I, ChatGPT
-
@Carnage I believe I've posted this video before (in the appropriate thread).
-
@Carnage said in I, ChatGPT:
Hmm.. Making the shit in my head into pictures seems like an idea...
Yeah, and we have a thread for ideas of that kind.
-
Some internal opinion from Google says they see themselves getting reamed.
Google "We Have No Moat, And Neither Does OpenAI"
Leaked Internal Google Document Claims Open Source AI Will Outcompete Google and OpenAI
-
@LaoC said in I, ChatGPT:
Some internal opinion from Google says they see themselves getting reamed.
Google "We Have No Moat, And Neither Does OpenAI"
Leaked Internal Google Document Claims Open Source AI Will Outcompete Google and OpenAI
Many of the new ideas are from ordinary people.
Those pesky ordinary people! How dare they step their unwashed feet on the sacred fields of The Great Googleplex.
-
@LaoC said in I, ChatGPT:
Some internal opinion from Google says they see themselves getting reamed.
The AI safety nuts must be shitting themselves over this. I haven't been following that stuff very closely but the timeline they give at the end of the post is pretty impressive for how quickly stuff is moving.
-
Amnesty International criticised for using AI-generated images
Group has removed AI images used to promote their reports on social media, including fake photos of Colombia’s 2021 protests
-
@boomzilla When the Guardian criticizes you for using misleading images, you're pretty far gone.
-
My AI is writing movie reviews for me, that I am submitting to a local website and getting top ranked reviewer rating in the process. FUN FACT, I haven't even watched all of these movies. :)
-
@jinpa said in I, ChatGPT:
@boomzilla When the Guardian criticizes you for using misleading images, you're pretty far gone.
I mean, it's not a very long step to go from directed photo shoots with models to just using completely generated images. It's be more stories if amnesty wasn't using generated images now that the tech is here.
-
@Carnage To avoid ambiguity, there's a difference between using directed photo shoots with models (e.g. an advertisement for a product) without deception and one where the intent is deception (e.g. an article where the picture appears to have been taken at the event in question but was not).
-
@jinpa said in I, ChatGPT:
@Carnage To avoid ambiguity, there's a difference between using directed photo shoots with models (e.g. an advertisement for a product) without deception and one where the intent is deception (e.g. an article where the picture appears to have been taken at the event in question but was not).
Yep, and Amnesty has been caught doing the latter. They also like doing poverty porn staged photoshoots.
-
@boomzilla said in I, ChatGPT:
Amnesty International criticised for using AI-generated images
Group has removed AI images used to promote their reports on social media, including fake photos of Colombia’s 2021 protests
Apparently, they made it clear those were AI-generated illustration images, their reasoning being "we don't want to publish actual photos, to avoid endangering people depicted".
Which makes some sense... but I don't understand how nobody said beforehand "wait, using fake-but-realistic pictures is a line we probably shouldn't cross".
-
@Nagesh said in I, ChatGPT:
My AI is writing movie reviews for me, that I am submitting to a local website and getting top ranked reviewer rating in the process. FUN FACT, I haven't even watched all of these movies. :)
Does it just give that warm fuzzy feeling of internetpointz or is there any other advantage to being a top ranked reviewer on a local website that would make it worth filling even more of the web with spam?
-
-
A (fake) pizza commercial (allegedly) generated entirely with A.I.
"Using AI technologies Runway Gen2, Chat GPT4, Eleven Labs, Midjourney and Soundraw AI, the creator was able to produce the background music, voiceover, graphics, video and even generate the script for the ad."
-
@Zerosquare said in I, ChatGPT:
but I don't understand how nobody said beforehand "wait, using fake-but-realistic pictures is a line we probably shouldn't cross".
It's because there's no line they wouldn't cross.
-
The AI revolution is powered by these contractors making $15 an hour
Two OpenAI contractors spoke to NBC News about their work training the system behind ChatGPT.
-
@Gern_Blaanston said in I, ChatGPT:
A (fake) pizza commercial (allegedly) generated entirely with A.I.
"Using AI technologies Runway Gen2, Chat GPT4, Eleven Labs, Midjourney and Soundraw AI, the creator was able to produce the background music, voiceover, graphics, video and even generate the script for the ad."
The last line caught me by surprise and made me audibly laugh.
-
I told you about chat bots creating spam emails for your inbox
.
Chatbots may also create whole websites with lots of "content", and some ads placed inbetween. It's cheaper than having typo monkeys creating such sites.
-
@BernieTheBernie said in I, ChatGPT:
Chatbots may also create whole websites with lots of "content", and some ads placed inbetween. It's cheaper than having typo monkeys creating such sites.
-
Thanks to @kazitor for mentioning HuggingChat as I had not wanted to give ChatGPT my phone number or other info, and thus couldn't try those things myself.
I did the test I'd wanted to do for some time, and asked it some basic domain-specific questions. Basically, the kind of things I'm asking university students when I give them a 2 days lecture on my main topic. So this is science, and relatively advanced science (university level and fairly specialised) but at the same time it's really basic stuff for this domain, and something that everyone in my line of work should be able to answer reasonably well without thinking too much about it.
The results are pretty much what I'd heard other people say, and what I'd expect given how the thing works (i.e. it doesn't "think," it just spews back combinations of what it can find in its training material). Basically it's what you'd get from a student who hasn't really studied, but isn't too stupid either. Some very convoluted answers that kind of revolve around the question, but never quite answer it directly, but do contain some of the keywords that must be mentioned, but does so in slightly off ways (such as mentioning as "new" developments techniques that have been used for at least 20 or 30 years). Not really wrong but not right either.
Not sure I'd give it a passing grade though if I failed it, it would be a close call. But there's no way I'd give it a "A" and I am in no worry about even a student using that to fake an assignment.
Basically, and as I expected, it is amazingly good at producing coherent speech, which in itself is already quite an achievement. But it is still nowhere near producing coherent content when asking it any sort of semi-specialised question.
-
@remi Did you try that in English or in French?
-
-
@BernieTheBernie English. I didn't even think about doing that in French, and I'd be surprised if it was any good. Not only the overall training material (and calibration work) is necessarily going to be more restricted, but when it comes to domain-specific knowledge there is relatively little of that in English as even French experts of that domain are likely to use English words.
I still have the chat opened and every now and then (
xkcd://compiling) I ask it a couple more questions. I am currently trying to get it spew (or at least mention) one key formula that absolutely everyone in that domain should have some awareness of (at the very least either knowing the name, or maybe that there is such a formula). I don't want to ask it specifically "what is the name of the formula that ..." and with this restriction, it has until now totally failed to produce that formula (or its name, description etc.). The answers are still never truly wrong though they are never right either, and totally fail to mention this equation. I think I would fail that part.I also failed a couple of times to add some domain-specific words to my queries and it seems to very quickly forget about that (which I assumed it would have known from the context of all my previous queries), but I'm willing to overlook that part as this may be a quirk of this specific chat bot.
ETA: just as I sent this post I managed to get it to mention the equation I was looking for! Though it mentioned something that is actually a generalisation of it, so probably not the version that'd be used in practice, but still, it did find it. But it also immediately spun into some extraordinarily complicated way to use it, which I can't even understand and is probably wrong (besides the fact that it's the wrong answer to the question, kind of like describing how you'd build a computer from scratch isn't the right answer to a coding problem even though it's not technically wrong per se).
ETA^2: I got bored of dancing around the topic and directly asked "what is the equation for ...," expecting an answer along the lines of "y = ax+b". It spawned about a screen of very complicated stuff that's maybe related to the most generic possible formulation of the problem (but I'm not even sure) and concluded what should have been a 1-liner with "Herein I omit details about how to obtain explicit expressions because it takes up many pages and can seem somewhat intimidating at first sight. Once all required material has been properly introduced, one may find simple closed forms under specific conditions easily readable without referring to specialized literature. Therefore, if you feel interested, please ask away questions about relevant mathematics as they come up throughout the course of discussions involving the topic."
-
@remi I've fucked around with it in Swedish and it's about as good at Swedish as English, so I'd bet that French is about on par with English. Err, I mixed it up with bing chat, the huggingface chat is a bit... Not very good at the Swedish part of the responses, but the information is on par with English.
-
@Carnage I just tried French (starting a new chat with the most basic questions) and... it's not good, not good at all.
I can clearly see that it's picking from a tiny training set because it is straight up lifting entire sentences from what must be a single website (including filler blog-style call-to-action stuff like "did this sound promising? if you'd like to know more you can contact me!"). It is also immediately straying away from the actual topic I'd asked it and answering a slightly different question, probably again due to the tiny amount of training material available. And finally, it has some sentences that simply don't make sense and reek of being machine-translated (from... English I guess?), which I can see as either attempting to beef up the answers by machine translating the question/answer in English, or using machine-translated stuff for training material.
-
@remi said in I, ChatGPT:
@Carnage I just tried French (starting a new chat with the most basic questions) and... it's not good, not good at all.
I can clearly see that it's picking from a tiny training set because it is straight up lifting entire sentences from what must be a single website (including filler blog-style call-to-action stuff like "did this sound promising? if you'd like to know more you can contact me!"). It is also immediately straying away from the actual topic I'd asked it and answering a slightly different question, probably again due to the tiny amount of training material available. And finally, it has some sentences that simply don't make sense and reek of being machine-translated (from... English I guess?), which I can see as either attempting to beef up the answers by machine translating the question/answer in English, or using machine-translated stuff for training material.
Cool, it's always fun fucking around with tech to break it.
-
@Carnage that also comforts me in the view that it's not close to take over any part of my job, except maybe the most tedious parts (such as writing marketese or boilerplate code), and that wouldn't be a bad thing.
-
It's also amazingly bad at even basic maths.
I asked it for one specific application example of the most basic of all basic equations, which is two terms that just have to be multiplied together. It did manage to find the correct equation to use, but it totally failed to take into account units (I used the customary units which are not the same thus a conversion is needed for the product). As I was trying to unravel how wrong it was, I noticed it even failed to put a dimensionally correct unit. But then I noticed it had also failed the multiplication, telling me that 2000*2.65 = 7.95e9.
I then asked it,
what is 2000 times 2.65?to which it said2000 multiplied by 2.65 equals approximately 5.30 million.It's getting less fun to play with it when it's so easy to break
-
@remi said in I, ChatGPT:
@Carnage that also comforts me in the view that it's not close to take over any part of my job, except maybe the most tedious parts (such as writing marketese or boilerplate code), and that wouldn't be a bad thing.
Yeah, I don't think LLMs will be coming for any advanced jobs anytime soon. And also, this kinda shows why I keep saying that LLMs probably won't work for full self driving. It's good at faking but can't actually understand.
-
@Carnage said in I, ChatGPT:
It's good at faking but can't actually understand.
Seems pretty obvious to me given what they do.
Still, I am very impressed by its capacity to produce correct sentences, the results are almost perfect (and at least as good as what the average human can produce). I can very well understand how people who can't look further than that can be afraid of it.
-
@remi said in I, ChatGPT:
@Carnage said in I, ChatGPT:
It's good at faking but can't actually understand.
Seems pretty obvious to me given what they do.
Still, I am very impressed by its capacity to produce correct sentences, the results are almost perfect (and at least as good as what the average human can produce). I can very well understand how people who can't look further than that can be afraid of it.
Yeah, it's all pretty obvious, but a lot of people, even those that should be capable of understanding, seen to fall for the marketing.
-
I’m firmly unimpressed with huggingchat from what I’ve tried to wrangle out of it, but without any experience of other contemporary models, I can’t say whether that’s a blandness common to all or just a symptom of huggingchat specifically.
-
I heard this morning (extracts of) an interview of Steve
JobsWozniak who I think made a good point by saying that those chatbots are going to be a boon for scammers.Given the ability to get plausible-looking text, I can very well imagine that the next iteration of Nigerian scammers will run with a chatbot, which means they will be able to run thousands of conversations in parallel and funnel the gullible ones up to where they give up money.
So in that sense, I can see AI stuff changing our lives, but not for the better, and not in the ways that most people are
ing about (i.e. not by taking over "real" jobs).
-
@kazitor said in I, ChatGPT:
I’m firmly unimpressed with huggingchat from what I’ve tried to wrangle out of it, but without any experience of other contemporary models, I can’t say whether that’s a blandness common to all or just a symptom of huggingchat specifically.
I think they added the trigger filter, because it's telling me it can't answer things because of topic far more now than when I tried it the first time. And that supposedly makes it quite a lot worse at all answers.
-
@remi said in I, ChatGPT:
I heard this morning (extracts of) an interview of Steve
JobsWozniak who I think made a good point by saying that those chatbots are going to be a boon for scammers.Given the ability to get plausible-looking text, I can very well imagine that the next iteration of Nigerian scammers will run with a chatbot, which means they will be able to run thousands of conversations in parallel and funnel the gullible ones up to where they give up money.
So in that sense, I can see AI stuff changing our lives, but not for the better, and not in the ways that most people are
ing about (i.e. not by taking over "real" jobs).It'll also be used for porn. And faking people saying things they don't.
The deepfake++ is going to be the larger problem, since judging the veracity in things will be a pretty hard problem, both for say, politics, but also in court.
-
I am trying to ask Bing for advice. It is... not going well. Like "what would be best of x and y" and the answer is always a variant of "the choice depends on the needs of the user".
So I asked if democracy or fascism is better and it went "I cannot answer that because I am designed to give objective answers" and refused to talk to me until I reset.
So then I asked the rhetorical "what's better, plague or cholera and it went first "I don't think either is good" and then I got it to continue with yet another "What would be least bad for you?", very helpful. Although third time's the charm and Bing decided that plague is the worst. Success!
-
@Atazhaia said in I, ChatGPT:
I am trying to ask Bing for advice. It is... not going well. Like "what would be best of x and y" and the answer is always a variant of "the choice depends on the needs of the user".
So I asked if democracy or fascism is better and it went "I cannot answer that because I am designed to give objective answers" and refused to talk to me until I reset.
So then I asked the rhetorical "what's better, plague or cholera and it went first "I don't think either is good" and then I got it to continue with yet another "What would be least bad for you?", very helpful. Although third time's the charm and Bing decided that plague is the worst. Success!
Yeah, they went overboard with the protective filter, it's borderline useless now for anything except pure data lookup. And those can't be trusted because confident lies.
-
@Carnage At least Bing does display what it searches for, and provides the sources it used from the results. So there's that.
-
@Atazhaia said in I, ChatGPT:
@Carnage At least Bing does display what it searches for, and provides the sources it used from the results. So there's that.
Unless you trigger the filter...
Last week, it could answer this:
Correctly too.
-
This post is deleted!
-
@Carnage said in I, ChatGPT:
The deepfake++ is going to be the larger problem, since judging the veracity in things will be a pretty hard problem, both for say, politics, but also in court.
Thankfully, we live in a post-truth society where nobody cares whether the events being reported actually happened or not.
-
@BernieTheBernie said in I, ChatGPT:
Chatbots may also create whole websites with lots of "content", and some ads placed inbetween
I think I forgot to put the ads on the wibbler :P
generating a lot of gpt content doesn't get you many users, by my experiments with toy projects at least
spammers are probably doing a lot more spammy tactics to get anywhere
or I'm just
-
@remi said in I, ChatGPT:
So this is science, and relatively advanced science (university level and fairly specialised) but at the same time it's really basic stuff for this domain, and something that everyone in my line of work should be able to answer reasonably well without thinking too much about it.
Can you give some more details, in the Lounge if necessary?
I'm just curious whether I could answer something that's on the one hand very advanced and on the other really basic, or maybe learn something new.Unless the answer is just "apply a Fourier Transform".
-
@Zecc said in I, ChatGPT:
@BernieTheBernie said in I, ChatGPT:
@remi Did you try that in English or in French?
Very likely.
I did not expect :@remi: to try it in German or Portugeese... 
-
If you are bored by your work and do not find new messages at WTDWTF, your
might perhaps be overcome, and you can read a lengthy article in PNAS on Creativity and AI:
"The muse in the machine"
https://www.pnas.org/doi/10.1073/pnas.2306000120
-
-
@Carnage said in I, ChatGPT:
judging the veracity in things will be a pretty hard problem, both for say, politics
Oh, that's an easy one. It doesn't matter whether it's written by an AI or a human; the veracity is zero.
-
Exactly. Think about it for a bit.
Nice-looking, plausible-sounding explanations that don't actually answer answer the question? And when they do, are complete bullshit?
We've basically created a perfect artificial politician. So much for the "I" in "AI".
-
Wendy's to begin replacing drive-thru staff with AI chatbots
“It’s at least as good as our best customer service representative, and it’s probably on average better.” After successful early tests, the fifth biggest fast food chain in the USA will start replacing drive-thru staff with AI chatbots next month.
They’re going to have a hell of a time field testing it in the Midwest and then trying to expand to the South. I can’t wait to see how it processes orders for “A Large Coke, a Sprite.”
AI is being used to generate whole spam sites
/cdn.vox-cdn.com/uploads/chorus_asset/file/24626885/537799082.jpg)
