general

[hollow ivy]

why easy?

[fickle venture]

Easy to understand in my opinion

[spark python]

https://youtu.be/DxRsjw20dIw I just want to know how this guy did this so good😭

[fickle venture]

I mean I haven't really understand everything right now but hopefully

[fickle venture]

What model he used

[spark python]

Thats what im asking id guess kling because I see the start and end frame transitions

[fickle venture]

Probably Veo 3.1

[spark python]

Veo has start and end frame?

[fickle venture]

Idk

[cold minnow]

Can somebody teach me how to make realistic photos like this

Pb me please

Thanks bro

I have

But how i make a beter model because the face switches a lot

Look here it looks different everytime

I use it on tiktok

[thorny schooner]

Ok I'm just going to say it I think mine browser is genuinely glitch you know how I was complaining about being forced out it just did that somehow even sooner in the group chat when I was just copying from that other group chat I managed to finish a a bit ago

[queen veldt]

Wait is this ai ???

Whole page like?

Damnn

[thorny schooner]

Anyone have any ideas for a fix cause genuinely I am kind of desperate cos I have tried different browsers resetting data logouts resets trying with or without blocks and more

[fiery gull]

#1397655624103493813

👀👀 🏳️‍🌈?

Sorry, Im joking

[thorny schooner]

I don't know why l find it funny that somehow my bug report is already just as filled as the last two other reports even though no one else has replied in it lol

[fair mural]

Hello

[fickle venture]

Anyone here got a Chromebook I found something interesting I wanna see

[thorny schooner]

Yeah why?

[toxic egret]

who choose C lol

theres no way its C

[keen beacon]

Sora 2 pro. Is so slow

[neat prairie]

When will imarenas update the charts?

[tulip kindle]

Hello, guys. What if Gemini 3 write Generating... and stop, don't write answer?

[halcyon nimbus]

either a bug or your rate limited

[tidal schooner]

said nobody

[empty stump]

scam

[edgy dew]

<@&1349916362595635286>

👍🏻

[jovial sapphire]

hi

anyone interested in

world models

here? like genie 3 etc

[strong cape]

hi

[jovial sapphire]

https://x.com/phth0nus/status/1995330605198856409

ً (@phth0nus)

I built Worlder🌎

It turns static images to immersive controllable experiences.

What do you think?

#ai #genie3 #coding #technology

X · Dec 1, 2025, 3:14 AM

Cool

[keen beacon]

Thats pretty simple just good editing

If you pay attention he only has the same background where all the characters are

The rest are cuts scene

[vivid coral]

Grok 4.1 thinking is a hallucinating fool, my gosh

[elder solar]

are there any news about gemini 3 flash?

[native yarrow]

gemini app is genuinely vibe coded

[pulsar rain]

Sometimes it talks like "trush me all of my friends do this". it's so bad it funny

[bleak lake]

Ai is pretty good at tutoring

If you did it right there's a pretty good chance you will not just pass

you are from poland? if you don't mind.

Zurek I had sometime ago was so good

[empty stump]

what's your method

[lucid geyser]

No

What

I’m just getting that impression

Wait what are summer finals

When’s summer

Not for like 5 months for me

[bleak lake]

that's peak, I've spent a week there it is historically and culturally rich with nice people

fr

[lucid geyser]

What month are they

Us

Seems like a long time to prep

Must be super hard

Is it like a college exam

I meant entrance

[keen beacon]

Something about nano is sus to me

It doesn’t seem very artistic

Like it has no style or flavor

[lucid geyser]

Which country is that

[keen beacon]

Post Soviet blocks have better education then in us

More rigorous

[lucid geyser]

U can say they’re biased but the us has top universities

[lucid geyser]

Oh I didn’t even know discord was allowed there

[lucid geyser]

No

Everyone has to take science, history, language arts and math

In hs

[keen beacon]

Most Americans couldn’t tell u where Poland is on a map lol

No geography

[lucid geyser]

Why u so offensive

[keen beacon]

I’m not

I’m from America

[lucid geyser]

Oh well yeah but American universities are good

For high school

It depends on the district

[keen beacon]

Cuz if money and funding

[lucid geyser]

Cause it’s not standardized

[keen beacon]

Ya

It’s not

[lucid geyser]

That sucks.

[keen beacon]

It used to be really good

[lucid geyser]

I think the only practicing the important skills might be more of a uk thing

[lucid geyser]

People want to leave it up to the states but

The federal government has to exert some control atleast

Also all these underqualified people making policies in fields they don’t know at all

[keen beacon]

Ya it’s sad how educated has been so degraded

https://filamtribune.com/spelling-and-vocabulary-problem-in-america/

[lucid geyser]

These colleges be crazy expensive too

Yeah this stuff gotta be addressed

[keen beacon]

Well the problem is federal funding too if u have the government give money they need results

So in order to make it look like it works

U have to pass more students that normally would fail

Aka dumbing it down

https://en.wikipedia.org/wiki/No_Child_Left_Behind_Act

Are there any ai programs

[lucid geyser]

Crazy

[keen beacon]

In Poland?

[lucid geyser]

It’s Russia

[lucid geyser]

Yeah I’ve heard of it but I don’t know the details or outcome

[keen beacon]

Not so simple

Russia and Poland have a long tedious history going centuries

[lucid geyser]

Nvm

[golden ocean]

[keen beacon]

It’s a complex relationship spanning many centuries

[lucid geyser]

Is it easy to leave?

[golden ocean]

[keen beacon]

From the news we get here seems like Poland is doing good

From all post Soviet blocks it seemed to really turn things around for itself

At least form what we’re told here

But from what ur saying about censorship and stuff that doesn’t sound to good

😭

[lucid geyser]

Nice

[keen beacon]

Especially blocking half the internet is crazy

[lucid geyser]

Done in china too or na

[thorny schooner]

I don't know if this is the right place to ask but what on Earth am I breaking here every time I tried to put this in it says violates of term or whatever

[keen beacon]

O.o

[thorny schooner]

lol

[keen beacon]

U can’t see why?

[thorny schooner]

Well then it's been way too inconsistent because no offense brother this is not the first time I posted something for the AI ( but seriously though I am way too proud of your reaction to it)

[lucid geyser]

Draft?

[keen beacon]

Make that a video

[thorny schooner]

I feel you even if for completely different reasons ( from America you know capitalist Society and that business and all) good luck )

[lucid geyser]

Good luck on your exams

Do most pass?

[spark python]

I know I just want to know what model he uses

[thorny schooner]

https://tenor.com/view/i-cant-even-talking-no-nope-gif-15448966

[keen beacon]

Makes it more interesting

[twilit folio]

Need a Indian editor

[lucid geyser]

Cost to attend uni?

[thorny schooner]

☠️

[lucid geyser]

Oh

American unis be crazy expensive

But if you’re low income you can get free tuition, housing and food

[thorny schooner]

Honestly I've been trying to think about what university I would even go to but that coin the idea hurts enough it makes me wince on some instinct 😓 \

Yeah I would recommend if you do american University probably saved up in case because certainly I my family can't leave but because no extend family issues but also kind of hard to save anything in America

[thorny schooner]

/ basically any lucky person of the majority ( cuz let's be honest low income is proportional here because low income is like the majority of Americans when if it's proportional to how expensive they are)

[lucid geyser]

But like if you don’t qualify private universities be like $90k+ per year

Crazy

[thorny schooner]

https://tenor.com/view/no-money-wallet-empty-cashless-gif-16030892

My wallet when it and l thinks about University

unfortunately both societies have their ups and downs and it doesn't help both of both are kind on the extreme or intense end when it comes to those unfortunate downsides ( for America social wise and russia well you'll be the judge at that but I don't think I need to say more)

[lucid geyser]

And Harvard got $57 billion

But still charges like 100k

[thorny schooner]

I did not need to be off of Discord and just saw that for that ( I know we are Americans are called wild but wow I'm very glad to hear the last parts because if I did not I would be very concerned)

I also hate sadly i have heard almost similar stories here in the states (for the schoolgirl parts if I remember correctly has been a while but I know now you have heard stories like that a while ago
)

[lucid geyser]

Crazy

[thorny schooner]

Fair enough it's just our kind of wild is not the good kind very systematic as I'm sure you have of it ( in fact I came I have to be careful to say I'm a classic liberal antitrust era type economic wise the only people I'm sure that are very tolerant as I am Progressive when it comes to institutes and rights but conservative when it comes to family values and stuff like that)

[thorny schooner]

Which now I think about it there's not really a term for that anymore in American politics the closest one I could think of is the one I use but that more sub-branch that's been forgotten over the years with the corporations now but for Europe I think I would classify as a social Democrat I believe

It's fine every country has its dark and bright side it's just America is very hypocrisy about it government-wise

Sadly true america just seems to be the main practicer of it cuz I don't think I have seen a more two-faced country and I'm saying that for my own country

[dusk dragon]

What's going on with nano banana pro?

It's not working

[spark python]

@echo aurora Why gemini 3 pro doesnt support video input? The normal gemini app has it

[thorn quest]

Ooo vidoe input in gemini 3 is crazy thanks brother for the info

[magic ravine]

Looks like Nano Pro is not working at all now.

[dusk dragon]

Yeah nano pro is down

[magic ravine]

[dusk dragon]

It's too bad

Worked for a second

Then stopped again

Generated this though

[magic sigil]

hi

[golden ocean]

true

[torn mantle]

https://x.com/nopainkiller/status/1995372634838876186

Zhipeng Huang (@nopainkiller)

whale seems upgraded , again ? @teortaxesTex 3.2-exp was updated on 9.29, no info about this update yet ...

X · Dec 1, 2025, 6:01 AM

deepseek updated or what

https://x.com/eliebakouch/status/1995427311861043606

elie (@eliebakouch)

Mistral seems to be about to release 2 new models: Ministral 3 and Mistral Large 3.

architecture wise it seems to be:

exactly the same as llama2/3 for Ministral, which is very similar (even if they add SWA) to the first Mistral 7B release 2y ago.
exactly the same

X · Dec 1, 2025, 9:38 AM

mistral getting ready too

we saw this stealth model already

[regal juniper]

Hello guys, anyone try Whisper Thunder?

[keen beacon]

Oh nice new screenshots

[whole sundial]

<@&1349916362595635286>

[torn mantle]

who did i tag

...

[torn mantle]

no

what we know is that its from kling

[keen beacon]

U sure?

Interesting, I was curious which model it was. Makes sense now that u mentioned it

[torn mantle]

yea i talked about this before

many were saying its from xai

but it was quite obvious it wasnt

@whole sundial why do u think Gehlo is related to that guy

[whole sundial]

nevermind, maybe discord's systems already blocked it

it is some person named sebu, they only managed to send it in 2 channels

[keen beacon]

https://www.microsoft.com/en-us/security/blog/2025/11/07/whisper-leak-a-novel-side-channel-cyberattack-on-remote-language-models/

https://youtu.be/_t1lN9cnL8w?si=UrImZ_gJO-yefXRB

[torn mantle]

so the solution is to add junk text

to mess up the timing and pattern?

why would anyone do that...

[keen beacon]

No in the encryption to make it harder to be pattern recognized or estimated probability of the likely hood token reconstruction of the actual encrypted tokens

[queen veldt]

Yes ChatGPT i mean the real mouse!

[keen beacon]

designed to destroy statistical correlations between the plaintext and the ciphertext.

[torn mantle]

thats what i said yea

thats the big idea

kinda of fingerprint of tokens

fingerprint ->
big packet
( short pause )
small packet
( long pause )
medium packet

size + timing

[keen beacon]

Avalanche Effect

[torn mantle]

and we add noise in between

like some random packets

im just asking why woulkd anyone like spy on someone's chat

like its not like you will send the ai your credit card or something

[keen beacon]

Cmon

Other sensitive information

Medical or other activities personal or business

https://tenor.com/view/listening-what-fbi-chris-redd-alex-moffat-gif-11784633

Only two groups that are really into that kind of info

[torn mantle]

apparently new deepseek update

[keen beacon]

Ye

Herd about it

It’s supposed to be good

[torn mantle]

hmm we will see

lol

[keen beacon]

Ya

[queen veldt]

We're only getting beter and better ai

I can't believe what we'll have i 3-5 years

[torn mantle]

Achievement: 🥇 Gold-medal performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI).

cool

[queen veldt]

I remember chatgpt 3

[torn mantle]

https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale

browsecomp is interesting to look at

so it can search the web better

seems like they improved a lot on agentic tool calls

uses more tokens too

[celest orchid]

yo @echo aurora when does november contest end?

[stray aspen]

There's a new deepseek finally

[spark python]

See we think the same way, I gave gemini 1 minute video then I got restricted in the app

[peak sapphire]

@echo aurora Nano banana pro started giving an error again "Something went wrong with this response, please try again."

[viscid cloak]

yea, just wanted to report that. nano banana pro high error rate.

[deft oak]

i hate is dmca suck

banned me i not care

f rules

[jagged cloak]

Same

[barren prairie]

But everything only R2 😂😂😂

[robust sluice]

not just me then.. wait too long for error

[barren prairie]

🙂🤝

[mint snow]

z image will come in lmarena?

[blissful patio]

i hate use lmarena new chatjpt

[fiery gull]

👀

[torn mantle]

dont bother

[fiery gull]

;-;

[granite igloo]

gemini 3.0 is hitting me this error "Something went wrong with this response, please try again." sent 3-4 texts in a row and it hit me back with this, the problem is the entire chat consists of a lot of study material and I can't just simply switch to a newer chat. A new chat will set me back a 3-4 weeks, what can I do?

[slate lake]

Good morning everyone. Does anyone know how to create this style of image? I don't know if it's cartoon, anime, I don't know, but I thought it was really cool, however I don't know the prompt or image style.

[keen beacon]

🤷‍♀️ good question lol

[alpine shore]

hello

[slate lake]

It's like a cartoon, characters with exaggerated features.

[keen beacon]

Ya idk it’s a tough one

[odd geyser]

Wow, pineapple never paid attention to my problem. Oh... Okay. But I'm going to lose my mind soon because I can't use the arena.

[maiden spear]

hi

[atomic lagoon]

Which one is coding for deepseek?

The newest model doesnt seem like anything crazy

Its barely better than its last model for coding

[torn mantle]

its more like they improved on agentic capabilities

[keen beacon]

Seems to be a trend

[torn mantle]

like you wont notice if you are just a casual coder and asking for like a UI thing

[atomic lagoon]

What would that mean

[torn mantle]

like tool call wise, it smarter at chosing which tool to use / how much to use

especially at search/browsing the web

when to use them too

also they released 2 models

v3.2 and v3.2 special

special is only available via api

[queen veldt]

WAKE UP

IT'S THE FIRST OF THE MONTH

[atomic lagoon]

Ohh alright

[torn mantle]

in my testing it got worse

i asked it in french and it started mixing english with it

it did get better at math tho

[fading rover]

Damn the way lmarena renders I love the output with nano banana pro unfortunately Its still 1k resolution .. is there a way to get the lmarena provide 4k with paid options ?? Tried hailuo to test 4k but it's not taking same prompts and output so I hope few other may have same prompt issues to render like lmarena output

[atomic lagoon]

I did see that

[atomic lagoon]

I feel like at this point they deserve to have a paid option on their site lol

[atomic lagoon]

Oh yea but their tokens somehow got even cheaper

[torn mantle]

lmarena are providing multiple services for free

[atomic lagoon]

So thats a plus

[torn mantle]

they cant just give you unlimited usage

they have their own formula for balancing the cost

[atomic lagoon]

Yes that and the cooldown on the rate limit isnt even that long lol

[torn mantle]

if the demand is high they will have stricter rate limits

[atomic lagoon]

Yea

I noticed they may have lowered the rate limit on opus because I've been using it fine with no rate limit

[fair roost]

why gemini 3 pro working sht slow

[torn mantle]

https://x.com/eliebakouch/status/1995457610368053508

elie (@eliebakouch)

very interesting table from deepseek v3.2 that compares the output token count on different benchmarks, dsv3.2 speciale version thinks much more than any other model, BUT since they are using sparse attention the inference cost will still be ok?

X · Dec 1, 2025, 11:39 AM

they are using something called sparse attention

idk what that is but thats what made the model cheap

[torn mantle]

hmm maybe

[atomic lagoon]

Ohhh alright

[torn mantle]

i only use gpt5.1 high on lmarena ngl

sometimes opus 4.5

usually before that i would use gemini 2.5 pro like for every task

i thought gemini 3 pro will replace gemini 2.5 pro for me but in some tasks it got even worse

https://x.com/runwayml/status/1995493445243461846

Runway (@runwayml)

Introducing our new frontier video model, Runway Gen-4.5. Previously known as Whisper Thunder (aka) David.

Gen-4.5 is state-of-the-art and sets a new standard for video generation motion quality, prompt adherence and visual fidelity.

Learn more below.

X · Dec 1, 2025, 2:01 PM

so its from runway

thats surprising actually

[fair roost]

imo gemini 3 pro > opus 4.5 (for backend)

[atomic lagoon]

Front end opus is better for sure. But I hybrid the AIs sort of

[torn mantle]

@keen beacon

[torn mantle]

backend??????

no way

maybe in frontend

[atomic lagoon]

I use Opus for the frontend, html and stuff like that and then gemini 3 for setting it all together

[torn mantle]

but in backend i find opus 4.5 much more reliable tbh

[atomic lagoon]

For a site that might be or an app

[atomic lagoon]

Really?

[torn mantle]

its good at cloning stuff

yea

[fair roost]

frontend opus 4.5 beats everything

[atomic lagoon]

I guess I have only used Opus for webs and UIs

[keen beacon]

Oh runway

[torn mantle]

yea and even for frontend anthropic shared this beautiful prompt that can make your website looks much much better

[atomic lagoon]

So this AI is image or video generation?

[torn mantle]

https://github.com/anthropics/claude-code/tree/main/plugins/frontend-design

[atomic lagoon]

This may be a stupid ask but is claude open source AIs like deepseek is?

[torn mantle]

https://x.com/trq212/status/1993786550656897491

Thariq (@trq212)

Opus 4.5 is our best model yet for design & vision.

Here are some of my favorite UIs we made with Claude Code's frontend-design plugin.

X · Nov 26, 2025, 8:58 PM

using that plugin

makes a big difference

[torn mantle]

video gen

[keen beacon]

I don’t know anything about coding but I can’t get one model to make a simple Sora video downloaded

[torn mantle]

no closed source

[atomic lagoon]

Ahh alright

I guess it makes sense though

[atomic lagoon]

Okay Okay

[keen beacon]

https://shortra.com/sora-remove-watermark?job=07a48f12-41be-40e4-b2d4-f222c2ed3e45#playground

Like this

[atomic lagoon]

How do you even use it?

[atomic lagoon]

Lmaooo I found one of those a while back and it worked greatly

[keen beacon]

There are tons

They work so good

I found the api documation

But idk how to make it work

[atomic lagoon]

Ohhhh nice

[torn mantle]

i think you just add the .md file to your directory and use @ to tag the file and ask it to use it

in cursor or vscode ide

[atomic lagoon]

OHHH

[keen beacon]

https://platform.openai.com/docs/guides/video-generation/sora-2-pro

[torn mantle]

nice

[keen beacon]

Ya they should be super easy to make

Pretty much can get the end point through open ai api for video

https://usesora.pro/docs

I tried to vibe code one but don’t know wtf I’m doing lol

https://019ad69f-72de-75bd-8258-1d5d8d4b638a.arena.site/

My poor attempt with no backend lol

[atomic lagoon]

LOL I mean its not the worst thing ever

What I've noticed is that AI always has a similar UI. Like you can tell if somethings AI generated ykwim

[keen beacon]

I have so many versions

U can make money of this liege the rest of these people

https://savesora.com/

This one is cool it shows prompt also

[supple sand]

deepseek v3.2

[supple sand]

anyone tried?

[keen beacon]

And u know damn well they are all vibe coded lol

[robust sluice]

Gemini is 100% error right now or its just me ?

[atomic lagoon]

Exactly, we all have so much money we can make by doing this LMAO its just for me idk how to advertise well

[atomic lagoon]

Mine is fine but I'm a plus user

[keen beacon]

You don’t need to advertise

Make it good effective offer it free

Then when they are all hooked

You hit them where it hurts

https://tenor.com/view/plutus-bank-card-bank-on-crypto-crypto-pluton-gif-5851022200991914938

Jk don’t do that 😭

[torn mantle]

@keen beacon https://github.com/yo-yo-yo-jbo/whisper_leak

this is the tool right

[keen beacon]

Yeah, I think that’s the one. I’m not sure if it’s the same one that they used in the research for Microsoft, but I’m sure they were very similar.

Same idea

[atomic lagoon]

LMAOOO

[atomic lagoon]

What is that?

Is it like a benchmark?

[languid crescent]

Is it just me or Gemini 3 is slow?

[torn mantle]

its a tool to spy on the llm chat traffic

[atomic lagoon]

Ahh alright

[torn mantle]

proof of concept* tool

to prove their claim that you can spy on chat even on secured network/encryption

[keen beacon]

https://www.amazon.com/Big-Story-Confidence-Man-ebook/dp/B004774D1E

[atomic lagoon]

Ohhhh

[languid crescent]

Umm, Gemini-3-pro is bugging for me, like it loads or doesn't generate anything.. it stucks on looping.

[atomic lagoon]

Not sure, doesnt do that for me

[languid crescent]

Like a simple "Hi" prompt takes ages to generate for me

[keen beacon]

Probably heavy load

Another tactic known to increase user demand

incentivize users to sign up for the higher tier subscription for faster speed and less rate limits so they can enjoy more of the product there addicted to with less hassle of waiting.

https://tenor.com/view/itch-crackhead-drugs-drugaddict-addict-gif-11128316954240258813

[atomic lagoon]

https://aistudio.google.com/apps/drive/1fHqoaBAoqm49TjBRtgTodQplfmYdJ5YD?showAssistant=true&showPreview=true&resourceKey=&fullscreenApplet=true

Since I dont have like

Claude pro or whatever

I just made it in AI Google studio

[atomic lagoon]

And the prompt seems nice

[atomic lagoon]

So basically with the whisper leak it can see updates and stuff or trainings I guess?

[keen beacon]

https://youtu.be/_t1lN9cnL8w?si=UrImZ_gJO-yefXRB

[prime mulch]

I planed to buy laptop but ram price increased 😭

[leaden cove]

<@&1349916362595635286> i have an problem guys please help me, on my pc lmarena doesnt loads at all and doesnt let me click on any buttons, i have an error in console it says Failed to load resource net::ERR_CONTENT_DECODING_FAILED

[keen beacon]

What browser are u using

[leaden cove]

brave

it was working perfectly fine yesterday

[keen beacon]

You have a vpn ?

[leaden cove]

no i tryed realoding the wifi and still same error

[keen beacon]

Clear Your Browser Cache and Cookies

[leaden cove]

lemme try

[leaden cove]

yooo it works bro thanks

[vivid coral]

New Deepseek 👀

[echo aurora]

[echo aurora]

I'm thinking of extending it. I'd like to see more!

[echo aurora]

#1372229840131985540

[spark python]

uhm, is it a different model that watches the video?

I think its still gemini 3

[echo aurora]

As in it has vision but for video files?

[spark python]

yeah thats what I dont know

[echo aurora]

Ah, because we don't have video as an available file upload option, yet.

[spark python]

will it be hard to add?

[calm lagoon]

@echo aurora

[patent bane]

please try again

[calm lagoon]

When you wrote — everything started working 😁

[warm zodiac]

It’s confirmed they’re adding ads to ChatGPT 😭

from Yuchen Jin on X

[echo aurora]

Sorry to see you get this error. We've been seeing higher than usual error rates with this model, but have been lowering it. I'll check in this morning to see what else we can be doing to make this more reliable.

[calm lagoon]

Thanks for your work ❤️

[leaden cove]

can u fire ur all moderators cuz u are literally the most helpful moderator lol

[echo aurora]

Lets not be rude please. Would note that the other moderators (lm_mod_#) are here for moderation purposes. They generally won't be able to assist with bugs/feedback/questions/etc.

[echo aurora]

I'm not sure how difficult it is, but overall we are very interested in adding more file upload capabilities. This certainly would make the platform better, and is something we'd like to do.

[vivid coral]

I'm sad to report the new Deepseek is a hallucinating fool. The app update I mean

[jagged cloak]

That wasn't good idea at all

[jovial sapphire]

What do you guys think?

[keen beacon]

How long did that take

[bright shard]

Nano Banana Pro is having a lot of errors again 😅😅😅😅

[jovial sapphire]

Instant

5 seconds

I'm building it right now

[bright shard]

It's amazing, how did you do it?

[final lance]

So... Uh... Who won the November contest?

[jovial sapphire]

Will release soon

[novel flame]

I want to see more. That is very cool

[echo aurora]

We're going to extend the contest.

[polar wharf]

💀 dang

[polar wharf]

<@&1349916362595635286> pls delete

[echo sinew]

Thanks for reporting

[polar wharf]

my pleasure

[jovial sapphire]

Thank you so much!

[fiery gull]

Dude, Gemini 3.0's metacognition is insane. I honestly thought it wasn't following instructions, but no, the work they sent me is complete garbage, and Gemini 3.0 Pro kept refusing to do it according to their instructions. I gave up and I'm going to do it the way Gemini 3.0 tells me to, lol

Gemini 3.0 knows that if I send it the way they're doing it, it's going to be a mess, lol

I was angry after analyzing the work, it was complete garbage

[bright shard]

@echo aurora Could you make it so that Gemini 3 Pro, in code mode, can accept images or TXT files for creating projects?

[proud bobcat]

any news on deepseek 3.2 being added?

[fiery gull]

4/10 ragebait

hmmmm, you're sure, but I'm liking the 3.0, but it can be better

serious? ;-;

[proud bobcat]

quantized

[fiery gull]

bruhhhh ;-;

now the opus 4.5 is better?

[proud bobcat]

not by much

quant 8 is retains 99% of smarts

quant 4 is like 95%

[fiery gull]

for my use in docs is an hard 50/50 but today I think is 40/60 now

[proud bobcat]

claude is awesome ngl just wish it wasnt so expensive

[fiery gull]

I can understand who said "gpt 5.1 is better that gemini 3.0 pro", but "gpt 5.1 is better that opus 4.5" HELL NO LOL

[proud bobcat]

5.1 is not on the SAME PLAYING FIELD as 3.0

5.1 blows

top 10 ragebait

[fiery gull]

in some case okay....... just okay... but opus 4.5 NO

[echo aurora]

We are looking into increasing the file upload types for all modalities. I don't have any news I'd be able to share, but will keep the community updated when I have news to share.

[fiery gull]

thx so much, I'll use the opus 4.5 back now :), have a good day

[calm lagoon]

❤️❤️❤️

[fiery gull]

lol

I'm using the gemini 3.0 because the aistudio is sooo good

deepseek 3.2v again?

ahhhh without EXP

okay >:), I'll test

[delicate frost]

yooo

[echo aurora]

[delicate frost]

the special one was burst

[hardy crow]

What is deepseek 3.2 specizl ?

[fiery gull]

is a deepseek 3.2v, but have focus in profissional use

[delicate frost]

its the best one

[whole swallow]

Is it a good model sir?

[echo aurora]

More info can be found here - https://x.com/deepseek_ai/status/1995452641430651132

DeepSeek (@deepseek_ai)

🚀 Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents!

🔹 DeepSeek-V3.2: Official successor to V3.2-Exp. Now live on App, Web & API.
🔹 DeepSeek-V3.2-Speciale: Pushing the boundaries of reasoning capabilities. API-only for now.

📄 Tech

X · Dec 1, 2025, 11:19 AM

[fiery gull]

go test youself

[echo aurora]

^

[delicate frost]

its even better than gemini 3

[atomic lagoon]

Well you got your news

[atomic lagoon]

No its not lol

Its barely better than its last model

[whole swallow]

🫩

[delicate frost]

look at the benchmark bruh

[atomic lagoon]

I literally did earlier today and it was barely even better than its own model

[zealous sparrow]

Benchmarks are not always one to trust.

It's your own testing to trust.

[atomic lagoon]

Its not difficult to test so thats fine

[delicate frost]

bruh, to your perspective though, im using it more for logic/proffesional question

[proud bobcat]

not really

still a crazy performant model though

[celest orchid]

hello pineapple please increase limit of characters , bc when you send him 100-character texts several times, the model breaks and the message isn't sent or when you've sent him about 20 messages, the messages aren't sent and you have to create a new chat, which isn't very convenient.

[atomic lagoon]

Gemini would still be better. I use AI to code and use it for my daily wants and that ranges from everything

[zealous sparrow]

The speciale model thinks too much imo

[atomic moon]

leave it to runway to drop the best video model in the world with basically no way to use it

[zealous sparrow]

But thats good because it double checks stuff

[zealous sparrow]

They will make it available soon..

[atomic lagoon]

LMAOO

[atomic moon]

even when it's available it's expensive. no way comparable to veo

[atomic lagoon]

I havent seen any of it yet

Veo sucks really bad though

[fiery gull]

gemini 3.0 still cooking in word.. good at least

[atomic lagoon]

I cant read any of this lol

[proud bobcat]

i wonder why

its almost like its a different language

🤔

[fiery gull]

I don't want you read it, Idc

you see the layout 🤤

[atomic moon]

is that an image or what?

[atomic lagoon]

Yea smart ass, that would lead to anyone else telling me what it is

[fiery gull]

calm down littler bro

[atomic lagoon]

"Littler" 😂 come on

[atomic moon]

littlest

[proud bobcat]

what is.... language....

hmm

[ocean vortex]

wdym

[atomic moon]

its a reading thing

[fiery gull]

still good in words 🙂

[ocean vortex]

It's the best model

[zealous sparrow]

speciale thinks too much and i mean it. It literally timed out

[atomic lagoon]

Not what the language is mister smart guy, what the site is about

[fiery gull]

the gemini is nerfed ;-;

[ocean vortex]

has it's quirks and weird things about it, but this is still the best model

[proud bobcat]

what is... a website....

[atomic lagoon]

Mine just wont send the prompt lmao

[proud bobcat]

:think:

oof

rip

[atomic lagoon]

You're just rage baiting

Get outta here

[atomic moon]

gemini 3 on the gemini website beats chatgpt for me for everyday things

the ui does need a refresh though

[celest orchid]

what the

[normal abyss]

nerfed?

[atomic lagoon]

You're the same way on the open ai discord, I'm sure it gets boring acting like that

[zealous sparrow]

exactly my point holy s-

[echo aurora]

Hello, if you haven't already, would you mind creating a post in #1343291835845578853 and provide a clear break down of the steps to take to reproduce the bug?

[ocean vortex]

Gemini website would be banned if I was to choose. It shits the bed so bad compared with aistudio

[zealous sparrow]

@echo aurora is this on the deepseek side of things or yours

[atomic moon]

2.0 is coming soon apparently

[wise gale]

wtf DeepSeek thinks it’s ChatGPT.

translate:
The code size will be very large. ChatGPT has limitations. But is it possible to generate all the code? We can try to generate it, but the response length may be limited. Even if the content is large, I can at least create the basic structure.

[celest orchid]

um

[echo aurora]

Hmm what do you mean?

[zealous sparrow]

we are not being fr

[fiery gull]

why bro, littler exist 🙃 , I thought I missed english lol

[echo aurora]

Thats odd, going to try and repro

[zealous sparrow]

it thinks too much for a hello

[fiery gull]

ahhhhh like another bad word

[normal abyss]

oh my god thats longer than GPT5 high when answering like a rly long question😭

[zealous sparrow]

I asked for a simple html 3d shooter bro

[native yarrow]

littler is not a word

littlest is not a word

[fiery gull]

ahhh, but html 3d is complex no?

[native yarrow]

[normal abyss]

ooh ive never tried that, how long would that take on other models?

[zealous sparrow]

It also hallucinates as hell

[native yarrow]

someone's ripping MAD butt in here

[atomic lagoon]

Good lord speciale talks forever lmao

[native yarrow]

DAMN

[echo aurora]

And ofc when I test this I get the same, and then when I go to get a recording it works just fine

[zealous sparrow]

@echo aurora

[celest orchid]

physmath language

[native yarrow]

so is deepseek just stupid or sum

[atomic lagoon]

Its too smart for its own good, taking a simple "Hello :)" out of proportion

[zealous sparrow]

watch it work for other models but speciale

[atomic lagoon]

But this is just a glitch

Has nothing to do with the hello

[celest orchid]

He doesn't like language, he likes numbers.

[zealous sparrow]

something's really wrong with speciale

[native yarrow]

hmmm so it is very smart with detailed questions or sum with a point

[fiery gull]

after opus 4.1, everyone like dumb lol

[zealous sparrow]

WAIT I THINK I KNOW

This is somehow set to Deepseek MATH

[echo aurora]

Btw I've flagged the issues with Deepseek-v3.2-speciale to the team.

[atomic lagoon]

🦊

[normal abyss]

looking at the model name, the correct term is special 😭

[native yarrow]

LOLZ

[zealous sparrow]

Have you considered it being set to Deepseek Math? It's doing math for everything.

[fiery gull]

LOLLLLLLLLL ;-;---;-;-;-;-;-7

[native yarrow]

bro omg someone is mouth breathing

in this bathroom

omg

hella hard

[atomic lagoon]

?

[echo aurora]

I don't think so as this is a response I got.

[zealous sparrow]

Hm. Let me test again.

[celest orchid]

We need to respond. The user said "hiii". Probably just a greeting. We can respond in a friendly manner. As ChatGPT

[echo aurora]

It's hit or miss. I had one test that was off, second test worked.

[zealous sparrow]

Still math hallucinating for me

[fiery gull]

I miss the ping, I'm laughing about the special deepseek 3.2

[native yarrow]

"you are chatgpt" ??

[zealous sparrow]

the model was born in hallucination world

[celest orchid]

xD

[atomic lagoon]

Yea its rough but it just got in lmao. When I asked it to code me something it basically overloaded and kept sending the same code over and over until it told me something was wrong

[native yarrow]

damn

deepseek gotta lock in..,.,

[atomic lagoon]

Deepseek when I catch you deepseek

[zealous sparrow]

Only speciale is hallucinating for some reason

[atomic lagoon]

I like it when AI gives good file structure

[fiery gull]

soo good

[proud bobcat]

ts actually fire?

[proud bobcat]

yeah id wait for a terminus update

[atomic lagoon]

I'm not sure, I gotta test the code it gives once its done

Yeaa

[proud bobcat]

remember that deepseek is made for both chinese and american users so it needs a balance

[echo aurora]

Yeah something is very odd.

[celest orchid]

Let's check this Einstein on HTML

[atomic lagoon]

HTML of all things lol?

[celest orchid]

yes

[fiery gull]

nah, my first try with deepseek 3.2v special it halucinate 2 (TWO) times

[atomic lagoon]

HTML Is possibly the easiest language to code, only issue is that it can be lengthy but all you mane

[proud bobcat]

let me try it ngl

[atomic lagoon]

I just wouldnt really call it a test

[proud bobcat]

even if 3.2 isnt revolutionary im still using it for roleplay personally

[celest orchid]

bc the chinese put a book on math in his brain

[fiery gull]

worse that 3.2v exp, I thought the next deepseek is better that this ;-;-;-;-;

[ocean vortex]

It is made to satisfy CCP, and then it is made for other markets - mainly west. 👀

[fiery gull]

I think this, and python tool too

[zealous sparrow]

Is this model ID speciale?
019adb32-bfb8-7885-8d85-85220995b7cf

[proud bobcat]

only on their api tbh

[fiery gull]

hmmmmmm I liked the fisrt prompt of deepseek 3.2v without special, but worse that sonnet 4.0

[proud bobcat]

id wait until other providers provide the model

[celest orchid]

something is very strange...

[zealous sparrow]

guys just ask it politely

[celest orchid]

omg

[echo aurora]

Problem solved.

[echo aurora]

[fiery gull]

have the ( . ) it hallucinate

[echo aurora]

Team is looking into.

[celest orchid]

bruh

bad ai for vibe coding

[proud bobcat]

right let me try

[ocean vortex]

Actually, on their API the censorship is less bad. It's the worst on their official website with input/output moderation and hardcoded responses not from the model

[proud bobcat]

ahhh i mean like when other providers start pushing it out

i dont mind the censorship tbh

i dont use deepseek for politics

its minimal imo

[zealous sparrow]

so putting "do not hallucinate math" fixes the issue

[ocean vortex]

It's insane levels of censorship and overfitting for politics. But obviously if you don't ask it any political questions you ain't gonna see it lol

[celest orchid]

Dickpeak revenge

[zealous sparrow]

I mean you just have to understand that in china you just cant have your AI say that

button works and 0 errors [speciale]
prompt: do not hallucinate math and make a cool website in html

[ocean vortex]

Almost everything that has to do with international politics and China, model is gonna respond with an overfitted text carefully curated by the AI lab, or to be more precise - CCP

[zealous sparrow]

I highly doubt this is speciale tho

[quartz light]

after 63 DAYS deepseek 3.2 released??? 🤣 🤣 🤣

[zealous sparrow]

I want to believe its MATH under DISGUISE!

[quartz light]

they released 3.2 exp 63 days ago 🤣

[proud bobcat]

at least its not BUILT IN

yknow

they keep it reserved to their app

[jovial sapphire]

loool

is it even good?

compared to g3

[fiery gull]

I thought the deepseek v4 and is a minimal a better model of kimi k2 turbo thinking

[jovial sapphire]

and opus

[ocean vortex]

You mean like it's gonna be the end of the world if it objectively answers politically sensitive question? Crazy

[proud bobcat]

deepseek is a good daily use tbh

[proud bobcat]

v4?

[jovial sapphire]

gemini 3

perfect for daily use

it's even good at cooking

[fiery gull]

I thought the name is v4

[jovial sapphire]

i cook with gemini 3

[proud bobcat]

yeah but- 2$ input and 12$ output is expensive

[jovial sapphire]

use gemini app lol

[proud bobcat]

ok

3 responses

[jovial sapphire]

nahhh

[proud bobcat]

deepseek unlimited

ez

[fiery gull]

use lmarena 👀

[jovial sapphire]

*well i pay for the app

[quartz light]

unlimited slop

[jovial sapphire]

100%

[ocean vortex]

Oh the overfitted answers are built in to the model itself. But with official website they go 1 step beyond to prevent it talking about certain topics altogether and don't take any chances

[zealous sparrow]

speciale is lazier with code than v3.2 thinking, but i think its because speciale is currently hallucinating math

[quartz light]

speciale might be special

https://cdn.discordapp.com/attachments/990348027422203969/1399757536206258196/togif.gif

[celest orchid]

oh

[proud bobcat]

did they pull themodel from lmarena???

im not getting it anymore

[atomic lagoon]

LMAO

[quartz light]

the only thing i like about 3.2 exp is that it beat the problem only gemini 3 could

[jovial sapphire]

wth is this test?

[zealous sparrow]

https://codepen.io/qdpuzepc-the-styleful/pen/NPNBMjO

[atomic lagoon]

Very likely because its sort of broken

[atomic lagoon]

What is this?

[jovial sapphire]

very bad

[quartz light]

"speciable"

[jovial sapphire]

compared to g3 or other top models

[zealous sparrow]

a deepseek v3.2 thinking html website when i told it to make a cool one

it aimed for a portfolio one

[atomic lagoon]

WAIT THATS INSANE

[fleet lintel]

deepseek v3.2 is actually good

[celest orchid]

💩

[low imp]

V3.2 is good at math

[proud bobcat]

i really hope it gets an update

3.2 terminus will be peak

[celest orchid]

im scared

[ocean vortex]

I will always praise the models that have no issues criticizing their country of origin and the lab that created that model - that's like the main indicator to see if the model is not tampered with to make it biased or politically censored

[zealous sparrow]

bro went own the rabbit hole

[quartz pike]

yall how good is deepseek v3.2?

[celest orchid]

good for math

[zealous sparrow]

https://codepen.io/qdpuzepc-the-styleful/pen/NPNBMjO
V3.2 thinking

[spark python]

Why cant I use it in arena?

[zealous sparrow]

This is like every site today so like

[ocean vortex]

Chinese models, unfortunately, fail this test spectacularly...

[low imp]

Tuff

[fleet lintel]

better than GPT and slightly worse than Gemini 3 pro . not sure about Claude comparison

overall, its a good model

[proud bobcat]

as long as its not built in its fine

[atomic lagoon]

Yea no this is crazy. Better than opus I dare say. Was this one prompt?

[zealous sparrow]

we cant fully test speciale because once it dropped on LMArena it started hallucinating math

[zealous sparrow]

yeah

[ocean vortex]

wdym "built in"?

[atomic lagoon]

Still, AIs are NOT making it this good in one prompt of "Make a cool website" LOL

[proud bobcat]

OH WOW LMAOOO

[native yarrow]

what does this website do i cant access it rn

[celest orchid]

bruh

[zealous sparrow]

@echo aurora can we expect deepseek v3.2 models on codearena anytime soon

[quartz light]

"But that could be a huge amount of code. I must be concise but functional. I can write maybe 500-1000 lines. That's doable in a response? ChatGPT has a token limit, but it can output up to maybe 4096 tokens? Actually the output limit for GPT-4 is higher, maybe 8192 tokens? I'm not sure. But we should aim for something around 2000-3000 lines? That's too many tokens. So we need to keep it short." - 3.2 speciale

[zealous sparrow]

still hallucinating

[atomic lagoon]

Well I mean, this is usual for deepseek. China is ran by Dictatorship and socialism so it would make sense

Try to say Taiwan is a country

[ocean vortex]

yeah... they made it refuse to answer (overfit), so it is forced to come up with bs excuse in reasoning 💀

[atomic lagoon]

Or something about taiwan

[proud bobcat]

so

western models basically have safety stuff (parameters) built in

chinese models often do not (open source)

so if you use them on a non chinese provider theyre great

[proud bobcat]

oh yeah of course

i dont really mind

laws law

[atomic lagoon]

Yea lol, no problem with me

[proud bobcat]

as long as its open source though and can be used freely on other providers and locally im good

ahem gpt oss huge refusal rates ahem

[zealous sparrow]

Im making an upgrade to the deepseek v3.2 thinking site, and it crossed 1k lines rn

[celest orchid]

Add deepseek v3.2 speciale to code arena, this is the most underrated vibe coder

[proud bobcat]

i think that the model is going through technical issues actually

[ocean vortex]

What I was talking about is very much built in. They train their models on politically skewed answers that praise one China policy disregarding the entire internet. This you WILL get with the open-source weights included because it's part of the model, it IS built in.

[zealous sparrow]

we are at 1.2k

[celest orchid]

iq overdose

[zealous sparrow]

anyone know what a glassmorphism is

[proud bobcat]

no one has had to abliterate deepseek yet

🤷‍♂️

[echo aurora]

We are planning to add, I will update the announcement made when added.

[proud bobcat]

if they are built in its far more minimal

[zealous sparrow]

You guys are the best, thank you.

[proud bobcat]

and its not weary on the performance

is ewhat i mean

[zealous sparrow]

https://codepen.io/qdpuzepc-the-styleful/pen/NPNBMjO
here is the site after the update

[proud bobcat]

western models build a huge amount of safety parameters into the model which lobotomizes it

[zealous sparrow]

the preview is false i updated it

[ocean vortex]

I mean I respect Deepseek very much. But they are forced to include and do certain things like every other Chinese AI lab which are not ideal

[proud bobcat]

as long as it doesnt lobotomize the model im down

love deepseek

[atomic lagoon]

Wait so the first one wasnt the most recent model?

[zealous sparrow]

wdym

All is still thinking i just asked it to make it more stylish

[celest orchid]

umm

[zealous sparrow]

Not going to use uh speciale as its hallucinating

[fiery gull]

but the glm 4.6 is soooooooooo ahead 😭

[proud bobcat]

its a little hallucinatory

i think

[zealous sparrow]

tf

[zealous sparrow]

im stuck with it hallucinating math

[atomic lagoon]

Well I mean it will see itself as a human, programmed like that to kind of have feelings like a human

[fiery gull]

the glm 4.6 is light years ahead of deepseek lol

[proud bobcat]

glm is so peak

[atomic lagoon]

Whats glm

[quartz light]

lmarena is the buggiest and laggiest site ive ever used which isnt a webgame

[fiery gull]

just be half of size and better

[ocean vortex]

not really

🥀

[zealous sparrow]

wdym

[fiery gull]

glm 4.6 just is the most CxB model

[quartz light]

look at the right

💀

[fiery gull]

BENCHMAXXX I IDC

[proud bobcat]

benchmaxxing unequal to performance

[ocean vortex]

glm 4.6 is such a random model to praise

[celest orchid]

Bro its not actual

[zealous sparrow]

checkmark just felt tired of standing bro

[ocean vortex]

nothing special about it tbh

[quartz light]

fr qwen 235 does NOT beat 4.6

[proud bobcat]

???

[fiery gull]

REALLY,

[quartz light]

uh

[atomic lagoon]

To be fair, I was saying deepseek newest model is barely better than the last one but nah this ones kinda crazy

[ocean vortex]

Kimi2 think or Deepseek

[fiery gull]

glm 4.6 is really good

[ocean vortex]

not this

[zealous sparrow]

thought for 1 second is crazy bro

[fiery gull]

is usable today

[quartz light]

kimi sucks dog

for coding

[quartz light]

this is better

see

3.2 better kimi

fr

[ocean vortex]

I think something went wrong

[proud bobcat]

i think theyre just having tech issues dw

i dont think the actual models will perform like this

[celest orchid]

I read his mind lol

[proud bobcat]

could be a rough launch

[zealous sparrow]

no like i think its an LMArena issue

[proud bobcat]

why are all requests to deepseek rerouting to clude for me???

[fiery gull]

bro this benchmark say the gpt 120b have ahead of OPUS 4.1, HELL NOO HELL NOO

[proud bobcat]

im having iussues on openrouter

[fiery gull]

this don't make sense ;-;

[proud bobcat]

120b BLOWS

[quartz light]

imo ernie 5 is good

[proud bobcat]

how did they screw up so badly that 20b outperforms 120b

i swear openai is a paper tiger

bloated datasets

[fiery gull]

serious? I don't tested so much the oss´s (I love this 3 sss)

[zealous sparrow]

the thinking model works fine tho

[proud bobcat]

yep

not even joking

[proud bobcat]

lemme see

honestly i wouldnt call 3.2 bad until tomorrow or 2 dayslater

[zealous sparrow]

Yeah its usually when models wear off

[proud bobcat]

it could be model degradation from high compute requests

[zealous sparrow]

prob why gemini 3 performs a bit worse now

[proud bobcat]

if 3.2 is peak at writing i will use it for rp mainly

[ocean vortex]

That's because it got destroyed on LCB, HLE and math. It's not their fault Anthropic made this a specialized model

[zealous sparrow]

It was OVERAnticipated

[proud bobcat]

i hope anthropic will make an optimized sonnet ngl

i love sonnet but god DAMN the price

[proud bobcat]

yeah no

nuh uh

[ocean vortex]

Anthropic models are not great on LCB

especially the older ones

4.5 is better

[proud bobcat]

i mean

120b scoring 88

is insane

reminds me of the stupid vibethinker model that claimed sota levels at 1b

used it on my phone and it started having a schizophrenic breakdown

[ocean vortex]

That model does score quite high and the LCB score is valid.

[fiery gull]

bro belive in benchmaxx in 2025

[celest orchid]

Guys deepseek v3.2 speciale is number 1 now🤯

[fiery gull]

worse that flat earther

[proud bobcat]

saturated

i believe SWE bench more

cause thats actual github issue resolve rates

[ocean vortex]

But I agree this may be the perfect case study for AA on how they could improve their weighted score. To penalize smaller models more. Though you can't say that OSS doesn't perform... It aces most of the metrics

[zealous sparrow]

accurate

[proud bobcat]

LMAO

please deepseek iron this out

😭

[zealous sparrow]

prob when they tweak the model to a good state it willbe low

[ocean vortex]

OSS is a bit like grok4. It performs good on most benchmarks but just doesn't hold up IRL. At least in my opinion. But you can't deny the scores it manages to get

[celest orchid]

We need to add a leaderboard for hallucinations

[weary cosmos]

can anyone build this please? thanks lol
https://cdn.sdappnet.cloud/rtx/ai_torrent_protocol.html

[zealous sparrow]

theres a hallucination bench

LB aint needed as deepseek will be tweake in prob less than 2 days

[atomic lagoon]

What is it

[celest orchid]

wth

[weary cosmos]

read it bro

[atomic lagoon]

AI Torrent Protocol doesnt tell me anything

What does it do

What is it for

Why use it

[torn mantle]

nice deepseek speciale was added

[celest orchid]

ai torrent protocol🐀

[zealous sparrow]

it's currently hallucinating

[torn mantle]

wym

[weary cosmos]

u know like torrent + ai models

[bleak fossil]

https://tenor.com/view/i-don't-know-idk-idk-about-that-gif-7336250873949261946

[ocean vortex]

official provider got ignore on OR

training on your data

[ocean vortex]

💀

To our knowledge, this provider may use your prompts and completions to train new models.

This provider is disabled, but it can be re-enabled by changing your data policy.

View this provider's privacy policy to understand its data policy.

OpenRouter submits data to this provider anonymously.

[zealous sparrow]

they need to train it to not hallucinate

[celest orchid]

yes

[left iris]

all video arenas seem to be down atm ?

[atomic lagoon]

Oh alright, but good luck getting AI to make that or anyone to make that lmao. If you mean torrenting AI models btw its not possible

[proud bobcat]

true

grok 4 was dissapointing

[atomic lagoon]

Hopefully grok can be good next.

[proud bobcat]

grok 4 fast and 4.1 though are good

[atomic lagoon]

I want grok good at coding

[proud bobcat]

not surprising

[atomic lagoon]

I'm sure it being more unfiltered it could make me a memory reader and spoofcall

[zealous sparrow]

I hope they unbrick speciale

i cant test the model

[proud bobcat]

i think 3.2 will make it into the top 20

and speciale (if it gets unlobotomized) maybe number 10??

[ocean vortex]

On top of that they are slow and the alternative on OR does not seem much faster. It's generating 22min now. This is very painful to test 🗿

[graceful warren]

The generation don't working in the serveur, please repport

[left iris]

servers down yeah

[zealous sparrow]

which model

[ocean vortex]

speciale

[zealous sparrow]

turron says that speciale is bricked on router too

[rocky fractal]

What's wrong with speciale besides giving an error on lmarena?

[ocean vortex]

If someone told me 2 years ago it's gonna be normal to wait 20min+ for a response from a LLM I would have thought they are crazy

[zealous sparrow]

speciale just seems bricked rn

thinking is the only one that works

[obsidian cargo]

deepseek speciale is pretty cool actually

[rocky fractal]

How so? They don't say speciale is for coding, just thinking

[obsidian cargo]

I think the problem with speciale is that it thinks so long that it times out or something

[zealous sparrow]

it gives you a math response to hi bro

[rocky fractal]

Oh that's just Lmarena no?

[obsidian cargo]

yeah

I gave it a game concept and watched it spend several minutes turning it into a GDD inside its thoughts then "something went wrong"

[zealous sparrow]

my brain aint braining, can someone explain what is this in a NES emulator, this is speciale btw

[rocky fractal]

Same, just opened a bug about it

[zealous sparrow]

speciale knows its sources

yeah speciale died when generating me the final code

[rocky fractal]

For front end, thinking gave nicer results than speciale

[zealous sparrow]

plot twist speciale is some sort of gpt-oss

[rocky fractal]

When speciale decided to work

[thorny schooner]

Honestly the security verifying is so weird for me because it seems very hypertensive where after a bit it athletes every other crop or sometimes even each and after each individual prompt

[zealous sparrow]

I asked speciale for pacman, WHY IS IT VISUALIZING THE MAP

[rocky fractal]

He's imagining it in his mind

Is hope they add this deepseek 3.2 thinking to GitHub copilot for free, it would be the best free model

[zealous sparrow]

speciale's way of working frightens me

Speciale wins award for weirdest model of the year

SPECIALE has been pulled

!

[celest orchid]

https://019adb89-203a-78a8-a718-fd2e2e7046e0.arena.site/ guys im maked a working model deepseek v3.2 speciale chat try it

[zealous sparrow]

it got pulled

[celest orchid]

try mine

[zealous sparrow]

@echo aurora Have you managed to identify if the issue is on your end or the provider's end?

[rocky fractal]

Did they remove speciale 🥀

[obsidian cargo]

its gone yea

[zealous sparrow]

yeah they pulled it

imo good

prob will get remade

[echo aurora]

Yeah we removed it for now. Team is exploring.

[left iris]

is video arena down for everyone ? or does it function for some ? for me it seems all 3 arenas are down

[echo aurora]

Hmm

[zealous sparrow]

Still unsure if the issue is on Provider's end, dang. That's one big issue.

[celest orchid]

https://019adb89-203a-78a8-a718-fd2e2e7046e0.arena.site/ im maked working pineapple test it

[rocky fractal]

💀

[zealous sparrow]

no its dead

[left iris]

shows generation failed for all the videos

[rocky fractal]

Wouldn't be surprised if western countries are attacking their servers again

[echo aurora]

Oh yeah I'm seeing a lot of errors in #video-arena-1

[proud bobcat]

a bit hallucinatory rn

[celest orchid]

no it's working

[echo aurora]

Our site appears to be working.

[rocky fractal]

Is it testable on deepseeks end?

Or app

[zealous sparrow]

bro it doesnt even send an API request

[celest orchid]

I made deepseek v3.2 from scratch

[zealous sparrow]

but you know that the API point you put in

was prob the LMArena one

that has now been pulled

[proud bobcat]

you can go on openrouter

[rocky fractal]

This kind of errors is common on Lmarena, GPT 5.1 high also tends to give the sorry couldn't do that request error

[proud bobcat]

or their api

[celest orchid]

mine is working

[proud bobcat]

i think its probably good

but

[rocky fractal]

It's free during release?

[proud bobcat]

no no

but

[zealous sparrow]

...

[proud bobcat]

reallly cheap

[rocky fractal]

I'll look into it ok

[zealous sparrow]

We are stuck with Thinking till speciale gets fixed

[torn mantle]

sorry but whats the issue here

isnt this what monte carlo is

?

stochastic simulation of N

[zealous sparrow]

the model responded absolute whatever to a hey

[torn mantle]

ah

[rocky fractal]

How hard do we have to cry for GitHub to add this deepseek to copilot

[torn mantle]

i didnt see 'hi'

lol

some api issues then lol

[cursive verge]

Hi

[loud cosmos]

Hallo

[echo aurora]

[loud cosmos]

Jjj