that I'm not skilled in financial markets (as for most Wikipedia readers). What do you think? Alenoach (talk) 17:19, 8 March 2024 (UTC)[reply]

To trim down the summary of the Patronus AI study, we can eliminate the list of different of SEC filings for starters (like I did for the Anthropic article). Specifically, what about the current summary of the Patronus AI study do you find confusing? I can take a stab at rewording it once I know what is unclear from the current revision. I didn't add the content summarizing the CNN article; I will look over it. --
CommonKnowledgeCreator (talk) 17:30, 8 March 2024 (UTC)[reply
]
I would say that the protocol of the study is well explained but inherently complicated. Most people have really no idea of the difficulty of using this retrieval system for SEC filings in the first place, so even with all these details, the result remains difficult to interpret (although 81% of errors for GPT-4 clearly looks bad). And it leaves a sense of confusion with the previous study by finder.com, a reader might think: one study says that ChatGPT is great, others say that it's bad, but the article doesn't make it clear enough why ChatGPT performed well or badly in certain circumstances. But I would like to also have the fresh perspective of some wandering Wikipedian if possible to see if it makes consensus.
If I had to shamelessly simplify it, I may write "GPT-4 was shown to frequently fail analyzing SEC filings (a type of financial statements), along with other AI models like Claude 2 and LLaMA-2." It's less factual, but it's relatively simple to understand and verify, and the details of the study are available in the reference. Alenoach (talk) 18:24, 8 March 2024 (UTC)[reply]

I think that we should write the article by using reliable sources trying to figure out what the truth is, and then say that, and if the sources disagree, reflect that in our writing. What we should not do is write "A study proved that it was good.[1] However, a different study later proved that it was bad.[2]"

The organization of this article, bluntly, makes no sense and seems like the product of many sentences being slapped in piecemeal over the course of years. This section is flatly bizarre -- it's in "criticism by industry". The subsection name is "financial markets". Are the financial markets criticizing ChatGPT? Here is a sentence from it:

An experiment by finder.com revealed that ChatGPT could outperform popular fund managers by picking stocks based on criteria such as growth history and debt levels, resulting in a 4.9% increase in a hypothetical account of 38 stocks, outperforming 10 benchmarked investment funds with an average loss of 0.8%.

What part of this is "criticism"? Here is another.

On the retrieval system version, GPT-4-Turbo and LLaMA-2 both failed to produce correct answers to 81% of the questions, while on the long context window version, GPT-4-Turbo and Claude-2 failed to produce correct answers to 21% and 24% of the questions, respectively.

Is this "criticism"? Why does this belong in this section?

The more serious issue, to me, is that this actually has nothing to do with the previous thing. The first citation -- that it "outperformed popular fund managers" -- is cited to a fluff piece from CNN. Buried in the fourth paragraph of that fluff piece -- the one that boasts 4.9% ROI -- is that "Over the same eight-week period, the S&P 500 index, which tracks the 500 most valuable companies in the United States, rose 3%". There is no link to more detailed information, i.e. what the stocks were, what firms it was being compared against -- literally all we have to go by is that CNN is quoting some guy as saying "some of the most popular investment funds in the United Kingdom". The website itself has a page about this -- note it's written by their "Head of Communications & Content Marketing", not an analyst -- that doesn't explain any of this stuff either.

It doesn't give us information like, say, how the model was actually being used: were they prompting it with actual filings? were they prompting it with stock information? news articles? if this were actual credible research there would be a paper explaining these things, but it is clickbait, so there's not. It's also specifically talking about the United Kingdom. The SEC operates in the United States. Moreover, there is not a basis to claim that this is being refuted by different articles, published later, and by different people, about a different country, which the sources have not claimed is relevant to this. What's the connection supposed to be between this and it "failing" to read SEC filings? As far as I can tell, nobody ever claimed or warranted or represented that it could read SEC filings, so it's difficult to see how this is a "failure". Have I "failed" to learn how to speak Xhosa (I have never attempted to learn Xhosa, and never told anyone I was)? Have I "failed" to bench press eighteen thousand pounds, etc?


Like I said -- if we can't be bothered to find actual research that attempts to analyze these things in a detailed way, we should not just be trawling around for random clickbait from content marketers and then pretending it's research in the voice of the encyclopedia. jp×g🗯️ 20:49, 8 March 2024 (UTC)[reply]

Retrieved from "https://en.wikipedia.org/w/index.php?title=Talk:ChatGPT&oldid=1217364733"