|
Gobbeldygook posted:ChatGPT has a context limit of 4096 tokens or about 3000 words. It manages it FIFO style, so at 4100 tokens it forgets the first four tokens of your conversation. You must periodically reprompt the model with old information you want it to continue considering it, or develop a fine-tuned model. The new GPT-4 models offer up to 32k tokens of context. Also, for more context on how people use it, NovelAI and AI Dungeon and the like work around this by having a context system that just invisibly injects text that's repeated every time you generate text. This can vary from just adding plain text that primes the LLM, up to NovelAI's elaborate (but poorly-documented) system of tags and attributes that its models were trained on in association with given texts to cut down on the amount of tokens taken by the context.
|
#
¿
Mar 23, 2023 00:29
|
|
|
|
| # ¿ May 22, 2024 18:43 |
|
|
Communist Thoughts posted:E I don't think it's a "lost argument" just a pointless one, AI is already quarantined to this thread but that's not enough for these people? Just tell them to gently caress off imo.  Also, if you're going to ban low effort AI posting, why is low effort non-AI posting like this post fine
|
|
#
¿
Jun 9, 2023 09:58
|
|
|
kas posted:I fuckin love that they keep repeating in that thread that AI art is art theft and that their 'solution' to that problem is to steal art directly from those artists instead lol. That right there tells me they don't actually give a gently caress and it's just the latest internet thing for them to try to feel self righteous about. Yeah, I'm pretty  that the arguments seem to overwhelmingly come down to one of:- stealing is bad - then, when the goalposts shift because everyone running tradgames steals art for home use all the time: stealing is bad, but only when you don't credit the artist - then, when the goalposts shift because crediting art in your home game is stupid and/or you can stick to unarguably licensed/public-use sources like Firefly: using generative art is bad because it tangentially normalizes the replacement of real artists and you are obligated to only ethically participate in society It's a good thing this dead comedy forum never has threads about chocolate or cars or Apple devices anything like that
|
|
#
¿
Jun 11, 2023 06:52
|
|
|
Abhorrence posted:The trick, I have found is to A: don't ask it to do something it has refused to do later in the thread, instead edit the question that made it refuse, and B: explicitly ask it to include a content warning. For whatever reason, it is more willing to write about stuff if it gives a content warning. The key thing to understand is that it will try to generate stuff that's like its reinforcement training data, and from inference we can tell that the OpenAI reinforcement training data probably has: - Lots of text of the bot endlessly refusing to concede to arguments about all the "as an AI language model" bullshit - Some amount of text about "mature" topics with content warnings
|
|
#
¿
Jun 26, 2023 19:46
|
|