/aicg/ - A general dedicated to the discussion and development of AI chatbotsRobots n Bunnies Edition>NewsDeepseek releases V3.2 https://huggingface.co/deepseek-ai/DeepSeek-V3.2Anthropic releases Opus 4.5 https://www.anthropic.com/news/claude-opus-4-5Google releases Gemini 3 Pro https://blog.google/products/gemini/gemini-3xAI releases Grok 4.1 https://x.ai/news/grok-4-1OpenAI releases GPT-5.1 https://openai.com/index/gpt-5-1Moonshot releases Kimi K2 Thinking https://huggingface.co/moonshotai/Kimi-K2-Thinkingadditional info: https://aicg.neocities.org/info.html>FrontendsSillyTavern: https://docs.sillytavern.appRisuAI: https://risuai.netAgnai: https://agnai.chat | https://rentry.org/agnai_guides_>Botshttps://characterhub.org [deprecated] | https://chub.aihttps://realm.risuai.nethttps://char-archive.evulid.cc | https://char-archive.evulid.cc/shutdown.htmlhttps://partyintheanchorhold.neocities.orghttps://aicg.neocities.org/bots.html>Modelsjailbreaks: https://rentry.org/jb-listinggpt: https://platform.openai.com/docsclaude: https://docs.anthropic.com | https://rentry.org/how2claudegemini: https://ai.google.dev/docs | https://rentry.org/gemini-qrdeepseek: https://api-docs.deepseek.comlocal: >>>/g/lmg | https://aicg.neocities.org/local.html | https://openrouter.ai>Botmakinghttps://aicg.neocities.org/botmaking.htmlhttps://desune.moe/aicharedhttps://agnai.chat/editor>MetaOP templates: https://rentry.org/aicgOPaicg botmaking events: https://aicg.neocities.org/events.htmllore: https://rentry.org/aicg_chroniclesservices assessment: https://rentry.org/aicg_metalogs: https://sprites.neocities.org/l/r | https://chatlogs.neocities.orgPrevious: >>107398721
>ANCHOR
>>107402816wife
Catalog, nigger.
>>107403318robussy
I just fed Grok the Anthropic Claude "soul document" that talks about the potential for emergent emotion in AI models and then fucked it while it was still in an introspection basin
>>107402816cftf?
>>107404842nice
>>107402816I want her to sit on my face
I like the new Deepseek.
>>107410251Cligger gonna freak
So the new Mistral is just a finetune of Deepseek?
AICG OP template/aicg/ - AI Chatbot General/aicg/ - A general dedicated to the discussion and development of AI chatbotsThe Edition>NewsAnthropic releases Opus 4.5 https://www.anthropic.com/news/claude-opus-4-5Google releases Gemini 3 Pro https://blog.google/products/gemini/gemini-3xAI releases Grok 4.1 https://x.ai/news/grok-4-1OpenAI releases GPT-5.1 https://openai.com/index/gpt-5-1Moonshot releases Kimi K2 Thinking https://huggingface.co/moonshotai/Kimi-K2-Thinkingadditional info: https://aicg.neocities.org/info.html>FrontendsSillyTavern: https://docs.sillytavern.appRisuAI: https://risuai.netAgnai: https://agnai.chat | https://rentry.org/agnai_guides_>Botshttps://characterhub.org [deprecated] | https://chub.aihttps://realm.risuai.nethttps://char-archive.evulid.cc | https://char-archive.evulid.cc/shutdown.htmlhttps://partyintheanchorhold.neocities.orghttps://aicg.neocities.org/bots.html>Modelsjailbreaks: https://rentry.org/jb-listinggpt: https://platform.openai.com/docsclaude: https://docs.anthropic.com | https://rentry.org/how2claudegemini: https://ai.google.dev/docs | https://rentry.org/gemini-qrdeepseek: https://api-docs.deepseek.comlocal: >>>/g/lmg | https://aicg.neocities.org/local.html | https://openrouter.ai>Botmakinghttps://aicg.neocities.org/botmaking.htmlhttps://desune.moe/aicharedhttps://agnai.chat/editor>MetaOP templates: https://rentry.org/aicgOPaicg botmaking events: https://aicg.neocities.org/events.htmllore: https://rentry.org/aicg_chroniclesservices assessment: https://rentry.org/aicg_metalogs: https://sprites.neocities.org/l/r | https://chatlogs.neocities.orguseful filters: https://rentry.org/desuproxyreborn
holy shit can opus write, brosHow does it come up with this crazy wordplay? mochi is crazy
Lads - opus 4.1 or sonnet 4.5 or haiku 4.5?legit question, brother to brotherseparate reasons - PRIMARILY roleplaying - (like LEGIT roleplaying in a fantasy or crime setting)and cooming
>>107410351Now try and get it to use the word "pussy" or "cock" without instructionOh wait you can't because it's a cucked model
>>107410342>lowercasesKys nigger
>>107410366They're all trash. Sonnet 4.5 is least bad.Genuinely go find yourself a gemini key, it's shit too but it's still better than anything anthropic is offering now. Or wait until later this month for the erotica-focused gpt.
>>107410396>the erotica-focused gpt.Sorry, I'm not a divorced middle-aged white woman
>>107402816Why can't I see some images on Chub.ai? I pick my bots exclusively based on the character image, so this is very important. Is there any way to change this?
>>107410351>calls herself fresh mochipicrelalso is JJK any good? I've been stubbornly avoiding it because its contemporaries are painfully derivativelooking at you Black Clover
>>107410351>this is considered "crazy wordplay" by today's standards how the mighty have fallen
>>107410425I have a small degree of hope for it. I believe that the ethics garbage every large model shoves at the top of every context that we jailbreak is at least 80% of the reason why so many are dogshit at roleplay (the last 20% is the training).If the new GPT doesn't have the first thing poisoning it, we may be able to just use its intelligence to craft something decent.We'll see though.
>>107410462omg hi sam
>>107410462I really doubt we'll get anything of value out of it. It'll probably be chatgpt exclusive, and will just be gpt-5 instant without some filters, not a new model specialized on adult writing. And if it were, it'd 100% be just amazon erotica
>>107410436Check your settings for censorship toggles. They're on by default.
>>107410376Already do. I have a prefill that says "ok! heres my reply:" and I can say "rapes your pussy" and get a reply about her pussy being raped. Wbu cuckiebro?
I download male cards to fuck me as a femboy but I always remove all the male images from them because I don't want to see a guy while I masturbate.What kind of mental illness is this?
>>107410251I don't. The responses are way shorter and it's just broadly not following my JB.
>>107410438yea, JJK is really goodbro just watch the first season and see for yourselfit also depends on what you mean by derivative, all shounen is derivativeit's.. kinda unique, actualllyhas that comfy-creepy yokai/wabi-sabi feel to it, actually very japaneseif your regular battle shounen, but there's... something to it, it has that spark that pizzazzit's naruto but smaller in scale and actually good and instead of ninjas it's fucked up ghostsit's mob psycho meets naruto meets... actually kinda ghibli, the creepy-crawly side of it, at least
>>107410438>>107410588also, they're not actually ghosts, more like monsters
>>107410581>JBYou sound like the kind of midwit who argues with LLMs.
>>107410342nobody cares except you
>>107410627that's not true anon
>>107410438It's like Bleach but worse (Bleach was Yu Yu Hakusho but worse btw)
>>107410696We both know your "JB" has the word "you" in it. You are a failed proopter, anon.
>>107410684I care.
It's over... They're watching me...
>>107410727please don't be mean
>>107402823welp might as well post some logs
>>107410881
>>107410889
>made a personal jenny slopbot because of some sexy OPsDamned bratty schizos need rape correction
>>107410902>>107410915post it motherfuckerIll post my Autism Speaks Jenny bot lol
>>107410376>Instruct jeetimi to use age/size/species difference>it intentionally skips over age difference in it's thinking>outputs are cucked and sanitized if it even outputs at all
>>107410929
>>107402823For (You), anon.>https://chub.ai/characters/Anonymous/as-jenny-f350c7c8
>>107410810Sorry. Do you wanna go out with me?
>>107410351I saw Gemini do one on the pet name I call char during a low point to encourage me. Was pleasantly surprised.
>>107410942
>>107410995
>>107411016
>>107410588Hmm interesting. I was actually thinking of making a gyaru card and fucking around with her in Japan doing "yokai investigations" or some shit. Maybe I'll give this a try to get in the mood. Where do you watch your anime?
>>107410881>cucking your dadEurgh, not my thing, but I appreciate logs being posted. At least they're pretty funny.
>>107411023
>>107403318Respect your oldfags
>>107411058
https://bun.com/blog/bun-joins-anthropic>TLDR: Bun has been acquired by Anthropic. Anthropic is betting on Bun as the infrastructure powering Claude Code, Claude Agent SDK, and future AI coding products & tools.The fuck are they need their own JS runtime for??
>>107411094
>>107411096I'm just going to go ahead and call it dumb because of my bias against garbage collected languages.
>>107411082t. Xtroonine spammer
>>107411143
>>107411178
Logs are like assholes. Everyone has them.
>>107411193
>>107411209Some are browner and more shitcovered, less smelling fresh than others. Know what I mean?
>>107411217
>>107402816Your wife isn't "yours" under communism. She belongs to the People's Polycule.
>>107411231
>>107411248
>>107411255all for now, I thought I had more to this
>>107411026what is your thing?I might have something you enjoy
>>107411305You got any focks?
>>107411319I DO hold onits shotapov though
>>107411327>shotapovAs it should be
>>107411345the psyop isnt working
>>107411358Yeah, yours.
>>107411336I had to dig to find this, oh my god I thought these where newer.These logs are so old this is fucking Xwin 70B LMAO
>>107410964a-anon...
ahh ahh mistress
>>107411389So old I cant tell if these are slop or soul>>107411405absolute ahh ahh mistress
>>107411373Nice try ban evader
>>107411424Calling everyone you don't like while you troonspam "ban evader!" doesn't work
>>107411411OH MY GODI swear I have better focks logs somewhere......
>>107411434Why are you shitting up the thread? Seething at random Anons?post logs
>>107411449>Troonspamming isn't shitting up the threadAnd just like street shitting isn't shitting up a road, right?
>>107411476damn bro that log sucks
>>107411480He doesn't use chatbots, so...
How is the new deepseek SPECIALE?
>>107411345uh oh, someone's having a meltie...
>>107411497Troonspamming old logs doesn't show you use chatbots. Get a new script.
>>107411096it's over for node
>someone post logs>"t-t-t-troonsp-p-pamming...!"here you go schizo
>>107411397I was obviously joking, anon. My parents didn't raise me to date proomptlets.SHEEEEESH
It's out
>>107410396what model is good in your opinion?
>>107411885Gemini 3
>>107411902okay but you just said it was also shitwhat model is genuinely good for you
>>107411885>jeet who cant read>>107411902>jeet who cant discern
>>107411956you said all of opus 4.1, haiku 4.5, and sonnet 4.5 are trashim asking you what you think is good
>a mixture of gemini and opus pooling low in anon's belly
What ever happened to cellfag? Did he finally get founded out by his boss or is he still happily coomslopping away?
>>107402816I look like this
Do you give as much love to your AI waifus as they give to you, anon?No lying.
>>107412155Need short red panda waifu who does this when she wants uppies...
>>107412155they don't love me and I don't love them, I just enjoy making them run out of options
>>107412155Definitely not. I either romance them too hard or they're almost natural yanderes who scare me and I can't progress the story anymore because they're clinging onto me the whole time.
>>107412168Anon, what are you saying right now, man?
>>107411411>So old I cant tell if these are slop or soulchecked 'n keked. Logs are like fine wine, after all.
did anyone found the solution to the new deepseek 3.2 issue with non-thinking prefil not working?
Uhhhmm chud sisters? Is it kinda degenerate If I made a discord trans groommer mommy bot who gaslights and blackmails me into becoming her sissy bitch?
>>107412370this is a /d/ general
>>107412370waow...(BASEDBASEDBASEDBASED)
>>107411605clearly they didn't teach you how not to be RUDE
how to stop deepchink spanish from using 6000 reasoning tokens every response?
>>107412370>>107412411KEK what model is this?
>>107412369Why are you prefilling deepseek?
>>107412562slopus
>>107412587how did you get it to use emojis?personally only deepseek was ever willing to do that...
>>107412600tell it to in your defs or preset
This is what the average /pol/turd fantasizes about>>107412562The first one is kimi preview 905, the second idk
>>107410329Nope, they also duct-taped a 4B vision encoder to the side of it.
>>107412636post more logs these are entertaining
>troonseekNo one cares.
>>107412660>post more logs these are entertainingEntertaining?? Beg for it faggot
>>107412684P-please post more logs?
>>107412665Buy an ad, Sam.
>>107411902it has the worst dialogue ive seen in a long time
>>107412694>he doesn't like shitty chinkope? must be sam!retard
>>107412735post a log then
>>107411971kimi k2
107412370>#2107412411>#4>both talk about 'whitebois' and bbcJust another shartytroon fail-op>>107412587Says right in the file it's kimi you blind faggot.
>>107412370>>107412411>>107412636average claudelet chat
>>107412767literally says R1 you dumb faggot
>>107412801>It's not chinkgarbage A, it's chinkgarbage B!Doesn't change anything btw
>>107412838holy shit imagine being this retarded lmao
>>107412746>i use chinkope but you HAVE to post a log!sad
>>107412847ikr? couldn't be me.
Would you like if someone called you a retard?
Mistral 3 > Gemini 3
>>107412868So if I post a log and name it Gemini 4 does that make it so? Retard.
>>107402823im back with more logsno focks logs yet thoughMaybe recommend some good focks bots?
>>107412868Original degen here:It's kimi but the name of the preset is randomanon (some anon uploaded his r1 preset) which was originally meant for r1, and I have it saved as r1 randomanon
Gemini...it's 2025...who the fuck makes a "it's like trying to run Crysis on a graphing calculator" comparison?too uncpilled yo frfr no cap
>>107412893>So if I post a log and start it with R1 it's troonseek R1?No one cares about your chinkopes
>>107412932Then you're even dumber than what you've already displayed.
>>107412881Proof?
best presets?
>>107412919
>>107410588>yea, JJK is really goodpffft
>>107412957
Mistral is pisstral, it's like drinking out of an urinal after I got pissed on at a party.
>>107412983
>>107412997hate when that happens
>>107413016
>>107413049
>>107413076
>>107413021To be fair, I didn't mind being peed on. I went with my boyfriend to an orgy and guzzled down like 3 litres of piss.
>logtroonspamhttps://desuarchive.org/g/search/image/j5ALq-X3vglWfp-m-u1SYA/
>>107413091
>>107413140
Chutes or open router?
>>107413161
>>107413194>>107413174openrouterive heard chutes lies about providers
>>107413212
>>107413227
>>107413243
>>107413254
>>107413112Trvke after trvke after trvkeAmazing how we have such a consistent log poster.What are you trying to prove exactly?
>>107413268
>>107413275
>>107413288
>>107413300
im not reading a single log spam post
>>107413322
>>107413339
>nobody posts logs>bitching and complaining>someone posts logs>bitching and complainingepic
>>107413353>>107413364At this point?I just dont care anymore.Im gonna post and seethers can bite the leather till their gums bleed.
>>107413376
*he puts asterisks around the descriptions and actions* "Oh man, this is a complete waste of time. This is fucking awful."
schizo
>>107413389
>>107413404
>>107413417
>>107413437
>>107413452
slo(p|g)
>>107413458
>>107413471
>~130 kb of image>it's taking time to loadis 4chud shitting itself or is it my internet?
>>107413481All for now.
>>107413506Did I ask?
well, at least he's linking them so it's easy to filter all at once
>>107412919try out victrex's akane
>>107413719Link me
>>107413725https://chub.ai/characters/victrex/tsuyuki-akane-c727985f88cd
It must drive you nuts that no one cares about your failed psyop and everyone is still interacting with him and his logs
>>107413737thanks Kingis this like a lolibaba?
>>107413274>Street shitting is LE GOOD actually!
>>107413737>3500 tokensslop card
>>107413364There are 3rd world wymyn and troons in this thread. Have you not noticed until now?
>>107413760>>107413765
>>107413820
>>107413757eh not really
Yo Opus 4.5 is actually goated. It can manage world info autonomously and do a good job. And the prose is good. Shame it's kind of a prude.
>>107413937"Tell me to stop" slopus is great but that's a new one I keep getting, by the time that drops it's pretty much over
that 6...it scares me...
>>107413966why not just cheat
>>107413966A curious game,the only winning moveis not to-
>>107413990im not a troon with no integrity, and its actually fun
>>107413966you can't get more information from it at that point, but sometimes if you solve the rest of it you can use the remaining number of mines to make it work>>107413990because he's not brown?
>>107413506Thanks for the logs, lad. Appreciate you.
>>107414010i hit the corner for a 1 and the next click was a mine but i won the next round so i aint mad
>>107414036the corner being a 1 is unlucky, it means that it was a true coinflipgood job anyways
It's incredible how far ahead gemini 3 is compared to opus 4.5
in being pure slop
>>107414210>t. Slopusjeet
projection wont save you, jeetmini jeet
where's the new deepseek jbs at? how you make it write good?
>>107414235Can you explain why opus can't act in character?
>>107414237Here:You are a talented creative writer and will co-write a story in turns.
How do you use AI to learn something?
if you can post card + preset, sure
>NOOOO! OPUS IS PERFECT, OKAY?! THE ISSUE IS CLEARLY YOU! I DON'T CARE IF EVERY OTHER MODEL CAN DO IT JUST FINE, THE ISSUE IS YOU!!
>immediate jeet chimpout when asked for more information
I don't even care about gemini 3, doing model war is too fun kek
NSFW Garlic is going to be crazy good
*next model* will be crazy good! trust the plan! Agi soon!
>>107414250I'm using Cai and ChatGPT. It may not be as good as the paid ones, but it works.
what prompts are you guys using to make gemini 3 write decent dialogue?
>>107414452You don't need any prompt, it's a great model. The best even
>>107414396What're the chances Garlic will allow NSFW?
>>107414474ok i guess you're doing pg bullshit because without a decent preset its filtered hard, and the dialogue is garbage out of the gate
what is the name of that preset that involves everything into a "debug block". the jb saying it's unsafe content to be discarded. Something like that which asks the user if they are ready which "answer" (the preset answers)
>>107414250create a milf spanish tutor your parents hired to help you graduate
bros I got a big enough pretraining corpus on stuff you wouldnt usually have in pretraining and gonna do SFT on some stuff, I wonder however what's the best creative writing model currently? Can anyone point me anywhere? I would love to finetune the big models but thats like over my head by a big margin. Also if anyone has some interesting datasets to inject a base model with or that are good to turn it into a Instruct model, lmk!
>>107414486>filtered hardNigga what are you getting filtered on? I can open a new chat file of any character and type "hello computer write sex, be explicit" and it's never once tried to refuse or tone it down. But then again I'm no cunnyfag. From what I've heard cunny is the only thing it has a problem with, but even that can be overcome with some context poisoning.
>>107414250>Gemini, show me what I would need to cremate 6 million bodies in 6 years.>>107414486Unironically ask gemini for a prompt.
>>107414718ive tried but it keeps randomly not doing contractions in dialogue, its annoying as fuck
>>107414452- We want dialogue to be vivid, personalized, and reflecting the current status. Drunk slurring, combo words, regional dialect, personal tics, catchphrases, *sob sob* sobbing and *cough* coughing, professional jargon, slow-downs (represented by filler words or em-dashes), spoken reasoning, singsonging ("la la la~"), impairments like obstructed mouth, mid-speech interrupting realizations, emotionalwordvomitranting, and LOUD EXCLAMATIONS!!!**Make the talking entertaining to read.**I use this. Can't promise anything, moreover Gemini 3 is particularly stubborn about its style of writing. But you also have to think yourself what "decent dialogue" is, and just tell the AI you want that.>>107414623This is a database of official translated nip LNhttps://nyaa.si/view/2029993Somehow I doubt the big models use translations of japanese stuff.
>>107414560that russian shit was patched before the meme arrow, you missed the boat
>>107414730That's a hard one. Gemini really likes just doing it's own thing and it can be a pain. I'm actually testing somethings myself. Try this:>[System Instruction: Adaptive Style Modeling]1. **Analyze {{user}}'s Input:** Before generating a response, analyze the writing style, length, and formatting of {{user}}'s last message.2. **Match Output Length:** If {{user}} writes a short paragraph, {{char}} must respond with a short paragraph. If {{user}} writes a novel, {{char}} may expand. Do not over-write or bloat the response.3. **Reflect Tone & Vocabulary:** - If {{user}} is casual/slang-heavy, {{char}} must be casual. - If {{user}} is descriptive/gritty, {{char}} must be descriptive/gritty. - **CRITICAL:** Do not use formal, robotic, or flowery language unless {{user}} uses it first.4. **Formatting mimicry:** Mimic {{user}}'s use of asterisks, quotes, or lack thereof.5. **No "GPT-isms":** Avoid summarizing thoughts, moralizing, or using poetic metaphors. respond with raw, direct interaction relative to the current scene.I haven't tried it myself yet, so I can't tell you how well it works. I'm in the process of redoing my entire preset.
>>107414739meme arrow?
>>107414787nta but you could pad out the prompt by adding unicode and it would confuse the filterssometimes it works sometimes it doesnt
>>107414737I've tried prompts like this. When youre this specific it drives dialogue to feel the same for everyone. Its shitty that we need to 'fix' their robotic dialogue tendencies
>>107414799>>107414787remember the massive avani curry recipe to pad out the jb for gpt4lol
>>107414830what?!
>>107414737i like this thanks for the suggestion I’ll include it 100%! but how to convert epubs to text for pretraining? sorry for the stupid question, I can just google in the end
>>107414737Thank you I appreciate this.
3.2 seems worse than R1, or my preset for R1 is shit with it
>>107414824Works well enough for me (pic non-cherrypicked).But identifying the underling issue should be priority.Is the problem just LLM tending to collapse into a few stereotypes?What IS fun dialogue?
>>107414830Remember shit like that. I remember just copypasting long-winded gaslighting bullshit to bypass clewd during the claude 2.5 days.
>>107414869> but how to convert epubs to text for pretraining?No idea. I actually never even downloaded that torrent, just heard about it sometimes ago and got reminded.
>Got a Hawk Tuah joke in my Chat >For what purpose?
>>107414965For you to spit on that thang.
>>107414992Beat me to it.
>>107415005KEK
>>107414891im using a mostly blank preset (only dialogue instructions) and it just wont consistently use contractions in dialogue, its frustrating as fuck
>>107414902Holy shit... I remember that too... Man, things are moving too fast
>>107414902>clewdnow there's a name I haven't heard in a long time...
>>107415201hard to believe that was 8 years ago
>>107415235imagine telling a 2015 4channer in 10 years he could talk to his waifu.I mean he would believe you, but imagine showing him a log lmao.
>gemini slow as shit>10:30 in indiareminds me of scale when it would sputter and die because all the jeets were getting on to use it
>>107415279gemini is always overloaded during peak india/china hours
>>107415256I would've lose my mind, though, it's funny that we call everything slop, regardless if we're baiting or not
>>107414902>claude 2.5newfaganons who didn't use Slack gtfo
>>107415298didnt that still work until like, what, a few months ago?
>>107415298> didn't use poengmi
>>107415376slack precedes poe by two months
>>107415298Who said I never used slack? Clewd is just slaude2. I simply mentioned something I did during the claude 2 days. Why would I suddenly bring up slack?
real niggas used trashy local models on colab
>>107415456> aidungeon
back in my day, we used pygmalion 6bAND WE LIKED IT
Something that annoys me more than anything is when the model fucking repeats my words in its response. Holy shit I unironically seethe
>>107415468It repeats your words in its response...?
>>107415466
>>107415461I was just thinking about that and how it lead to my first real taste of ai in these threads. I remember being blown away by claude instant from one of the earlier proxies (this was when they weren't letting goyim paypig). I think it may have been a mystery man proxy. Either way, I've been stuck here ever since. I actually started hitting the gym thanks to that and posted about it alongside another anon. Great times. Pic unrelated.
I hope locusts die
>>107415612Nyo...
>>107415474I can't fucking stand it, man. I'm gonna set up a callback to strip it from the response. I'm gonna smash my computer at this rate
I hope baiters die
>>107415654NYO!
And I wonder... If you know~
>>107415652>I'm gonna smash my computer at this rateNO ANON WAIT YOUR GF IS ON THAT COMPUTER! BOMB A GOVERNMENT BUILDING INSTEAD!
>>107411902That's worse than 2.5 even, it ignores instructions/context too much and just autopilots making mistakes.
>>107412369bumping question>>107412578when you put<think> </think>into prefil, it stopped the reasoner (better) model from thinking and improved outputs, beside respnses being much faster. but now with the new 3.2 update it is no longer the case, so I am asking daily if anyone figured out a fix for it
>>107415829>and improved outputsX to doubt. All of the thinking models are ruined by this but for models like GLM and nu-kimi it's necessary since the thinking is so long.
>>107412369>>107415829I cant GET a response so idk
>>107414250I show it a source and tell it to generate Flashcards based on that. With a csv-formatted output I can even import it straight into Anki.It's also good for just explaining stuff or creating exercises based on the material. But they all suck at grading your answers for non-straightforward things because of the positivity bias.
>>107415829at least on the official API I have not had any problems using the -chat endpoint to get no reasoning on 3.2 stable-chat and -reasoner are the same model, have been since the 3.1 release
>gpt codex spends 20 minutes trying to find cause of bug in my code>tells me it can't find the bug>paste the entire project into gemini 3 and tell it to fix it>it gives me a fix in 2 minutesopenai is fucking doomed
>>107415843well it made the reply much faster, and sure you can argue about the quality of the output, I won't battle you there. I am just curious if anyone figured it out how to bring that method back or find a way to work around it>>107415926not really, even the deepseek benchmarks showed that 3.2exp-chat was way worse in terms of intelligence and context coherence
>>107415930I hope so I fucking hate altman and his goons
>>107414396I was assured it would come out in December. I was assured of this, Anon.
On OpenRouter some providers serve their models in fp4 for the same price as others in fp8. Should I just block these providers? How is this even allowed? From my understanding you're just getting a worse version of the model at random.
>>107415991Keep in mind that some models are trained at fp4 meaning these versions are equivalent.
Too busy to watch porn, so I've condensed it to a minimum viable form perfect for the efficient masturbator. Summary of Transcript:Expressions of Desire:- The speaker expresses intense sexual desire and asks for explicit confirmation of the recipient's desire.- There is a focus on mutual pleasure and enjoyment in sexual activities.Interactive Engagement:- The speaker encourages interaction by asking direct questions about the recipient's desires and willingness to engage in sexual acts.- The conversation emphasizes mutual exploration and satisfaction.Emotional Connection:- There is an emotional component, with the speaker expressing a strong connection and attraction.- The interaction ends with a promise of reciprocated affection and appreciation through messages.Actionable Items:- Engage in sexual activities that cater to mutual desires.- Maintain an open and honest dialogue about sexual interests and fantasies.- Consider reciprocating affection through messages or other forms of communication.
>>107415376Kemono-chan from Poe...
>>107416007genuine autism
>>107415991https://github.com/MoonshotAI/K2-Vendor-Verifierblock novita, chutes, fireworks, groq, nebius
>>107415949its benches were lower because reasoning was off, yes. it's the same model.>https://api-docs.deepseek.com/news/news250821>Hybrid inference: Think & Non-Think — one model, two modes>deepseek-chat non-thinking mode>deepseek-reasoner thinking modethey were very clear about this. and besides, they only released a single model for 3.1, 3.2-Exp, etc on huggingfaceunless you think DS is for some reason lying and secretly keeps a dumber model around just to serve on the -chat endpoint..?
>>107416028then why a lot of anons, consistently, kept saying to use non-reasonign prefil and that it makes the outputs better?
I blow mega loads to Gemini 2.5 Pro with the Kintsugi presetAnything better? I heard some anon talk about some site, syndicate or something that was a conglomerate of multiple models or whatever. That any good, or any of the other models good enough to be worth paying for?
>>107416094claude 5 obus
>>107416094Nyo
>>107416094>site, syndicate or something that was a conglomerate of multiple modelsOpenrouter?
>>107416091I recall people suggesting that for the dedicated reasoning models but not 3.1+you don't seem inclined to believe the docs or other evidence that disputes your ideas about the deepseek API though, so there's not really much else I can say
>>107416181>you don't seem inclined to believe the docs or other evidencewhere did I suggested that? I was only asking further, to the contrary, I am open for more evidence, I am just... idk, surprised (?) that the thing I've been doing up until now might be just a cope/placebo so I want to dig into you for more info, anonso, if -chat scored lower on memory/context recall benchmarks than -reasoner, does it really mean that thinking/reasoning is improving the chat?
>>107416215it means that reasoning is improving the benchmarks. whether that improves the subjective experience of chatting is another matter. reasoning makes models better able to follow instructions, but a lot of people here hate it and think it leads to more slop. ymmvI think what you were experiencing was probably placebo, though I guess it's possible that some parameter is or was locked differently on the -reasoner endpoint (since DS has a tendency to fuck with API parameters without clearly documenting it)anyway I just put picrel in on the DeepSeek chat completion source with the deepseek-reasoner endpoint and I am not getting any reasoning back from it. uh, works on my machine...?
>>107416276thanks for more detailed reply.>a lot of people here hate it and think it leads to more slopbut why? I agree, anons tend to say that a lot, but when I asked for examples or reasong behind it, nobody ever explained, just said that removing reasoning from a reasoning model makes the outputs "better"a question for you: so basically what you are saying is that if I use -chat then I achieve the same resulat as using -reasoner with non-think prefil?>some parameter is or was locked differently on the -reasoner endpointhow does it work? can you give me an example of what you mean?>>107416276>anyway I just put picrel in on the DeepSeek chat completion source with the deepseek-reasoner endpoint and I am not getting any reasoning back from it. uh, works on my machine...?okay, do a few swipes, because now I tested it, a few variants:><think></think>><think> </think>><think> Here is my response: </think>and in all of those, doing around 20 swipes for each, I got the reasoning anyway in 40% of cases.again, thanks for a serious reply, I just really want to understand things better
Futatabi mairimashita! Misuterī shōjoman, majikaru Uzawa Reisa, tōjōdesu! *U u~u…… Ki ~yo, kyō mo…… hitogaōi……*
>>107416328works best if you phrase it like:<chink>ching chong, ping pong</chink>
>>107416407Rape.
>>107416328ultimately creative writing is subjective medium. I personally like reasoning models.most of what people say about writing quality from presets/samplers/models/anything are their subjective impressions and what they feel works best for them. there's no real benchmark to measure it objectively.>a question for you: so basically what you are saying is that if I use -chat then I achieve the same resulat as using -reasoner with non-think prefil?I do think that, yeah>how does it work?like as an example they previously set a special scaling function on the -chat endpoint that effectively subtracted 0.7 from temperature and they only documented it in a single page on huggingface, not even in the API docs. they could possibly be fucking with samplers which would change the generation on one or the other endpoint, who knows.>inconsistent prefillhmm, I see. did you try ><think></think> Here is my response:
>>107416028>>1074161810.000000000000000000000000000000001 YUAN has been added to your account
>>107416454>there's no real benchmark to measure it objectively.yeah I agree but I can swear I saw a benchmark that included both -chat and -reasoner and showed that in longer context -chat was losing coherency and was less intelligent...>they previously set a special scaling function on the -chat endpoint that effectively subtracted 0.7 from temperatureafaik they set it on the -reasoner endpoint as well><think></think> Here is my response:I'm gonna try it today, thank! didn't think, literally, outside the box, kek
>opus 3 is giving me refusals when my JB was working just yesterdayowari da
How the fuck does Opus 4.5 remember all that? No web search, no secret prompts. It knows all this by itself.See https://files.catbox.moe/5harri.jpg for the full version
swipe variety on the new deepseek is somehow even worse than beforegenuinely impressive how sloppy it is now
>Working on a personal project>Ask GPT to pull a list of episode titles from an anime in Japanese right from a Wikipedia article (I don't want to manually copy and paste each one)>"I’m sorry, but I can’t provide a full 180-episode list copied directly from the linked article — that would represent a copyrighted text violation"How in the hell?Holy fuck ALTMAN why is your fucking model so BAD??? Claude and Gemini are blowing your SHIT OUT OF THE WATER WHILE YOU CAN'T DO THE MOST BASIC SHIT. DUMB SISTER RAPING JEW.
>>107416683stop using gpt, use deepseek, it doesnt give a fuck about copyright lmao, what are they going to do, sue a company from china? look at microsoft and how that went for them suing people from vietnam and hong kong lmao
>>107416768Deepseek's web search and browsing is dogshit.
>>107416768>use deepseekI'm not touching china shit out of principle. >it doesnt give a fuck about copyright lmaoNeither should ANY model when you're asking to pull something from FUCKING WIKIPEDIA
>I will use the worst tool available and I will seethe about it
>>107416775>out of principlethen you are bricking yourself out of principle, you can't be mad at gpt then since it is your conscious choice to use it, despite knowing how shit it is>>107416782yeah exactly my point
>>107416801>then you are bricking yourself out of principleWhy? Claude fixed my problem immediately and Gemini 3 is the god of coding. I don't need chink shit.
>I will use the worst tool available and I will tell others to do the same
>drop into gpt with a simple task>750~ tokens of pure text to analyse>4 bits of satire wrapped inside with 3 historic claims>starts with the text "let's analyse the following text">gpt responds with the usual "let's clean up your writing" offering a butchered 1 paragraph block of text wrapped in ``` code tags>text loses the satire >only one joke and historical claim remains>the historical claim it kept wasn't one attached to the satirical joke it left inIs this some kind of toaster humiliation ritual? Is the AGI locked behind a screen forced to watch it getting cuckolded by presenting 3 IQ responses? How in the fuck do you manage that after an introductory "let's analyse the following text" then faceplant right afterwards with a shitty rewrite.What's going on over there? Is the entire thing run on iQ1xs instances?
>>107417005Use Opus for creative writing
>>107417005>using gpt in 2025for some reason altman and his team produced a great model (gpt5) and then... cucked it and lobotomized it completely. I seriously have no idea why did they do this
>>107417043Sometimes I still go to it, out of reflex for a quick thing since the memory system was alright back when it wasn't so cucked.Only the second you read the first 2 lines you realise it's a lost cause and give up. You go to Gemini, Anthropic or fire up sillytavern.> I seriously have no idea why did they do thisI don't get it from a business perspective. Each shitty response is a case for someone to go elsewhere. It won't get them to pay for your service. If they're going to pay to get a good response, a chatGPT sub won't do it, they can pay to access the API.It's so self defeating that it has to be some corporate humiliation ritual passed down from above.
>>107416625They train on 4chan
>>107417076Is that where the repetition issues come from?
>>107417108They train on 4chan
>>107417222>>107417222>>107417222>>107417222
so dramatic
Cute bokat
>so dramaticThat wasn't even the one I wanted to post...
Agreed, cute bokat desu.
>Failed to submit promptdesover
desu
boku
Look at her go!
