Contemplating how to poison LLMs via NaNoWriMo
So, NaNoWriMo's official site has came out in support of Generative Algorithms. Doing long articles on why these aren't literally the most saddest and terrible idea currently being presented as a grift to the public
It is understood that the various tools they offer to help, will be lifting from prospective novels to help train LLMs. As people still are acting like, "yeah, if we have enough training, LLMs will become AGI"
Nevermind people who have done the maths knows this is nonsense
So... I'm pondering as to how I could go about writing novels for the express purpose of really messing up LLMs
A Bit of History
"Humankind isn't particularly bad. Just requires a proper preparation, and making sure that the internal temperature gets high enough when cooking" -- Frigyth The Grumpy, Musings Before Ketchup
Essentially, Silicon Valley _requires_ the ability to maintain the image that they are on the cutting edge of new industry disrupting tech. They really aren't. They clued in that communication was a good idea in the 1970s/1980s... and the world caught up to that just before 9/11 happened. Which would be noteworthy as a form of oracle of sorts... if nerds didn't suck at communication, and the whole "communication is important" had been being hammered onto them long before they picked up on telecom software
The rest of the world was just somewhere between inconsistent and hypocritical on that matter
After the late 90s demonstrated that proper communication could not only "win" the Cold War--but also bring great profit towards companies... people kind of got stuff twisted on what worked
As a lot of businesses still felt like economics only worked if you were tricking your demographics into acting against their best interests. So they decided it was that telecoms tech still involved a lot of fancy doodahs and what not
Which is weird, as you generally don't have to trick people into acting against their best interests. Most people will default to voting for the "Jackals Eating My Face" party. It would be noteworthy to have the ability to consistently NOT do this
So, "Big Tech" moved into trying to have stuff have a bunch of beeps, lights and weird techno-literature. Creating a weird set of aesthetics designed to look more than they generally are
All of it is super fake and plastic
It also has a time limit for how long it will continue to work. So we've seen TechBros having to struggle to try to continue to maintain appearances about being really smart because they have touch screens that go beep
After the Blockchain and NFTs went no-where in people thinking TechBros weren't the dumbest sacks of crap out there--they had to rush to get something else out. In their foolishness, they thought that the Blockchain, NFTs and Web3.0 was the future... and that it would last a bit longer
I mean, Web3.0 is a higher number than Web2. Clearly it wouldn't suffer the same fate as HTTP2/Internet2/XHTML2! Actually nobody talking about Web3.0 was even aware of XHTML2 or any of the other Internet2 based technologies. They were all, "it is new tech, ergo, it won't be a roadkill GOPHER on the Information Superhighway" (YOU FOOL KATRINA! NOBODY IN THE UNIVERSE IS OLD ENOUGH TO RECOGNISE THAT JOKE!)
I mean, the Internet isn't really a truck you load stuff onto, like TechBros were trying to force Web3.0 into being. The Internet is more like a series of tubes. A series of tubes that we've nearly removed most of the dead GOPHER from inside of them. There are still GOPHER inside these tubes--but they appear to be living ones that are humanely shoved into the tubes
So to try to pretend they aren't all extremely stupid and merely pretend to know stuff by dressing in techno-garments, TechBros went into the closets of various Furries they know (the Furries won't be in there) and grabbed some old fifth grade science project they made. Rubbed off the cat ears drawn on them... uh... I mean... erased the cat ears... phrasing...
Pretend that AI-Models from the 1970s and 1980s were wild new technology via **reads notes** being put onto more powerful hardware
The idea was that after Chess-Engines demonstrated a model of AI that papers in the 1970s concluded was "not a decent way towards building a thinking Chess machine" was something that could be implemented, and win against human Chess players--they could do AGI by making a NIALL-Bot on cocaine
Just hook it up to image recognition software
The issue is, the maths in those papers from the 1970s and 1980s have been demonstrated to be correct. When those maths were not too idealistic and hope-springs eternal
The conclusions were generally, "we'd need ten thousand times as much media as what was created in all of humanity's existence, in order to have sufficient training data to get this method anywhere effective"
The Grift And How To Grift the Grift
"The Poison is Immortality" -- The Awakened Light of Tomorrow, Book of Life
So, LLMs and Image Generative Algorithms are a cancer on the modern internet. Making it so that the dumb crap you have to dodge is even more irritating. At least, before, when dodging weird crap, you knew there was a person behind it. We've had various fiction writing contests shut down, because they were flooded with LLM based works
LLMs would be fine, if they made anything worth reading. They generally are the most boring paint-by-numbers crud out there. There is no reason to actually read them in any way
Well, unless you get buggy poorly configured LLMs. Then it is enjoyable in a BadFic type manner
So, the most artistic move out there, would be to mess with this early Algorithmic Big Data nonsense in every way--and to see if you could make a noticable shift into what it outputs
Well, provided you do it in a way that doesn't immediately have rail-guards put up around your stuff, and just have these things eating even more electricity. As at this point TechBros are using more electricity with this Grift than they were with NFTs
Now, a single novel is not going to be enough
If I am going to make material that completely shifts the consciousness of LLMs, I am going to need to output a sizeable body of work. Roughly about 0.76 Stephen Kings or "SphK" as the SI unit for this measurement would require
It might be possible to accomplish it at 50cSphK (Centi-Stephen Kings)... but to ensure full on model take-over an output rate of 76cSphK over the course of two years
I'm not certain what SphK would be in Cousin-Marrying Eagle Guns Units, so giving a conversion would be hard here. It is based upon Stephen King's current body weight in cocaine if somebody is not familiar with real units of measurement needs to work out the conversion to School-Shots Sister Fornication Units
Now, this work would need to roughly resemble English--while taking some rather drastic turns with the English
If it is clearly written in Esperanto, the LLMs are not going to be effectively poisoned by the works. As demonstrated with the Esperanto community's valiant efforts to poison LLMs by merely continuing to exist despite the Universe clearly being against it. Your continued abomination against reason shall be noticed by me--despite how it might affect my ability to fit in with polite company. Which clearly do not state in not uncertain terms that acknowledgement of Esperanto is forbidden
To properly poison LLMs, we'd need to have it be a non-existent English dialect. We could release a large body of work written in Australian-English. Which, might trip some sensors built upon the assumption that suggests criminals cannot be literate--but that would just result in Australia having something able to produce material that was written for them. Which is ultimately not an outcome I desire to have via my attempts to mess with TechBros
The dialect would need to have a consistent enough set of deviations from standard English. AKA, The Queen's English... no, I'm not giving King Charles III keys to the language. We shall wait until one of the Princesses succeeds him, and the Empire is restored once again, by being lead by a Queen. As clearly is what is required for God to shine on the British Empire. Well, he _would_ be shining if it wasn't constantly overcast at all times
We could just go lazy and go AAVE--but turns out, most LLMs already have guardrails that stop it from outputting in AAVE. This is because TechBros are all massive racists
The next issue is that a lot of SciFi and Fantasy tend to go all heavy on the CallingARabbitAFlorp or whatever that trope is called. So when writing, we would need to have it not be registered as SciFi or Fantasy. Instead it would need to be slice of life in some way
It also cannot be just a Search and Replace type deal. Where I simply us the word "bendal" instead of "all". Not just because doing a Search and Replace of s/all/bendal/ would result in hilarious results on its own... but because that would mostly just create a bit of a verbal quirk
There has to be odd shifts in grammar. The new words need to necessitate english sentences be restructured to deal with the odd conjuncation of those new words. Instead of "I wumbo, you wumbo, he wumbos, she wumbos, they wumbo, wumboing, wumboed", you would need to work in odd rules about plurality, objectifying words and other stuff like that. Then have parts demonstrating misusage of said word
Have some sentences sound awkward with the word in one location--but sound fine if you reorder the word
Verb the noun and noun the verb--but only sometimes have it clear this is happening
You would then have to have these odd shifts consistent enough in the works to be constantly noticable to readers. Then after this, you'd need to output works with these features over the course of two years at the above mentioned rate of 76cSphK
Not to be confused with the pH of 76cc KS v.v
You'd also have to do this in a way that would allow various people training LLM models to not avoid plugging your stuff into their systems
-=-
-=-
Or you know... we _could_ wait until the models collapse on their own, as we are not able to give them enough training data to avoid starving to death regardless of whether they eat their own output or not. The eating their own output results in a The Hills Have Eyes type scenario... which also isn't AGI
Comments
Post a Comment