|
BoldFace posted:Rather than saying "I made the bot", I should have said that the former Google and Meta AI engineers at https://character.ai/ made the bot and I simply asked it to behave in a certain way. Chatbots have come a long way since the days of the Cleverbot. Many people often feel awe how pretty the Midjourney and Stable Diffusion pictures look, but, at least for me, the real innovation is how they can take instructions in natural language and transform them into something else while keeping the original meaning intact. This thing is eerily good. Seems to properly learn behavior from just a simple description too! (I didn't do anything except write the top line as guidance)  edit: you can pit AIs against each other. I am now making the AI vtuber shark play the haunted mansion survival game. ymgve has a new favorite as of 02:50 on Oct 21, 2022 |
#
¿
Oct 21, 2022 02:16
|
|
|
|
| # ¿ May 12, 2024 22:29 |
|
|
Carthag Tuek posted:https://www.whichfaceisreal.com/index.php pro strategy: look if the reflections in both eyes match. if they don't, it's AI. also blurry nonsensical backgrounds means AI
|
|
#
¿
Nov 15, 2022 17:30
|
|
|
ChatGPT does not know anything about character shapes:
|
|
#
¿
Dec 3, 2022 15:30
|
|
|
I tried making it do an interactive knock knock joke, but it just preempted me with the full conversation. Also it doesn't understand the punchline to its own joke.quote:> Can you do a knock knock joke? You start. edit: it doesn't do math well but it seems to be consistent quote:> what's 1925 + 7235? Retrying the question in a new chat gives different answers, but none of them are entirely correct. I wonder what kind of process leads to this. ymgve has a new favorite as of 21:08 on Dec 16, 2022 |
|
#
¿
Dec 16, 2022 20:56
|
|
|
ikanreed posted:GPT, is as the name describes, a transformator. It is a big matrix that you multiply a text string by, then shift and multiply again with later in the string where the results of the previous multiplication are also input into the multiplication(which is how it keeps context going forward) Yeah, I know the basics and know it doesn't actually try to do real math like Wolfram Alpha or something, but I found it curious that when I asked for A * B and got wrong answer C, doing C / B gave back A even though the math was wrong. Also that multiple chats gave different wrong but always close-ish to correct answers. Ariong posted:I hope one day we are able to find a more efficient way to get computers to do math. Do you really think we should trust computers with the power of MATH? ymgve has a new favorite as of 06:57 on Dec 18, 2022 |
|
#
¿
Dec 18, 2022 06:54
|
|
|
while I agree with some of his reply, I do have an issue with his stance that great songs must be borne out of suffering. please don't feed the myth that the troubled artist is the only valid form of artist.
|
|
#
¿
Jan 17, 2023 05:55
|
|
|
AI doesn't do what you want? threaten to take a virtual character's points away so it "dies"! https://twitter.com/venturetwins/status/1622243944649347074
|
|
#
¿
Feb 6, 2023 21:45
|
|
|
It is weird tho, didn’t the script containing the AI «rules» that someone managed to trick the bot into leaking end with a line telling the bot the current date and time?
|
|
#
¿
Feb 14, 2023 03:42
|
|
|
Looking back, it seems like that "The current date is XXXX" instruction to the AI was hardcoded, so it actually caused the issue in the first place: This was discovered last week and the AI is still instructed that the year is 2022
|
|
#
¿
Feb 14, 2023 17:01
|
|
|
I dislike quotes like that, because it makes it sound like robots doing boring tasks and robots doing art are equivalent problems with equal complexity. In reality, most boring jobs are drat hard to automate, and it turns out creating art and text is «easy» because the tolerance for mistakes is absurdly high. Everyone rightly yells when a Tesla AI fails to see a child, but nobody cares that much if the picture generator thinks humans have six fingers on each hand.
|
|
#
¿
Feb 23, 2023 15:32
|
|
|
credburn posted:I am a cashier. I'm only slightly more useful than a fully automated self-checkout machine. Your worth is that you can see when someone tries to sneak something through without putting it on the belt, or when someone puts the barcode of a cheaper item on a more expensive one. Once checkout machines can do that, bye bye job. (Which is also another retort to "why aren't AIs doing the boring poo poo" - you think capitalism isn't fully on board with automating your boring job? The only reason it hasn't happened yet is that the tech is hard to develop.)
|
|
#
¿
Feb 23, 2023 16:39
|
|
|
Evilreaver posted:My job is unautomatible That doesn't seem to be an issue if they confine the AI to a non-networked server, but don't tip them off to that. (The more likey outcome is that they will eventually say "gently caress privacy" and give the job to some cloud AI anyway)
|
|
#
¿
Feb 23, 2023 16:50
|
|
|
uber_stoat posted:hah, a while back i read an article where someone had the robot write a Nick Cave song and Cave wrote this long eloquent post about how awful and soulless this poo poo is. inspired by that, i asked the robot to provide me with a discography of Cave's band the Birthday Party. at first the robot told me there was no such thing and that perhaps i was thinking of the children's party song. i notified it that it was incorrect and it apologized, providing me with a list of songs, all of which were completely made up. this is poo poo you can look up on a wiki, but it just farted out a bunch of garbage. thought it was funny. reminder that chatgpt can NOT look up anything on a wiki, it only knows what was in its training data
|
|
#
¿
Feb 28, 2023 04:58
|
|
|
ultrafilter posted:Unless OpenAI is doing something very different from everyone else, most of Wikipedia is included in the training data they use. but the training data isn't accessible in pure text form to the AI, it is condensed into a mush of weighted word connections, so more obscure info gets lost. its like compressing a jpeg of the internet a hundred times at the lowest quality edit: also I'm sure the poster probably had bad luck or got the AI in a suboptimal state, because I asked "Which bands have Nick Cave been in?" and then "What's the discography of The Birthday Party?" which gave me a seemingly correct but incomplete and somewhat mangled categorically list of The Birthday Party releases. double edit: it also gave me a correct but incomplete discography when I asked "Can you give me a discography of Nick Cave's band "The Birthday Party"?" in a new session ymgve has a new favorite as of 05:34 on Feb 28, 2023 |
|
#
¿
Feb 28, 2023 05:27
|
|
|
It does sound very odd that they would have a special case for ascii art, calling an external program and risking a million security issues just for that single use case. It's more likely that the whole of the art piece falls into the "window" of GPT's memory, and it has learned that those character sequences at specific intervals (in this case, in a line straight down) is "better". edit: I can't make it produce a single legible ascii art word with the same prompt so I guess they got super lucky ymgve has a new favorite as of 23:24 on Mar 24, 2023 |
|
#
¿
Mar 24, 2023 23:15
|
|
|
Carthag Tuek posted:do note that openai constantly fucks with the backend using underpaid african manual labor. whenever something hits social media, like if some prompt makes the rounds, or if a certain one spikes, the workers have to fix that to maintain the illusion of ~true ai~. why would they gently caress with it so it makes worse responses with the same prompt? your theory that they call external programs on the basis of a few keywords is more absurd than chatGPT somehow finding some ascii art "good scoring" in the conversational context (and considering my own experiments, let's not ignore the possibility that the original tweet dude was just faking the conversation for clout)
|
|
#
¿
Mar 25, 2023 19:45
|
|
|
Another thing that these "takedowns" of AI often do is say "well this version of ChatGPT can't perfectly simulate whatever I want it to do, which means future versions will never do it either" I'm skeptical how far things can go before the computational requirements for improved models become overwhelming, but dismissing the whole future field due to the current state of ChatGPT is just dumb
|
|
#
¿
May 26, 2023 05:19
|
|
|
 edit:  Well...yes, but... ymgve has a new favorite as of 19:52 on Oct 11, 2023 |
|
#
¿
Oct 11, 2023 19:42
|
|
|
Is it possible to make something without lyrics? https://voca.ro/1gck2Gw1O5D1 edit: https://voca.ro/1mAZuq383uvz best one so far: https://voca.ro/1eCZkBdVmsMt ymgve has a new favorite as of 15:01 on Nov 17, 2023 |
|
#
¿
Nov 17, 2023 14:31
|
|
|
|
| # ¿ May 12, 2024 22:29 |
|
|
Wasabi the J posted:The lyrics make me think remi wolf ngl its from this Haverchuck posted:"its really amazing how the actions of goons and mods are so freaking far beyond the pale that its actually impossible for a bystander to believe."
|
|
#
¿
Nov 17, 2023 17:20
|
|