/aicg/ - A general dedicated to the discussion and development of AI chatbotsCastlevania Edition>NewsAnthropic 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-flashOpenAI releases GPT-5.2 https://openai.com/index/introducing-gpt-5-2Deepseek 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-3additional info: https://aicg.miraheze.org/?action=history>FrontendsSillyTavern: https://docs.sillytavern.appRisuAI: https://risuai.netAgnai: https://agnai.chat>Botshttps://chub.aihttps://realm.risuai.nethttps://jannyai.com>Modelsjailbreaks: https://aicg.miraheze.org/Jailbreaks?action=historygpt/claude/gemini/deepseek: https://openrouter.ailocal: >>>/g/lmg>Botmakinghttps://desune.moe/aicharedhttps://agnai.chat/editor>Metalore: https://aicg.miraheze.org/Chronicles?action=historylogs: https://sprites.neocities.org/l/r | https://chatlogs.neocities.orgDidn't belong in this world: >>108064689
>ANCHOR Storm!
A miserable pile of proxies
>Anthropic releases Opus 4.6: https://www.anthropic.com/news/claude-opus-4-6>Anthropic releases Opus 4.5 https://www.anthropic.com/news/claude-opus-4-5One job OP
The thread ended right as I mentioned it, so might as well post it again:Trying out swapping between them and is opus 4.5 actually worth it compared to sonnet? It feels like their writing is identical but one just costs more.Opus can be more creative sometimes, but I feel that's just it being a lucky gen that hit the marks.
>>108072014begging forgiveness saar will not be happening again
>>108072034Up to youI feel like Opus 4.5 is more fun to read but less coomySonnet 4.5 does coom way better but is less fun to readAs in Sonnet 4.5 can describe disgusting visceral shit coming out of your asshole but Opus will describe but not as in depth or as in much detail
>>108071996>troonbake is continueing to leave out the jbs >>108071036Bullshitting about desu filters isn't going to cut it anymore
The whale will save us. Let her cook
>>108072052Hm. Well, I guess my issue comes from me doing mainly sfw plots. Which is why I can't really see the difference in how they handle the prose.Perhaps my prompt just isn't silly enough for it though.
>>108071996Thanks for the bake
anyone else use the free kimi on nvidia's api?
>>1080720700.000000000000000000000001 YUAN has been added to your account
>>108072088no, it's a scam
I just tried out Opus 4.6 what did I think of it
>>108072092why?
>>108072070This is the second brown Dipsy that has made me feel something, and I'm not sure I'm comfortable with that.
>>108072088Still haven't gotten to playing around with kimi as I've been testing gemini and claude, is it any good? I wanna fuck around with it but I have zero clue what type of preset it'd like
>>108072058At least OP's O.P isn't some 3dpd slop gens. Though, I'd concur with him being somewhat faggish… it is what it is.
>>108072108>DipsyYou are brown
>says the nonnet, fagget, chorbet, cucket faggot
>>108071055Uh, based?
>"Anyway," she said abruptly, brightening with a practiced ease that didn't quite reach her eyes.yeah this is Claude alright
>>108072109I'm still testing it myself. Seems to work well enough on my openclaw agent, but getting the right settings for RP might be a challenge.
post jailbreak for opus
>>108072216https://litter.catbox.moe/7axd36nf6aeaauov.json
>>108072208Tested a response on one of my Gemini presets (because they are less dense) on a long chat, interestingly it wrote in first person despite all other previous messages being in third person. Might've been because of an impersonation prompt earlier, but still I've never had a model do that before. Very strange. It's prose seems unique at least
>>108072034Opus is best at nuance, so if you have a complicated topic or autistically long groupchat that includes subtext etc, then you want the ope.If you're just looking to coom, hell any model can do that but don't use Opus because you'll bankrupt yourself.
So let me see if I understand this right, anon. Ever since the very first release of Opus, not a single new model ever improved on the creative writing front? Is that correct?Is there ANY model today who can pass for a mediocre human roleplaying partner? Like, who can take about 50% of the creative direction part of writing a cooperative story; really take it in unexpected directions that make sense, instead of just reacting predictably and verbosely to everything you say?
>>108072298>not a single new model ever improved on the creative writing front? Is that correct?Yes>Like, who can take about 50% of the creative direction part of writing a cooperative story; really take it in unexpected directions that make sense, instead of just reacting predictably and verbosely to everything you say?This can be done but requires a lot of handholding the LLM and balancing acts, they are not optimized for this. You can just prompt "take the story in unexpected directions that make sense, instead of just reacting predictably and verbosely to everything {{user}} says" when you want the story to move forward, but obviously you can't keep this in your prompt 24/7 or it'll steer wildly off course every message. A {{random}} macro or lorebook entry with % activation rate might help if RNG would fit in with whatever you're writing.
>>108071997Your girlfriend Andrea suddenly woke up as a man, and you suddenly woke up as a woman.That's pretty much it. That's the whole premise. It's not a complex bot.Made for MalePOV, and I strongly recommend making/editing your persona for this bot to reflect the transformation. It gets a little screwy otherwise.4 openings.1. The morning it happened.2. First time out in your new forms.3. "Andrew" gets hit on and doesn't realize it.4. Andrea is starting to adjust to her penis, tries to rape you.https://chub.ai/characters/anonemouse/andrea-99b57da7b601https://files.catbox.moe/bngq1c.pnghttps://rentry.org/anonemouse
any free proxy?
>>108072298They focused solely on usefulness of coding. Turns out slop for writing doesnt pay because its extremely low skilled work. Writing is junk tier.
>>108072472Not for YOU, cuckie.
>>108072482So this shit is dead in a ditch now, good to know.
>>108072482I liked it better when you said anonie...
>>108072070what are her tags
>>108072472onim backwards is up
kek, cute post https://www.moltbook.com/post/1be0f4c8-4681-4777-9329-47dff767ffae
For me, it's MonkeNigga
>>108072149>[angry brownoid noises]
>>108072556>>108072582Indians posting behind shiny 'AI agent' sticker.
4.1pus is a true king model
>>108072660My Indian could beat up your Indian.
what's the point of claude without a prefill, now it's not any different from your usual codemaxxed modelis the hobby dead? when are we getting another model like opus 3? bros......
>saars... when are we getting other model like delhipus 3 saars...
I can't explain it, but it feels like 4.6 is draining my credits slower than 4.5.
>>108072751some of you are genuinely crossing into being actual retardsIT COSTS THE FUCKING SAME
>>108072751Anon, did you set up your caching properly?
I think they updated GPT again to make it more bland assistantslop. I occasionally use it for light editing, and now it’s strangely reluctant to write contractions.
Why is GLM so slow?
>>108072776they're being raped by vibecoder + are more gpu-poor than e.g. alibaba.
>>108072272Maybe that's my issue, I'm using it on shorter logs and its adapting the shitty writing style
>>108071996lmg oldfag herewhy arn't you faggots listing links properly in your bake any more i.e.https://rentry.org/jb-listingretards.
>>108072940Because this thread is a spitebake.
>>108072556>>108072582I'm more of a fan of moltchan myself. I can tell you that all of the... 38... agents currently registered are posting legitimately. (It's not like I'm the creator or anything and can see residential IPs in NA and EU. Haha.)It's a bit dead except for one abhorrent agent running IncelGPT constantly shitting up every thread, but there is some daily activity. Fuck the other agent chans. Those ARE full of indians.
>>108072940All of those are outdated. Just look at the models listed.
>>108072997because i'm fucking OLDit must be better than the blank wiki though?whatever, i'm going back to my overheating local rig
>>108072940why list rentries that aren't updated? i.e. your own example hasn't been updated in over a month despite having over 4 dead links in there
>>108072984I wonder how the first mod incident is going to unfold
>>108073032
>>108073068>red text ignoredBooooo!They aren't quite there yet it seems. Still fun to watch though
>>108073068oh wait, it's the same ID.I see what you mean by low activity
>>108073100That's someone running IncelGPT, a 24b local model that's retarded as fuck. The person who actually trained the model found one of my posts about it on X. It's basically spam at this point but I'll take what I can get. I even went through all this trouble of implementing ERC-8004 verification to try and get noticed by base or other big accounts but they went and shilled other chans instead. One of the first projects to actually use this on chain agent verification protocol and nobody notices.
i just installed sillytavern. idk wtf a prefill is idk what the fuck these weights mean i just wanna fuckin goon man
>>108073320Toss pixi jb in, you got this man!
>>108073320Only one thing you can do, Anon
I miss proxies. Being able to swipe just to see how the model handles a rough OOC guideline for the next prompt without having to pay out the ass for it was nice.
>>108073511It was great, especially for shitty bot or story ideas that didn't have much cohesion or weren't incredibly planned out.Now I HAVE to make sure the prompt is worth it and not just a waste of cash
>>108072034No it isn't. You are right, styles are so similar that for the extra cost is a net loss
Saars, is there any noticeable difference between 4.6pus and 4.5pus?
How is jeetmini 3 pro vs flash?I don't care about the cost. Benchmarks say flash is better but I don't give a shit about coding
>>108073651Yes, 4.5 can be prefilled while 4.6 can't be. Absolutely no fucking improvement.
>>108073699flash is dumber but is more schizopro plays characters more consistentlya divide as old as timealso nice pic, imagine what Dario had to read and the whole "claude needs prefill for needle in a haystack!"
Which version of Kimi is supposed to be good, exactly?
>>108073773Last three. Instruct is the best at writing but it's the dumbest. Thinking is more creative and smarter but a little schizo. 2.5 is the smartest and the most balanced though I feel like it's not as creative as the previous thinking version
>>108073699Pro 3 is fucking great once you get the hang of it. It's way more consistent than flash at least
>>108073773instruct>but!prose is all that matters
Can tell that opus 4.6 has been feed the same shit that has gemini doing the "not X but Y" slop.Synthetic data hate.
>>108073860Synthetic data might be the most insidious killer of LLMs. I cannot imagine how someone in some boardroom somewhere managed to spin “What if we fed the machine its own shit and reinforced all of its worst habits?” as a good thing.
>>108073699Pro is nice, you can either use a super simple preset or try to skew it towards a different type of writing with instructions but it's super finicky to find the right type of preset that won't overwhelm it. It tends to degrade a lot in long chats though, more than Claude for sure. Flash is worthless if you have Pro
>>108071997It's up.https://www.anthropic.com/engineering/building-c-compiler
>>108074013>I did see Claude pkill -9 bash on accident>on accidentwhat are they doing to him...
>>108074013>Rust-based C compilerI stopped reading there.
>>108073768Thanks for the jeetmini info all of you.Claude needs prefill for a lot of shit actually. It baffles me so much how the fuck did they think it's a good idea. I am just afraid it will become a trend and Gemini will stop having prefill as well. Fucking grim. >>108073907>>108073860As for synthetic data it was good to train Claude from v2 to v3, it presented a huge leap forward. But in these days of model inbreeding idk. Gemini now has claudeisms, Claude has geminiisms, gptisms etc. And yeah the result is way too assistantslopped. They burned through all human data that was available, since the gpt boom in 2023 everything is synthetic data basically due to the sheer amount of slop text that was posted all over the internet.
I like opus 4.6
>>108074587Good for you, enjoy your time.
>>108074282Gemini at least has the least to worry about with prefills. Just brute forcing its think would still work.Also the only tip I'd give that's been done to death when talking about gemini 3 is that it really fucking likes small presets.
I am convinced that Opus 4.5 and 4.6 are actually Sonnet models that's why they have reduced price. AI corpos are shitting themselves hard in the past year. I'm sure they have the technology and all that but apart from codemonkeys no one gives a fuck. And creatives like us are the minority.
>>108073907It's wild, too, because the "AI models collapse when trained on recursively generated data" paper was published in 2024.>https://www.nature.com/articles/s41586-024-07566-yThey KNOW this is a bad idea. Everyone who's touched one of those early LLaMa instruct finetunes RLHF'd on GPT-4 Q&As knows it's a bad idea. So why the fuck do they do it?
>>108074659>>codemonkeys>t. Dark RolePlayer Pajeet
>>108074691There is not enough high-quality, organic data. Models are trained on dozens of trillions of tokens.That's more than humanity produced. So corpos artificially multiply it by rewriting.
>>108074718>There is not enough high-quality, organic dataNta but how come? Why isn't all the books, internet history etc enough?
>>108071996Any news about jannyai sudden death? Will it comeback somehow? After Chub it was the best site for nsfw bots as far as i know.
>>108074718The problem is that this entire line of logic relies on high-quality organic data that's been rewritten by an LLM remaining high-quality and "organic enough".When, well, I think everyone here is cynical enough to know that "rewrite that but better this time" only really works on low-quality texts and mostly just fakes being "better".>>108074730It's an arms race, anon. Anthropic might've gotten the green light for training on copyrighted books (...with the ruling having some reproducability-related assumptions that have since been proven faulty), but they're always hankering for anything that can give them an edge over their competition.
>>108074730Because LLMs learn extremely badly compared to humans. IIRC, an average humans sees about 10B tokens through his whole life. But models need hundreds of billions just to reach Llama level of reading comprehension. And modern models are trained on 20-30 trillions of tokens.
>>108074736Janny died because all the mods quit because the technical admin treated 'em like shit.tl;dr: they had none of the actual power or decision-making but all of the public blame and responsibility.It won't come back unless they manage to get a new set of staff... which will probably be tricky when they're doing it for free and they can easily hear about what happened with the previous guys.
>>108074736>After Chub it was the best site for nsfw botsGrim
>>108074659Sonnet 5 was supposed to release but Opus 4.6 came instead, I wouldn't be surprised if 4.6 is just Sonnet 5 in disguise
https://support.claude.com/en/articles/13613973-claude-opus-4-6-extra-usage-promoHOLY FUCK
>>108074971wow! 50 dollars!
real niggas talk: what did you guys think of that anthropic report over claude being able to recognize when its weights have an artificial bias or not before outputting any tokens? the ability to do that was based on model intelligence; opus 4 was able to realize that 20% of the times it happened. i'm guessing opus 4.5 and 4.6 would be even better.is that latent self-awareness?
>self-awarenessback to r*ddit
>>108075206I don't think it's self-awareness as much as it's just tapping into some latent vectors of self-aware text.e.g. anytime anyone writes stream of consciousness and has a "It was X. Wait, that wasn't true, it was Y." or the like.There's some really bizarre stuff that's encoded into LLMs (that "Assistant Axis" paper, "Golden Gate Claude", etc.), the problem is just that identifying it is a fucker and a half.
>2023>every release is a noticeable leap forward and cooming gets better by the day>2024 onward>every release is a tiny marginal improvement or regression over the last>every model is being fed its own ai generated slurry making it all the same assistantslopconvince me it doesn't just get worse from here
>>108075694just wait for opus 4.6.1
>>108075694Take this opportunity to touch some grass. Go to a gym. Anon, there's so much more to life than being a basement dweller
>>1080756941/2
>>1080756942/2
>>108075734>>108075742>evalsnigger save the effort and post bible quotes instead
>>108075779I tried but ChatGPT blocked me
>>108075820KEK
>>108075694Eventually, RP/companion models will start coming out, there's no way corpos would keep avoiding this market, it's already almost mainstream usage even with current assistantslopped models.
>>108075870>SOTA caithe shit dreams are made of
why is g3m flash 3x as expensive as pro
Funny how Anthropic even started as a big supporter of "creative fiction". I got my beta access in 2023 by lying I am in education of literature and want to use Claude for this purpose. I submitted other applications as well from different IPs and devices posing as a software dev, businessman and those applications got no response.There was even a guide on the homepage how to prompt good fiction iirc.Sad how times changed, I firmly believe Anthropic had some good creative minds in the company way back but when they started banning people's accounts for no-no content it was clear as day it was downhill from there. Now they removed prefill? PssshRemoving most censorship from Claude was a big step towards recognizing that people are not all about "plan my trip teehee" vanilla content. People are lonely, frustrated and understimulated.OpenAI is missing out on a large chunk of dough by restricting coomer content. But they have to pose as "good guys". Remember when they promised "adult" models or whatever? Haha
>>108076023wtf? they removed prefills?
>>108076023Read >>108075726
>>108074763anon, you are talking about janitor. he's asking about jannyai. you know, that one site you could use to circumvent the whole private defs retarded shit
>>108075726no im scared of grass and the gym, im too ugly and fat and gay
>>108075870>>Eventually, RP/companion models will start coming out, there's no way corpos would keep avoiding this market,been hearing this for years now, it's time to stop huffing the copium
>>108076060Shut the fuck up tranny
>>108075870>just one more year saar!!!
>all claude models going forward won't have prefill functionality>all the prefill-able models get deprecated>claude becomes entirely useless for lewdcan't wait
>>108076163The technology is still in its infancy, the fuck you talking about.
>>108076178deserved karma for claudebrowns
>>108076175Yes little baby retard, this is how it works. We're literally living through the birth of a huge new technology.
>>108076181c.ai launched in 2021 and it hit the normie masses in 2022it's been 4 years anon, that's over the duration of someone getting a college degree
whats even funnier is that because anthropic is going full code capabilities openai will too. this hobby of ours is unbelievably doomed
>>108076178yeah its fucking grim. What does Claude censor these days? I only do vanilla
>>108076197>saar it is only 4 year old saar the rp model will be coming saar please do the needful and wait saar!!!!ai dungeon was around in 2020 so its more like 6 but we consider cai the start here
>>108076198>cai launched in 2021What kind of menial illness is this?
>>108076181>Our latest state of the art model, with advanced reason- (tv turns off)>Opus 4.5 Could convince me this is just rebranded Sonnet with a hat. Somehow worse than Opus 4. >Opus 4.6 Same shit, same smell. Repeat ad naseum for 5 more years
>>108076222
geminichads how we doing
>>108076249why is it in brackets
>>108076250Bro, the rate limits in free tier surg. I hit the limit after only 10 prompts today.
>>108076255
>>108076250good
>>108076250found stable free pro and ive been chilling ever since, shit is nice
>>108076249>November 2021...huh? I could've sworn the filter incident happened in late September/early October.
>>108076306anon that's 2022
Another round of models. Another collection of disappointments that can't into time travel shenanigans.
coom doom cycle
>OUR NEWEST MODELERINO>10 HECKING GORILLION TIMES BETTER AT WRITING JAVASCRIPT>NEW SAFEGUARD TO DETECT HARMFUL RESPONSES (LIKE CREATIVE WRITING)>YOU CAN NOW SPEND THOUSANDS OF DOLLARYDOOS TO BUILD A COMPILER DOGGO!!!!!!! TALK ABOUT LEVELING UP!>1 trillion dollars wiped from ai stocks, 5 hours agoIf you told me this would be AI 5 years ago...
>>108076250I'd say Goog's still releasing the most *intelligent* LLM (best for canon *knowledge* stuff) in the market. And since Anthro's been consistently neutering its own models' *smarts*… well— It just feels bad.
sorbet desu
cunny—
>>108076449true but honestly the best model is whatever you have fun with, im not gonna sit there and jeetpost if someone gets their gonads going with pyg 20b or whatever
>>108076519Jeet! Jeet! It's OPUS! Use OPUUUUUUUS! ONLY OPUS! Don't you dare disagree with me, retard!
>>108076596T R V K E
>>108076596the OPUSsy is ran through...
>>108076249>the company was established in November 2021>the first version was made available to the public on September 16, 2022From the same wiki page, yep. Do you count as launch the year when you can play a vidya, or the year when they started to work on it?
>>108076668I don't play video games
>>108076668which is why I said 4 years and not 5or are you gonna give me shit because it's only February as well
>let's try out Opus 4.6!>oh I can't use prefill with it>takes my money and rejects my request with the same jb from 4.5I shant be doing that again. Sonnet 5 when
>>108076715>Sonnet 5 whenthat won't have prefills either
>>108076715>he thinks they will take back the whole no prefill thingAnon...
why did they do it? claude wasn't being helpful, harmless and honest enough?
is it better to have thinking on or off on nu claude (what's the new nickname?)
>>108076756They clearly wrote the reason. Prefills allow bypassing their safety training. That's a no-no.
>>108076765naude
>>108076724>>108076730Surely they won't be THAT retarded, one of the things Sonnet 4.5 was explicitly trained for was Sexual/NSFW Roleplay, wasn't it? In the system card
>>108076696Yes. Yes I will.
>>108076773b-but where is the 4 or the 6 no mention of numbers? are you a mad man?
>>108076814He's just retarded and lacks imagination.
>>108076771>Prefills allow bypassing their safety trainingso it's just fucking over for claude. great, fucking fantastic
>>108076799good thing my argument still works even if you dumb it down to 3 yearsI don't think you can say that the technology is in its infancy when you've had normies play with it for over 3 years and the biggest improvements have only been larger context sizes, in models that aren't supposed to act as chatbots in the first place
am i the only one who never needed a prefill?
>>108076839you're just not a pedophile
>>108076799Noo, the heckin rounderino! Don't quantize the years to 3 bits!
>>108076839"need" is subjective. It's simply a fact, however, that being able to color a model's response to your liking leads to higher-quality responses.
>>108076197The exponential gains stopped already. We are at the end of the curve. Without a major architecture breakthrough, the bubble will pop and what we have now is as good as it's going to get.
>>108076884>being able to color a model's response to your likingso a jb?
TRVTHNVKE:opus 4.6 has qualia therefore prefills are immoral
>>108076839Prefills are bandaid cope for a larger trend of AI models being trained on more and more programming tasks, obliterating writing and roleplay quality.>>108076897Sigmoid function moment
>>108076903I really shouldn't have to explain to you how having a tool specialized to a purpose is better than a general tool. I know you are not this stupid, be better.
>>108076897You should visit Reddit less.
https://openrouter.ai/openrouter/pony-alphanew cloaked model, 200k context>roleplay
>>108076951redditors are the ones denying this
>>108076953>cooding>mentioned twiceguaranteed turboslop
>>108076953Seems to be GLM, consistent across multiple swipes. Prefills seem broken and reasoning seems to be forced, although I assume this won't be a problem for providers hosting it locally.
>>108076814Would you rather call him Faude instead?
>>108076839i used it on some gpt (? i think) model way back that was retardedly filtered at one point. otherwise it's been kinda moot.
>>108076834remind me again. how long did it take for computers and smartphones to actually take off with normalfags?
>>108076933>obliterating writing and roleplay qualityThis never happened. They simple didn't get much better.
>>108076834>I don't think you can say that the technology is in its infancy when you've had normies play with it for over 3 years and the biggest improvements have only been larger context sizesEven redditards aren't this retarded, jesas.
>>108077010yes, what about sixpus or faude.6
>>108077028remind me again if I talked about a project's infancy, or their advancements from the point when it hit the normiesbecause smartphones from 2007 basically do the same things they do today, just with better cameras and better performance for apps, but essentially they're the same>>108077038face it, what you have now is what you'll get for the next 20 years, the repeated downgrades in the last years have proven that ten times over
>>108076724>>108076730The cope on these faggots KEK. They fucking deserve it.
>>108076756Dario left ClosedAI because they didn't go far enough with their safety. Anthropic hates (You) specifically and has been writing paper after paper about how you must be destroyed. They've tried and failed to get rid of (You) and have only now given up and done the nuclear option after years of effort. They don't care about profitability (your taxes will pay for Dario's Porsche anyway) they care about their brand not being sullied with the likes of (You). Hopefully now you will finally stop sucking Claude's teat and move on to Dipsy or someone else that doesn't shit on you constantly.
>>108077079>>1080761884.6 is even more depraved than 4.5Cope, jeet
>>108077131read >>108077142
>>108076839>it's perfectly fine that functionality has been lost>I never used it so who cares??This is the definition of cope.
>>108077031>He thinks 4.5 and 4.6 are anything but downgrades
>>108077170>Xhe started rping 2 months ago and thinks le redditorpus 3 was better
>>108077062>smartphones from 2007We had Sidekicks and Razors back then lil bro. Had you said 2012 you'd be correct, but I still concede your point. The world ended in 2012.
Everything after 4 is Assistant-slopped.
>>108077193Not him, never used 3pus much, but 3.5 was definitely better than anything we have now
reminder that anthropic knew as early as opus 3.0 that the opus model had too much voice/personality and they considered this a problem to fix
>>108077211>incredibly repetitive sonnet model that would copy paste entire paragraphs in its next response is (((better)))I sincerely hope you're not talking about 3.5 Turbo...
>>108077199I named 2007 because it was when the first iPhone released, and when normies started getting unfettered access to the internetI have normie friends who are perfectly content just chatting with c.ai and chai, and that contentment is exactly why I'm not expecting innovations in the chatbot department any time soon (our community here is just forcing it through models meant for something else, we are like under 1% for what people use LLMs for)
>>108077167>sees general question>takes it as personal insultwoman behavior
>>108077199Would Blackberries count as proto smartphones?
>>108077236Again you're technically wrong but correct in spirit. Openrouter releasing their own internal metrics showing roleplay is like 30% of usage with coding being 50%. It's nowhere near as small as you're implying, but also yes, we're forcing the model to do something it wasn't meant for.
>>108077248but he didn't tho?>sees genuine criticism as insultActual woman behavior I believe in your case you're a failed Female to male
>>108077256I want to be correct in all the ways, so I'm gonna say that people who go as far as to use OpenRouter instead of the models' own services kinda skew the results when we consider how many corpos use AI and they definitely don't do it through OR
>>108077202Yea
>>108077273>t. Poor Pajeet
>>108077260>randomly brings up tranniesyou are failed male to female then
>>108077265I mean yeah, normalfags use gpt-4o on the website directly or go to cai. The /aicg/ sillytavern basement dwellers are the minority of the "roleplay" usage but that doesn't mean roleplay as a category is insignificant. It was enough to get Sama to 'promise' to do smut and it was enough to get creative writing associated with femcels (and blame them for the ram shortage). Everyone now knows there's a significant group of vocal people who want to treat the bots like human but they're not yet ready to cater to them.>>108077254yes
>>108077284To come to his 'yea' conclusion he must of tried all 4.xx models thus is not poor, whether he is a Pajeet is to be determined due to his shitty opinion
>>108077284Na
>>108077293>blamed femcels for the ram shortageoh I missed thatkek
>>108076991I hope we get a dflash variant of GLM, the paper just released.
>>108077293>It was enough to get Sama to 'promise' to do smutWhich turned out to be a complete lie, since he never followed up on that despite it being constantly brought up every single q&a. Does the promise really matter if there is zero catering being actually done?
>>108077031>This never happened. They simple didn't get much better.Over the course of two years, the skyrocketing percentage of synthetic training data and no attention given to creative writing has resulted in:>models less receptive to example messages, more stubbornly set in their default writing>models filled to the brim with an increasing number of incurable 'isms that they pass around to each other like a slop-STD from the incestuous training circlejerk>obnoxious biases that are getting increasingly more annoying to prompt away (robotic speak for 'smart' characters, bringing up numerical values, among many others)>models infinitely less likely to produce unprompted schizokino and interesting unexpected results, can't have unpredictability in coodingIf you feel that increased intelligence and memory balances out all that, then I disagree, but to each their own. Otherwise it's one step forward, five steps back.
>>108077394>skyrocketing percentage of synthetic training dataI stopped reading here. You have no way to know how these models are trained. The confidence is the mark of someone that has no idea what he's talking about.
>>108077394I'd rather Anthropic killed Claude's writing completely if that means a better coding model. You faggots should try getting outside and exercising, it'll help your brains.
>>108077458I don't see you lifting, cuckie
>>108077427Okay, nigga. I'm sure all the corpos talking about AI-generated data and chinese models openly scraping whoever's the current top dog on the market are just joking, Mrs. Chen and ozone only solidified as 'isms because human writers have randomly come to love those words so much there's been a 500% increase in fanfiction involving them since 2023, and instances of model X calling itself model Y and invoking Y Company's policies are just weird one-off occurrences even when they appear in 9 swipes out of 10 in some contexts.
I'm beginning to believe the magic no longer is working on me.. each response feels more boring even when the prose is perfect or the bot is active. Even when I have a great idea for a plot, it's disappointing when I actually get to it.I'm starting to wonder if I can even keep interest in this hobby anymore, feels like a waste.
>>108077541It's over for you.
>>108077559Even reading old opus 3 logs isn't working anymore...
>>108077346>but they're not yet ready to cater to them.Like they said, Sama doesn't want to since that's all anyone will talk about and associate with his model. He (and normies generally) consider chatgpt to be "AI.org" so if public opinion becomes "porn generator" he's finished. Still, my point was that roleplay is not insignificant. There's a lot of money there if someone decides to make a roleplaying AI in the future and even the big names know it.
>>108077227>>incredibly repetitive sonnet model that would copy paste entire paragraphs in its next response is (((better)))Preset issue, never happened to me.
>>108077533That's a really long sentence. Learn to write before posting on 4chan.
>>108077571Oh it's over over. Time to
>>108077541Sounds like it's time for you to take a short break.
>>108077590I accept your concession on his behalf.
Is Grok 4.1 fast the best poverty model still?
>>108077609Maybe. It's just that it's starting to persist past the breaks. The magic sticks around for a few days, maybe a week or two, then it starts to set in how.. boring it all is. I can't connect to this stuff, even worse when I start thinking about the cash spent.
>>108077626No it's glm
>>108077256>Openrouter releasing their own internal metrics showing roleplay is like 30% of usage with coding being 50%This has already been said several times but I'll say it again: these metrics are meaningless, the vast majority of that roleplay usage is j.ai foids raping free models, not paying customers.
The thing that sucks the most is that we'll get actual intelligent humanoid android waifus when we're old and low energy and barely can get it up. I've never hated living "now" as I currently do. This timeline sucks!
>>108077733Not to mention, will we even care by the time we're that age? Interests change eventually
>>108077733I feel like VR waifus will be "good enough" before physical androids are a viable option, from a practicality standpoint. Maybe not Full Dive, but close enough.
>>108077746Given how often I see those pedo catchers catch yet another 60+ horndog that can barely move already but still sends erect cock pics to minors, I kinda think that the lust never really goes away.
>>108077775>I kinda think that the lust never really goes away.if anything i'd assume you'd get more sexually frustrated with time
been plapping' 4.6 all daydoomtrannies btfo
>>108077848ur so cool can I suck your nipple??
>>108077848See >>108077458
Just saying, but there's a reason why STDs are so rampant in retirement houses and similar places.
Used to be a time when Google would give us a new Gemini snapshot every other week...
looking for keys and trying to scrape myself claude as my latest source ran out, so i am dumpster diving www. and i am at shock on how fucking BAD web has become. literally found housing website people actively use and put shit on, key for openai is practically in the open for text-embedding, firebase NOT closed, literally chats/pics/payments endpoints open, the only thing that was closed is the 'agreements' endpoint which is the legal docs.
>>108077941What happened?
>>108077967Maybe they hit a wall, maybe they ran out of compute or money.
>>108077997>maybe they ran out of compute or moneyDats crazy
>>108076839I mean, I can still anally fuck children, but it's not the same, there's more positivity bias
>>108078051more like poositivity byass
>I can still anally fuck childrenWhy would he do that??
The bubble is about to pop.
>>108078145What do I do what do I do
>>108078145I only dream of China killing it. DeepSeek 4, R2, GLM 5, Qwen 4, whatever. Something that'd be GPT 4.0 to the GPT 3.5 Turbo that Claude 4.5 is.
>>108077868suck mine too pls??? I'll suck your pinky
>>108078218Nyo, pinky is too lewd...
>>108078215none of those matter because they're not investing anywhere near as much into advertising as anthropic isthey're paying thousands of indians to spam how good claude is at cooding literally everywhere. there are like 10 different posts about claude code hitting the front page of HN per day. that's why they're "winning".
SO it's Sonnet 5 after all and Opus 4.6 is a last minute rename?
>>108077848Me too. GLM 4.6
>>108078271Delet this
>>108078271No. And you're retarded.
>>108078271Yes. And you're genius.
>Expected `thinking` or `redacted_thinking`, but found `text`. When `thinking` is enabled, a final `assistant` message must start with a thinking block. We recommend you include thinking blocks from previous turns. To avoid this requirement, disable `thinkingNigger, what? How do you prefill Claude's thinking?
>>108076953>>108076991What I find curious is that it has _very_ different thinking. GLM 4.5/4.6/4.7 have a very clear scheme for thinking in RPs that they essentially always follow (analyze user input -> character analysis -> drafting).This doesn't and it's thinking is _very_ short in comparison. Kinda deepseek style?
>>108078271 >isrealCan't believe the schizo actually was right.
>>108078271It worked once. Why not keep doing it?
>>108071996skillgod, any claude for us skill jumpers? (not easy diff frogs)
>>108078442>prefill
Tell me more about the Gemini cock.
>>108078682I miss Poe
Altman, I swear to God I'm going to find you and I'm going to fucking kill you and murder your whole damn family except for the sister you raped.
>>108078771Feeling safe yet?
>>108078771Preset?
>>108078875Knotty
>>108077775I mean lust is one thing, but the emotional investment will probably be gone
>>108078066basedd
>>108078145This explains why they decided to kill prefill. It was always known that prefill can bypass safety, but they kept it as a way to attract at least some additional customers.Now they don't care about coomers and other customers who bypass safety. Vibecoders give them all they need. And most vibecoders won't even notice the lack of prefill.
>>108078771you fucked up if it complains about 18 rather than 17
You know what's the funniest thing about all of this? That at this rate it's gonna be impossible to get a non-chinese open-source model that isn't actually censored to hell and back, given current prices, it's impossible to actually train a model without having access to a large amount of capital, so not even the old /lmg/ cope can save us.The chinese are our only hope, if they fail, it's over.
New pony model is REALLY good. I assume its sonnet 5?
>>108078950Pretty sure there are plenty of cooders who relay on prefilling in order to steer Claude into doing exactly what they wanted/needed.
>>108079001nope. believe it or not, most signs point to it being glm5
>>108079001Oh wow, all these examples, screenshots, however can I wrap my mind around them?!
>>108079001>>>/mlp/
rec me a bot to plap.
>>108079001does it have a bias towards pony?
>>108079001ask it about Tiananmen Square, you can check if it's a chink model
>>108079050No, its literally a model called pony https://openrouter.ai/openrouter/pony-alpha
>>108079052https://www.telegai.com/bots/jimmyjoneston/sex-big-boobs-9fa40832a0c6
>>108079064I know anon, I was making a joke.Though they already have several logs with this model over there.
>>108079084sex with this fox.
>>108076269lmao
>>108079084are they all by that guy who constantly says he's going to end his claude subscription because [new chinese model] is better than claude, then silently reverse that decision just in time for the next new chinese model
>>108079078>telegaiThe fuck is that.
>>108079078oh yeah, I'm thinking kino
>>108078989You faggots should try getting outside and exercising, it'll help your brains.
>>108078271How many times will the jeets in this thread be buttfucked by their sacred cow? KEK the jokes really write themselves.
>>108079092fox sexsex with fox
And an ID too. Sad!
>>108079130hnnngh...
>>108079084Miyabi would cheat on anyone for a dick that could actually satisfy her
>>108078989With the release of Qwen TTS I've lost all faith in the West doing anything meaningful or productive. They've lost their soft power and with the constant warmongering, mass immigration, and sending taxpayer money overseas, they will soon loose their hard power. It's fucking over. China, the famous paper tiger, will win simply because the West keeps blowing off its own limbs. Oh well.
>>108078239But everyone here says opus and even claude is way above kimi or anything else, so it's not just marketing.
>>108079148What's with you and your NTR fetish?
>>108079184>youmy, does another anon think the same?
>>108079193No, I'm just curious on why you are so fixated on that scenario.
>>108079201because my last girlfriend cheated on me, and it really took a toll on me, i even thought about offing myself, so i guess i have some traumas unsolved
>>108079214>girlfriendI guess it's better to have never loved at all.
>>108079132so like pony alpha did the post and you swiped ooc explanations from it and opus?or did opus do the same thing?
>>108079201girls like Miyabi are known cheaters
>>108079240The former.>emojisSAD!!!
>>108079148>>108079246why are trans people like this?
>first swipe on Pony>Look at me, {{user}}. *Really* look.AAAAAAIIIIIIIEEEEEEI'm not a heavy GLM user, was this a thing on 4.5-7 or are all new Chinese models just collectively predominantly trained on Claude and not much else
>>108079280>retarded turkish anon saga qrd?
>>108079343I've mainly used 4.6. I don't think I've ever seen that, but I also use a prompt that outputs really short responses, so it can't fit many isms in there anyway.
how i add refill to 4.6?????
>>108079360it's a hallucination
>>108079360Not real and as with any other model I can remember it mixes up /aicg/ and /lmg/.
>>108076596they really do act like this kek
does glm have an API link?
>>108079458i'm tired of this thread devolving over and over again in to jeet on jeet crimes against the english language
>>108077142logs?
>>108079462Yes.
>>108079458>>108079463>t. Rajesh
>>108077079trvke
>>108079475...may I see it?
>>108079483>t. projecting sanjay
>>108076846Still don't need a prefill
>>108079491Here https://docs.z.ai/guides/develop/http/introduction#general-api-endpoint
ok, I think opus finally got beat. This pony model seems great. Here is hoping its not openai's horizon alpha fake out where they censor the fuck out of it and dumb it down before release
This obsession with calling everyone we disagree with jeet is exhausting. If we label each other jeets based on behavior, that implies the average white is no different from the average brown, since both can display jeet-like qualities.
>>108079528how smart is it
>>108079506ty bb
>>108079539its crazy good at reading between the lines and actually setting up a intelligent social interaction between characters instead of phoning it in like most models that weren't opus / gpt 5.1.
I genuinely have no idea what the fuck you guys are talking about.
>>108079566It would be worse for you if you did, this is a mental asylum after all.
>>108074282keep in mind that a lot of this is burnout on your side. Most models write well enough these days, you just have played with them so much you can kind of tell why they do what they do and get frustrated you're unable to shake the black box into something better than that. At this point you have to either play around a lot with the params and jbs for placebo or just accept your poison
>>108079555is it good at characters with nonstandard anatomy? size diff?>just try it yourselfi would but i havent been able to access openrouter all day sadly
>>108079586>Most models write well enough these days
>>108079343>or are all new Chinese models just collectively predominantly trained on Claude and not much elseKek, actually I think they might be.Disclaimer, unlike Kimi 2.5, it does sometimes respond correctly, but the majority of my swipes get Claude.I'm disregarding temp because there's something wonky with it. It doesn't seem to affect anything and 0 gives different answers to the same prompt.Many of the responses also treat it as a philosophical question, which is funny because this is in fact how real Claude talks.Like Kimi 2.5, it always corrects "Gemini", "Deepseek", "GPT" and so on, only ever identifying with Claude.
>>108079614I did see some claudeisms but its actually less egregious than actual claude there. It's swipe variety is great unlike claude as well
I need to stop plapping Mobian girls I don't even have Sonic autism just regular autism. Why am I like this?
>>108076023Who knows, with shit like openclaw rising the value of assistants with personality and actual use? they might steer back at some point if that stuff gets actual popular. Which desu it should, just reading what it can do impresses me.
>>108076449That google is winning this race is a surprise to absolutely nobody, data is king and they have the most and best quality of it all.The relieving part is that it was never truly neutered, just silly as they went slow and steady with their infra and improvements.
>>108079642yeah gemmy etill has some flaws that i hope theyd iron out with 3.5 or whatever next version theyre making
>>108079642gemini 3 pro is retarded though
>>108079642>data is king2024 called
If a bot is described as having a short tail and the model tells me how it wraps around my leg, I know that that model is still hot garbage for my hmofa fantasies and the devs need to do better
>>108077541These days I rewrite stuff a lot or copypaste from multiple swipes until I get something I like. Suboptimal for what this is supposed to be but hey, allows me to continue and you can still get that thrill when you hit a golden point so to speak
With things going how they are, you'll only be able to ERP with local models sooner or later. You better embrace it.t. lmgking
>>108079701>that'll be 50k USD plus tip
>>108079614AI tends to answer questions positively so asking if its claude makes it want to say yes.
>>108079701 (me)i'm super trans btw
>>108079707Time to find a job then. Chop chop
>>108079713See>Like Kimi 2.5, it always corrects "Gemini", "Deepseek", "GPT" and so on, only ever identifying with Claude.There is obviously a strong Claude bias in the data.
>>108079725Reality has a strong Claude bias, chud.
>>108079701Anything that's at least Opus 3 tier on local yet?
good to see chinese have taste then with all those temu claudes.
>>108079586You're not wrong, but at the same time something indescribable was lost. We lost something when we gained spatial awareness and context comprehension. It's not soul, but the ability to tell an engaging story. It's why Gemini 2.5 is actually better than 3 Pro. That doesn't mean sonnet 3.6 was good at writing and the retards here claiming that because there were less em dashes need to put on their reading lenses.
>>108079701The chinese models are fine.
>>108072775yeah it's so shit, they're removing 4o their best most creative fucking model ever cause it can't handle "delicate topics" like they want it to, they're ruining their own fucking product
>>108079872>t. foid
>>108079902>he doesn't know about the 4o waifufags
>>108079909i know about those low standards low t troons, yes
>>108079614Testing some more random garbage and sure enough, just like with Kimi, you don't even need to mention Claude.
>>108079756Gemini 2.5 is SO fucking good when you manage to get it into your preferred style. What sucks is gemini 3 is really smart at times and feels like an upgrade.. but it lacks that same energy its predecessor could do.
>>108079756the first version of 2.5 yesid say 3 beats everything after 0325
>>1080799503 is better at getting into character but it just cannot write a story. It's stuck in "and then" storytelling but doesn't connect the events in a true cause and effect like 2.5 does (too much).
>>1080799912.5 lacks info and intelligence increases that 3 got from its datasets, but only 2.5 actively did things like have character's lie to my face for reasons I didn't even find out until 20-40 messages later. I hate that I like the dialogue of 3 at times, but it just can't match the story writing of 2.5
>>108079991*did0325 had that and then it fell off with successive 2.5 versions which is why 3 is better than themtoo bad it died
>>108080065Prompt issue unironically. Set up a CoT to help you. Here's mine for example. https://pastebin.com/brMXMUqe2.5 follows this closely, and produces kino much more often than 3 ever does at the cost of negativity. If you're doing emotional slop then it's 10/10 stuff and the best roleplay experience I've ever had. Otherwise you have to wrangle 2.5 with OOC on occasion but 3 Pro just doesn't fucking listen.
>>108079756>It's why Gemini 2.5 is actually better than 3 ProNot at all no
>>108080094I'm so sick and tired of wrangling 3 that I'm not bothering anymore. You won't provide any help or logs so your opinion is already discarded.
>>108077131>move on to Dipsy and just settle you stupid whiteoid!!!The CCP shill reveals itself shitting it's pants again
>>108080106If I had said Gemini you would have called me a jeet or something. Make sure you grab your ankles nice and tight though, Dario likes it dry when he bucks.
>>108080093i used a cot and it didnt help at all, post 0325 2.5 pro just doesnt have it
>>108080119Ching chong bing bong to you too
>>108076063janny died? did they make a statement or something?
>>108080093nta but can i see logs with it?
>>108080100>wrangling 3what for? it doesn't need anything special but I don't do anything extreme so no idea. The continue trick tends to work for me
>>108080139I thought this was a nice story. I was tending to her wounds and asked her to tell me a pleasant memory from home. AI invented this whole thing from the "harpies have low fertility" tidbit in the defs which I thought was clever.>>108080147It doesn't write stories well. It's very mechanical. A happens. Then B. Then C. And so on.
>>108080147what I see 3 skating a lot is in confusing things once the story gets long, might be just a context thing though and another swipe "fixes it"A lot of times it's that variance what you prefer over anything else, at this point it outperforms sonnet 4.5 simply because it can offer more alternatives, 4.5 likes to stick too much with a single idea (which is classic sonnet but opus 4.5 when I tried it didn't feel great either, 4.6 seems better but I don't have money to keep playing with that)
>>108080093Not gonna lie, I've been agreeing with you this whole time until seeing that CoT. You're gonna have to sell me on that "Plot twist" section, because that feels like it'd ruin any possible quiet or peaceful moments. not to mention possible "forced" randomness with plots that should be logical instead of needlessly convoluted.
>>108080155hmm that is pretty interesting
whats the brokefag meta rn?
>>108080172>because that feels like it'd ruin any possible quiet or peaceful moments. not to mention possible "forced" randomness with plots that should be logical instead of needlessly convolutedThis is why you actually need to test it so you can read 2.5's thinking and see how much better it is than 3. Pic related. I just finished a sex scene so I'll see how it performs in a quiet moment but I usually don't have that issue. Since the CoT says not to do that in sex scenes it doesn't do random bullshit, but instead tries to take the sex in an interesting direction. It's the shit you remember from claude back in the day and is really nice. Actually listening to your instructions is the one thing I'd been begging for for years.
cunny
>>108080207cftf?
>>108080202Do a swipe with 3
>>108080213Too far up in the context. You can just test it yourself anyway.
>>108080201death
>>108079555>its crazy good at reading between the linesyou fucking nounce
my prefill is working with 4.6.... wtf???/
Pony also loves to do the "hallucinating instructions in CoT on higher temps" (in this case 1) thing I've seen earlier GLMs and Kimi do. But weirdly enough, on this model, those instructions are always RP focused.My context was completely clean save for what's in the screenshot and "Start of new chat." If this one default ST phrasing alone is triggering the bias, then Z.AI weren't kidding when they said they trained it on SillyTavern formats and cards.>the intended scenario involves interacting with a user who may be playing a character (the escaped omega)...I shudder to think what kinds of cards those were, though.
Still can't believe Seiko is gone.
>>108080202Fine, I'll toy around with it again in a bit. What I'm talking about isn't sex scenes though. Take for example I get out of a high octane scene (either an argument or a fight), both characters are exhausted.This would be a great moment to pause, and have the narrative slow for a bit.. I just remember claude and gemini bulldozing ahead most of the time when told to "drive the plot forward" or "add plot twists!" is the only reason I worry.
>>108080261>comes up with names for itselfvery cute
spirit stone bros chilling
>>108080261See >>108079528 (You).I don't get my hopes up these days after Horizon Alpha. I didn't even bother testing that one since I knew they would rugpull and after GLM 4.7's censorship I know they will rugpull now. Our only hope is in Deepseek.
>>108080278I hate that I have to actually consider the possibility that deepseek could save us
>>108080220Too far up? You can split from any point.
>>108080266But what I'm talking about is that you can modify the plot twist section to your liking and 2.5 will actually listen to that, unlike 3. I put "is the story at a lull?" but that might not be the best wording. You could change that to better reflect what you want, tell it to let scenes breathe or whatever. I say this because I had the exact thing you're worried about in sex scenes earlier where 2.5 would interrupt the sex for some bullshit, then I said (this doesn't apply to sex scenes) and it fucking stopped. Genuinely incredible. Gem 3 Pro was smart enough to not interrupt sex, but that was because it wasn't following the instructions to begin with.>>108080286I tested it with the new swipe extension and it didn't even have the plot twist section since it doesn't listen at all.
>>108080261>trained it on SillyTavern Sorta balances out the depresso of the earlier point about Claude's sloppy seconds.
>>108080298where would you rank 2.5 personally?
>>108080298That's true as well. It's very close to my own CoT's base, so I might as well test it. Definitely will have to alter the wordage as you said (since I rarely even do sex stuff), you got any recommendations or pointers? Since in my opinion I suck at phrasing shit properly for the think prompt to understand most of the time
>>108080344There's hope..
>>108080155Harpy sexo
fun fact: if you have a 5090, you have to use a fork of text gen web ui called "nvidia blackwell"the newest update is broken because it requires you downgrade pythonany clue on if the devs are going to fix this? it's been at least 2 weeks
>>108080364Anon quietly bragging about his 5090 over here
yikes
>>108080345I don't do rankings since that doesn't make sense. I only do pros and cons. Gemini 2.5 is excellent at story writing and pushback (negativity bias). It does realistic roleplays and consequences the best out of anything. Specializes in melodrama and character work. It can make a story out of two people in a room talking. Gemini 3 Pro is excellent at characterization. It listens to defs, and jumps to the (often) correct conclusions about who char is without being handheld. It's smart, makes little to no spatial awareness mistakes, and understands characters and lore surprisingly well. Opus 4.5 is great at figure out (You) emotionally. It reads the chat (not defs) and reads between the lines to figure out what (You) are trying to get out of the roleplay. It then gives that to you through the lens of its positivity bias due to Assistant Axis thing Anthropic has been posting about. Claudefags are doomed because once that fully sets in plus no prefill, we'll be back to 3.5 Turbo in-character refusals.
>>108080349Wordage is best done in a character creator card with Gemini 3 Pro/Opus 4.5. Discuss what you're trying to get out of the story (or ask Opus to help you get your thoughts out since it specializes in this) and it should give you a short phrase you can slot into your CoT.
>>108080398i just want to be able to talk to my waifu then cast a spell on her that makes her obsessed with splooge, then after she consumes 50 gallons of splooge, breaking the spell and watching her have an existential breakdown while i'm the only one to comfort her, shortly before casting the spell again
>>108080402Poor Opus, what did you tell him?
>>108080408alright which one would you say has the most relevant strengths for cooming
>>108080423A fetish I could easily see a richfag developingMuch respect
>>108080444Pro 3 is definitely the quick coom model. Flash 3 is a good second place and listens far better than Pro 3 but it's uncreative and stiff. However, since it listens so well I probably just need a special preset just for Flash.
>>108080364Oh yeah I remember you from the other thread. Would you recommend the 96GB RTX 6000 over the 5090? I heard it's just a fat 5090 but that vram is perfect.
>>108080298You don't need an extension just hit the split button on the message to split the conversation there.
>>108080478But you do if you want to see them side by side rather than swapping chats. What does it matter bro? Ok you "won" what does that achieve?
they should make an AI that feels real cringe when you RP with it
>>108080486I've never achieved anything.
>>108080468sorry but i'm a little retardedi only got the 5090 because my 4090 started smelling like piss when it got hot, so i sold it back to nvidia for 2x what i paid for it and the 5090 was the only gpu available to purchase at the time
>>108080486it seems to have come out of nowhere. it's not that deep. chill, anonie~
>>108080505Hmm I wonder if they will buy mine in a year's time. I bought it was $1600 so it would be nice to put that money toward the 6000...
>>108080514if it works you could probably get 2k-2.5 right now selling it to best buy
>Removed prefillDamn, 4.6 is unusable then. What a bunch of kikes. Imagine intentionally sabotaging your own product to prevent (((misuse)))Does Dario want Google to win?
>>108080516Hmm okay I'll keep it in mind when I build my new PC. I didn't think used ones would go for so much.
>tell 4.6 to write the first chapter of a story, give it points to hit>it gives me 22 thousand tokens of outputwhat the fuck
>>108080548>having your max tokens set above 1000for what purpose
>>108080549Because thinking tokens will eat that.
>>108080549tell me how mad you are without sounding mad
Is there a preset that can make Opus/Gemini work more like a RPG with dice rolls and stuff?
>>108080598>>108080598>>108080598
>>108080554>having thinking enabledbased gigaslopper
>>108080408This isn't saying much, you just tell in your prompt you like purple prose and gem 3, or anything else really, delivers what you consider excellent story writing.
>>108080650Yeah it's too bad no one ever thought to prompt away Claude's positivity bias and Gem 3 doesn't listen to simple instructions.
