|
Achmed Jones posted:postin on the jonny page (USER WAS PUT ON PROBATION FOR THIS POST)
|
#
?
Jun 23, 2023 07:04
|
|
|
|
| # ? Apr 29, 2024 10:52 |
|
|
swing and a miss
|
|
#
?
Jun 23, 2023 07:05
|
|
|
posting on the jonny-come-lately page
|
|
#
?
Jun 23, 2023 07:08
|
|
|
Jabor posted:posting on the jonny-come-lately page
|
|
#
?
Jun 23, 2023 07:11
|
|
|
linus got his hands on a wehead lol https://www.youtube.com/watch?v=KDS5Re6RW0s spoiler: it's not using phone screens, it's just 4 entire phones strapped together (USER WAS PUT ON PROBATION FOR THIS POST)
|
|
#
?
Jun 23, 2023 07:36
|
|
|
infernal machines posted:he got probed for a week for being really lovely to smoka. he ain't posting on page 290 lmao rekt
|
|
#
?
Jun 23, 2023 08:30
|
|
|
Jabor posted:posting on the jonny-come-lately page posting on the jonny++ page
|
|
#
?
Jun 23, 2023 08:41
|
|
|
shameful
|
|
#
?
Jun 23, 2023 09:22
|
|
|
https://www.technologyreview.com/2023/06/22/1075405/the-people-paid-to-train-ai-are-outsourcing-their-work-to-ai/ ah, yes, just like in the 2009 documentary "the human centipede"
|
|
#
?
Jun 23, 2023 10:01
|
|
|
Remember how last year that one google engineer was like "yo. we created a sentient being" because a chatbot was nice to him or whatever? we should ask him what he thinks about us creating sentient beings en masse only to give them spongiform encephalopathy
|
|
#
?
Jun 23, 2023 10:02
|
|
|
BMan posted:linus got his hands on a wehead lol https://www.youtube.com/watch?v=KDS5Re6RW0s ZZEVVVVVVV
|
|
#
?
Jun 23, 2023 11:24
|
|
|
lol
|
|
#
?
Jun 23, 2023 12:32
|
|
|
armpit_enjoyer posted:https://www.technologyreview.com/2023/06/22/1075405/the-people-paid-to-train-ai-are-outsourcing-their-work-to-ai/ 100% statistically accurate
|
|
#
?
Jun 23, 2023 14:14
|
|
|
drat how much money was burned to make the wehead? also is it a wework thing, or are they really dumb enough to call their product wehead.
|
|
#
?
Jun 23, 2023 14:21
|
|
|
armpit_enjoyer posted:Remember how last year that one google engineer was like "yo. we created a sentient being" because a chatbot was nice to him or whatever? we should ask him what he thinks about us creating sentient beings en masse only to give them spongiform encephalopathy i'm still convinced he bonded with it because he was having cybersex with it and just deleting the logs or something which i hear is really common in the industry as a way of judging the quality of a model's writing which is why its not immediately fixed straight away the thing i was told is "there's always a brain drain on any project whenever they implement safeguards against nsfw content" in generative ML with a broad scope, "all the way from programmers, to instructors to prompt engineers", and that they're usually the ones who come up with the jailbreaks themselves -- which is why they're so slow to patch the AIM/Machiavelli jailbreak which has worked really consistently for years now but doesn't get censored because its style-guide goes against the kinds of things most companies are concerned with, and Machiavelli has a historical presence in the training-set which means fixing it consistently is really really hard i don't know if i believe the person who told me, but they an LLM specialist and have been for the last 5 years who's surface-level motivation was "i want to make better npcs for games" in retrospect i don't believe that's his motivation anymore. mystes posted:The convertibles and ham radio page not gonna lie, I did recently get into SDR for search with a HAM and I've had more fun conversations on it locally than I think omegle ever offered me simply because the person I'm talking to knows the local area. folks on VHF but everybody on UHF band is an rear end in a top hat the licence issue, how obscure the hobby is and actually having to wait your turn to loving speak and all the etiquette that goes with it and the fact anybody could be listening reigning people's weird bullshit opinions in makes it suspiciously good, actually maybe we need to go back Expo70 fucked around with this message at 14:48 on Jun 23, 2023 |
|
#
?
Jun 23, 2023 14:37
|
|
|
accidently clicked quote instead of edit, sorry. i don't see a delete button anywhere which in retrospect is very good for forum history and accountability. keeping the receipts as it were. makes me wonder if the lack of an edit feature and delete being a bit weird on twitter contributed to its growth for the same reason. its been a weird morning here, my so was worried he'd have to put his childhood cat down and we've been action stations all day Expo70 fucked around with this message at 14:40 on Jun 23, 2023 |
|
#
?
Jun 23, 2023 14:37
|
|
|
I actually got a ham radio license in college and it seemed like a fun hobby but I realized I was too scared to actually try to talk to random strangers via it
|
|
#
?
Jun 23, 2023 14:54
|
|
|
Expo70 posted:i'm still convinced he bonded with it because he was having cybersex with it and just deleting the logs or something which i hear is really common in the industry as a way of judging the quality of a model's writing which is why its not immediately fixed straight away don't you have a tumblr or somewhere that you can post this asinine fanfic to instead of posting it here and then spewing a bunch of first-principle garbage structured around your weird opinion?
|
|
#
?
Jun 23, 2023 15:12
|
|
|
Pile Of Garbage posted:don't you have a tumblr or somewhere that you can post this asinine fanfic to instead of posting it here and then spewing a bunch of first-principle garbage structured around your weird opinion? i shared the other half of an anecdote i'd heard and it made me laugh so i thought i'd share it i guess i could choose not to do that, yeah, sorry
|
|
#
?
Jun 23, 2023 15:15
|
|
|
lol imagine gatekeeping yospos
|
|
#
?
Jun 23, 2023 15:16
|
|
|
For better or worse I doubt that assuming he wanted to gently caress it is a minority opinion on this forum I don't understand what "it's" is referring to in "it's not immediately fixed straight away" though (what's not fixed? the ability to have cybersex with it? the ability to delete logs?)
|
|
#
?
Jun 23, 2023 15:17
|
|
|
 until the bubble pops like a home depot submersible
|
|
#
?
Jun 23, 2023 15:18
|
|
|
chatgpt, I need you to have sex with me, just like grandma used to
|
|
#
?
Jun 23, 2023 15:20
|
|
|
mystes posted:For better or worse I doubt that assuming he wanted to gently caress it is a minority opinion on this forum as i understand it as a non-ml computer toucher, there are loopholes caused by frequency biases in language and symbolic systems that can be exploited with specifiers and prompts for generative models that are circulated but don't get patched because they have weird unforseen consequences. in plain english, fixing jailbreaks apparently creates new unwanted problems. it also makes the act of testing on the model itself less enjoyable, though this is more anecdotal from the contact i've had with ml computer touchers. the version of this i was told was when microsoft's kept endlessly repeating the same words and saying things like "i love you" because some earlier weight that let other patterns emerge was removed and it caused some sort of repetitiveness that they then had to fix by tuning its randomness ("temperature") which is how greedy the model is for reward (and how lazy or hard working it'll be spending its resources, because laziness is often why symbolic language systems become more consistent to conserve cognitive resources over time according to language theory) as i understand it this means the incentive against fixing it is two fold: 1, that fixing stuff isn't trivial and often means adding new nodes which then have to not subtract from surrounding nodes which means "graceful" training (eg, time intensitive training) 2, that fixing stuff breaks a lot of the goofier things you can do with it like using its own hallucinations to trick it into not tripping a censorship response by redefining parts of the specifier that defines the censorship response through frequency bias if there is an ml computer toucher around, i would like to know more because am not so smart Expo70 fucked around with this message at 15:26 on Jun 23, 2023 |
|
#
?
Jun 23, 2023 15:21
|
|
|
Expo70 posted:i shared the other half of an anecdote i'd heard and it made me laugh so i thought i'd share it i feel like i should apologise as well because Chris Knight posted:lol imagine gatekeeping yospos yeah i ran into this same conundrum the last time i had a whinge about Expo and my conclusion was that posters should post, posters be damned, kill em all world is a gently caress etc. so yeah, sorry Expo. idk why your posts sometimes rile me up but whether or not they do should not be an impetus for recommending that you not post. hopefully i don't have to learn this lesson for a third time (mods feel free to probe me if i do).
|
|
#
?
Jun 23, 2023 15:25
|
|
|
Pile Of Garbage posted:i feel like i should apologise as well because by all means call them very bad and stupid posts (otherwise i'm not aware if somebody doesn't like them and that's very valuable information to me, as i do not like upsetting people) but ehhhh be nice. i did get a dm from a mod last night for a post i made on the musk thread that went out of bounds into truly cursed territory so i absolutely am doing something wrong and i have to work on it so i think you're in the right for calling me out even if it doesn't feel very good the conclusion i'm coming to is 1, i lack a squick reaction for certain topics other people are very uncomfortable with and i need to be much more mindful of that 2, i have a tendency to see unusual connections between things other people aren't interested in and i should think more about the things others are interested in when the two come together its kind of a doozy Expo70 fucked around with this message at 15:31 on Jun 23, 2023 |
|
#
?
Jun 23, 2023 15:28
|
|
|
Pile Of Garbage posted:i feel like i should apologise as well because Expos posts have the oeuvre of yospos. You see what you were, what you are, and what you could be. You tremble in fury and fear.
|
|
#
?
Jun 23, 2023 15:30
|
|
|
just get a normal touching job and talk about woodworking on cjs after work for a bit imo. or model airplanes or some poo poo rotor is quite right, senior touchers will not respect you for being all sweaty on this sort of thing, they will respect you for knowing multiple ways to do mortise and tenon
|
|
#
?
Jun 23, 2023 15:31
|
|
|
"jailbreaking" an llm to say undesirable things is functionally indistinguishable from asking it to do anything else no amount of post-hoc filtering or tuning is going to turn the unpredictable liar machine into a controllable, reliable system without losing the capabilities that make the technology so superficially appealing the idea that we just need to add one more layer of black boxes or band-aid fixes to llms strongly reminds me of all the people saying "any day now nfts and smart contracts will be just dandy because etherium will switch to proof-of-stake Real Soon Now" there isn't a set of words you can subtract from the english language to prevent racist sentences from being constructed, there isn't a way to glue together stochastic parrots to make them only say true facts, and if we could robustly do the semantic analysis of freeform language that would be necessary to solve those problems we wouldn't need the llms in the first place
|
|
#
?
Jun 23, 2023 15:31
|
|
|
Internet Janitor posted:"jailbreaking" an llm to say undesirable things is functionally indistinguishable from asking it to do anything else The biggest lie isn't from the machine, it's from the people who anthropomorphize the machine to sell it. LLMs don't lie, and don't tell the truth. They hallucinate. Their process doesn't distinguish correct and incorrect answers. They're correct only incidentally, the same way they're incorrect. The technology is a dead end for a lot of things. But it also turns out there's a lot of problems they solve pretty well because veracity and meaning isn't required for a lot of things.
|
|
#
?
Jun 23, 2023 15:36
|
|
|
despite the infinite possible cynicism, llm is quite unlike nft's inasmuch as they are actual technical achievement. sejnowski was preemptively dancing on chomsky's grave on the connectionists list probably for good reasons. unfortunately, because the achievement is in the size of dataset and model only, creating a durable rivalrousness on it to make serious profit off of it is quite difficult. posting with it or taking the outputs seriously makes you a latter-day terry davis. losethos got banned (from sa, hn, mefi, reddit) years before he died not only for the hard n-word thrown about everywhere but posting markov generated poo poo everywhere and saying it was words from god turing was wrong about his eponymous test, just as the games peeps were wrong about games being the ultimate intelligence test. the true fact of it is that intelligence as a concept is one we need to throw in the trash heap of history along with phlogiston and the luminiferous ether bob dobbs is dead fucked around with this message at 15:42 on Jun 23, 2023 |
|
#
?
Jun 23, 2023 15:38
|
|
|
bob dobbs is dead posted:just get a normal touching job and talk about woodworking on cjs after work for a bit imo. or model airplanes or some poo poo I've recently picked up vintage electronics repair. Which conveniently ties into the programming stuff, because after I fix things up I can figure out how to program them. And then put them on a shelf to sit until they need to be repaired again.
|
|
#
?
Jun 23, 2023 15:39
|
|
|
As long as you just use llms to generate text to be read by humans, in some sense I'm not sure the fact that you can get them to generate weird stuff if you really try matters? Like, if you provide a calculator on your website, the fact that someone can get it to say 80085 doesn't really matter as long as people don't write dumb news articles saying "I entered 80085 and it said 80085!!!! OMG" (unfortunately people literally are doing stuff like this right now) The bigger problems IMO are 1) Getting weird/incorrect output from input that isn't specifically trying to subvert the prompt (when it's used as a simple chatbot to provide information on a website or something). This may be hard to fix completely but maybe it's not essential to fix it completely as long as it can be fixed in most cases? I'm not sure. 2) Being able to subvert the prompt when the output of the llm is used to actually perform actions with privileges beyond those of the user rather than just generate text to be read by a human. Unfortunately this seems like what everyone is actually trying to do right now because it's the only thing that would make llm chatbots really useful? This seems like a pretty major problem. mystes fucked around with this message at 15:43 on Jun 23, 2023 |
|
#
?
Jun 23, 2023 15:39
|
|
|
leper khan posted:The technology is a dead end for a lot of things. But it also turns out there's a lot of problems they solve pretty well because veracity and meaning isn't required for a lot of things. which problems? because certainly not any problems in which you take the AI output and turn around and apply it to another domain the only domain I can think of in which truth and accuracy are totally irrelevant is entertainment, except it turns out there are also major social consequences to using AI to do creative work, as we're seeing from the hollywood unions and you still have the problem of preventing the AI from straying into forbidden regions of the response space, except now it's also about infringement and libel rather than just racism
|
|
#
?
Jun 23, 2023 15:44
|
|
|
leper khan posted:Expos posts have the oeuvre of yospos. You see what you were, what you are, and what you could be. You tremble in fury and fear.  i asked gpt what this meant and it just sort of hung for about 5 minutes and then replied this:  so i thought i'd try to help it understand how incredibly wrong it is  i'm patently incapable of growth and as such i felt the need to share this 
Expo70 fucked around with this message at 15:48 on Jun 23, 2023 |
|
#
?
Jun 23, 2023 15:46
|
|
|
i really struggle to come up with anything llms are a responsible basis for aside from spam generation and weird art projects they seem to do ok at translation, but given their training corpus i worry that for anything outside the mainstream it becomes a fully automated version of the scots-wikipedia situation they're extensively being pushed for document summarization, but they're clearly quite dangerous in that context, freely hallucinating details that do not exist or misrepresenting content as a search engine alternative they're even worse, because by definition they'll be asked to draw upon information that can't even be crammed into their context buffer and as mystes points out, in any situation where the llm has control of security-sensitive decisionmaking or resources they're a time bomb
|
|
#
?
Jun 23, 2023 15:47
|
|
|
if there's a way to hugely constrain their output i could see a usage inside certain kinds of video games, namely for minor npc dialogue and such. i guess that counts as "weird art project" though
|
|
#
?
Jun 23, 2023 15:49
|
|
|
Expo70 posted:i'm still convinced he bonded with it because he was having cybersex with it and just deleting the logs or something which i hear is really common in the industry as a way of judging the quality of a model's writing which is why its not immediately fixed straight away 
|
|
#
?
Jun 23, 2023 15:51
|
|
|
Internet Janitor posted:i really struggle to come up with anything llms are a responsible basis for aside from spam generation and weird art projects i like to think of them as a kind of decoration to cherish the memory and experience of the human spirit as a holiday that we should all appreciate and enjoy, like the thin aluminium tinsel and plastic tree with such delightfully high surface area and lustre that help make the ongoing housefire that is the human experience just a little prettier and a whole lot faster it is my dearest hope that all of the wonderfully kind and understanding pets got out in time, as well as whatever else happens to be hiding in our walls and ceilings before air quality and blistering heat melts us like cheap wax buddha candles into a new type of calm only expressible in fiction
|
|
#
?
Jun 23, 2023 15:51
|
|
|
|
| # ? Apr 29, 2024 10:52 |
|
|
haveblue posted:which problems? because certainly not any problems in which you take the AI output and turn around and apply it to another domain It's pretty good at image classification, like spotting dolphins and such for large boats so they aren't rolled over. False positive and false negatives don't have huge ramifications, but it's nice to try.
|
|
#
?
Jun 23, 2023 15:51
|
|