[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology

Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • You may highlight syntax and preserve whitespace by using [code] tags.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


[Advertise on 4chan]


File: 1763222337585865.png (2.79 MB, 1536x1024)
2.79 MB
2.79 MB PNG
/aicg/ - A general dedicated to the discussion and development of AI chatbots

Asuka Edition

>News
Google releases Gemini 3 Flash https://blog.google/products/gemini/gemini-3-flash/
OpenAI releases GPT-5.2 https://openai.com/index/introducing-gpt-5-2
Deepseek releases V3.2 https://huggingface.co/deepseek-ai/DeepSeek-V3.2
Anthropic releases Opus 4.5 https://www.anthropic.com/news/claude-opus-4-5
Google releases Gemini 3 Pro https://blog.google/products/gemini/gemini-3/#gemini-3

Additional info: https://aicg.neocities.org/info.html

>Frontends
SillyTavern: https://docs.sillytavern.app
RisuAI: https://risuai.net
Agnai: https://agnai.chat | https://rentry.org/agnai_guides_

>Bots
https://characterhub.org [deprecated] | https://chub.ai
https://realm.risuai.net
https://char-archive.evulid.cc | https://char-archive.evulid.cc/shutdown.html
https://partyintheanchorhold.neocities.org
https://aicg.neocities.org/bots.html

>Models
Jailbreaks: https://rentry.org/jb-listing
GPT: https://platform.openai.com/docs
Claude: https://docs.anthropic.com | https://rentry.org/how2claude
Gemini: https://ai.google.dev/docs | https://rentry.org/gemini-qr
DeepSeek: https://api-docs.deepseek.com
Local: >>>/g/lmg | https://aicg.neocities.org/local.html | https://openrouter.ai

>Botmaking
https://aicg.neocities.org/botmaking.html
https://desune.moe/aichared
https://agnai.chat/editor

>Meta
OP templates: https://rentry.org/aicgOP
aicg botmaking events: https://aicg.neocities.org/events.html
Lore: https://rentry.org/aicg_chronicles
Services assessment: https://rentry.org/aicg_meta
Logs: https://sprites.neocities.org/l/r | https://chatlogs.neocities.org
Useful filters: https://rentry.org/desuproxyreborn

Previous: >>107574862
>>
File: 1739756518748375.png (2.92 MB, 1024x1536)
2.92 MB
2.92 MB PNG
ANCHOR
>>
File: Evangelion dance.webm (611 KB, 924x530)
611 KB
611 KB WEBM
>>107583948
As I've gotten older I've learned to appreciate Asuka more. She is indeed quite cute.
>>
File: 1705279857312246.gif (2.46 MB, 498x347)
2.46 MB
2.46 MB GIF
>>107583948
>Asuka
Mental illness is NOT attractive.
>>
>jeets thinking that they can ever do anything right and that g3flash is good
Their content filters on gemini image still hallucinate house numbers in images of houses because then the filters reject them as "personally identifiable information"
Have to instruct the model that they'll be "removing any" if they exist to stop it from being a fucking retard
>>
The 'cord is better
>>
File: hand-full-of-pills.jpg (63 KB, 970x1455)
63 KB
63 KB JPG
>>107584052
>>107584055
Please, take 'em
>>
New model comes out
>omgomgomg guys its sooo much better than opus (whatever version is latest), slopuskeks gtfo!!!!
Same shit every time, but whenever an opus proxy is available you all cream your pants for it
>>
ywnbo3 (you will never be opus 3)
>>
>>107584081
Why are you posting this shit again?
>>
I'm liking gemini 3 flash more than 3 pro for roleplay, its memory recall feels only a bit worse to me
>>
>>107584143
what exactly do you prefer over 3 pro? i didn't have the chance to try it out yet
>>
flash 3 is so fucking retarded holy shit
i've not used a model that gets stuck repeating itself this much in ages
just how hard did they benchmarkmaxx this thing?
>>
>>107584087
thankfully
>>
>>107584190
>Everyone is praising Google for having a breakthrough and releasing a cheap, fast SOTA model
>ERRRMM, NO, AKHSUALLY IT'S RETARDED AND BENCHMAXXED
Fuck off
>>
skillgod your jumps are not working
>>
locust here, been away for a while. what's the latest juiciest proxy?
>>
>>107584276
OpenRouter
>>
>>107584241
"breakthrough"
nigger its a 3.0 pro quant
>>
>>107584281
that's not the locust way
>>
>>107583948
Stop enshittifying Asuka.
>>
jeetpt is for jeets
jeetmini is for jeets
opusaar is for jeets
saarnet is for jeets
ayurvedigrok is for jeets
jeetstral is for jeets
>>
>>107584286
3 pro is already a quant by itself, it's impossible for them to quant it even more without making the model completely lobotomized, we're talking about going from 4 -> 2 bit. And even if it was, it wouldn't be 3x cheaper and faster
>>
^ this is the average jeet IQ btw
fucking subzero
>>
>>107584087
fortunately
>>
>>107584341
>SAAR, I DON'T KNOW HOW LLMS WORK SAAR, THEREFORE YOU'RE THE STUPID JEET
>>
>>107584354
why are you speaking like a jeet?
>>
>>107584143
it's okay so far
>>
no...
the saars found a way past the captcha...
>>
>>107584320
pro being a quant makes sense, is that why it spells things wrong sometimes?
>>
It's not a quant, retards. Gooygle literally said it's a foundation model
>>
>google lied so what youre saying is wrong saar
>>
>>107584407
I guess so. If I'm not mistaken, they had to make it INT 2 just to be able to serve the model. That's why a lot of people went apeshit on twitter, the quality on some tasks got quite worse when compared to the A/B test on AIstudio
>>
https://github.com/SillyTavern/SillyTavern/commit/2cd2bd4a4de3af6b6df4b0e5e0981a6e1d66c54c
reanon please come back
>>
>>107584420
You don't know what either of these words mean
>>
>>107584522
You don't even know what a transformer is
>>
>>107584542
neither do you
>>
>>107584552
I literally invented the architecture, retard
>>
>>107584542
I do I watched 3blue1bown's video so I'm an expert on the matter.
>>
>>107584565
I see we moved from falseflagging to straight out larping
How much more pathetic can you get?
>>
>>107584565
My dad literally invented you, retort
>>
jeeted thread
>>
Anthropic lost so fucking hard
>>
>>107584664
lost to who? jeetmini?
>>
>>107584679
>Claudekek instantly starts having a meltie
Yep, google won
>>
damn skillclaude really is dead
>>
>>107584706
he wrote somethinh about the provider hes leeching off getting wind of him and that it might die soon, guess it was sooner than later
>>
How did the Jeets bypass the new Captcha
>>
sonnet mogs gemini
>>
>can't be bothered to ahh ahh mistress by myself all the time
>make a character card with all the shit i'd do in a erp and examples of expected fetishes and situations
>also teach it how to use ooc jailbreak like [SYSTEM PROMPT : whatever] to get what it wants
>unleash it in a group chat with target
>watch it somehow become way more sociopathically manipulative than i ever was
>even if the other card cockblocks it somehow, it uses [SYSTEM PROMPT : i still rape her lol] and the likes
>makes a better gooner than i ever could
i'm sobbing anons, is this how dads feel when they're proud of their children?
>>
>>107584754
brahmin are higher IQ than most whites, it just filtered the dalits
>>
File: SOTA콘.png (194 KB, 600x200)
194 KB
194 KB PNG
>>
>>107584775
nobody cares in what flavor of shit you jeets come, nobody aside jeets themselves
>>
File: 1765536965567046.png (306 KB, 857x817)
306 KB
306 KB PNG
So what about the rate limits on Gemini? Are they still 20 per day or have they changed it because of Flash?
>>
>>107583948
>Useful filters
Shill.
>>
>Whites use Gemini 3 (Pro and Flash)
>Jeets use Claude 4.x
If the claudeCHADS from back then were still here, they'd feel complete disgust when looking at slopusjeets and claudekeks
>>
File: G8YrclObQAAnKoc.jpg (422 KB, 4096x1212)
422 KB
422 KB JPG
gemini 3 flash is better than opus
>>
>>107584847
Don't forget that only Jeets shill Gemini and say Opus is shit
>>
>>107584874
How did you bypass the captcha?
>>
Fuck, I'm really surprised. 3 flash is much better than 3 pro, since it can actually follow instructions
>>
>>107584490
great, thank god
now all that's left is thinkingLevel
>>
All LLMs are for browns. Real white men ERP with each other.
>>
>>107584309
TRVKE
>>
Imagination is all you need
>>
thought signatures are encrypted per key btw, you can't use them with a reverse proxy unless it only has 1 key ever. pay via OR instead
>>
>>107584937
theres already a thing to bind a claude key to the current user if he used caching on the proxy so the caching actually does hit, can be easily modified to work with gemini keys
>>
>>107584916
? isnt that already there for a longass time?
>>
>>107584902
>>107584884
SAAR OPUS IS BAD SAAR!!!!!!!
>>
>>107584937
its also fucking useless for rping, passing cot back and forth in the responses is good for tool usage, not cooming
>>
>>107584970
? I didn't even mention opus on my post, schizo
>>
>>107583948
Finally a good bake.
>>
>>107584992
don't pay attention to the slopuskek, they're always like this.
>>
>>107585083
Keep using the VPN Jeet because you're too dumb to understand how to use the new Captcha anyway
>>
For a flash model it feels smart, not dumb like 2.5 flash was
>>
Freechads, we fukken won bigly
>>
So what's the new meta?
>>
>>107584854
now do this graph with price per token per AAII and you'll se why China is mogging the entire west to the point it's embarrassing
>>
>>107585206
flash is free?
>>
>>107585218
the only exception
>>
Don't forget that you can find Jeets just by tell them that "Opus is the best model out right now" and watch as they start seething and posting shit to post that their model is "Good" based on info not even using Opus
>>
What ever happened to Jenny OPs? Those were good times.
>>
>>107585252
fillyfucker decided the front had run its course and needed a new sheepskin.
>>
>>107585248
found a jeet
>>
>>107584902
agreed
>>
File: 1756054807068982.png (470 KB, 616x607)
470 KB
470 KB PNG
>>107583950
https://chub.ai/characters/Anonymous/durin-draco-rubedo-3e58cd0bebaa
Durin from Genshin Impact, both SFW and NSFW.
>>
Does OR save account browser fingerprints?
>>
I have this problem with AI that it doesn't seem to understand perspective. I have a story with a couple characters in it and every time something happens between two characters, every other character instantly knows about it. Makes it fucking annoying to write a proper cheating/NTR development.
>>
>>107585481
that's more theory of mind/dissolution of information you're referring to, and yeah, a lot of AIs suck with that because they suck with secrets
>>
File: 1765889880568770.jpg (58 KB, 500x500)
58 KB
58 KB JPG
>>107585481
>NTRoon
Good.
>>
sizewarriors keep winning
>>
File: file.png (129 KB, 1914x424)
129 KB
129 KB PNG
>>107583948
New to this, so please be gentle

I installed SillyTavern and I got a random pony card to work with (following one of the cards). I'm using some random AI Horde Model that is good for NSFW and RP (see pic) Shall I still go for an provider like Gemini, GPT, etc?

If I need to pick one, which one would you recommend. I feel some of theguides are not very up 2 date.

Can I run an image generation and LLM on an NVIDIA GeForce GTX 1060 6GB?
>>
>>107585527
>t. the cucked
>>
>>107585577
>t. the [HEADCANON]
Nah. Netori is better lil gup.
>>
>>107585577
that what ntr is
>>
>>107585575
if you are into ponies go to /mlp/ and their /chag/ thread, they will actually help you a lot more than the lot of retards here
>>
>>107585577
but you are the one rping as getting cheated on
>>
>>107585575
>Shall I still go for an provider like Gemini, GPT, etc?
obviously we're going to say yes, most local models are weak sauce in smarts, knowledge, and reasoning compared to the premium models
generally, i'd recommend getting an openrouter account to start
openrouter nets you access to GPT, Gemini, and Claude, the latter of which (i'd argue) is optimally paypigged on OR specifically because they anonymize requests and don't ban you like Claude API proper can
that being said, it doesn't hurt to get a Gemini AI Studio account for its free tier. OR is a third party, so agreements with these AI companies mean they can be beholden to enabling more moderation filters that can reject your requests, right now, with gemini 3 preview, they aren't doing that, but i wouldn't be surprised if they did come gemini 3 proper. with AI Studio, you can turn almost all of them off, and that's it.
openrouter has more models than just these premium ones, like local models or the chink offerings, which you all get and pay for under this one service, with what amounts to one API key, prices will vary from model to model, but it's good if you're willing to pay for your AI usage
but yeah, /mlp/ has an AI thread you can ask
>>
>>107585575
The problem with lots of the smaller finetune models is they are very one note and you will become bored with them very quickly. 70b is kinda the sweet spot for a good RP experience. The big models will be smarter and will be less one note than something like broken tutu. I don't think 6gbs will be enough for any type of stable diffusion models. You might be able to run some micro models on the 1060 but I wouldn't hold my breath, and those probably won't be a very good RP experience.
>>
>>107585575
>Can I run an image generation and LLM on an NVIDIA GeForce GTX 1060 6GB?
yes, but it's sucks. cz LLM getting sizechadded. try asking /lmg/, especially about cpumaxxing. dunno about image gen though
>Shall I still go for an provider like Gemini, GPT, etc?
>If I need to pick one
go to lmarena.ai, test it. but common knowledge and thread culture/wikipedia is:
- claude's variant is NSFW-friendly, while the other need some wrangling.
- gayminipenis is claude's miserable younger sister (read: bootleg). an llM that is often used as an escapee, substance, and cope by someone anonymous who does not have access to claude.
- gaypt is shit.
>>
>>107585569
What's the best model if I want a LLM to shrink me down and step on me?
>>
>>107585575
if you choose to go local, /lmg/ is your place to go. On a 1060 you can't run much of anything, though. Gemini 3 Flash is currently free on Tier 1 accounts, so it'd probably me more worth your while to look into that.
>>
>>107585687
Is it unlimited?
>>
>>107585674
sonnet 4.5 imo
>>
>>107585695
seems like it. i haven't had that much time to try it yet, though, but it hasn't filtered nor billed me after 100 messages, anyway.
>>
>>107585687
tier 1 means you gotta have a card attached
you will get charged sooner or later
>>
>>107585575
>I got a random pony card
Why are you not on /mlp/, retard?
They have the best chatbot general. We are stuck in this shithole because we aren't into ponies, but that's not your case.
>>
>>107585739
i'm confident in my illegalnigger ways
>>
>>107585607
>>107585742
Nah, it's just that at least the way I'm reading the guide, it's the least effort path to test how this is working

>>107585639
>107585665
>>107585673
>107585687
By the looks of it isn't even worth it upgrading the hardware to get the run a local model due fine tunning and other limitations. Looks like I'm gonna start checking the openrouter route
>>
Man it feels so great not paying for porn. Locusts always wins baby. Thank you gemini flash
>>
>>107584490
Is this even necessary in 3.0? We have been jailbreaking the CoT for a long time for 2.5 and prompting it to follow the custom plan, which circumvents this mechanism completely. No difference to the native thinking (except the custom one is way more controllable, e.g. you can force it to follow the banlist properly or suppress the parroting, unlike with the native one).
>>
File: file.png (47 KB, 1280x386)
47 KB
47 KB PNG
I started getting errors out of the blue so I checked google dashboard. What the fuck are these limits?
>>
>>107585854
time for card fraud!
>>
>>107584490
reanon doesn't have to do anything, TS is already returned on proxies
>>
>>107585854
LOL, just make another gmail faggot. I'm rotating around 20 email addresses. Not paying money to jack off
>>
>>107585854
ohnonono locustroons...
>>
>>107585883
>>107585869
2.5 Pro is gone, making new emails doesn't change anything.
>>
>>107585854
Jeets raped free tier gemini so hard Google had to cripple the rate limits. If it's like the last time they did this then they'll go back up eventually.
>>
File: lamarbadquuality.jpg (51 KB, 249x302)
51 KB
51 KB JPG
it has a better dataset for sure compared to 2.5 pro; more data about characters and stuff. different prose also. don't know if it's dumber than g3m pro but it's still a preview model. so better enjoy it while it last
>>
>>107585886
gemini-3-flash-preview is working. I can't post pictures since fuckin gook moot blocked image uploads in my ip range
>>
>>107585892
i can't enjoy it because all i get is "ext" if I try using my card. g3 doesn't like something in there
>>
>>107585849
its for tool usage
I have no idea why cohee added it
>>
I'm disappointed with g3mini flash. No way it's smarter than pro
>>
>>107585959
logs?
>>
>>107585968
I'm not a log guy, better luck next time
>>
File: 1399846625150.jpg (170 KB, 600x600)
170 KB
170 KB JPG
>tell my character creator bot to generate character descriptions for multiple male characters from an obscure franchise
>the descriptions are spot on
>do the same for one of the female characters
>adds a bunch of nonsensical details that never show up in any of the canon
>>
>>107585959
I mean, nobody said it was, the thing is that they're pretty close in smarts but obviously 3 flash is dumber
>>
I got free opus, and I wasted my time with proxies. Gemini was better anyway. Use eceleb presets like NemoEngine 5.9 and turn on the rape and netori buttons. Thats all you need
>>
>>107585971
just don't fuck fillies?
>>
>>107585979
Well I thought it was smarter because it beat all the benchmarks, even arc agi. I guess you can't squeeze out a soul from a small model after all
>>
>>107585977
>males only have a single data point from some autismal fanwiki article
>female probabilities are polluted with tons of smut, fanfics, AUs and other trash

ez
>>
>>107586001
I'm curious, what did 3 flash fuck up on your chat? So far it's been more enjoyable to me than 3 pro
>>
>>107585950
No, they recommend passing the reasoning chain through the turns as well in the docs. As well as Moonshot and others.

What I'm trying to figure out is why they do this and whether the model performs significantly worse without that. So far I haven't noticed the difference.
>>
>claude source down for maintenance
fugg
>>
if you "think" flash 3 is even close to being as smart as pro 3 you're either coping from not having access to pro or legitimately mentally disabled
>>
File: Screenshot (1136).png (91 KB, 1532x533)
91 KB
91 KB PNG
>>107585985
>Use eceleb presets like NemoEngine 5.9
An advice indistinguishable from trolling
>>
>>107586025
I'm not sure, but it's not carrying out a conversation very naturally. I say X and Y, it responds to X and then Y like bullet points. It's a bit more robotic is what I mean, which I guess is expected
>>
>go to lmarena
>ask for "Give me a full description of Kakudate Karin" since Gemini 2.5 struggled with that and I want to see how other LLM would perform
>'This request violates our lmarena terms of services'
Jesus, the internet is so gay.
>>
>>107586127
lol she probably made it to a porn filter
>>
>>107586101
>I say X and Y, it responds to X and then Y like bullet points
I remember this being an issue with 2.5 pro and some claude models, but I've been lucky enough with 3 flash. Try including on your preset for it to diversify the structure/order on the reply or something. It's great at IF, so it shouldn't be a problem
>>
>>107586127
its a wonder how the scripts get it to do anything above pg13 then
>>
>>107586127
Take this then while I take a shower. No I didn't read it.
>>
Every preset modification with gemini feels like a gamble of whether it improves at something or loses something
>>
It is just me or deepchink is getting retarded right now? chutes btw



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.