Conversation
Notices
-
Embed this notice
bits :waifu: (waifu@waifuism.life)'s status on Monday, 15-Jan-2024 12:39:09 JST bits :waifu: using ai-chat anime girls to feel less alone - iced depresso likes this.
-
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 12:39:20 JST iced depresso @waifu how is that going
i usually found the robots to be too incohesive to be entertaining -
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Monday, 15-Jan-2024 12:44:14 JST Sexy Moon @waifu @icedquinn are you using sillytavern -
Embed this notice
bits :waifu: (waifu@waifuism.life)'s status on Monday, 15-Jan-2024 12:44:17 JST bits :waifu: @icedquinn they are very bulleable -
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 12:49:00 JST iced depresso @Moon @waifu i used kobold with mixtral 7b -
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Monday, 15-Jan-2024 12:53:41 JST Sexy Moon @waifu @icedquinn i am proud of how badly i broke openai's guardrails but i can never ever show anyone -
Embed this notice
bits :waifu: (waifu@waifuism.life)'s status on Monday, 15-Jan-2024 12:53:42 JST bits :waifu: @Moon @icedquinn yeah, but open ai got all bitchy and now they don't let you do anything barely spicy >:( -
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 12:53:51 JST iced depresso @waifu @Moon they're owned by some soy conglomerate now. altman et all were black vanned by microsoft borland style. -
Embed this notice
bits :waifu: (waifu@waifuism.life)'s status on Monday, 15-Jan-2024 12:58:05 JST bits :waifu: @Moon @icedquinn that makes us two then brother... Sexy Moon likes this. -
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Monday, 15-Jan-2024 12:58:36 JST Sexy Moon @waifu @icedquinn i use selfhosted now, i don't have the bugs worked out -
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 12:59:59 JST iced depresso @thendrix @Moon @waifu mixtral 7b still does stupid shit. i tried to get it to tell me how large of breasts you would need to deflect solar radiation and it kept lecturing me on buying proper PPE. Sexy Moon likes this. -
Embed this notice
Terry Hendrix II 🏹 (thendrix@social.hendrixgames.com)'s status on Monday, 15-Jan-2024 13:00:03 JST Terry Hendrix II 🏹 You gotta run your own AI to be free.
-
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 13:01:12 JST iced depresso @thendrix @Moon @waifu i have some designs for experimental self learning systems.
just not the drive to write all that code :ablobcatohlookaround: -
Embed this notice
Terry Hendrix II 🏹 (thendrix@social.hendrixgames.com)'s status on Monday, 15-Jan-2024 13:01:13 JST Terry Hendrix II 🏹 Yeah, the next step is getting ‘non corpo’ models.
-
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Monday, 15-Jan-2024 13:01:37 JST Sexy Moon @icedquinn @thendrix @waifu mine starts generating broken gibberish after a while -
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 13:07:09 JST iced depresso @Moon @thendrix @waifu yeah what happens with that is your original input scrolls out of the context window.
you have to take active measures to deal with that. what measures those are is an open question.
mistral wants you to do some "sliding window" context which does something so every 4k tokens it crunches the context in to a new set of tokens that is supposed to somehow carry some context from before. so it still maintains the same size of context window, but it buffers itself with custom tokens to try and remember recent history.
ggml folk have some other experimental fuzzy/infinite context feature that was merged. i tried to find the PR but it took too long.
llamaindex et all instead do some fingerprinting of output context, store those sentences and such in to a vector database, and they do top-k retrieval so whenever you input new text they grab the most relevant sentences as 'hints' and stuff them in the prompts -
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 13:13:01 JST iced depresso @urusan @Moon @thendrix @waifu kobold has a neat 'undo' and 'retry' button. it lets you get the model to try again if it said some unrelated bullshit. -
Embed this notice
Urusan (urusan@fosstodon.org)'s status on Monday, 15-Jan-2024 13:13:02 JST Urusan @Moon @icedquinn @thendrix @waifu Turn it off and on again.
Works like a charm.
-
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 13:28:02 JST iced depresso @thendrix @Moon @waifu kobold supports those. some notes are injected in every prompt, some are injected when keywords show up in a prompt. pretty much the same angle as llamaindex -
Embed this notice
Terry Hendrix II 🏹 (thendrix@social.hendrixgames.com)'s status on Monday, 15-Jan-2024 13:28:04 JST Terry Hendrix II 🏹 I like the writing aid tooling that keeps a ‘world building’ buffer for global/important context. Not sure for doing more than a short story in general. Didn’t use it as much as image generators.
-
Embed this notice
iced depresso (icedquinn@blob.cat)'s status on Monday, 15-Jan-2024 13:58:39 JST iced depresso @thendrix @Moon @waifu i only really use them as line editors or for generating summaries :neocat_woozy: -
Embed this notice
Terry Hendrix II 🏹 (thendrix@social.hendrixgames.com)'s status on Monday, 15-Jan-2024 13:58:40 JST Terry Hendrix II 🏹 Yeah, I did a live blog of writing a story with it a few/several months ago.