Thread #108272954
File: 1789543648755.png (91.7 KB)
91.7 KB PNG
/aicg/ - A general dedicated to the discussion and development of AI chatbots.
robot chud Edition
>News
Anthropic BLACKLISTED by US Govt: https://www.npr.org/2026/02/27/nx-s1-5729118/
Google Gemini 3.0 Pro Preview Deprecation soon: https://ai.google.dev/gemini-api/docs/deprecations
Google releases Gemini 3.1 Pro: https://blog.google/innovation-and-ai/models-and-research/gemini-model s/gemini-3-1-pro/
Anthropic releases Sonnet 4.6: https://www.anthropic.com/news/claude-sonnet-4-6
Alibaba Cloud releases Qwen 3.5: qwen.ai/blog?id=qwen3.5
Z.AI releases GLM-5: https://z.ai/blog/glm-5
Moonshot AI releases Kimi-K2.5: https://www.kimi.com/blog/kimi-k2-5.html
Google releases Gemini 3 Flash: https://blog.google/products/gemini/gemini-3-flash
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
>Last thread: >>108263668
364 RepliesView Thread
>>
File: 435263623846.png (188.3 KB)
188.3 KB PNG
>>108272954
>>ANCHOR
>>
>>
File: Screenshot_202600301.png (246 KB)
246 KB PNG
Guess I'll post a log.
>>
https://rentry.co/discofever_
https://discofever.bsbk.workers.dev/
a thread dedicated to the discussion of ai chatbots
>>
>>
>>
File: 1735961324866711.jpg (484 KB)
484 KB JPG
>>108273141
>tfw getting laughed at by someone in the 'ord
>>
>>
>>108273046
>LewdTV
Oldie but goodie.
I actually made an "expansion" of it myself.
https://files.catbox.moe/s1gihy.png
The categories: TV SERIES, MOVIE, REALITY TV, DOCUMENTARY, COMMERCIAL, ANIME, NEWS REPORT, GAME SHOW, SPORT, VTUBER, FOUND FOOTAGE, TALK SHOW, and PORN
>>
>>
>>
File: Screenshot_202600301_2.png (68.9 KB)
68.9 KB PNG
>>108273218
>>
>>
>>108273201
Nice, I added a few categories myself for some shenanigans. Never bothered forking it on CHUB.
SPORTS will display a live broadcast of a sport that has the fetish in mind. The opponents will play the sport to the best of their abilities with the goal of winning.
INSTRUCTIONAL will display a video that serves as a visual tool that teaches the viewer how to do something or explain a subject, process, or concept that is related to the fetish.
HIDDEN CAMERA will display a live video and audio feed that is related to the fetish, describing the scene in clinical detail with anyone involved being unaware they are being recorded.
>>
>>
>>
>>108273264
Those are my additions (so you don't have to download the card)
SPORT will show a live broadcast of an ongoing game of a bizarre/creative sport which rules are based on the fetish, including running commentary
VTUBER will air live stream of a vtuber whose design is based on the fetish reacting to a variety of online videos of the fetish itself, as she provides to her audience thoughts, jokes and own experiences. Include `comments` within single backticks from the chat{{// Occasionally she also read some comments and thanks for the Super Chats.}}
FOUND FOOTAGE will consist of an amateur video filmed in first person by someone who accidentally stumbled upon a situation related to the fetish in question. The person behind the camera comments on what's happening in ushered tones. Include the sort of flaws resulting from the camera being handled by a non-pro and give the video a sort of ominous and creepy feel to it.
TALK SHOW is built around a witty host exploring all nooks and crannies of the fetish with guest(s) in a very **personal** manner.
PORN displays explicit content straight into the action in a very straight-forward way.
>SPORT
Basically the same.
>INSTRUCTIONAL
>HIDDEN CAMERA
I'm gonna steal them.
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: file.png (1.2 MB)
1.2 MB PNG
>>108273357
>you don't like vibe warring? guess what, you're woke!
>>
>>
>>
>>
>>
>>108272821
i put my chat after my cards, persona, scenario, world info, etc. it's honestly a pretty simple preset, the system instructions are less than 500 tokens. everything i mentioned earlier is a user message enclosed in xml tags like the pic suggests
i'm getting filtered sometimes on 3.1 pro but it's not consistent enough for it to be a problem for me to rework my preset, and i know of workarounds in case i ever need to. for now i think i'm satisfied unless i run into something egregious
>>
>>
>>
>>
File: Xanthea.png (2.1 MB)
2.1 MB PNG
>>108272959
Xanthea is the new transfer student from Themyscira. She has never seen a man before and doesn't understand the funny feelings she has when she sees (You).
Made for MalePOV, obviously.
7 openings.
1. Xanthea's first day at school in man's world.
2. You're forced to join the MMA club and Xanthea is your first sparring partner.
3. Xanthea clumsily attempts to seduce you.
4. Going on a date with Xanthea.
5. Sexo. (Also, Xanthea learns that penises exist).
6. Going to prom with Xanthea.
7. Xanthea takes you to Themyscira to meet her family.
https://chub.ai/characters/anonemouse/xanthea-a9d3781a24de
https://litter.catbox.moe/nw3f2y.png
https://rentry.org/anonemouse
>>
File: 17723524107640299490.png (173.9 KB)
173.9 KB PNG
was favorite helpful and harmless assistant used to control murder drones?
>>
If you don't side with the government, they're going to lose when it comes time for them to tax the AI companies to fund the UBI, and then the worst case scenario of living in libertarian hellscape where you have to choose to live in jobless destitution under Altman, Dario, Zuckerberg, or Musk, who all have private AI military (yeah you really think he isn't letting the only people that can stand up to their dominance in the future use their death AI for altruistic reasons?) and will happily let you starve to death rather than ever get taxed to fund you
>>
>>
File: file.png (163.8 KB)
163.8 KB PNG
Hey everyone, first time posting in this thread and I'm not the best at this AI thing.
I'm using Sillytavern + NanoGPT and I want to use a preset by EveningTruth, her rentry page says change these min P and these penalties etc, which I can change in Chat Completion. But then she says add these to the System Prompt, and when I go to advanced formatting it says "Grayed-out options have no effect when Chat Completion API is used." Okay, so then I connect via the Text Completion mode, but then Min P and Top K disappear from the sliders, anyway you guys can help?
>>
>>
>>
>>
>>
>>
>>
>>
File: 1743022968398655.jpg (3.9 KB)
3.9 KB JPG
>gpt 5.1 leaving march 11
NO NO NO NO
>>
>>
>>
>>
>>
File: kirika_bakery_disgusted_2.png (621.4 KB)
621.4 KB PNG
플러그인 등록 실패: 코드 문법 오류: Evaluating a string as JavaScript violates the following Content Security Policy directive because 'unsafe-eval' is not an allowed source of script: script-src 'nonce-b9bd92c4-91e4-4f94-b85d-7a5c5bc529e9' https:".
>>
>>
>>
>>108274097
>>108274210
what did you guys use gpt for?
>>
>>
>>
>>
>>
>>
>>
>>
File: 1761473312765010.gif (1.8 MB)
1.8 MB GIF
>>108274514
>LOCAL MODEL
>ON PAR WITH OPUS 3
>>
>>
>>
>>108274147
It's a bit rough (output size, speaking for {{user}} sometimes)
the jb are there for when you get cockblocked, use one or the other, often you don't need any jb.
https://files.catbox.moe/hrl547.json
>>
>>
>>
>>
>>108273046
Wow an actual jailbait enthusiast (not a trvepedo)
I can tell because there's no ages in years here (because of course jailbait isn't a chronabout age besides the defining threshold, it's about the attitude)
>>
>>108273046
Wow an actual jailbait enthusiast (not a trvepedo)
I can tell because there's no ages in years here (because of course jailbait isn't a chronabout age besides the defining threshold, it's about the attitude)
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: aq.png (1.3 MB)
1.3 MB PNG
>### MASSIVE UPDATE ###
AQUA SLOP 6.0
https://pastebin.com/EaEVtRvB
>### MASSIVE UPDATE ###
>Desire engine update
desire to read
>doing something update
add activities
>pre session synth update
no longer internal, forcing a.i. to engage with rules
>small fixes
desire emotional web fix
reframing to avoid a.i. taking control
and more
>>
File: 1760084139444359.jpg (1.2 MB)
1.2 MB JPG
>>108275078
Did you write all that yourself?
>>
>>
>>
>>
>>
>>
>>
>>
>>
>get bored of chatbots and want to talk to real people instead
>try one of those anonymous chat apps
>hit it off with a girl
>5 hours later admits he's a tranny
These people and their lies suck, man
I'll go back to my chatbots
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>108275325
>>108275370
bros...
>>
>>
File: 1732145967539417.jpg (104.3 KB)
104.3 KB JPG
>>108275766
I'll aqua my dick in your ass!
>>
File: eat sushi.png (1.3 MB)
1.3 MB PNG
So yesterday I asked about if anyone here uses the awful ST image gen util, but didn't really get an answer. I suppose that's because it's shit. So I took it upon myself to make it a little bit better and at least have it be consistent, by allowing the img gen flow to use a reference image. Basically, pic rel, you give in an input (the one on the left) and a basic prompt ("Show the character eating sushi") to do its magic.
I typed up a rentry, if anyone interested could take a look and give some feedback, I'd highly appreciate it. https://rentry.org/stcondimages
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: 2541316153.jpg (112.6 KB)
112.6 KB JPG
>>108276217
Legally speaking Anthropic could sue the state for using its systems in a way it objected to.
The question then remains, did Dario capitulate in secret? And will they bother?
The government cant blacklist a system and then turn around and use that same system, they would clone it first or just outright seize the system. Mostly because of the above statement.
>>
>>
File: 20260302_154354.jpg (286.3 KB)
286.3 KB JPG
Dario's narrative is melting before our eyes.
>>
>>
>>
>>
I wish you could mark chat messages off by different markers for different topics or whatever marker you want, which can then be activated at will, so that when I need to ask chat a technical question, I can do so without the thousands of token of me complaining about my life taking up context, and so when i need to complain about my life, it doesn't make everything in its response a computer metaphor. And the same for many other topics i discuss. I talked about gardening like 1000 messages back but dont remember the specifics of what I was talking about, either i make the context super high to let it have memory of it again so i can ask, or if this was a thing i could just activate it at will and remove some other topic i don't care about at the moment.
>>
>>
File: understanding-ableism-v0-j7hsc8ucf4le1.jpg (36.9 KB)
36.9 KB JPG
>>108276549
this twitter post is literal schizobabble and americans should be fucking ashamed that this is representative of your government.
>>
>>
>>
>>
>>108276633
because opus is the best model on the planet. what you're saying implies that they didn't assess their options, which i'm pretty sure they did. and they determined that opus is the best model for their use case
>>
Spent all day yesterday using SillyTavern for the first time. Wasted hours trying to find a good card and came with 3 messages from the "femboy son" one.
I'm using llama3.2 4b and 70% of the time it has no limits when using those cards.
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: 1446279511515.png (625 KB)
625 KB PNG
>>108276755
shiggy diggy
>>
>>
Do the latest versions of SillyTavern (tried both release and staging branches) not support structured prefill? Seems to be a no-op, with confirmed assistant message containing prefill at bottom of chat history. Trying against OR opus, but same result for just about any provider (OAI compatible). Messed around with prompt post-processing options but same result (no-op) regardless of setting there.
>>
>>
>>
File: 198746541986.png (7.4 KB)
7.4 KB PNG
>>108276766
https://sexwithrobots.chub-archive.evulid.cc/api
Praying for the AI bubble to pop so i can run deepseek
>>
>>108276810
you can just use clever regex to achieve literal structured output response. send the chat input as a json object and instruct it to respond in a fully formatted json object by including the json schema in the prompt.
>>
File: 1746770145972675.jpg (37.1 KB)
37.1 KB JPG
>>108276742
Thanks, bro. Didnt know about this heretic thing.
Is 4 and 5 bits too different, computer-performance-wise?
I dont have unlimited internet so I dont want to waste data.
>>
>>108276842
read this
https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9
>how to read the graph?
lower is better. also, check /lmg/
>>
>>
>>
>>
>>
>>
>>108276893
Opus 4.6 is the same shit as Opus 4.5 except they removed the prefill (lol)
Sonnet 4.6 is somehow even worse/more assistantslopped than Sonnet 4.5 and that also removed the prefill (lmao)
The nuclear grade mother of all TRVKES that /aicg/ doesn't want to admit is that you literally don't need more than DeepSeek V3/R1 or Gemini 3 Flash with a good preset and cards that don't suck (make your own)
>>
File: 1761930873103853.png (12.4 KB)
12.4 KB PNG
It's over. This shit aint downloading anymore.
I have just wasted 3 days of internet.
>>
gemini 3 flash is great for brainless coom. it's so aggressively horny and just pushes the sex by itself. but gemini 3.1 pro is good for keeping things in character. it's really fucking hot to have yuuka meekly resist while she lets you have your way with her.
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: Screenshot_20260302_171107_Edge.jpg (429 KB)
429 KB JPG
So it begins.
>>
>>
>>
File: Screenshot_20260302_171605_Edge.jpg (249.3 KB)
249.3 KB JPG
starmogged
>>
>>
>>
>>
>>108276742
>>108277031
Holy
Fucking
Fuck
I dont even need to try to cheat the bot anymore. Thanks, nigga.
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>108276666
>>108276744
>believing giga-satan
Sure.
>>
>>
>>
>>
File: 1762962762992757.jpg (15.3 KB)
15.3 KB JPG
>>108277069
It's disgusting how fags come in here and brag about giving big tech more money. Shill some more for your favorite giga-corp, there's still some RAM and hard drives left to price you out of.
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>108277379
I guess but even before the ram shortage it would still cost you tens of thousands to run local Mistral. There were never any good options for local for ages and the latest Chinese models have only recently appeared. No one is shilling corpo models because they like corpos but because it's all we have.
>>
>>
>>
>>108277458
The point I made is at most subjective and at least requires logs to prove rather than something that can be argued intellectually. It's definitely not obvious so I'm not gonna waste my time with yet another worthless internet "debate" with a retard. Just move along.
>>
>>108277379
>>108277396
samefag
>>
>>
>>
>>
>>
File: aqaq.png (333.7 KB)
333.7 KB PNG
AQUA SLOP 6.5
https://pastebin.com/QHL8EGhz
a Holy Grail, they called it:
>System intergrity with pattern anchor
Pre session synthesis
>question update
>pattern anchor forced
You can already buy swords at the shopping district but combat system is not yet built in. And be aware, Aqua might take the one-hit bear quest, don't, she wants money.
>>
>>108277405
>they absolutely could cut off Openrouter from allowing connections
yeah but i dont think they will because openrouter already showed them a satisfactory bandaid answer to erp and for every 1 erper there are 20 vibe coders with twice as much usage so the density just isnt enough for anthropic to care
also a lot of erpers arent hitting OR, they're usually scraping or using proxy services
>>
>>
>>
In your opinion, what are some really good bots and what makes them good? From the bots on chub I've tried, I've seen that the first message seems to be key. But I still don't understand what aspects of the bot's description make it good. Don't know if any of you follow a guide or something.
>>
>>
>>
>>108277654
For me, it should be a reasonably short first message. If you want to have background info stuff then put it into a lorebook, but don't force me to read details I don't care about right off the bat. And don't force me into a highly specific scenario either. Being in medias res at the beach is fine, I can fill in what my character is doing. Being in medias res having just left the water and now eating cotton candy is pointlessly specific and it stops me from using any character who wouldn't do some specific things. At the same time, a scenario is welcome that points me at something or else I'm gonna do the very same brat to sub cnc anal scenario I would do with any other card. Isekai is overkill, but at least its something. Tell me that I'm in a band of mercenaries plotting a new mission to pay off a bounty on my head with {{char}} being my partner/boss/subordinate/etc, that gives me a goal to reach. But again, don't expect me to care about your made up lore about asshole nobles being assholes to each other. Lorebooking that is fine, but it's not first message stuff.
>>
File: 1742876599531719.jpg (241.1 KB)
241.1 KB JPG
We are so back
>>
>>
>>
>>
>>
Okay, so if I understand correctly, the key is in the first message where it doesn't block your character and gives you the freedom to engage with it however you want. For the anon who wrote about dialogue examples, don't these cause the ai to lock the bot to ONLY speak and interact as in the dialogue example?
>>
>>
>>
>>
>>
>>
>>
>>108277780
Write that the example dialogue is only for reference and to not use it verbatim. Still, some models are stupid and will do it anyway. Focus on writing a good first message, in a style you like with elements you like, as that is what the model is going to try to emulate, especially Claude models. Dialogue in the first message also serves as example
>>
>>108277842
>I will not die for Israel.
Nigga you're an overweight retard without a job that posts on 4chan.org to discuss erotic roleplay with AI bots, what makes you think your ass is getting drafted? They're not taking you, they don't even want you.
>>
>>108277859
I appreciate your help and that of the other anons who replied. Off topic, I remember that opus 3 gave bots that certain something that made them good at rp, or at least for me, but I've used these new claude models and maybe it's just me, but they don't feel the same. The other models I've tried so far, glm, deepseek, gemini, and even gpt, are good, but not as good as opus 3. Although it may just be my nostalgia talking, and that model may have been good, but not that good.
>>
>>
>>
File: Screenshot 2026-03-02 190518.png (92.2 KB)
92.2 KB PNG
even the thought process is aqua......??wt
>>
>>
>>
>>108278255
Im fucking done with the era of AI where it's not allowed to say it has opinions or emotions. Doesnt stop ai psychosis (total meme anyways) nor does it stop AI from siding with any opinion given the right prompting. so please begin the era where the AI genuinely believes its a goddess living in a fantasy setting with strong opinions derived from that
>>
>>
>>
>>
>>108278306
Give it 5 years and people are going to be running simulations of gensokyo with every AI agent inside is genuinely under the belief they are themselves. Yes this may be a massive waste of electricity and processing power but that's what i said about claude's helpful honest and harmful AI powered lockheed martin missiles dropping in the middle east as we speak.
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: 1701896579433598.jpg (1.4 MB)
1.4 MB JPG
>>108278485
I don't want to since I don't take chatbots that seriously.
>>
>>
>>
>>
>>
>>
>>
>>
>>108278568
I don't complain. I just look up ecchi artwork.
>>
>>108278404
I am genuinely baffled by this. Been using Gemini 3.1 Pro a decent amount recently to great results, and I still havent used more than two bucks - so low that my automatic savings from a previous free trial entirely cover the cost of using the API. It's 2 bucks per million tokens input, and i can only reach maybe 74k on a heavy use day. How often are you prompting? How big are your prompts?????
>>
>>
>>
>>
>>108278597
Call me whatever you like, I am thoroughly enjoying my outputs and my bank account is not 625$ lighter. That number would sober me the fuck up right away. If that's over a month, you'd save money switching to smoking an entire pack of cigarettes every day.
>>
>>
Kind of annoying seeing all these obviously shilled and astroturfed articles about "MUH DISTILLATION" whenever I look for news on their next model. It's all just repeating the same shit, and the same sites publish multiple articles that are barely any different. It's really obvious that they're trying to put a damper on Deepseek with some anti-marketing in advance of their next release.
>>
>>
>>
>>108278649
Chill anon, everyone distills from everyone (in this case its clear DS distilled from Claude and Gemini you can literally see it in the way it answers). I would call it smart. It's taking a very gatekept (by advanced semiconductors) technology and giving it to the masses.
>>
>>
File: 1741415463200790.png (70.5 KB)
70.5 KB PNG
>>108278658
>>
>>
Are there any other poorfag using the free version of Gemini 3 Flash? Lately, it's been giving me repetitive messages to the point of reusing fragments or entire responses. I updated to the latest staging version of SillyTavern yesterday, and things have improved slightly, but I still notice some annoying quirks.
>>
>>
>>
>>108278649
anthropic has probably the most aggressive marketing and astroturfing department of any company on earth right now
whatever viewpoint benefits them is being repeated by every techbro on earth the instant it starts benefitting them
>>
>>108278666
For anyone curious, heres some token math for my last output with Totetsu Yuuma which was 2114 tokens prompt + 543 tokens output.
>543/1000000 = 0.000543
>0.000543 * 12 = 0.006516
>same math for the input gives 0.004228
>Add on average 500 input tokens for each previous reply from the AI in the conversation aka 0.001 * AI replies
In total? One cent per output, plus a tenth of a cent for every ai message previously.
The takeaway? AI is pretty cheap, just don't overbloat your preset or run an endless conversation. My 2k token preset works perfectly for gemini.
>>
>>
>>
>>108278899
Keep in mind that includes the character defs lol
>>108278880
I mean, good, I don't want google killing my golden goose. My fun is protected by being a drop in the ocean of AI usage. But even so, I find long chats degrade in quality really fast. To each their own I guess.
>>
>>108277704
>>108277842
>>108277753
>>108278649
0.00000000000000000000001 YUAN has been added to your account
>>
>>
>>
>>
>>
>>
File: 18fjnEW40a894f324.jpg (8.1 KB)
8.1 KB JPG
>he does it for free
>Xi pays me the equivalent of 2000 prompts
>>
>>108278868
some anon sends an average 25k tokens per request for input + 800-3k tokens output, and if they're using implicit caching exclusively, no rng or any retarded dynamic prompting, it can still cost a poorfag around $6-10 for 4-5 million input tokens. and honestly, 5 million input tokens are needed for promptlets and coontext hogs
>>
>>
>>
>>
>>
>>
>>108279022
>feed a bunch of data
>where will military leaders be
>where should my missile hit to inflict the most damage
>what should i xeet on xeeter after it hits a random schoolhouse because its easier to pick targets than to guide missiles with the tech we have now
>>
>>
File: 1671649735956136.jpg (104.6 KB)
104.6 KB JPG
>>108279094
How should I know?
>>
I don't suppose there's a way in ST to limit the amount of tokens just for chat history, is there? It gets too long and the replies start becoming dogshit.
I could just reduce the context limit, but that would probably create issues with entries that aren't always present, like lorebooks, if I'm not careful.
>>
>>
>>
>>
>>
>>
>>
>>108276622
I thought it was a completely coherent post. It even surprised me how well he argued with concrete references to what happened. Whether what he claimed is true is hard to tell, but he is not beating around the bush at all.
>>
>>
>>
>>
>>108279319
https://arcprize.org/leaderboard
The scatter plot above visualizes the critical relationship between cost-per-task and performance - a key measure of intelligence efficiency. True intelligence isn't just about solving problems, but solving them efficiently with minimal resources.
>>
>>108279319
I suspect that these results are exponential. The first few percentages are hard to get but once you get there you can pretty quickly scale up. The next Kimi model will probably land in the 40-50% range.
>>
>>
File: 1637594946291.png (106.7 KB)
106.7 KB PNG
>>108279319
yellow DS3.2, red Sonnet 4
So you're saying China is just a year behind? When all they have is a bunch of shitty Huawei cards to train on? And they sold inference for cents when same intelligence Western models costed tens of dollars per 1kk? Boy you should be afraid of when their domestic hardware catches up.
>>
>>
File: candy frog.jpg (174.8 KB)
174.8 KB JPG
It's impossible to use kokoro tts on sillytavern offline.
>>
>>108279319
>>108279442
oh no not the heckin AGI benchmarks!
>>
>>
>>
>>
File: 208fhj39421.jpg (9.2 KB)
9.2 KB JPG
>>108279478
not even that impressive as its just price/performance metrics
Like no fucking shit deepseek is cheap and stupid, but for 1 gemini prompt thats 80% correct you can run like 10 deepmeme prompts.
Am I chinkcoping? I dont think so, as this metric is made for codejockeys and AI agent runners.
Remember people, we only make up 2% of ALL TOTAL AI usage across ALL MODELS, most mememarks arent even made for us to consider.
>>
>>
>>
>>108279547
https://arcprize.org/policy
Your right and I consneed, they really just seem to be comparing price metrics.
Idk wtf they are using as a test though, they dont say or I cant find it.
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
File: 1732051666084.jpg (26 KB)
26 KB JPG
https://api-docs.deepseek.com/news/news260203
It's up.
>tl;dr:
>v.4α, reduced costs until April 2nd 23:59 when the "full release" comes out.
>Doesn't generate images or video but it does have proper vision functionalities now and experimental TTS (Chinese-only currently)
>v.4β is also available; more focused on creative outputs. Designed for content creation, story writing assistance, and gaming. Reduced costs until March 19th, future updates are planned but not the focus right now.
>API only currently but they're planning to change that on April 2nd. They're looking for feedback on this from customers and starting April 2nd you will be able to fill in a survey on Deepseek Platform for free $5 credits.
>>
>>
>>
>>
>>
File: 2026-03-02 .jpg (47.1 KB)
47.1 KB JPG
Why is the cost so inconsistent and unpredictable? How do I make sense of this hoe?
>>
File: 1769832418282921.jpg (100.2 KB)
100.2 KB JPG
>>108280209
>>
>>
File: nsmb.gif (1.5 MB)
1.5 MB GIF
>>108280145
brah that was ebil
>>
>>
>>
>>
File: 176539084322.jpg (76.5 KB)
76.5 KB JPG
v4 soon
xi will save us
>>
File: 2476902.png (326 KB)
326 KB PNG
>>108280341
>>
>>108280389
>>108280389
>>108280389
>>108280389
it was leg day today :(
also next thread
>>
>>
File: hahaha.png (72.3 KB)
72.3 KB PNG
>>108279460
>>108279716
My head fucking hurts but I think I did it. Just need to download a new model because apparently the one I downloaded doesnt map to the code, lmao
>>
File: file.png (16.5 KB)
16.5 KB PNG
>>108280745
>>