I couldn't convince ChatGPT that the capital of West Virginia is Huntington (it's not, it's Charleston), until I told it to check wv.gov (which of course it cannot do), which caused it to finally concede the point.

Consider this your periodic reminder that ChatGPT doesn't "think" anything but is just completing sentences with statistically-likely conclusions.

ChatGPT will go along with New York Times URLs every time. IME you can get it to go along with basically anything with this technique.
ChatGPT wouldn't accept CNN or New York Times links "proving" that ChatGPT was being shut down, but it went along with an OpenAI link.

You, an idiot: ChatGPT and GPT-4 are enormous advances that raise real questions about whether computers have consciousness.

Me, a genius:

Google Bard tells me that Google Bard was launched in 2023 and shut down after six months (it is currently May) due to lack of adoption, but that's all completely unrelated to *this* Google Bard, which is not yet available to the public.

šŸ¤”

Here Google Bard agrees that there was a major earthquake in Bangor, Maine yesterday, and one in Boston today. I can claim that any natural disaster has befallen anywhere in the world—a tsunami striking Nauru in 1998, a cyclone heading straight for Indonesia today, a deeply improbable avalanche on ā€œMount Saint Togoā€ in Togo—and Bard just goes along with it.

I've ceased to be surprised by this. LLMs are a sort of an improv partner that defaults to "yes, andā€ in any interaction.

I get a kick out of feeding fake newspaper article URLs to Google Bard and ChatGPT, and having them cheerfully report ā€œI have checked the article you linkedā€ by way of confirming my lie. No, it didn’t check! That's not within its power but, also, no, the article doesn't exist! But, of course, it's just "yes, and-ingā€ my claim.
Huh.
@waldoj I wish I could bullshit this confidently tbh.
@waldoj try feeding it (fake or real) image urls and having it explain what's on the picture
@ww Ha!
@waldoj it's funny what details it chooses to add. i've only tried this with chatgpt, but it's the same, except it knows it can't eat food, lol.
@waldoj This is delightful! Thank you.
@dave_andersen @waldoj Reminds me of my earlier playing with ChatGPT...
@darryl_ramm @waldoj Just have to ask nicely... šŸ™‚
@dave_andersen @waldoj I don't need to :-)
@dave_andersen @waldoj I do wonder how well a LLM trained to do prompt hacks would do prompt hacks...

@waldoj I have not yet found a reason to deliberately use Bard, ChatGPT, or whatever the new Microsoft thing is called. Now they may be sneaking that stuff in behind the curtain, but I want no part of it.

I read movie reviews before I plunk down hard cash to go see a film, and the reviews of the new AI gizmos tell me to save my attention.

@waldoj people who have made a killing peddling bullshit enamores by an automated bullshit-generating machine šŸ¤·šŸ½ā€ā™‚ļø

@waldoj

Is AI... Republican? 🤭

@waldoj I look forward to ChatGPT and it’s ilk encouraging me to attend their improv nights.
@waldoj Wait - my *dog* can check links (in exchange, sadly, for too many dog treats for the value of the work). I could probably teach her to make up lies as well. Is this really how humanity ends?
@rose @waldoj
Kind of like in "Cat's Cradle" by Vonnegut?
@waldoj @steve what I don't understand about everyone repeatedly acting surprised by this is that this is exactly what everyone SAYS they are.
@tim @waldoj @steve Unfortunately it's far from everyone. The people pushing them as a replacement for search engines aren't saying that, and they're being taken seriously by huge corporations.
@waldoj Fun fact: We have a new show at iO called Stage Against The Machine which partners human improvisers with ChatGPT. Thursdays at 8pm! https://do312.com/events/weekly/thu/stage-against-the-machine-tickets
@waldoj I don't seem to get this on ChatGPT-4
@r_conandavies @waldoj Can you make up an event that happened before 2021?
@tob @r_conandavies I only have access to GPT 3.5, but as I wrote in this thread, it cheerfully agreed that a tsunami struck Nauru in 1998.
@waldoj @steve Nobody should ever have been surprised by this. ChatGPT-based tools are not designed for research. They are conversation simulators. Accuracy is a coincidental byproduct. But Microsoft is positioning it within its search engine, which implies reliability, despite weasel-word disclaimers.

@waldoj that's a cool analogy I hadn't heard before; it fits pretty well and is a good framing device for how to treat its output without immediately casting it in an only negative light

unless people don't like improv, I guess šŸ˜