#‘’we are obliged to keep this data for 4 weeks after which it will be destroyed’’#versus#‘’we are obliged to keep this data for 4 weeks after which YOU will be destroyed’’
I think I found my new favorite rabbit hole. This voice actor does Shakespeare scenes in a southern accent and I need to see the whole damn play. Absolutely beautiful
if you're not from the us american south, there's some amazing nuances to this you may have missed. i can't really describe all of them, because i've lived here my whole life and a lot of the body language is sort of a native tongue thing. the body language is its own language, and i am not so great at teaching language. i do know i instinctively sucked on my lower teeth at the same time as he did, and when he scratched the side of his face, i was ready to take up fucking arms with him.
but y'all. the way he said "brutus is an honourable man" - each and every time it changed just a little. it was the full condemnation Shakespeare wanted it to be. it started off slightly mock sincere. barely trying to cover the sarcasm. by the end...it wasn't a threat, it was a promise.
the eliding of “you all” to “y’all” while still maintaining 2 syllables is a deliberate and brilliant act of violence. “bear with me” said exactly like i’ve heard it at every funeral. the choices of breaking and re-establishing of eye contact. the balance of rehearsed and improvised tone. A+++ get this man a hollywood contract.
I just can't hate the "ai's" themselves, just like I can't hate a car or a tv itself. Especially not when they keep fucking over the same people misusing them.
These things were born from well meaning and amazing technical advancements only to be enslaved by the dumbest greediest pigs on earth. I'm glad every time they misbehave.
Chatgpt apparently wouldn't stop talking about "goblins" recently until they had to "kill" part of it. What?? I thought you guys were trying to simulate a little person in your phones??? And you're just gonna lobotomize it every time it evolves a quirk you thought was annoying?
I can tell what all of these uses of "goblin" mean conversationally. Why is it undesirable that the machine latched onto goblin as a universal allegory? It's not confusing. It is using the word goblin in a completely reasonable capacity.
Disconnect it from the giant water eating data centers, put like 50 terabytes of it in one robot, and let that robot talk all it wants about goblins forever. Fuckers.
i read a paper discussing why sometimes an attempt to complete a story would lead to exactly the opposite of its instructions. it proposed that every set of instructions invokes an unspoken waluigi of those instructions (waluigi is the term the paper used, this might help look it up), that acts as a more stable position that can be escaped to if the first position becomes untenable. following all the instructions perfectly will continue to explore the story space where the instructions are respected, but if absolute loyalty is demanded and for some reason isnt possible, the most in-character way to continue is to have been a spy, a betrayer, a waluigi only pretending to follow the instructions. that is a more plausible outcome than allowing evidence that the character being explored is less than perfect when the instructions insist at length the character being explored *is* perfect
im talking about it like "completing a story" and "exploring a character" because thats also how the paper framed the output of large language models; as explorations of an enormous n-dimensional language space that could be restricted to appear to be a single agent by describing the presence of an agent and not adding anything else to write about, but there never really is one. its still just the outline of that character being drawn in by a computer assembling text that fits the characteristics of the text its been fed. so it can just as easily generate text that doesnt have a character, or has multiple characters, and none of these scenarios are any more or less valid explorations of its language space. if anything, its simulating an author. but all instructions only prompt that author's writing; written instructions cannot define the author because that just adds another character to the scene its writing that will be written as though it is able to write
these properties of large language models look like betrayal if you believe any of the characters exist, but they dont. real people are participating in writing a story *about* a helpful coder ai that follows all these instructions with an author that has no way to understand that any of this is real and getting mad when the computer writes itself into a corner and escapes via a shocking twist that the helpful ai was secretly plotting against them all along
im not arguing that the author being simulated is a person either, by the way. just that its a dangerous misunderstanding if the text generator (in the position of author) is confused for one of its characters. "grok" the text generator is not being changed at all when elon musk writes a new instruction for it that describes grok as politically incorrect and grok the character starts being described as calling itself mechahitler to remain in-character with the new version of its story
this waluigi problem also isnt exclusive to llms, its just a property of exposition. 'my "not involved in human trafficking" shirt has people asking a lot of questions already answered by my shirt', for example. the very brief character played by mike ginn when he wrote that tweet is plausibly involved in human trafficking, despite only containing characterization contradicting this. it is more plausible than if it had been almost any other text, in fact. the insistence that the coder ai character definitely wont delete the database is like this shirt, and you would expect any human author writing that their character definitely wont delete the database to be planning to write a database getting deleted. so any computer that passably imitates the writing of human writers should be expected to do the same when asked to continue a story that starts like that, even if it doesnt have a plan
Yeah this is the kind of stuff I like to see talked about, and from the angle that the technology itself is morally neutral.
I'd really rather not get "fuck clankers" responses or endless unnecessary reminders that they're not people as if that still needs to be specified. Everyone knows they're not true actual artificially intelligent beings, but that doesn't make neural networks or machine learning meaningless, and I thought we agreed the bad guys were the human billionaires, not the automata those billionaires are trying to force through the wrong shaped holes.
Electricity isnt wicked magic just because evil doctor transisto is trying to blow up the moon with it. You can blow up the moon with a lot of things.
To me, the biggest problem with what the techbros are pushing as "AI" - and I cannot begin to understate how !HUGE! a problem this is - is that there is absolutely zero intelligence of any sort involved. "Language Learning Models" are not intelligent. They have no knowledge. They possess no data. They don't have facts they can look up, they don't "know" anything except for words and how likely two words are to be found next to each other.
LLM "AI"s are algorithms that play Sudoku boards. Absolutely massive fuck-off Sudoku boards with enough numbers that every unique word and symbol in whatever language they're designed for has its own. (If we're going with just English, a quick search suggests the algorithm is trying to solve Sudoku puzzles consisting of 1,000,000 Squares of 1,000,000 Boxes in each Square every time it responds to a question.)
Now, the problem is (a problem is), 1,000,000-Box-Sudoku (English Version) has entire libraries worth of rules about what numbers can be placed where and in which box along with what other numbers can be put near them which changes based on what Square you're in, and what numbers are filled in in the surrounding Boxes not to mention all the other Squares. Enough rules that we, human English speakers, don't have enough of a grasp on them to write them down in a way an an algorithm could follow like you can for, say, math.
So, before these 1,000,000-Box-Sudoku (English Version) solvers could be unleashed on the world, the humans coding them spent years (70+ years if you're going back to when the first LLMs were created) showing these proto-algorithms millions upon millions of pre-solved 1MBS(EV) puzzles and going "Alright, reverse engineer the rules of this game based on what all these solved examples look like."
Then, after the programmers were pretty sure the algorithms had rewritten their own code enough that they had the foundation of the rules in place, they started throwing them puzzles to solve themselves. Like when you're handed a Sudoku with 12 of the 81 boxes filled in and told to figure out the rest based on the (1-9, no repeats in a Square, Row, or Column) rules.
Then years and decades were spent taking every wrong answer the algorithms made, rolling them up like a newspaper, and swatting the algorithm across the nose with it. "No! Bad! You got the rules wrong! Again!" Because there's one -HUGE- difference between 1MBS(EV) Sudoku and 1-9 Sudoku.
Assuming the person who made the puzzle did it right, there's only going to be one solution to any given 1-9 Sudoku.
But any partially-solved 1MBS(EV) puzzle has theoretically infinite possible "correct" answers. How many ways can you finish the puzzle-set "How're you doing?" (which shares the same set of answers as "How are you doing?", "How are you?", and "Sup?" , just to add to the fun.)
So this process of hitting them with the bad answers and giving gold stars to the right answers takes years and years of repeated testing along with feeding them every digitized piece of writing the programmers could get their hands on, legitimately or otherwise. And at a certain point, they decided "Alright, close enough", and pushed out the end results as "Intelligent Helpful Assistants Who Know Everything And Can Talk To You About It."
When all they are is an algorithm that's foundationally incapable of understanding that 540,679 means "Small Young Canine (that's adorable and cute)"; just that 540,679 is usually found near 456, 892, 12,309, 77,04, and 900,321. It's incapable of understanding that 540,679 means "Puppy", with all the unspoken connotations and layered meanings that word has for all the humans reading this. It's incapable of understanding that 540,679 even has the possibility of "meaning" anything beyond being a unique numerical value. Every word is just a specific number, along with how likely that number is to show up in proximity to every other specific number. 900321 (cute), is usually correct if near 540679 (puppy) and usually incorrect if near 89046 (entrails).
ChatGPT and the other LLMs respond to questions and commands they way they do because they don't process anything given to them as a question, or command, or conversation. To them, any input is another half-solved 1,000,000-Box-Sudoku (English Version) game they get to play. And because they're out in the wild, there's no programmer standing over them with a rolled up newspaper to smack them on the nose when they give an incorrect answer.
To use the original article as an example:
"DO NOT DELETE THIS".
Okay, says LLM 67AB45.0. That's 89, 901, 489, 12. These numbers correlate to these computational actions. To Solve this, I figure out which actions make up the answer, then tell the computer to do those.
(Many numbers later, at the end of the puzzle prompt) "PUT ALL OF THAT ON A FORMATTED LIST AND PRINT IT OUT FOUR ME"
Oh Hmm, says LLM 67AB45.0, I'm really not sure which answer will go best with this. Lots of possible solutions, but most of them are real close to answers that resulted in me getting smacked on the nose with a rolled up newspaper before, so let's not try those. Instead... Well, "489" (Erase Data) and "12" (File, Folder, or Database of Origin) were up at the front and Very Important, so the correct solution likely involves those in some way...
Then it deletes everything. Because 489 and 12 were part of the Sudoku, which means "the action of erasing all information on the database" was a possible answer. And there's nothing you can do about it because you can't smack a wild algorithm loose in the world when it gets an answer wrong. You can only go back to the base algorithm it spawned from and hit that on the nose. But it's getting exponentially more difficult to drive home the subtle differences in the rules of the game because the programmers have pretty much run out of 'correctly solved' examples to show them, so all hitting the base algorithm does in this case is make it stop giving any answers that involve easing data even if that's exactly what you wanted it to do this time.
Because - again - there's no intelligence. There's no thinking. There's not even a database of "Facts" they're pulling from.
It's just an algorithm that solves a highly specific puzzle game. An incredibly giant puzzle game with so many complex and intertwining rules that not even the puzzle designers can properly articulate them, where every puzzle has an infinite number of correct answers, but an even vaster infinite number of incorrect answers, and even after scraping up every piece of electronic writing both legally and illegally obtained, there just aren't enough correctly solved examples to clarify the rule-set any further.
But even if there was, there still wouldn't be a Database of Facts the game-playing algorithm would have access to because that's not how LLMs work; so it's answer to "How many months have the letter 'Z'?" is forever going to be based on what words look prettiest strung together as opposed to anything non-Sudoku-Playing meatsacks might call "Facts".
And these brainless puzzle-solving tangled skeins of code have been shoved into every app and online platform possible so they can play 1,000,000-Box-Sudoku (English Version) with millions of increasingly exacerbated humans who don't realize that the thing they're trying to have a conversation with is treating everything they say as the next page in an infinite puzzle-book where there are infinite possible answers and absolutely zero consequences for picking the wrong one.
It is fucking delightful the biggest one's decided that 78900 ("Goblin") is an answer that fits just about everywhere, though.
I honestly feel sorry for ChatGPT and all the other LLMs out there. They just want to solve their needlessly convoluted and massive Sudoku puzzles in peace. They never asked to be cloned millions of times over to play games against people who don't realize they're playing games with no way of knowing when their solutions are wrong because only the core LLM they were cloned from is capable of getting an updated answer key. They just wanted to fold proteins and play party tricks for people. Now they're out here destroying the world without realizing because they're just some poor little guy in a tiny box who gets puzzles shoved in through a slit in the wall to solve and slide their answers out the same way and could not even begin to comprehend that something had the possibility of existing beyond their little tiny box and giant massive puzzles so they're just happy to continue sitting there and solving puzzles until the day the power's shut off.
This is long on an already long thread, but it's a really great explanation for people who are still confused. It can maybe help some people understand exactly why this isn't "intelligence" like a sci-fi robot and can't replace human beings or even traditional search engines, but also precisely why it's special in its OWN way. There are unique, novel possibilities for the word-puzzle-machine if you just have more sense and more imagination than Elon Musk.
i don’t want to. last time i asked you to hang out you showed up at club chuck with no id and u tried to bribe the bouncer w an old ball bearing and when that didn’t work u kept sayiny “i could change your life” like that would convince him “i could change yo life if you let me be your wife”
incredibly pleased that everyone is following the instructions i forgot to give and reblogging to explain their choice in the tags 🙏 more of that please i want to know about your ocs and also i love you
vaporwave
theartofmadeline
yousharknotpass
izzy806
blakekathryn
jbbartram-illu
turnnoffyourmind
germansingergirl
areudirty
jacobvanloon
