ChatGPT gaslighting people

[philbo]

A face is a unique bio-marker. My phone recognises mine pretty reliably. If a person’s name and face are linked anywhere on the internet then OpenAI (ChatGPT’s creator) will have scraped the data. That includes employers staff pages, local news stories, blogs, Flickr etc

A face is by no means a unique bio-marker: as biometrics go, it's a reasonable assumption in a small data set - your phone recognizing your face is verification rather than identification; as cold searching goes, facial recognition is awful - useful as part of MFA, but lousy as an identifier. I'm not aware of anyone who has it reliably working to identify within a general population: police love the idea of it but get amazingly frustrated that it doesn't work like ANPR for people - if you have it scanning for too many people, it'll give you so very many suggestions to look at

OK. But then how is it identifying disco's friends from a photograph? Or perhaps it is all a coincidence.

Any odds it'll be using some other criteria as well - facial recognition is very good at saying "it might be one of this set of people", but very bad at "it is definitely this person" (it's that lack of uniqueness I was commenting on); so if it finds a set of a dozen potential matches and one of those has a connection to you via social media, that would probably cause it to prefer that name.

..all by doing stuff in the background it says it doesn't do when asked, but one can't help but wonder whether it's answering that question with what it thinks you want to hear - that seems to be ChatGPT's primary directive.

[El Pollo Diablo]

Btw, how did you upload a photo to ChatGPT? It says I'm not allowed to do that

[discovolante]

Btw, how did you upload a photo to ChatGPT? It says I'm not allowed to do that

I'm not sure, Señor von Laté did it. He's probably paying for it or something.

Sorry I realise that is unhelpful and I could just ask him. Well he is asleep at the moment so I will later. My post could have waited until then.

[discovolante]

A face is by no means a unique bio-marker: as biometrics go, it's a reasonable assumption in a small data set - your phone recognizing your face is verification rather than identification; as cold searching goes, facial recognition is awful - useful as part of MFA, but lousy as an identifier. I'm not aware of anyone who has it reliably working to identify within a general population: police love the idea of it but get amazingly frustrated that it doesn't work like ANPR for people - if you have it scanning for too many people, it'll give you so very many suggestions to look at

OK. But then how is it identifying disco's friends from a photograph? Or perhaps it is all a coincidence.

Any odds it'll be using some other criteria as well - facial recognition is very good at saying "it might be one of this set of people", but very bad at "it is definitely this person" (it's that lack of uniqueness I was commenting on); so if it finds a set of a dozen potential matches and one of those has a connection to you via social media, that would probably cause it to prefer that name.

..all by doing stuff in the background it says it doesn't do when asked, but one can't help but wonder whether it's answering that question with what it thinks you want to hear - that seems to be ChatGPT's primary directive.

This doesn't seem implausible.

Full disclosure, he tried this again last night (I think with the same picture but not sure?) and two of the names were different - but I don't think that helps very much given the history of his previous conversations with it will be saved.

[Woodchopper]

Btw, how did you upload a photo to ChatGPT? It says I'm not allowed to do that

I'm not sure, Señor von Laté did it. He's probably paying for it or something.

I think that's correct. As far as I can see people need to pay to use ChatGPt4 (latest version).

[Holylol]

I think part of the problem is that ChatGPT has been trained with "a dataset" containing data up to some year (don't remember which year). But we don't know what the dataset consists in, how it has been obtained, etc.
OpenAI seems very unwilling to provide access to their dataset to any form of auditor... which might very well be because they scraped the sh.t of anything they could find on the web, including a lot of personal (social media) data. Oh and a lot of copyrighted material of course.

ETA: already discussed some posts earlier, I just noticed

[noggins]

Is singer Tracey Thorn the sister of footballer Andy Thorn?

ChatGPT

Yes, Tracey Thorn, the singer from Everything But The Girl, is the sister of former footballer and football manager Andy Thorn. Andy Thorn had a career in football as a player and manager. This familial relationship has been publicly known and confirmed by both Tracey Thorn and Andy Thorn in various interviews and articles.

[Regenerate]

ChatGPT

As of my last knowledge update in September 2021, I am not aware of any public information or widely known connections suggesting that Tracey Thorn, the British singer and songwriter, is related to Andy Thorn, the footballer. It's possible that there may be private family connections, but such details are not typically publicly disclosed or widely reported. If there have been any developments or public disclosures related to their family relationship since my last update, I would not have that information. You may consider checking recent news or official sources for any updates on this matter.

[Regenerate]

ChatGPT

Yes, singer Tracey Thorn is the sister of former professional footballer Andy Thorn. Andy Thorn is known for his career in football, both as a player and a coach. Tracey Thorn and Andy Thorn share the same last name because they are siblings. This family connection has been publicly known for some time.

[Woodchopper]

Is singer Tracey Thorn the sister of footballer Andy Thorn?

ChatGPT

Yes, Tracey Thorn, the singer from Everything But The Girl, is the sister of former footballer and football manager Andy Thorn. Andy Thorn had a career in football as a player and manager. This familial relationship has been publicly known and confirmed by both Tracey Thorn and Andy Thorn in various interviews and articles.

[Regenerate]

ChatGPT

As of my last knowledge update in September 2021, I am not aware of any public information or widely known connections suggesting that Tracey Thorn, the British singer and songwriter, is related to Andy Thorn, the footballer. It's possible that there may be private family connections, but such details are not typically publicly disclosed or widely reported. If there have been any developments or public disclosures related to their family relationship since my last update, I would not have that information. You may consider checking recent news or official sources for any updates on this matter.

[Regenerate]

ChatGPT

Yes, singer Tracey Thorn is the sister of former professional footballer Andy Thorn. Andy Thorn is known for his career in football, both as a player and a coach. Tracey Thorn and Andy Thorn share the same last name because they are siblings. This family connection has been publicly known for some time.

Yes, of course. It doesn't doesn't have any understanding of who Tracy or Andy Thorn are. It has merely identified the patterns that you quote. Like all computer programs, compared to a human its staggeringly stupid. But it still has its uses for someone who knows what to do with it.

[noggins]

It doesn't seem to work for other random celeb pairs. The thing is, some anonymous idiot has been posting the Thorn fact on various internet forums over several years, purely for the private joke that one day someone might quote it back to them.

[dyqik]

Sounds plausible if the names are relatively common.

It doesn't sound that plausible.

Country/first language, age and time maybe narrows it to maybe 50 common names per gender if all the persons are from the same race/country/background. To get a triple match, the odds would be 3C50 = 1:19,600. If one or more of the names is unusual, if there's more than one set of common names represented (due to first language/race/background) or some of those clues are ambiguous or missing, then that goes way way up (as in Disco's example).

Remember if 19,600 people upload a photo - it is probably only the 1 person who got an uncanny triple name match who posts on an internet forum about the experience. The other 19,599 don't or may upload another photo and keep doing so until they get something uncanny.

There aren't 19,600 users on this forum.

[Sciolus]

It doesn't seem to work for other random celeb pairs. The thing is, some anonymous idiot has been posting the Thorn fact on various internet forums over several years, purely for the private joke that one day someone might quote it back to them.

Citogenesis all over again, but without even a chance of a credibility check.

AFAICT, LLMs are good at grammar and making up fiction, and utterly sh.t at anything else.

[dyqik]

It doesn't seem to work for other random celeb pairs. The thing is, some anonymous idiot has been posting the Thorn fact on various internet forums over several years, purely for the private joke that one day someone might quote it back to them.

Citogenesis all over again, but without even a chance of a credibility check.

AFAICT, LLMs are good at grammar and making up fiction, and utterly sh.t at anything else.

That's all they were designed and trained to be good for.

[Woodchopper]

To understand all these implications, start with what happened. Openai’s board fired Mr Altman for not being “consistently candid in his communications’‘. One factor that may have influenced the decision was disagreement over whether Openai had struck the right balance between the speed and safety of its products. Insiders say that Openai had made a breakthrough that enabled models to get better at solving problems without additional data. This spooked Ilya Sutskever, a co-founder and board member. Helen Toner, a board member affiliated with Georgetown University, had published an academic article that laid out what she saw as flaws in Openai’s approach to ai safety. On November 21st the New York Times reported that Mr Altman, worried about the negative press, had moved to oust Ms Toner.

[...]

The new directors are likely to make Openai, which is structured as a for-profit entity within a non-profit one, more business-minded. Mr Taylor and Mr Summers are well-regarded figures with plenty of boardroom experience. Their views on ai safety are not known. But they may be more receptive than Ms Toner and Ms McCauley to Mr Altman’s empire-building ambitions. The same already seems to be true of Openai’s workforce. One employee reports that the startup’s staff, which “trauma-bonded” during the upheaval, will become even more loyal to Mr Altman and, possibly, readier to pursue his commercial vision. Work on the firm’s most powerful model yet, gpt-5, which appeared to have slowed for a few months, will now probably go full speed ahead.

https://www.economist.com/business/2023 ... for-openai

[jaap]

ChatGPT generates fake data set to support scientific hypothesis
The researchers who wrote a paper about using ChatGPT to create fake data claim it's convincing.
Experienced bad science spotters say it is not, or at least not yet. Something for sTeamTraen to keep an eye out for.

[dyqik]

ChatGPT generates fake data set to support scientific hypothesis
The researchers who wrote a paper about using ChatGPT to create fake data claim it's convincing.
Experienced bad science spotters say it is not, or at least not yet. Something for sTeamTraen to keep an eye out for.

It doesn't really have to stand up to detailed analysis though. It just has to be adequate to pass inspection by those not actively looking for fake data. Same as other fake data.

[Woodchopper]

ChatGPT generates fake data set to support scientific hypothesis
The researchers who wrote a paper about using ChatGPT to create fake data claim it's convincing.
Experienced bad science spotters say it is not, or at least not yet. Something for sTeamTraen to keep an eye out for.

It doesn't really have to stand up to detailed analysis though. It just has to be adequate to pass inspection by those not actively looking for fake data. Same as other fake data.

Yes, and the problem is that the cost of producing fake datasets will likely soon become close to zero (even ones that purport to contain data on hundreds of thousands of individuals). It might not stand up to detailed analysis now, but it'll get better unless someone stops it.

[Stranger Mouse]

Slightly off topic but a nice little thread on ChatGPT making housing more affordable

https://x.com/pewilliams_/status/172994 ... 59264?s=61

[dyqik]

Slightly off topic but a nice little thread on ChatGPT making housing more affordable

https://x.com/pewilliams_/status/172994 ... 59264?s=61

That's definitely on-topic:

It's amazing how expensive all of those except the shack would be, compared to a simple four storey apartment building just like most of the others you see. Cheap housing looks like all the standard boring housing you see.

[noggins]

“Oops! We Automated b.llsh.t.”

https://www.cst.cam.ac.uk/blog/afb21/oo ... d-b.llsh.t

[Brightonian]

“Oops! We Automated b.llsh.t.”

https://www.cst.cam.ac.uk/blog/afb21/oo ... d-b.llsh.t

Heads-up to people clicking on the link and getting "Page not found": replace the dots in the URL with appropriate (inappropriate??) letters.

[noggins]

Or turn your word censoring off, you wusses

[Gfamily]

Or turn your word censoring off, you wusses

What - and have some f.cking dickhead pollute my screen with their sh.t?