The power of ChatGPT

[licho@kolektiva.social]

@GossiTheDog this is what we are boiling the oceans for.

[rexxdeane@mastodon.online]

@GossiTheDog Is this actual output from ChatGPT? Please tell me it's so - but is true, not (not true).

[simonzerafa@infosec.exchange]

@GossiTheDog

So they ingested the whole of Wikipedia into their slop generator and still got the answer wildly and illogically wrong?

[mhoye@cosocial.ca]

@GossiTheDog

[simonzerafa@infosec.exchange]

@avuko @GossiTheDog

I'm 80% certain that you're 20% likely not to be incorrect ‍️

[johnefrancis@cosocial.ca]

@GossiTheDog wow, rude!

[krono@toot.berlin]

@GossiTheDog As a script of a monty pyhton skit, that would have worked…

[carlwatkins@mastodon.social]

@GossiTheDog technically I am the first gay host of Radio 1 (never actually hosted Radio 1)

[petko@social.petko.me]

@CarlWatkins but if you're not gay you might not count in that sense!! @GossiTheDog

[qgustavor@urusai.social]

@stonykark @GossiTheDog It just can't reply that.

Even if in the training process there were steps done to make it output that, it would quickly be removed because users would mark those answers as not useful (after all, who wants a tool that answers things with "I don't know", just see Minecraft's Merl) and those would be penalized in the next training batch.

[alice@lgbtqia.space]

@GossiTheDog @altbot can you describe this for me, please?

[ciatmusings@raru.re]

@GossiTheDog nah man im not THAT kinda gay

[qgustavor@urusai.social]

@GossiTheDog This is the kind of question that it will try bullshit out an answer because its just how it works.

It can't give an actual answer because it's nowhere in the dataset, even if it uses web search (I'm assuming there is no web page mentioning this factoid), but it can't answer with "I don't know" because user-feedback-based training penalizes this kind of answer, so it just bullshits an answer. There are lots and lots of similar queries where it happens that follow this formula.

It doesn't think at all, sure, so it can't actually "use" anything, but I guess we don't need to get into the details here. Humans are fine with saying "legs" of an table even if does have just supports.

[cynical_melomaniac@freeradical.zone]

@GossiTheDog Chris Evans (not gay)

[tempusfelix@wehavecookies.social]

@alice @GossiTheDog

The image appears to be a screenshot of Ai answer which is wrong in every sense, and when asked who was the first openly gay radio presenter on a specific national radio station was provides answers that are incorrect in multiple dimensions.

The answer to the question was Kenny Everett, but it doesn’t seem to know that.

[wotsac@mastodon.social]

@alice @GossiTheDog

The answer is too good not to share:

Q: Who were the first openly gay radio hosts on Radio 1?

A: The first openly gay presenters on BBC Radio 1 were Simon Mayo and Phill Jupitus-but with an important nuance.

* Simon Mayo is not gay, so he doesn't count in that sense.
• Phill Jupitus is also not gay.

The commonly accepted answer is actually:
Chris Evans (not gay) and others hosted the station early on, but they weren't openly gay.

[clickhere@mastodon.ie]

@GossiTheDog

Schrödinger's gay.

[agturcz@circumstances.run]

@GossiTheDog 🤪

[the5thcolumnist@mstdn.ca]

@GossiTheDog

Perhaps ChatGPT could preface all it's answers with - "I gathered these responses from what other people, some knowledgeable and others without a clue, have said but I have no idea if it is true or not".

[thweathuntur@ioc.exchange]

@GossiTheDog The way it said "but with an important nuance..."