Aww shit, here we go again EditionLast thread: >>497218312AI Dynamic Storytellers are games that use text & image-generation AI to create limitless, open-ended, interactive fiction.Last Thread: >>496920186 Read First: https://i.imgur.com/lwWTTp9.pngWarning: https://rentry.org/AntiwhineMkIII▶Ways to WriteAI Roguelike — Paid; local or remote, multi-model: https://store.steampowered.com/app/1889620Featherless — Paid; local or remote, multi-model: https://featherless.aiKoboldAI — Free; local or remote: https://rentry.org/cixpvn93NovelAI — Paid; remote SaaS: https://novelai.netOpenRouter/TogetherAI — Paid; remote SaaS, multi-model: https://www.together.ai ; https://openrouter.ai>Chat Models — e.g. AetherRoom, Claude: https://rentry.org/chataidsv3 ; >>>/vg/aicg>Local Models — e.g. Command R, Mistral, Qwen: https://rentry.org/cixpvn93▶Tools & ResourcesWiki: https://aids.miraheze.orgPrompts: https://aetherroom.clubNAI Guide Collection: https://rentry.org/guide-collectionOther/Historical: https://rentry.org/aids-miscellany>AID: https://rentry.org/remember-what-they-took-from-you▶/aids/ EventsThemed Prompts: https://aids.miraheze.org/wiki/Theme_FridaysThemed Decks: https://aidsrentfree.github.io/decks>Write prompts for this week’s theme, Underwater Caretakers Friday! Results: >>497180762 ▶News>(10/02) Storypad: Novel-style UI reconstruction https://gitgud.io/mir/storypad ; OpenAI introduces Realtime API https://openai.com/index/introducing-the-realtime-api(09/25) Meta releases Llama 3.2 https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices(09/23) NovelAI releases Llama 3 Erato https://blog.novelai.net/muscle-up-with-llama-3-erato-3b48593a1cab(09/20) NovelAI announces Llama 3 Erato’s release window, new textgen samplers, and the removal of CFG https://files.catbox.moe/ct12an.png
Now that the dust has settled, how is Erato?
(crosspost by the way)>>497485178>>497485056i'm with you two about the "8K". yes it was that maddening due to the short number of the context; achieving the average word count of a novel, about 70 000-100 00-ish, which necessitate in excess of "32K" and more, would considered a "novel" co-wrote with ai. although, i have this one curious thought: does the number of context affect the server/clusters' speed and response?
>>497495791More context = more space to store all the text = more expensive to run
stories?
Yes
>>497495251dogshit
Now that the dust has settled on the already settled dust.
>op was posted six hours ago and only 7 replies.Dead hobby.
I'm a newfag cooming at home using koboldcpp. I'm having a good time with Mistral 7B Instruct (no fucking clue if all of that is relevant), but it's already over a year old and I'm wondering what I might be missing. Are there newer, better models that run on potatos, or has it all shifted to giant model paypigging?
>>497543818try magnum 12b if you can run it or lama 3 8b stheno.also this is the real thread >>497493950
>>497546498This is the real thread.
For people complaining about context length, you can always increase it with rope scaling (assuming you're running a backend that supports it and you're able to manually plug in the command line arguments for the server yourself). You can get a model with native 4k context up to 16k but it does break down and get a little stupid around the ~12k mark
>>497543818Check out Fimbulvetr or Silver sun. They're 11B but are able to outperform 13-20B models according to some. My machine's capable of running those bigger models but I stick with Fimbulvetr anyway because it gets me almost instant responses for similar quality.If you want ways to reduce how fat the model is so you can fit it on the GPU, you can get models with heavier quantization and imatrix quants. Should reduce the size of them without too much quality loss unless you go below Q4 quants.Once you're less of a newfag you can abandon koboldcpp and try setting up llamacpp for even finer control.
>>497495251This