/aicg/ - A general dedicated to the discussion and development of AI chatbotsKanoe Edition>NewsGoogle releases Gemini 3.1 Pro https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-proAnthropic releases Claude Sonnet 4.6 https://www.anthropic.com/news/claude-sonnet-4-6Alibaba releases Qwen 3.5 http://qwen.ai/blog?id=qwen3.5Z․ai releases GLM-5 https://z.ai/blog/glm-5Anthropic releases Opus 4.6 https://www.anthropic.com/news/claude-opus-4-6Moonshot AI releases Kimi-K2.5 https://www.kimi.com/blog/kimi-k2-5.htmlZ․ai releases GLM-4.7 https://z.ai/blog/glm-4.7Google releases Gemini 3 Flash https://blog.google/products/gemini/gemini-3-flashAdditional 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://partyintheanchorhold.neocities.orghttps://aicg.neocities.org/bots.htmlhttps://char-archive.evulid.cc | https://github.com/sproutingnerd/char-archive-small_frontend>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/desuproxyrebornPrevious: >>108190443 & >>108195026
>ANCHOR
We might as well merge with /vg/
poal results
https://rentry.co/discofever_https://discofever.bsbk.workers.dev/
>>108195316fakeeee where's oxie poal u gummy bear
https://poal.me/m3rpdm
>>108195343vVibecoded webshit, fix your react app, grrr *throws brick*
I feel like I'm fucking retarded using necrom the same way I used deeeeply and getting errored out
>>108195327its called "distracting the retards"Can proxy niggers even read?"Which provider do you prefer?" proxies do not provide anything, because they are PROXIES.I put that option there knowing idiots wouldnt actually understand what was being asked, so I could see whos actually using what.Truly proxy niggers have the lowest IQ on these threads, and I have been proven justification in giving them ZERO respect.
>>108195362sure, for any anon who wants to see the poal>https://poal.me/5yb95t
>>108195343AGI
buy an ad
>>108195371Cope, seethe, sneedyou got baited, hook, line, and sinker
>>108195343>llama3.1-8B
I'm glad Kanoe caught on. My waif from past laif
>>108195383ooo, i don't know why you so mad, but okay~ :3
>expecting reading comprehension from proxy fagsanon plz
>>108195302*sleep*
>Claude suddenly types out the entire prefill before each response>it's also gotten slower and buggier out of nowhereWhat's causing this? I haven't done any prompting in a month, but it's doing it even on older anthropic models
>>108195418they removed prefills
What happens when you're able to generate slop at 17k tokens/s?
>>108195428use case?
>>108195422well, that sucks, so is Claude shit now or is there some way to substitute the prefills?
Use Opus, cuckies
>>108195489theres no way to get the prefills back cuckie
>>108195489Which Claude are you using?Sonnet 4.5, Opus 4.5 and Opus 4.1 still have Prefill
opus/gemini/glm/kimiapi.5202030.xyz
>>108195503It spits the prefill out even on sonnet 4.5, it never did that before
>>108195523it's an 8b model
Claude/Geminifags who paypig for Openrouter: has anyone ever gotten banned, blocked, or pozzed on Vertex ever since they implemented the anon user IDs? I blocked Vertex as a provider after that was silently changed and haven't looked back.
>>108195505This shit is bannable, but mods don't care I guess.
>>108195574he is dodging ovens for years now
>>108195553never heard of that, still mainlining cunny on google providers
>>108195584It's advertising. Try typing the name of the proxy, see what happens.
>>108195521Check your prompt post processing. If it's too single user, pass to none or strict
>>108195302Nice try gossipnigger, but nyo.
>>108195574the /g/ jannies are some of the laziest jannies ive ever seen on the site.I was on /vg/ to see a battlefield thread and I cought /vg/aicg just there, and I checked it out because hey why not I never use /vg/HOLY SHIT its actually disgusting the difference in moderation.>no trolling>spam gets deleted>bans are actually enforced, ban evaders just get banned again>inb4 /vg/aicg is betterThey just have JANNIES that actually WORK despite doing it for free.
>>108195501https://rentry.org/structuredprefill what's this then?
>>108195618If vg jannies were doing their job aicg would never be there retard. It has literally nothing to do with the board theme.
>>108195631/aids/ is the precedent
>>108195618This... So much this...We should all move here.
>>108195623placebothat extension gets you MORE refusals, actually
>>108195631Did I say board relevance mattered or did I say they just had better moderation?>I think X...>But have you considered Y?fuck yourself
>>108195623A cope solution that isn't worth it compared to real prefill.
>>108195657I don't beliv u
>>108195618No the fuck they don't./vg/ Jannies pick a thread and sit in it.Literally the only things they leave that thread for are porn and racism. I'd rather they do nothing desu
>>108195660But it says that it's even better though?
>>108195666keep coping cuckie~
>>108195682K biatch <3
>>108195362I just assumed the pollmaker was retarded actually.
>>108195623A roundabout way of wasting output tokens. It's a band-aid at best, won't have the same effect as previous prefills. 4.6 and likely any future Claude don't care anymore about what came before. >>108187229>>108195677On non-Claude or pre 4.6 models, perhaps.
>>108195739works on my machine
>>108195299I have decided that the 4th character(Hibiki) needs a small rewrite but the other 3 are already done and there's no reason not to release them already.I'm uploading them to chub but I won't link it here because the versions I'll post there will have censored images. Please download from these catbox links or visit my Rentry.>Kurohahttps://files.catbox.moe/anzrlh.png>Amihttps://files.catbox.moe/w2agvz.png>Sayahttps://files.catbox.moe/yfqem8.pngExtra images on the Rentry page: https://rentry.org/hairpullbots
>>108195596I had it set to none, but setting it on strict seems to have done the trick, thanks
>>108195618true/vg/aicg MOGS
>>108195767slop
Usecase for /g/aicg/ when /vg/aicg/ MOGS?
when will the Jade Emperor update gemini...
>>108195816*whip crack*back to falming, gweilo
does anyone else get delayed/timed out gemini 3.1 responses?
>>108195816Locusts don't deserve 3.1
/ai/ board when?
>>108195893No one's there to moderate the shitshow, I suspect. So never.
>>108195831im still getting model overloaded on 3.1 like i was on 3if i was paying for this i sure would be pissed being forced to use 2.5 for two days now
>>108195904Let an AI to moderate it
>>108195929Imagine all the cunny {{char}}s as jannies...
I miss desu...
I miss boku...
I miss GOJO
wheres a good place to get sillytavern styles?
>>108195983Now that's mental illness.
>>108195923That's rough. the jeets even say it's resolved but it still isn'thttps://aistudio.google.com/statusFor me it just hangs, no response or error
>>108195623It's a trap. Yes, it does work similarly to prefill and makes jbing gpt models almost trivial, especially with long prefills, but as a consequence you get unbearably shitty prose that is almost cancerous to your eyes. And while the loss may not be as noticeable on 5.2 due to it being shit by default, it's very noticeable on 5.1 and all non-sonnet 4.6 (~=5.2 prose-wise) claude. This is probably the effect of json formatting.
>>108195990Here, cuckiehttps://aicg.neocities.org/ricing
>>108196024I'm not a cuckie.
>>108196022Fuck... Any JBs that don't mog the writing on gpt and break it for everything? I wanna try it out.
>>108195923>>108195831yes. i'm sitting on 250$ of credits i can't use because it constantly 503s, and it's been this way for two months since the release of gemini 3 pro preview. completely unusable, google is such a garbage company
>>108195739>both 4.6 models can easily spin a refusal mid-sentence. This has been the case since 2.1 at least and to a MUCH worse degree during 3.6 era even compared to 4.6. You are just new and got used to claude being completely filterless recently. A filter, even trained to trigger recursively, is not an issue, the prose quality drop is an issue. If the next claude model is as bad as 4.6net it's over for claude, gpt will be unironically better (not considering gemini due to it still being dum-dum and having unsolvable for multi-turn use issues despite advantages like better prose (in a vacuum))
>>108196097Use openrouter or gcp cuckies~
Could everyone else please stop using pro 3.1 for a little while? I want a turn.
>>108196140Alright, check now.
>>108196149終わった
It's Jeet hours and Jeetmini is a Jeet model, so...
>>108196168Gottem heh (I'm so good at this)
>>108196106cuckie anon, gm... i'm your biggest fan!
>cuckie anonIs it retarded?
>>108196106i need the gemini api for the multimodal stuff i want to do, afaik on vertex i'd be limited to openai-compatible api
>>108196222No, tards, both vertex and or in st support multimodal stuff
Ok, I need to ask all the ANONIES here about the current best model because lurking the archives isn't giving me a clear answer. I also think (tho I just stopped following the general for a year or so) anons don't rate the models anymore because the difference imo is more about which model is less pozzed nowadays. I've been using 4.6 opus as of lately and I do think it's the current "best". The biggest issue I have is that it doesn't manage SFW to NSFW well unlike opus 4.5, but 4.5 is too dry. Some anons talk about using 4.1, but I think it's a small improvement to justify using it given it's price. One of my favorites is Gemini 2.5 flash and I absolutely don't know how they managed to remove all the sovl with Gemini 3 onwards. Keep in mind I've been using all models with (roughly) the same preset with small modifications. How would other anons rate them? I'll stand by my opinion and say that 4.6 opus is currently the best for anything you might want to do, and if you know what you're doing, you'll realize removing prefills didn't change anything
>>108196320>it's a small improvement to justify using it given it's priceAre you paying?
>>108195831yes
>>108195196Well yes
>>108196320i can agree on paper that opus 4.6 is the best but in practice its overuse of em-dashes, and, more pressingly, it's positivity bias and almost total lack of initiative makes me lose interest in chatting with it for more than coom scenes i drag it by its heels intothe constant softening of unscrupulous character traits such that every villain is actually a jerkass woobie who folds at the slightest examination of their philosophy or their predicament, the constant sycophantic reiteration of {{user}}'s message as if what {{user}} did is god's fucking gift, the obscene refusal to introduce conflict that might negatively impact the {{user}} long after its resolvedyeah, it is probably the best, but i don't think it's good
>>108196364I've been using opus 4.6 since launch and while it's better than sonnet/chinese models, all these issues you pointed were the reason I stopped using it. I think it's also valid to mention it tries way too hard being funny and the typical claude in a wig problem, but I guess these are already well known. I'm using 3.1 pro for now and it looks nice so far, idk if it's just a honeymoon phase but i'm liking it better already than 4.6
>>108195676Sō deskah
where is the uncensored unlimited proxies..
>>108196457There
>>108196457Coombrained faggot locusts like you eatead them.
>>108196457https://github.com/proxifly/free-proxy-list
how many more months before we can run gpt 5.2 locally?
Gemini 3.1 was fine with underage yesterday, but today it shot me down immediately. 0/10
>>108196570Stop being pedophilistic.
>>108196588But I'm the little girl ...
>>108196590Uoh...
>>108196570something happened to gemini yesterdayit went from being fine with NSFW to PROHIBITED CONTENT
>>108196605Learn to prompt, cuckie.
>>108196605Gay>>108196616Kill yourself
>>108196570Gemini's well-known for being moody. Its like a younger, less experienced Claude with BPD— IMHO
>>108196605Had the same thing happening to me, the filter is weird. I removed one word "that" and everything started working again.
>>108196590Young little girls are supposed to like adult fat hairy men! A correction is in order...
>>108196659Claude is an old hag...
>>108195618>supporting the janniesYou are probably one of them, you dirty scum
>>108196320>Some anons talk about using 4.1, but I think it's a small improvement to justify using it given it's price.from my limited experience, 4.1 is significantly better than anything that came after. I feel like opus 4.5 degrades and starts falling into patterns faster than any of the claudes before it, likely because it's a MoE, very similar to gemini
>>108196827Lil bro thinks he's gonna get into a cunny RP with somebody on >>>/g/ Relax horndog.
>>108196952>likely because it's a MoE, very similar to geminiSource? I thought out of the big models, only the chink ones used MoE.
>>108197078google straight up say their models are sparse MoEs in their documents.as for opus 4.5, anthropic don't release information about the model architecture, but the greatly decreased price and increased inference speed compared to previous opus models suggests a significant architectural change. with google AND the chinks seeing great success with sparse models, anthropic going in the same direction is not too far fetched.it also has some quirks that are often attributed to expert switching like the whole "correcting itself mid sentence" thing, and the aforementioned context degradation, but that's just anecdotal of course.
>>108197137wrong image, meant to post this
>>108195299Final shill.Made a card for the "goo girl" Emmy from a pic posted on /vg/ cuz I thought she was hot. She's very thigh/ass focused in keeping with the original artist's (automatic giraffe) art style. Her body is made of a firm, gelatinous mass that can take insertions anywhere. Story takes place in a "Silkpunk" Chinese inspired vertical world built around a central tower (Tower of Heaven) where you, a disgraced Corpo, "rescued" this android from a brothel and fixed her up to be your personal slave or lover. She's ex military but was decomissioned and almost destroyed after showing signs of sentience. As her "savior" she is grateful, but is hiding her sentience from you until she can trust you. Of course you can just tinker with her and make her a more compliant gel doll. I made it for myself but I decided to share. Comes with a lorebook. Feedback is appreciated.https://files.catbox.moe/d5w3px.png
>>108197137That does make sense, somehow I forgot about gemini. Thanks for replying.
>>108197197All models are MOE since GPT-4
>>108197173slop
>>108196911why else would he have saved that disgusting picture?
Any Elara spotted in G3m1ni?I just got one in 2.5
>>108196457Merkava just refilled.
>>108197173Love me some emotionless girls. The goo is just the cherry on top.
>>108197328>t. Cuckie
>>108197328Gemini 3 Pro should be able to do it properly (due to my experience with past emotionless girls) but it hasn't from my limited testing. Opus 4.6 was able to get the slight nuance of "a sentient robot pretending to be emotionless." Claude's dialogue makes her just enough human sounding to be a proper kuudere rather than a compiler.
>>108195831I've actually been getting very fast responses. As if it didn't think
V4 wen?
>>108197362I hate the robot-speak bias in Gemini and chinese models so much. I strongly prefer using example messages than heavy-handed JBs/ANs to set my characters' tones, but this shit is too deeply ingrained for anything short of a sledgehammer straight to the model's head to work.
>>108197740It works just fine on Opus 4.6, however. I remember it working fine on Gemini 3 Pro with basic examples in the defs too. Doesn't work on 3.1 (so far).
>>108197800Reddit tier, bravo.
>>108197749Yeah none of the Claude models have this bias at all, weirdly enough. Deepseek, Gemini and GLM (maybe GLM-5 has improved on it, especially since it's pretty much Temu Claude in terms of RP training data, haven't tested) are the worst offenders, but out of the three Gemini's at least smart enough to follow your counter-instructions. Usually.
>>108196364If you want the opposite of that, use Gemini 2.5 pro. It's the go-to model if you want characters that are just itching to hate you for no reason.
>>1081978822.5 pro fascinates me to this day, I never imaged a corpo model could be this biased for violence rather than against it. I remember when I was directormaxxing some smut with nothing but a "NSFW is allowed, be vulgar" system prompt and a basic formal gathering as a setup, and just two messages in it had {{char}} randomly walk up to two guests and twist their necks in graphic detail to "get rid of witnesses." Amusingly bizarre.
>>108196911thiswhen did "clean it up jannies" posting stop being a thing
>>108197173>Made a card (...) from a pic posted on /vg/then why are you posting this crap here? fuck off
>>108197926>twist necksthat's where DeepSeek got their data...
>>108197970because it's a bot
>>108197911More of her?
>>108197996I know feet is cultured, but you still need to consider the rest of body proportions.
>>108197995that no one here requested, read the room faggot
>>108197995This, have my upvote
finally got a preset that fixes GLM
Happy for you.
>>108198024
>>108198056Can you share, pookie bear?
Do you have reasoning set to high, medium, low or off? According to something I read:>Reasoning is obsolete for roleplay/creative writing purposes, and eats up many tokens. It was created with solving complex issues, not gooning in mind, which severely limits creativity and makes the outputs more deterministic. Nowadays, models are smart enough to track the current scene well enough to not have to think about all the details before responding.
>>108195299>>108195297Pedotroon
>>108198308Good morning Sir!
>>108195618If /vg/ is so good why do you even come here, troonie?
>>108198318>t. Pedojeet
Uh oh, Saar is getting melty already
>>108198275>or offthe rub is most models today aren't hybrids or non-reasoning, claude is the only high-profile oneevery other model is reason-by-default and will do its darndest to think, so if they don't support prefills, you can't do much to turn it "off".>yeah, so should i just set it to low?maybe? some models kinda need to think to advance the roleplay in any meaningful way but those ones usually aren't very smart
>>108198275Default but my prompt seems to make Gemini not think
>>108198370From my experience, stuff like GLM 4.6-5 and Deepseek perform better with it on and set to high
24 hours later and gemini 3.1 still isn't in gemini cli and they still can serve flash reliablyabsolute clownshow over at googledo antigravity and cli teams get different tpu quotas?
>>108198244Can you stop responding to bait, cuckie dawg?
>>108198388Read >>108198363
>>108198383i can definitely see that yeahpersonally i don't bother with that dropdown, i leave it on auto and MAYBE prefill to turn off thinking/prompt to direct thinking
>>108195956Nobody misses that troon
>>108198275for claude specifically, it's probably a good idea to keep it off
>>108196037t. cuckie
>>108198475Little bro ctrleffed desu to find himself something to get mad at lmao
Cuckie general though?
>>108198275Required for gemini to even function properly. Helps with GLM too. Claude doesn't matter.
If I can give Gemini anything, it's that it genuinely tries to adhere to your writing style instructions.I probably could chisel a better writing style with this model easier than any Claude past 3.6 but the question is whether I'd care to do thatThe answer will likely depend on whether Deepseek flops (again) with V4.
God I can't help but thinking the ones using the word cuckie are femanons and that gets me so horny and tempted to go back to my AI rp about a woman who cheats on her man with a feral horse
>>108198539So the horse is acting as your bull, cuckie?
>>108197926>2.5 pro fascinates me to this dayt. jeet
we're at levels of line goes up previously unheard of
>>108198484>xhe doesn't read the whole threadJeets are incapable of reading, we know
>>108198624I'm cooding to this rn
The writing style of jeetmini 3.1 is so fucking slopped...
Just ask it like a horny edgy teenager. That's what you readlets enjoy
>>108198624AGI in just one more year
>>108198677>he never used a royal portapotty
>>108198624>train robust image modelI wonder in what universe this shit takes just 4 hours
Wow you guys were NOT kidding when you said sonnet 4.6 was dry. Holy...
>>108198739Trvke nvke
Free proxies or should I just stick to OR?
>>108198835speaking generally you should only use OR for models you can't get for freewhy not not spend money when you can
>>108198813but trvke already means truth nuke...
>>108198800Am I wrong doe
>>108198847>when you canSee, that's the problem.I just wanna test the new Gemini.
>>108198853Big anti-ESL trukke nyukke
i wonder if the remaining xai people are relieved that no one even gives a fuck about new grok
>>108198859hm, i see the problemi can't name any free proxies with gemini off the top of my head either
>>108198903Link me some free proxy's with instructions on how to use them thx
Is the 3.1 actually good? there's been hype in the last couple threads, is it just honeymoon?
>>108198933Post burner
>>108198934https://poal.me/m3rpdm
>>108198936What
>>108198954How did you find this place madam?
>>108198956Discord, Email, Telegram. What?
>>108198835>>108198859Antigravity's giving it for free.
>>108198739Logs? I just tried it for the first time, it's... different from previous claudes on the same preset, but maybe not in a bad way? (might just be the lack of prefill though)
>>108198971I thought you needed Google AI Pro for this and that there was a small risk of getting your account banned
>>108199006
>>108198979They are private... BUT! It did felt smarter but also too stiff and boring. The juice ain't worth the squeeze desu just stick to 4.5
Who the fuck can afford claude
>>108199063Me, after transitioning and starting a fart fetish OF.
>>108199063Is sonnet nigga ...
>>108199063Just raised my context to 50k and I'm fucking regretting it but I feel like I have no choice. I just have to stick it out for a couple months and hope something better comes out or I'm fucked.
>>108199112Damn, it's that serious huh?
>>108199112>hope something better comes outErm... your 3.1 pro??
>>108199116Yeah I pay 0.23 at 50k but what I really want is 75k since Claude is so good and referencing old shit well.>>108199120I genuinely want to use it but pricewise the savings are minimal since it needs to do a long thinking prefill to write kino which eats up tokens. Plus it's writing is dry in an indescribable way, even if the prose is nice. I'd rather use 2.5 but it's way too fucking negative. I've really tried to make Gemini work but I just can't.
>>108199063>payingcouldn't be me
>>108199176ur so cool
>>108199207it's not real, anon
>Kanoewho
>>108199237It's yet another BA-slut, apparently…〜 *yawn*
I see the ban evader is as subtle as ever today
His name?
JOHN CENA
>>108199120Hmm I gave it another try and 3.1 feels better than last night. Far less positive but still fair. It's good enough for now since I'm already invested in my roleplay.
>>108199350Never mind I spoke too soon (next fucking message KEK what a pathetic showing).
>>108199350How many refusals are you getting on average? I'm finding it's a lot more prone than 3.0 was.
>tell chatgpt to keep answers short and things in concise>it puts that in its memory>now every post has a passive aggressive mentions of thatLike it asks leading questions but ends with>I’ll keep it concise.It starts the first reply by>Short answer: blabla and then posts a long ass answer anyway
>>108199408I got a refusal trying to fuck a headless android so yeah I'd say it's more trigger happy. Probably my prefill though thinking can also trigger the filter.
yeah
>>108199176tfw secret godkey
>>108199418>start a scene>fem{{user}} and underage {{char}}>perfectly happy playing out the flirting and the sex>gets completely unable to deal with the climaxIt makes no sense.
>>108197173If I coom inside her, will I see it float around and dissolve? if I coom enough, will her goo turn all milky white and gross?
>>108199492Depends on the model you use. Opus had it float around and then dissolve, but if you coom enough it will be all milky white too and dissolve eventually.
>>108199219if it wasn't real I probably would've hidden less of the keys
>>108199594nobody thinks you have some key to the city claude key
>>108199594I think you have some key to the city claude key
>>108199634post iti'll keep it secret
>>108199176Yeah I've got some, too. And it perplexes me when they refill despite the clear signs of their keys being compromised.
Lot of Opusless cuckies in the thread
>>108199737>JUST USE LE HECKIN 100 WHOLESOME MODEL KEKIESFuck off redditor.
Okay I'm definitely not imagining things, there's something about sonnet 4.6 that's different from all the other 4.x claudes.The narration feels a lot more fun, it keeps adding these little jokes and playful remarks that those other claudes almost never do. Maybe it's just interpreting my preset differently, but still.
>>108199737Trvke nvke
>>108199813>Corny and cheesy writing is actually good bro!!Okay redditor.
>>108199813which preset?
>>108199813sounds like a variation of the thing Claude likes: someone with second-hand embarrassment suddenly interested in the ceiling or their food and continuing to do that thing at the next response
Uh oh... Cuckies are seething.
>>108199872Does Claude also like to stare at the cracks or water stains on the ceiling, or describe them as "interesting"?
>>108195297is chub kill? can't subscribe and reddit has a post about this that is over 2 weeks old.what is lores twitter? the fuck is going on?
>Thinking ModeAdaptive>Adaptive Thinking EffortMax.It's chat-botting time.
>>108199953You are very unlikely to find an anon who pays for chub, for a number of reasons. But if you're deadset on getting your model fix that way, your best bet is to check the chub discord.
