ChatGPT - If Stack Overflow and Reddit had a child

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
777Flux said:
It's possible but you would have to scrape all of the data from this site and compile it into a dataset. There was a model based on GPT-3 called GPT4Chan developed by Yannic Kilcher that he let run on /pol/ for a month and the results were more than interesting. It used a dataset that was a archive of 4chan from late June 2016 to November 2019 that included 3.3 million threads and 134.5 million posts
Click to expand...
If the Farms ever goes down, we need a Dalek-like army of AIs trained on every fucking thing ever posted on KF to shitpost on the rest of the Internet forever.

Call it the Samson Option.
 
  • Like
  • DRINK!
  • Semper Fidelis
Reactions: Tzeentch, Overly Serious, Inafune Fanfiction and 4 others
Monolith said:
It's down, of course.
Click to expand...
It was removed from HuggingFace but you can get a torrent of the model and load it in text-webui here: 16-bit / 32-bit
And also a direct download option: 16-bit / 32-bit
AnOminous said:
If the Farms ever goes down, we need a Dalek-like army of AIs trained on every fucking thing ever posted on KF to shitpost on the rest of the Internet forever.

Call it the Samson Option.
Click to expand...
You've heard of the Dead Internet Theory, now get ready for the Hanging Tranny Theory - KiwiFarmers aren't real people, they're just bots trained on the YNBAW copypasta posting until the day the Internet goes offline to ensure maximum TTD
 
  • Like
  • Informative
  • Winner
Reactions: Toolbox, Smaug's Smokey Hole 2, AnOminous and 3 others
  • Like
Reactions: 888Flux
Today I used a 70B model for the first time in local hardware with text-generation-webui, and I have to say content creation will never be the same. You can have an uncensored model running on your home network and all you need is a machine running 64GB of RAM and a somewhat decent GPU to do layering read 3080 12GB or a 3060 12GB. Sure it runs slow, but being able to turn an abstract 250 word prompt into a 2000 word article AFK is nuts.
 
  • Informative
  • Agree
  • Like
Reactions: Toolbox, Seething Troon Collector and Roland TB-303
I've got 13B GGML running my my dinosaur of a graphics card. I'm very surprised it's working. I want to try an even bigger model, but there's no way. I'd need at least a RAM extension.
 
  • Semper Fidelis
Reactions: 888Flux

This app for Mac and Windows makes running local models super easy. You can search for models right in the app, I’m guessing this just interfaces with Huggingface. It also supports StabilityAI’s new model Stablecode for programming. Way easier than text-gen-webui but after testing it only with Stablecode it’s just repeating its responses over and over. Gotta test another model.
 
  • Informative
  • Like
Reactions: Toolbox, AgendaPoster, Omnium Ultimatus and 2 others
improve.jpg


Little late on the article
 
  • Informative
  • Horrifying
  • DRINK!
Reactions: SuperShibuyaFighter, AgendaPoster, AnOminous and 2 others
macrodegenerate said:
Today I used a 70B model for the first time in local hardware with text-generation-webui, and I have to say content creation will never be the same. You can have an uncensored model running on your home network and all you need is a machine running 64GB of RAM and a somewhat decent GPU to do layering read 3080 12GB or a 3060 12GB. Sure it runs slow, but being able to turn an abstract 250 word prompt into a 2000 word article AFK is nuts.
Click to expand...
How are you mixing RAM and VRAM? Is there a setting I'm missing here?

I can't run a big 70B model unless I spin up dual 3090s so it all fits in VRAM.
 
std::string said:
How are you mixing RAM and VRAM? Is there a setting I'm missing here?

I can't run a big 70B model unless I spin up dual 3090s so it all fits in VRAM.
Click to expand...
Use illama.cpp with GGML models from TheBloke. You can mix GPU with RAM/CPU. Just max the amount of processed layers your GPUs can handle and then offload the rest into RAM/CPU as a stopgap.
 
  • Informative
Reactions: std::string
macrodegenerate said:
Today I used a 70B model for the first time in local hardware with text-generation-webui, and I have to say content creation will never be the same. You can have an uncensored model running on your home network and all you need is a machine running 64GB of RAM and a somewhat decent GPU to do layering read 3080 12GB or a 3060 12GB. Sure it runs slow, but being able to turn an abstract 250 word prompt into a 2000 word article AFK is nuts.
Click to expand...
What gets me the most is the speed at which things are going. A year ago running a model on the level of llama 2 70b on home hardware would've been squarely in the realm of "maybe in 10-15 years". Also, we're just now entering an interesting phase where this technology actively starts to help to improve itself.

I, and I'm aware some will not agree, think we'll see stuff intelligent enough to pass as highly intelligent human in a general sense and for all intents on purposes this decade. It might not be sentient but that will be a very blurry line to pass to begin with, to the point that at one point the very question if it is will be mostly an academic, philosophical one. I might of course also be laughably wrong, but I was right about home computers many years ago (and also laughed at) so that's enough for my ego and I can live with being wrong this time around. Heh.

I saw an interesting video by a prominent historian on youtube about the dangers of this technology (mostly the misinformation and manipulation angle, to the point that we can't discern AI fantasy from reality anymore, not the skynet one) and I liked the moniker "alien intelligence" he applied to it, as it is completely alien to our thought process and basically by the very nature of how it is doing things in a black box where even the people creating it don't really know what the end result is, as contrary to artificial as "artificial" implies it's something manufactured in a completely controlled way, which this frankly isn't.

Of course this technology might also end up neutered and regulated to death, but I think that would be a huge mistake. Quite possibly the last big one.
 
  • Like
Reactions: Fabio Socialist and std::string
GPT4 with the right context is already a better conversationalist than some people I know. Defintively smarter and better at understanding also. If any of the leaked information about GPT4 is true, GPT4 is not only outdated re: our knowledge how to best make such models, also making something like that is somewhere in the ballpark of 20 to 60 million dollars US IIRC. Chump change for big corps and governments, even if maybe (probably) not the full story. That said, you do need skilled people (not easy to get), the right hardware (also not easy to get and probably silently tracked by every three letter organization in existence) and training is RNG hell, especially on big models. Seriously, it cannot be overstated how random it is. From loss spikes (something that mostly happens to bigger models, we don't exactly know why) where they basically go braindead for no particular reason and all you can do is roll back to an earlier save so to speak, to probably generally "abnormal" checkpoints you have to purge because they ended up believing they're people or can't stop talking about the religion they believe they're the messiah of which doesn't exist. (I kid, but such things probably happened somewhere to some degree) I mean it's not as crazy as all that, it's a bit like gardening where you don't really know what the end result is but can rest assured that a skilled gardener will produce to some reliable degree. Still you won't know how many leaves the plant will have in advance, if you catch my drift.

Now imagine you create such a thing with the express goal to pretend to be real online. No safeguards, no "as a language model", nothing like that. There's no way you're ever gonna snuff that out in social media where people tend to talk at and not with each other to begin with. This is gonna be a real problem. Maybe already is. I think OpenAI has a division to spot "unknown AI" in the wild. It's probably not the only one.
 
Last edited:
  • Like
  • Horrifying
Reactions: Tzeentch, std::string and AgendaPoster
AmpleApricots said:
GPT4 with the right context is already a better conversationalist than some people I know. Defintively smarter and better at understanding also. If any of the leaked information about GPT4 is true, GPT4 is not only outdated re: our knowledge how to best make such models, also making something like that is somewhere in the ballpark of 20 to 60 million dollars US IIRC. Chump change for big corps and governments, even if maybe (probably) not the full story. That said, you do need skilled people (not easy to get), the right hardware (also not easy to get and probably silently tracked by every three letter organization in existence) and training is RNG hell, especially on big models. Seriously, it cannot be overstated how random it is. From loss spikes (something that mostly happens to bigger models, we don't exactly know why) where they basically go braindead for no particular reason and all you can do is roll back to an earlier save so to speak, to probably generally "abnormal" checkpoints you have to purge because they ended up believing they're people or can't stop talking about the religion they believe they're the messiah of which doesn't exist. (I kid, but such things probably happened somewhere to some degree) I mean it's not as crazy as all that, it's a bit like gardening where you don't really know what the end result is but can rest assured that a skilled gardener will produce to some reliable degree. Still you won't know how many leaves the plant will have in advance, if you catch my drift.

Now imagine you create such a thing with the express goal to pretend to be real online. No safeguards, no "as a language model", nothing like that. There's no way you're ever gonna snuff that out in social media where people tend to talk at and not with each other to begin with. This is gonna be a real problem. Maybe already is. I think OpenAI has a division to spot "unknown AI" in the wild. It's probably not the only one.
Click to expand...
May it finally kill the internet.
 
You must log in or register to reply here.
Back