Artificial intelligence still has some way to go

i used chatgpt to c0de today lol (asked it typescript a react component) it worked enough

A student handed in a chat gpt generated paper that consisted almost entirely of hallucinations. It was on a short poem and the paper seemed to be talking about a totally different poem, even concocting fake quotes.

saw some teachers talking about ai fake quotes in papers on twitter, jfc

whatever happened to hallucinating for real, maaaaaaaaaan

I felt like I was hallucinating when I read the paper and it was discussing a completely different poem from the one I remembered. And plausibly too, internally the paper made sense. Then I remembered this AI stuff and was like aha so this is what’s going on.

why does it make up quotes lol thats not a good feature

If you say “write an essay about x” it doesn’t. I think you need to specify “with quotes.”

with real quotes

you would think quoting would be about the easiest thing it could do

It’s not because it’s based on text prediction. It’s always just approximated what you would expect it to say. Very weird technology.

In this student’s case, the bot talked about themes present in many of this poet’s other work. Alas, they weren’t relevant to the poem they were assigned to write about.

https://i.dailymail.co.uk/1s/2023/04/07/15/69594635-11949507-image-a-79_1680879209616.jpg

Since Canvas Turnitin introduced an AI calculator, I've busted six students for using ChatGTS, two of whom earned a 100% percentage; when confronted he said he got lazy.

*ChatGPT

damn

Chat GPT "Write a lame excuse for handing in an essay as your own that was entirely AI-generated"

Start regularly recording your parents, elders and loved ones. With enough transcript data, new voice synthesis and video models, there is a 100% chance that they will live with you forever after leaving physical body. This should be even possible by end of the year.

— Pratik Desai (@chheplo) April 8, 2023

i watched a play about this concept a few years ago. i believe that in the end, we learned that it was creepy

this was a fucking Black Mirror episode ffs

its funny how this ai conversation is playing out over chatbots, a technology thats trained to look like talking that doesnt really care if what its saying is true or not, its hard to talk about ai in a wider sense because its very technical and because it doesnt really mean anything, so we just talk about the chatbots that make up fake quotes when student use it to write a paper for them, lol

Last month, Google released a chatbot, Bard. Weeks earlier, ethics experts recommended blocking the launch, saying Bard was dangerous and inaccurate. Their director changed the document, removing the recommendation and downplaying Bard’s risks. w/@KYWeisehttps://t.co/L2OUZO0ASd

— Nico Grant (@nicoagrant) April 7, 2023

I asked #ChatGPT for the average age of US soldiers in the Korean War. It began with a confident "22". I asked for sources. It eventually confessed to getting it from a MASH episode. pic.twitter.com/ywyRozlH9I

— Gene Zubovich (@genezubovich) April 11, 2023

how long until we have AI generated "I made ChatGPT do this thing" articles.

it doesn't actually know where it gets information, right? the second answer is just predictive text based on the two questions

I did a spot check and one of the articles it cites was attributed to the wrong institution, and another was a book published in 2010 so it's unlikely it'd have parsed the text

it's just regurgitating things related to the korean war

yeah its a chatbot not a search bot its trying to chat not give perfect info, im sure someone is attempting to address this issue but right now its just really unreliable

like idk why "it's citing M.A.S.H." is the punchline here, it's just making shit up on the fly

because its funny but yeah prob not true

brb I'm filing some patents about "factbots"

Can't wait for fact checkers to move to AI fact checking, so that we get Politifact saying JFK's head "just did that"

im c0ding patentbot

damn, outclassed again

this technology could never really vet sources unless it was programmed to only look in certain places. this would make it unable to function, because it requires a vast dataset to internalize enough pattern recognition to create syntactic sentences that respond to the topic. as far as i understand it, deep learning is a dead end for producing artificial general intelligence and this is a good thing.

in general i think the a.i. revolution is a lot of hype.

i think it will have many uses it already has some but of course the disrupt everything put everyone out of a job stuff is just silly and of course the idea of creating a general intelligence on a computer is nerd science fiction poisoning

it is fn hilarious that the thing thats blowing everyones mind currently lies constantly, just like a real person

is it trained on wikipedia? seems like it would be but maybe not

prob i imagine they just have a webcrawler soaking up whatever they can

i think it could put a number of people out of jobs. it can't run a marketing department by itself or whatever but trash copy that the department doesn't care about could be written by it, idk.

the real effect, i suspect, will be that the internet becomes even more overloaded with unreadable SEO drivel.

yes, wikipedia is part of the dataset for the major LLMs aiui. also lots of digitized books and Common Crawl: https://commoncrawl.org/

openAI won't reveal their dataset heh probably because they think it's their secret sauce

my understanding was that includes basically every publicly available piece of text

xp also copyright and privacy law violations galore!

but I don't think they're that secretive about their sources, they just claim they "filtered" them and are vague about *that*

both the image and text generation stuff is uncanny for sure and impressive superficially, but none of it has any real value. there is always something sterile and "off" about it, even before you know that it was not produced by a person.

no offense to anyone in this thread who happens to be a bot.

Go forth and sin no more.

i think it could put a number of people out of jobs. it can't run a marketing department by itself or whatever but trash copy that the department doesn't care about could be written by it, idk.

― treeship., Wednesday, April 12, 2023 11:39 AM (one hour ago) bookmarkflaglink

doing something that no one cares about is prob not the most secure job in the first place, but regardless most office jobs arent real anyway so im not sure why people are worried about the computer stealing them when they havent already been stolen by space

in the interest of social stability and in lieu of better alternatives, we need to preserve the bullshit jobs

a huge part of marketing/website writing is learning to internalize the very inconsistent and weird preferences of some particular VP of marketing at the client and all the weird brand language stuff they prefer (and don't prefer), often you have an audience of one.

that's true. and so far a.i. cannot really do that. when they try to imitate famous authors it never works. they have no understanding of the difference between style and content, so invariably if you ask it to write something in the *style* of james joyce it will say something like "i have ambivalent feelings toward my home country, ireland, and my religious education in the roman catholic church."

i talked about this upthread with the gpt-4 version of jonathan swift that was supposed to impress us. the diction was of the 18th century, and also other little details like random capitalizations of common nouns. but there nothing of his actual style in the writing -- no irony or anger. he just described his beliefs and works the way wikipedia would.