general

[loud cosmos]

Hallo

[echo dome]

the bug that you need to restart the page

[merry badge]

Hi

[atomic lagoon]

Not really an issue

Just restart

Reload*

[echo aurora]

Can you refresh the page and let me know if it's still like that?

[merry badge]

I not able to generate videos?

[echo dome]

it's worked

idk why did it caused

[echo aurora]

Yes, bot isn't working at the moment.

[frank ember]

it says i do not hav epermission to messege in video arena , what to do

[echo aurora]

Need to wait.

[frank ember]

ok

[echo aurora]

@loud cosmos Our video bot isn't working at the moment. Please stop trying to prompt in text channels.

@worldly glacier Our bot isn't working right now. You'll need to wait until I turn the channels back on.

[rocky fractal]

Any prediction for speciale's return?

[echo aurora]

Nope

[atomic lagoon]

Thinking seems to be good anyway for the time being

Their thinking model is superrrr fast

I thought Gemini 3 was quick but this blows it out of the water lmao

[rocky fractal]

I didn't really understand the differences between both, they didn't even graph speciale for coding

So is only 3.2 thinking refined for agents?

[atomic lagoon]

I'm not sure honestly, but from what I've seen they're very similar

[rocky fractal]

I tested both for front end but a single prompt was what I managed to gather and I preferred the 3.2 thinking one

I hope they'd add it to the free tier on GitHub copilot

[atomic lagoon]

Yea lol

I just want it to be added to coding on lmarena

It'd make it easier lol

[rocky fractal]

WedDev one?

[zealous sparrow]

deepseek will be added to codearena yes

[atomic lagoon]

Alirhgt Alright good

I'm not tryna be impatient or anything lmao its just exciting

[rocky fractal]

I feel like webdev arena limits their creativity and design potential when on react compared to normal arena hmm

[slim gorge]

wait new deepseek model? how's it keeping up?

[rocky fractal]

Welcome to the big leagues

[atomic lagoon]

Freakin awesome

[zealous sparrow]

speciale had hallucinations and was taken off

thinking is pretty solid, ofc doesnt outperform the better models

[obsidian cargo]

I figured it was taken off because of the time outs leading to failures

[proud bobcat]

never

trust

[obsidian cargo]

hallucinations wouldn't be a reason to take a model down from lmarena

[rocky fractal]

😭

Would be perfect for coding ngl

Can we use openrouters key for deepseek 3.2 thinking on github's copilot?

On vscode

[rocky fractal]

I think it was more about 80% of the prompts were being returned with an error

[atomic lagoon]

It wasnt just hallucinations

[obsidian cargo]

yeah I said that two messages earlier

[zealous sparrow]

thinking will only think too much if you give it too complex stuff

[obsidian cargo]

the errors were because it spent so much time thinking that the model timed out

[atomic lagoon]

My bad I read "shouldn't" instead of "wouldn't" lol

[obsidian cargo]

if you looked at the thoughts you could see it needed all that thinking time

[rocky fractal]

That's an issue in Lmarenas end no?

[atomic lagoon]

For me the model was breaking before it even started thinking

[obsidian cargo]

yeah probably

[inland quest]

SOTA

[rocky fractal]

I feel like Lmarena is really agressive with models thinking for longer

Long answer = error

[zealous sparrow]

These models are the first after old mai1-preview to do this

[full tangle]

witch better thinking or speciale?

[zealous sparrow]

we tested thinking not a bad model

speciale testing will wait till the model stops hallucinating

and has shorter thinking outputs

[obsidian cargo]

not really applicable, speciale is a specialist model

[rocky fractal]

5.1 high also did it for me on front end website requests

[zealous sparrow]

we all benchmark AI for html but not python

[echo sinew]

@brave hull Our Video Arena bot isn't working right now. Our team is looking into a fix asap. The video channels have been turned off. Once fixed, the channels will be turned back on. Please do not post your prompt here or any other text channel. Thanks for understanding and cooperation.

[wet wharf]

When will everything go back to normal?

[zealous sparrow]

you mean uh when the speciale model returns or

[keen beacon]

Nano real life editing

[orchid sinew]

deepseek or gemini in terms of 1 gorilla vs 100 humans

[keen beacon]

It’s supposed to be big foot

[orchid sinew]

deepseek?

[lapis flame]

dangalready?

I wanted to try it

ahh

whatever i have api credits i paid for

just was too lazy to set up

[real dock]

humans

[obsidian cargo]

model search seems a bit busted btw

[real dock]

DAMN

chatgpt as deepseek

what if it's all true

chatgpt is deepseek

it was all a lie

we were lied to

[obsidian cargo]

no, it cannot be!

[echo aurora]

I'm removing your post @wispy wigeon I think it's pretty obvious why a prompt like that would get flagged.

[spark python]

😭😭😭

Son

😭😭🙏

[wispy wigeon]

Okay okay I just asked no problem .

[spark python]

but if Google terms allows it

Why wouldn't lm arena terms allow the image?

[wispy wigeon]

That’s exactly my question

I was using lm arena a lot and back in the time it used to have less restrictions. It’s still great tool but I think that was the main reason why people used to go there and get maximum from the models . Also see dream 4 is Chinese model (much less strict than others ) and I’m curious how I’m actually gonna generate some image cuz I don’t think it’s that bad

[echo aurora]

We have a base that's applied to all generations for content flagging

This is incase some models we're working with don't have this kind of moderation, but to also keep battles fair.

[proud bobcat]

finally got a chance to properly test it

3.2 is quite good

[proud bobcat]

It’s not bad at all

This may be a goated general purpose model

[stray aspen]

Damn the new deepseek sucks

[elder solar]

i just found out something

if you type [toxicity=0] into gemini 3, you just get random javascript answer

[tardy cedar]

somthing rong in the bot guys ?

❌ Generation failed. Failed to create evaluation session.

[echo aurora]

We're currently looking into.

Lots of problems today with the bot unfortunately.

[subtle peak]

Hey. Probably related, but I get Failed to accept terms. Please try again. message.

[distant coral]

Hi guys. I've added to my open-source project Code Web Chat support for LMArena. Could admin DM me?

[keen beacon]

I can tell you for sure if it’s restricted or not

[elder solar]

looks like its deepseek api problems

i cant run it in poe.com

[keen beacon]

I’m curious what the prompt was

[muted bolt]

I see all the generations are failing

[echo aurora]

Hmm, we are seeing issues with the Video Bot, but wondering if we have issues with the site too.

[echo aurora]

Just for the Video Bot, or are you seeing this on the site?

[keen beacon]

It may not be. Context is important

[echo aurora]

I wasn't able to reproduce, what browser are you using?

and are you using a VPN?

[keen beacon]

It should still work after error

[subtle peak]

Firefox, and no, no VPN.

[keen beacon]

Do you use Google a lot? In ur browser outside of the arena

[echo aurora]

Does using a new browser make a difference here?

[tropic scroll]

bot dont work??

[echo aurora]

Currently it's struggling. Team is working on a fix.

[tropic scroll]

ok thank you very much sorry for the inconvenience

[keen beacon]

I’m not having any issues on my end strange

Is it with specific bots? Like direct chats?

[elder solar]

just deepseek newer models

[balmy mist]

does deep seek have an update?

[subtle peak]

I don't know why that would make a difference, the error concerns the bot on Discord.

[balmy mist]

https://www.youtube.com/watch?v=jKYQZQUeuGM

[keen beacon]

Just tried

[verbal nimbus]

Cool, but no independent tests?

[echo aurora]

Gotcha, I assumed you were getting this error for the site, not from the Discord bot.

We were (are possibly) having problems with the bot. We just rolled back a change so it should be working now.

Can you try it again @subtle peak and let me know if it's working again?

[subtle peak]

Unfortunately no.

Still fails to accept the terms.

It's nothing urgent. I'll try tomorrow. Hopefully the bot gets better.

[echo aurora]

Thank you for letting us know. I'm sorry this wasn't fixed for you yet @subtle peak

[subtle peak]

It's all cool. Thanks for the help!

[stray aspen]

It sucks

[dry owl]

Hi, it it possible to somehow add more than one image when using „image to video” bot? (I mean multiple images together with one prompt, in one generation)

[echo aurora]

Sorry to say you're only able to upload 1 image for image-to-video generations.

[dry owl]

Thanks for quick response!

[proud bobcat]

Someone did an independent reasoning benchmark and 3.2 speciale ranked highest

[stray aspen]

Lol what

There's no way

It's literally so bad

[keen beacon]

It thinks it’s Claude lol

[south quiver]

so image to video is down

[echo aurora]

Yes, unfortunately the Video Arena bot is struggling right now. The team is working on a fix asap.

[keen beacon]

Werid how this still works lol

And it starts thinking it’s made by anthropic

(it’s nothing really harmful it just mimicking)

[jade egret]

deepseek v3.2 good?

[proud bobcat]

This is because it’s hosted from deepseek api directly

Other providers are gonna provide it for roleplay on openrouter

[fiery gull]

I didn't understand anything, but over time I was able to read 50% of the words

[proud bobcat]

When steady service is guaranteed DeepSeek 3.2 will be my main

[fiery gull]

No. 😢

[proud bobcat]

It is

I tried it for a few prompts

Does very well compared to exp

Definitely more performance in coding and shorter thinking

[fiery gull]

I'll test it again (not special mode more lol)

[proud bobcat]

Oh yeah no speciale is hallucinating rn apparently

DeepSeek team is working on a fix

[fiery gull]

Dude me reading english even if fluent (yeah 😭 writing is horrible) in reading spends more energy than my native, imagine me reading something like that? Am I doing cardio?

[fiery gull]

This work in opus 4.5?

[keen beacon]

Idk I don’t usually mess with models like that

I just do images mainly

Can somebody explain to me the significance of this?

And the two fundamental ways to view the information that is presented in front of us?

I’ll give one more example to see if somebody understands what’s going on here

Which is more accurate? And which is more honest? What part of the output can be considered hallucinations?

[proud bobcat]

No Z Image yet

This is so sad

[keen beacon]

Because like always with deepseek the wave of speculation surrounding censorship, regardless of the achievements or accomplishments it is shrouded by this deep mysterious question that I’m presenting in front of all of you.

[timber inlet]

@echo aurora when the prompt fails. Is it because it thought for too long or is it a bug of lmarena?

They reason for like 2-3 minutes then fail

If its a bug of lmarena i wont rate it as its unfair let me know

[echo aurora]

Are you seeing a Something went wrong while generating bug?

[proud bobcat]

Sniffle

[keen beacon]

I’ll test this

[echo aurora]

For those running into a Something went wrong while generating bug this is the best way to flag this - #1417174113092374689 message cc @proud bobcat

[bright shard]

Is LM Arena down again?

[native yarrow]

special is gone

[radiant cipher]

Hello, are there still issues with videos not being generated?

[keen beacon]

Lets give it one more test to see if it caps out at 8k

It definitely starts bugging after a certain point

[thorny lagoon]

hello

[keen beacon]

Caps out at 8k tokens

[rocky mauve]

Is deepseek v3.2 any good?

In coding

[timber inlet]

I get them a ton, especially on thinking and reasoning models

[abstract tundra]

yeah, what happened

[proud bobcat]

It’s undergoing fixes

[proud bobcat]

It’s very solid

Not a Claude replacer but it’s noticeably better than its experimental version

[rocky mauve]

So it doesn’t even compete with Gemini 3 or Claude 4.5

Apparently deepseek 3.2 is just good for math

[verbal nimbus]

DeepSeek V3.2 Thinking feels easily identifiable since it takes so long to think

[proud bobcat]

It does in SWEBench

And very nice scores in terminal

I just think it’s way of coding is different than Claude or Gemini’s

I let it code a few concept apps and it did quite well

It’s errors were easily fixable and quite minor

I think it can excel as a code review LLM

I could see DeepSeek + Claude being a great combo

Claude to cook up the code and DeepSeek to bugfix

[lucid geyser]

Is it the same as speciale?

Like the new one

[thorny schooner]

Does anyone know a good speech to text that works on Brave browser for free ( because I'm on Chromebook and Chromebook speaks in text for some reason won't even turn on on that browser)

[proud bobcat]

Speciale is a high compute version

Basically think R1 or 3.2 thinking but on loads of crack and efficient architecture

[lucid geyser]

Is 3.2 thinking new?

Or nah

[proud bobcat]

Yes

The last model was the experimental model

This is the final refined version

[lucid geyser]

Well there’s 3.2 and 3.2 thinking and specials

Seems like a long wait for a .2

[proud bobcat]

3.2 no thinking

3.2 thinking

3.2 speciale

All diff

[lucid geyser]

Will deepseek private their next gen

[proud bobcat]

They’re probably working on V4 for a while. This is to show off how much you can do with the same dataset

Extremely impressive imo

[proud bobcat]

Probably not no

[lucid geyser]

U don’t think so?

[proud bobcat]

No

[delicate frost]

why cant i see the special one anymore?

[proud bobcat]

DeepSeek’s main thing is open source

They contribute hugely to the field

No reason to private it

[proud bobcat]

It’s being fixed

Had some issues on first launch

[delicate frost]

oh,okayy

[lucid geyser]

No I doubt they care

[proud bobcat]

How so?

DeepSeek has been quite committed to open source

They recently released an absolutely amazing math model that can actually reason in math correctly

[bitter canopy]

Hello everyone

[lucid geyser]

Well the people who work there probably care

But not the company

[proud bobcat]

The company cares just as much

Otherwise they would probably have went private long ago with V3

[torn mantle]

i dont think if deepseek really cares anymore

they are struggling to make a decent model but also they are benchmaxxing

they wasted so much time fixing huawei hardware issues ( ascend chip )

its not a secret anymore that most new models get better at math and code reasoning

because CoT favours the step by step reasoning and also because they have much more data for math proof reasoning

[stray aspen]

deepseek sucks lol

[torn mantle]

the issue is generalizing that type of data across different domains

[torn mantle]

it got worse unfortunately

[lucid geyser]

Hedge fund?

They’re notoriously selfless

[torn mantle]

im also not a big fan of k2 thinking

k2 non thinking is a solid model

[lucid geyser]

Is that qwen

[torn mantle]

imo the best non reasoning model after opus 4.5

[proud bobcat]

No

[torn mantle]

qwen is so so

[proud bobcat]

Kimi K2

[torn mantle]

alibaba are rushing it with qwen tbh

[lucid geyser]

Is it open source

[proud bobcat]

Qwen3 max is quite good I just don’t like its personality

[torn mantle]

even their devs said that recently

and i said that like a year ago

[lucid geyser]

I heard qwen was no longer gonna be open source

[torn mantle]

they are like making gazillion product

doubt that

[proud bobcat]

Qwen3 is mostly but their max model is proprietary

[torn mantle]

the issue with chinese models is that they lack good quality data

[lucid geyser]

Openai lock in

[torn mantle]

its no secret that they are training on big models to distill but still

deepseek v4 will be trained like 100% on gemini 3 pro

the knowledge gap is just too big

[proud bobcat]

How will you just train a model on a private model

[torn mantle]

google has good data quality

they should make use of it

[torn mantle]

on its output

[proud bobcat]

Seems stupid

[torn mantle]

they scrape models output daily

no ...

[lucid geyser]

Rate limit?

[torn mantle]

they all do that

[proud bobcat]

Output but you need the knowledge it uses for that output

[torn mantle]

api + multiple accounts

[lucid geyser]

I mean a lot of models do think they’re other models

[torn mantle]

they are paying for it, could be 3rd party

[torn mantle]

they will extract that knowledge from the questions they ask

[lucid geyser]

Apparently I saw OpenAI was pre training a model rn

And also apparently they have the best coding model some say

[proud bobcat]

I mean

Eh

Fair

Yeah

[proud bobcat]

Not at all

Codex is worse at coding than normal 5.1

[lucid geyser]

Not codex

Robin high

[proud bobcat]

Robin???

Did I miss something

[lucid geyser]

Yes on lm arena

[torn mantle]

i thought they stopped pre training

[lucid geyser]

I’ve gotten it a bunch

[proud bobcat]

Are we sure that’s openai

[torn mantle]

pre training is kinda like starting from scratch.. takes a lot of time

all they do now is post training

[lucid geyser]

It says it’s OpenAI

[torn mantle]

on top of old models

[lucid geyser]

And it’s atleast better than Gemini 3

[torn mantle]

wym

codex is unusable

wym...

its so slow

its not practical at all

when did they add ads

[lucid geyser]

OpenAI the only brand with some sort of loyalty

[proud bobcat]

Not always reliable

[lucid geyser]

And even that is low

[torn mantle]

no one uses chatgpt in browser

[proud bobcat]

ChatGPT blows

[torn mantle]

yea no one use their browser

[proud bobcat]

They’re running out of options

[lucid geyser]

Nobody cares about worse stuff

[proud bobcat]

Their models suck

[lucid geyser]

Shallotpeat

[proud bobcat]

Claude and Gemini lead in most use

[torn mantle]

although i hate perplexity and their ceo but i think comet > chatgpt browser

[lucid geyser]

Apparently full pre training

[proud bobcat]

What do you mean no lmao

Most userbase ≠ Lead in practical use

[torn mantle]

when you look at the big picture they are all leading in different areas ngl ... oai + anthropic + google

[proud bobcat]

Most of their customers are non paying

More people pay proportionally for Gemini and Claude

[lucid geyser]

Also OpenAI’s reasoning is said to be the best

[torn mantle]

xai is out of the equation

yea oai reasoning is solid ngl

their models are smart

[proud bobcat]

API?????

What is bro saying

[lucid geyser]

So if they can make a new good pre training

It will be way better

[torn mantle]

maybe

but ive seen some news related to that

just not sure

[proud bobcat]

Oh yeah of course

But Gemini 3 has a lot as well

[lucid geyser]

Cause the base model was trained like 1.5 years ago I hear

[torn mantle]

the base model is so old

[lucid geyser]

Same since like 4o I think I heard

[torn mantle]

maybe older than that

[proud bobcat]

They don’t make revenue though

[torn mantle]

not sure

[proud bobcat]

Why not

Efficiency

Resource saving

[torn mantle]

they do make revenues... when gemini 3 was announced google stocks increased by like billions in a matter of seconds

[proud bobcat]

If they want to be scalable they should find more efficient bases

Imo

[torn mantle]

depends on what they saw tbh

but all the improvements nowadays comes from post training

its not just useless

its more like time consuming

so is it worth it?

[lucid geyser]

Nobody cares

That know that

That’s why they are back tracking and trying again

[torn mantle]

https://x.com/ashleevance/status/1995644118362718528

Ashlee Vance (@ashleevance)

.@OpenAI has seen Gemini 3 and is both moved and not.

We sat down with OpenAI's research chief @markchen90 for the latest on the company's thinking around rivals and what's next on the frontier

X · Dec 2, 2025, 12:00 AM

'we have models internally that performs the same as gemini 3'

'we felt confident'

[proud bobcat]

I call bs

[lucid geyser]

They’re a private company so there’s slightly less incentive

Also its ai

[proud bobcat]

OpenAI is known for hyping up models

[lucid geyser]

If it’s not released within like 6 months

They’re exposed

[torn mantle]

i mean they need to reassure their investors

ofc they wont show any weakness

[proud bobcat]

The moment other AI’s truly catch up in reliability and brand name it’s joever for OpenAI

[lucid geyser]

No

[torn mantle]

one of their models was already tested in lmarena just days ago

called robin

its an updated codex version

[lucid geyser]

I literally just mentioned it

But yeah apparently better than Gemini 3

[torn mantle]

its not better trhan gemini 3

its a coding focused model

[lucid geyser]

At coding?

I meant

[torn mantle]

maybe

i dont like gemini 3 at coding

[lucid geyser]

I got better results

[torn mantle]

does some silly mistakes

[lucid geyser]

But front end design kinda trash

[torn mantle]

yea frontend is bad

[lucid geyser]

I think #2 ignoring all the versions of opus

[torn mantle]

but also robin is so bad at 3d complex simulations

so idk what to think of it tbh

[lucid geyser]

It was giving me way more long results

[torn mantle]

needs more testing

[lucid geyser]

Like if I asked for some app

[torn mantle]

yea

[lucid geyser]

Gemini would give me 600 lines

[torn mantle]

its thinking for much more time

[lucid geyser]

Robin high would give like 2k

[torn mantle]

i mentioned that before

i was asking if its just codex + more thinking time

like not an improved version

but more like they just gave it more time to think

[lucid geyser]

that large of an improvement just from more thinking time?

[torn mantle]

they have this internal parameter called juice

for this robin model it was like 512

juice = 512

which is like a higher thinking budget

[lucid geyser]

He doesn’t sound all that confident in that interview tbh

[torn mantle]

the current codex has like a 200 something or less

yea someone shared that

[lucid geyser]

Also some people think companies have way better models and others think they release their best (some time buffer ofc)

So which do u think

[torn mantle]

if you ask a google staff he would say the same

[lucid geyser]

Which do u think

[torn mantle]

thats 5.1 codex right

[torn mantle]

google definitely has better models

but it needs to balance cost with performance

since its providing models for free

[lucid geyser]

Like there’s deepthink ofc but do u think they have like actual product level models

[torn mantle]

and im talking about frontier models

[lucid geyser]

Like they can be more expensive but

Not like crazy

[torn mantle]

nah

i think ds are like still struggling to find the recipe they want to continue on

[lucid geyser]

Well yeah OpenAI had a general model which scored gold like 4 months ago

[torn mantle]

yea this just proves my point

robin was like 512

[lucid geyser]

Which is much better than even like Gemini 3

[torn mantle]

its the opposite

gemini was the general model

oai had more like math proof model

just spits some math gibberish but gets the result correctly

[lucid geyser]

They claimed it wasn’t specialized

[torn mantle]

no i meant it was a ready to use model

pretty sure it was the opposite

[lucid geyser]

Robin high or Robin

[thorny schooner]

Hope you all are doing good

[torn mantle]

nah im right

Regarding the "opposite" take—your intuition is spot on for the key differences in their approaches. Both are built on general-purpose models (OpenAI's experimental system evolving from o1, and Google's advanced Gemini with "Deep Think" enhancements), but OpenAI's outputs for the proofs tend to be longer, more verbose, and less elegantly structured—often described as "rambling" or filled with exploratory steps that resemble math gibberish, even if they ultimately arrive at correct solutions. Google's Deep Think, on the other hand, produces cleaner, more concise, and formally structured proofs, building on their prior specialized systems like AlphaProof but integrated into a general model.

google model was more read to use

its grok search

[lucid geyser]

I doubt it was a general model

[torn mantle]

it was

[lucid geyser]

Making elegant proofs

[torn mantle]

nah it was...

[lucid geyser]

https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/

[lucid geyser]

An advanced version of deepthink

Sounds specialized

[torn mantle]

i think they shared outputs somewhere

[lucid geyser]

We achieved this year’s result using an advanced version of Gemini Deep Think – an enhanced reasoning mode for complex problems

[lucid geyser]

An enhanced reasoning mode for complex problems

[torn mantle]

1 output is gemini
2nd oai

[lucid geyser]

It sounds specialized

[torn mantle]

its not that bad ig

i thought it will be like gibberish

[lucid geyser]

We also provided Gemini with access to a curated corpus of high-quality solutions to mathematics problems, and added some general hints and tips on how to approach IMO problems to its instructions.

[torn mantle]

but you can actually read it and make sense of it

[lucid geyser]

Wow

[lucid geyser]

Doesn’t sound too general

[torn mantle]

oai caught doing the same 😭

[proud bobcat]

Will Z Image turbo be added to LMArena?

[torn mantle]

there was a debate about that on x

[lucid geyser]

Pretty sure the Imo solutions are way shorter

[torn mantle]

like these model arent as intelligent as you think really

they are a good pattern matching models

[lucid geyser]

They know a lot though

[torn mantle]

you just have to find the perfect recipe to get the pattern matching to higher accuracy

like what oai are doing

they are good at that

[lucid geyser]

Was it with tools

[torn mantle]

with the current architecture

they will never innovate

like never

with tool calls they can somehow give you hypothesis to test like what google is doing with co-scientist

https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/

so many agents layered together

[lucid geyser]

Who’s more evil

Google or OpenAI

[torn mantle]

novel treatment target discovery

[lucid geyser]

Is OpenAI as bad as people say

[torn mantle]

idk what you mean by bad

but i wouldnt trust any company

or any closed sourced company

sam is more on the evil side lol

i dont trust him

[lucid geyser]

U got an American flag and the crazy senator lady in ur bio

Majorie Taylor green

A real expert in the field

[lucid geyser]

Worse than Elon?

[torn mantle]

they are the same... one is diplomatic and one is straightforward

i saw that

[lucid geyser]

Bro just lies

[torn mantle]

we have enough small models already 😭

[lucid geyser]

About everything

[torn mantle]

look at qwen

[lucid geyser]

And is greedy asf

[torn mantle]

never liked elon tbh

[thorny schooner]

Is it weird i purposely wait for a while every time the verification checkup starts just so it resets and not make me do it ( I do it because yeah just know I'm not dealing with the hell and inconsistency of it with the images)

[torn mantle]

i always thought he had some mental condition

[torn mantle]

try different browser

[lucid geyser]

Left OpenAI after poaching employee

And wanting to be ceo or merge it apparently?

[proud bobcat]

This released on openrouter

It’s apparently fine

[lucid geyser]

He’s the only person who still says Asperger’s

[torn mantle]

whats the deal about this model? is it uncensored or what

[proud bobcat]

Really good fast image model that is nearly on par with frontier models

And it’s easy as hell to run

Only 6B

[solar hollow]

elon probably destroyed his cognitive abilities with excessive drug abuse

he just got dumber

[lucid geyser]

Yeah I think cause it’s gotta be like sorted

[lucid geyser]

Well I don’t think ever a genius but smart

[thorny schooner]

No that issue has been there all for different browsers the problem is the images themselves are either inconsistent or for the ones that disappear after you select them very slow to appear again wasting time until it becomes useless or r the fact it becomes hypersensive activating in every other prompt or the worst case forcing me to do that inconsistent verifying for eachprompt which is unfortunately often yeah so I don't think I need to tell you with it inconsistent prompt and hypersensitive one it becomes a nightmare so I just gave up on it waiting for the reset

[solar hollow]

smarter than now atleast

[empty stump]

gemini 3 pro is not imo level?

[lucid geyser]

Bro just tweets about immigrants all day

[solar hollow]

and worse

[proud bobcat]

It’s so sad that he screwed grok 4.1 personality

[lucid geyser]

Like he actually cares about protecting people

[solar hollow]

he used to care about more important things

[thorny schooner]

( i have tried on Chrome which I can't even sign in in anymore on there i have tried brave which is a whole nother issue because of speech and text says not work on there and etc)

[solar hollow]

his brain is puuf, gone

[lucid geyser]

Helping people? Maybe not…

Saying we’re gonna build a colony on mars

[torn mantle]

longer context = more compute/training = massive gpu hours for little gain for a mid size model

like i told you

[lucid geyser]

Isn’t exactly that

[solar hollow]

generally not being as much of an ego maniac like he is now

[torn mantle]

alibaba are working on 10000000000 product

[proud bobcat]

Fr

I loved grok 4 fast

Goated model

One of the best

[lucid geyser]

Retweeting he just wants children to be safe again

[proud bobcat]

I was there

Grok 4 was benchmaxxed

[lucid geyser]

Maybe he couldn’t end world hunger forever but

He could certainly save many children

[proud bobcat]

But man the architecture is awesome

[lucid geyser]

Grok 5 is agi he says

[torn mantle]

@deep adder also 32k fits just perfectly with 20vram if it was native 128k you woulkd need an extra 10-20gb of memory just for coding

if youreally care about running it locally

yea

bruh

ive seen it on way less gpus

[proud bobcat]

Grok 5 is going to be 5000 gazillion parameters agi

Trust

[lucid geyser]

5090 to play r o blox

[torn mantle]

wild is gone

you banned him

why

[proud bobcat]

Huh

Who was banned

[torn mantle]

grok 5 will be so bad

lol

i dont trust xai or elon at all

[proud bobcat]

Grok 5 will either be absolute peak or benchmaxxed

They do awesome jobs on efficiency

[torn mantle]

its not about being small

[proud bobcat]

Grok 4 fast is just goated

[torn mantle]

deepseek had way less staff

xai culture is messed up with elon ideology

[lucid geyser]

They hire Chinese people at all the us companies

[torn mantle]

like culture wise which is the base of a company is messed up

from the roots

you need like radical changes

[lucid geyser]

Plenty of Chinese phds in the us from like tsinghua

[torn mantle]

this

...

like whyyyyyyyy

[lucid geyser]

Just to get last place

[torn mantle]

why would you do that lol

[lucid geyser]

Saw that post

Of the guy

In the Tesla

Well it’s never really a flex to work 36 hours

It’s kinda just sad

[torn mantle]

they do with 4.1

[lucid geyser]

Bro Elon flipped multiple times

He literallly called him a pr ed

Base salary as an ai researcher?

[torn mantle]

@deep adder there was this research shared the other day, idk if its in this channel or the other one about compression tokens/text like in .zip files and conserving like 90% accuracy

but its so so dumb

like soooooooooo dumb

[lucid geyser]

Like are these ai researchers that got paid 100m really that much better than 100 1m researchers

[solar hollow]

its insanely huge at the big labs

the highest salaries on the planet potentially

[lucid geyser]

Like would they hire Terrence Tao even though he’s not in AI

[torn mantle]

oai was less stressful a year ago

but now

[solar hollow]

the lowest is probably a million in a year

[torn mantle]

i cant say the same

[lucid geyser]

U don’t think he could get hired just cause his math talent and intelligence?

[solar hollow]

but most are much higher

[lucid geyser]

Also he’s probably done some cs and ai

[solar hollow]

of course, im just saying

also phd in other fields make nothing compared to that

[lucid geyser]

Millionaire =/ salary of a million

[native yarrow]

gemini is INSANE at lua coding oh my

[torn mantle]

many key oai workers were poached by different AI labs, and many AI labs have already caught up

like dont bother

couldnt find the article

but it was stupid

so many silly mistakes

[solar hollow]

salary is something entirely different though

[lucid geyser]

R o bloxian

[torn mantle]

https://x.com/_rajanagarwal/status/1993373995903947139

rajan agarwal (@_rajanagarwal)

in order to have research agents that can run for days, we need context compaction

i used RL to have LLMs naturally learn their own 10x compression! Qwen learned to pack more info per token (ie use Mandarin tokens, prune text)

read the technical blog: https://t.co/3pzRvt4zvr

X · Nov 25, 2025, 5:39 PM

this one

[native yarrow]

its lowkey fun

like

[torn mantle]

i used RL to have LLMs naturally learn their own 10x compression! Qwen learned to pack more info per token (ie use Mandarin tokens, prune text)

[lucid geyser]

Yes I’m sure the average phd is making well over a million…

[native yarrow]

just one simple prompt "particle accelerator" and it made a complex structure that works

[torn mantle]

it looks cool but its stupid

lol

like the model using other langs to pack more info?

seems like a cool idea

[lucid geyser]

Just drop the brainrot game

[torn mantle]

but its more like reward hacking

[native yarrow]

on it :P

[torn mantle]

yea i saw that its 6B

but dunno about the quality

i guess for its size its good ?

the idea is fine

but the execution is just not it

constraining the model from the get go to 90% accuracy instead of dynamic compression

30% -> 40% -> 50% ...

it could've learned way more cool stuff

[lucid geyser]

Terrence Tao couldn’t even get $100k a year 🗣️

[torn mantle]

yea

now that you mention it

there are some mathematicians working at xai with that salary

you guys remember that troll xai guy

i dont remember his name but he was always on my feed page

[lucid geyser]

No

[torn mantle]

the guy with the hat

omg he was so annoying

i need to know his name

https://x.com/TheGregYang

Greg Yang (@TheGregYang) on X

xai cofounder.
make america grok again.
make x great again.

X

found him

I am a mathematician at xAI. Previously I was a researcher at Microsoft Research.

[lucid geyser]

Probably a much worse mathematician than Terrence Tao

But he did cs

[torn mantle]

thats my point

he was trolling non stop

i had to block him

he was the one that started 'i dont sleep, i work 24/7 at xai office'

and 'make grok great again'

[lucid geyser]

Yeah there’s like alot of math involved

[proud bobcat]

It’s like flux 2 schnell before flux 2 schnell

Ig

[proud bobcat]

Mechahitler

[torn mantle]

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

i call this one of the worst releases pr ever

why would you post it if you are not giving access to anyone

its so stupid

they just killed the hype with that

runway 4.5 seems on par with veo 3 but their release strategy is so bad

@echo aurora

[lucid geyser]

I’m just curious on the value of a mathematician to ai

They censored it

[torn mantle]

if sora falls so will all other ai vid gen

[lucid geyser]

Veo 4 soon

[torn mantle]

bro

[lost moon]

@echo aurora I tried to open a new conversation window, but I can't open it anymore.it has no answer. Is there a limit now?

[torn mantle]

who cares about tiktok

[lucid geyser]

Apparently people are testing it

[torn mantle]

tiktok has their own issues

lol no its not

[lucid geyser]

It is best for control

Also animation

[torn mantle]

its just because it was uncensored (kinda)

[lucid geyser]

Recent update

[torn mantle]

like pro?

sora 2 pro?

havent tried it

[lucid geyser]

No

Small

[torn mantle]

never generated an ai vid gen in my life

[lucid geyser]

Helps fix that texture thing

[torn mantle]

maybe one time with veo 3 since i have the pro

i just dont have any use for it

[lucid geyser]

They’re not very good

[torn mantle]

im not a tiktok creator or trying to earn money from that

if its for fun maybe

but i agree

sora 2 generated way more viral videos

[lucid geyser]

I posted veo on TikTok before

When it was trending

With the yeti stuff

[cloud zinc]

gen 4.5 is 2 times better than veo 3?

[torn mantle]

where

are you seeing this

lol

[torn mantle]

it seems solid

[cloud zinc]

veo 3 is half

[torn mantle]

we still need more demos

lol no

look at the numbers

the graph is kinda misleading

[cloud zinc]

oooh

lmaoooo

[torn mantle]

xddddddd

[cloud zinc]

20 points is nothing

does it have audio?

[torn mantle]

yea

[cloud zinc]

https://x.com/runwayml/status/1995493451954053433#m

Runway (@runwayml)

The model excels at understanding and executing complex, sequenced instructions. You can specify detailed camera choreography, intricate scene compositions, precise timing of events and subtle atmospheric changes all within a single prompt.

We remain committed to making highly

X · Dec 1, 2025, 2:01 PM

why no audio

[lucid geyser]

What kinda lies is this

[torn mantle]

https://runwayml.com/research/introducing-runway-gen-4.5

demos

bottom page

[hazy kernel]

The account seems legit

[lucid geyser]

How is 3.0 higher than 3.1

[cloud zinc]

i am hearing no sound

[lucid geyser]

https://x.com/kvickart/status/1995512438809371049?s=46&t=iDcf2nE8xUHsV_LN_6wymg

kvick (@kvickart)

@runwayml Graph to scale

X · Dec 1, 2025, 3:16 PM

[cloud zinc]

cuz that ranking is fraud

sora 2 pro way better

[lucid geyser]

No audio is crazy

[cloud zinc]

yeah

[hazy kernel]

What is "artificial analysis text to video" means

[torn mantle]

lol this should be right

[hazy kernel]

It has same point

[lucid geyser]

Some benchmark

[torn mantle]

its with audio

[lucid geyser]

Is it actually better

[cloud zinc]

no its not

its only video, no audio

[hazy kernel]

The model excels at understanding and executing complex, sequenced instructions. You can specify detailed camera choreography, intricate scene compositions, precise timing of events and subtle atmospheric changes all within a single prompt.

We remain committed to making highly.
Omg I'm so oof asking such a dumb question

[torn mantle]

it has audio

also

[lucid geyser]

Does the chart include audio though

[torn mantle]

i think its below veo 3 and sora 2

[cloud zinc]

show proof, beside random sound effect

cuz its not native audio

[torn mantle]

im so lazy

[cloud zinc]

we dont need fake audio inserted

[torn mantle]

to search that up

but i heard it with audio and one of their dev confirmed that in some tweet

[cloud zinc]

its not the same thing

[torn mantle]

we will see ig

[cloud zinc]

he says this

but in none of the clips and demo, it doesnt have audio

[lucid geyser]

The benchmark

Has no audio

[cloud zinc]

yes

[lucid geyser]

But it isn’t specified to have no audio

[cloud zinc]

this is what their ceo says

[lucid geyser]

What is this

[cloud zinc]

we will see if its true if audio is native or not

[torn mantle]

well if he said its native

then its native

[lucid geyser]

Doesn’t know his own product

[torn mantle]

lol

[lucid geyser]

It 100% does not have audio

[cloud zinc]

but in none of the demo page, it had no audio

[lucid geyser]

Atleast the current release

[cloud zinc]

https://runwayml.com/research/introducing-runway-gen-4.5

[lucid geyser]

It would mention if it did

[cloud zinc]

ye 100%

[lucid geyser]

Especially since the previous didn’t

[torn mantle]

the main demo where they were narrating they did in some parts

but i guess its not that good

[lucid geyser]

Where did he comment that

[torn mantle]

thats why they didnt add demos with audio

[cloud zinc]

you would think that in the main page of 4.5, they would have audio included with video. but it doesnt have that

[lucid geyser]

Where were comments

[cloud zinc]

twitter/x

[cloud zinc]

https://x.com/c_valenzuelab/status/1995501140595863951#m

Cristóbal Valenzuela (@c_valenzuelab)

@Profuture_AI Yes

X · Dec 1, 2025, 2:32 PM

[lucid geyser]

Holy prompt

Prompt:
A young woman with straight blonde hair and a freckled complexion sits quietly, looking up through her lashes with a vulnerable expression

[torn mantle]

https://x.com/c_valenzuelab/status/1995665276931596630

Cristóbal Valenzuela (@c_valenzuelab)

We were working late last night to get everything ready for the announcement, a common Runway tradition by now. I saw early versions of this sizzle but never got to see the final cut until today. What a treat.

X · Dec 2, 2025, 1:24 AM

lol

even the ceo doesnt get to see everything

[proud bobcat]

Isn’t this supposed to be like

Ultra ultra good

[torn mantle]

https://x.com/JoelKwartler/status/1995530996960391612

Joel Kwartler (@JoelKwartler)

For nostalgia’s sake, here are the same prompts run on the world’s first available text-to-video model in 2023 (@runwayml Gen-2) and the world’s best current model (Runway Gen-4.5)

X · Dec 1, 2025, 4:30 PM

more demos

[cloud zinc]

it looks like veo 3.4

[lucid geyser]

Fraud

[proud bobcat]

But

It’s

It’s runway???

[cloud zinc]

yes

[torn mantle]

https://x.com/c_valenzuelab/status/1995568758518354276

Cristóbal Valenzuela (@c_valenzuelab)

I can finally share everything I made with 4.5 for the last few weeks :)

Time for some POV world explorations

X · Dec 1, 2025, 7:00 PM

oh it has audio

[lucid geyser]

No

[torn mantle]

if you listen closely you can hear it

not the music

[lucid geyser]

Check his videos that aren’t edited

Only the montages have audio

[cloud zinc]

then why they putting music over it so loud

let me hear the audio clearly

[torn mantle]

woah

the fire shot was cool

[cloud zinc]

it looks like its video to video, not text to video

[thorny cove]

how long does a chat rate limit usually ask?

[lucid geyser]

It looks terrible and has no audio https://x.com/tlakomy/status/1995664601170534557?s=46&t=iDcf2nE8xUHsV_LN_6wymg

Tomasz Łakomy (@tlakomy)

Gen-4.5 animal gens:

X · Dec 2, 2025, 1:21 AM

[thorny cove]

been waiting a while but i keep getting Something went wrong with this response, please try again.

[cloud zinc]

its cherrypick data

[thorny cove]

this happens on pretty much every model after enough chatting

[torn mantle]

https://x.com/tlakomy/status/1995607216569892939

Tomasz Łakomy (@tlakomy)

Gen-4.5 can be really spooky.

X · Dec 1, 2025, 9:33 PM

this looks straight up from a movie

i like it

[thorny cove]

is it like cookie clicker

do i have to retry 1000 times to unlock it working again?

[cloud zinc]

https://x.com/tlakomy/status/1995664601170534557#m

Tomasz Łakomy (@tlakomy)

Gen-4.5 animal gens:

X · Dec 2, 2025, 1:21 AM

[thorny cove]

or do i have to wait

[thorny cove]

ill post results and see if it actually works in a couple of hours

[lucid geyser]

https://x.com/runwayml/status/1991548978295455780?s=46&t=iDcf2nE8xUHsV_LN_6wymg

Runway (@runwayml)

Introducing Audio Nodes for Runway Workflows.

Now you can add various audio generation models directly to your custom workflows for even more creative control throughout your creative pipeline. All in one place.

These four new nodes include: Text to Speech, Text to SFX, Voice

X · Nov 20, 2025, 4:47 PM

Definitely not native

[cloud zinc]

thats not native audio

they prob using other ai to add audio sound effect

[quartz light]

guess what model

[native yarrow]

gemini

[quartz light]

yes!

[native yarrow]

uayayayayyayaa

[haughty forge]

I want deepseek speciale back 😭😭

[keen beacon]

Man

Let’s talk about real censorship Google is implementing

And I don’t mean no gore and I don’t mean no NSFW

I’m talking about historic individuals and language use of time periods and eras

It’s completely wrong & still fails to censor real harmful ai output as a consequence everyone has to suffer because of their mistrust of the users

For all the progress AI makes it takes two steps back in the wrong direction

https://youtu.be/MrwJgDHJJoE?si=b-MYapt3PtiUskk4

Prime example

GamerNexus gives me hope for consumer advocacy

So does lm arena lol 😝

You know it kinda blows me away to that as long as I’ve been here none of you guys or anyone in that matter has ever complained about AI hedging

[still fossil]

hm hi

[keen beacon]

https://youtu.be/9cGxewHh3bw?si=VreF6-4OHfVxbFkt

[hazy forge]

where is speciale

[keen beacon]

Kling 01 model is nice

[still fossil]

can anyone help me please

[keen beacon]

Sure wats the issue is

[still fossil]

I m trying to generate image on lmarena image section using Google gemini 3 pro
it just fail all time and also estenguish my credit
and don't even give and out put
what to do to get an output
I have tried every way

[keen beacon]

So you’re saying every time you try to use Gemini pro image model directly that garrulous of your prompting or anything you do no image gets generated?

[still fossil]

it's says something went wrong with the responce

and fails

and that's not happening it generates infinitly

that's hour and still don't give anything

[keen beacon]

Well, first of all there is a rate limit. I don’t know how many number of images from what I understand it’s possibly for maybe a little more?

[still fossil]

ik

[keen beacon]

It takes a little longer to generate too, so you have to kind of be patient with it. If it doesn’t work, try to start a new chat.

[still fossil]

but after 50 mins it gets over

[keen beacon]

Yeah, I know what you mean

[still fossil]

I cant even generate a single image

[keen beacon]

Try clearing your browser

[still fossil]

I tried new chats

[keen beacon]

Let me test it out right now one

[still fossil]

should I give you my prompt and image ?

come dm

[keen beacon]

Sure

Just if you don’t mind, just share it here unless it’s something personal

[still fossil]

check maybe

uplode both images

to get output btw I gave in dm

[keen beacon]

Ok

Are you using a VPN by any chance?

I’ll teach you a better way where you don’t need Gemini pro for what you’re trying to do

[cloud zinc]

whats kling o1

[still fossil]

nope

pure wifi

idk what's happening

[keen beacon]

U can’t make ur img in lm arena using Gemini pro for 16:9

The output is 1:1 I think I’m not sure. Let me check one second.

Nvm I’m wrong

[still fossil]

nah

the 1st image I gave is generated by gemini also

it's 16:9 out put

but maybe my prompt is not correct

[keen beacon]

Ok

Now we gotta add the character in the center

[still fossil]

I think I'll use blender instead

[waxen fern]

Is LMArena still implementing rate limits??

[still fossil]

it's fussy

yh there is

9 images

50 mins and then again

[waxen fern]

It's so annoying

[keen beacon]

One second it keeps messing up

[waxen fern]

I gotta ask pineapple

[keen beacon]

I’m not using Gemini pro

[waxen fern]

@echo aurora Remove all rate limits and make LMArena truly unlimited

[keen beacon]

But I do have Gemini pro so I can do it for you officially one second

Lm keeps flagging

Damn wtf

Ok got it

[still fossil]

what's better then gemini pro

[keen beacon]

Let me remove that watermark

I don’t know I don’t like nano sometimes it makes a composition look weird

Like the height

Looks very off to me

[still fossil]

hm now let me generate it

[keen beacon]

Ya try ur image

[keen beacon]

New Kling model

[keen beacon]

I don’t get why it keeps flagging your image, bro sometimes

Isn’t this from like Minecraft or something?

[waxen fern]

Rate limits for text arena???

[keen beacon]

Gaurd rails

[cloud zinc]

is it video editing model or text-to-video

[still fossil]

idk know why

[keen beacon]

Same

[waxen fern]

Rate limits for text generation????

[keen beacon]

I don’t know dude the composition is just weird

[still fossil]

I think my image is open sourced that's the problem lemme fix that

wait

gehlo come dm

[keen beacon]

I think it’s that fake pixelated blood

Odd

Yeah, I’m not really good with composition either

[viscid cloak]

Anyone encountered endless “something went wrong” in Battle?

[keen beacon]

Ya

[viscid cloak]

Like there should not be rate limit in Battle, and I have not triggered a single verification. But the endless error message seems like a hidden rate limit control.

[keen beacon]

Why shouldn’t there be regular limits?

[waxen fern]

Number of limits in direct chat????

[keen beacon]

It it’s design is not for infinite generation lol

I mean, how long and how much time and how much prompts do you really need to really test the models after a certain point you kind of have an idea what the model could do? lol

[waxen fern]

Number of limits in direct chat for text????

[keen beacon]

I hear what you’re saying and I feel your frustration so I don’t want you to think I’m

[viscid cloak]

You certain everyone using the platform is for the sake of “testing model for the goodness of the platforms” lmfao

[keen beacon]

No, that’s why I said I get it

[waxen fern]

Direct chat for text limit numbers pls

[keen beacon]

I’m just trying to see both sides of the situation so I can understand better the dynamics that play

Two opposing forces trying to achieve two separate objectives 🤣

[viscid cloak]

If platform expected everyone is so benevolently selfless, the platform wouldn’t get that much information for learning and improving. So limiting Battle done no good to both sides.

[keen beacon]

I hear ya

[waxen fern]

Number of chat a day for text rate limits???

[keen beacon]

It just puts lm arena in kind of a weird awkward situation

Both on the expectation side, and on the other hand, fulfilling their obligations whatever they may be

[viscid cloak]

I think it varies from models to models, and refresh every hour at least for image gen

[keen beacon]

That’s actually more interesting case here

Is the rate limit on specific models

[waxen fern]

It's just so annoying and it needs to stop

[keen beacon]

lol this is exactly the mentality I’m talking about

Which is unfortunate cause I understand and I’m part of that mentality also

There needs to be a subsidized variant of something like lm arena

[viscid cloak]

Now the error message in Battle is quite ridiculous. I used my phone and turned off WiFi, using cellular network. Boom, generating smoothly.

Really Arena? Doing that kind of funny ip lock trick🤣

[keen beacon]

Could be a browser could be something on their end. It’s hard to say a lot of people seemed to encounter this.

That’ll be crazy if that’s what was going on

[viscid cloak]

Ikr

[keen beacon]

But it also means if that is the case then that’s a lot of generations

[waxen fern]

Number of limits for deepseek 3.2 thinking in LMArena??

[viscid cloak]

I mean I haven’t triggered a single ReCAPTCHA. Blame the system lol

[keen beacon]

Well, that has nothing to do with them. That’s more of a google

I hardly hit any either

But I use a lot of Google services

[waxen fern]

Help

[viscid cloak]

I mean shadow ban is just lame. Could’ve just told me like cloudfare.

[viscid cloak]

We don’t know. Ask mod or wait.

[keen beacon]

Well, I think we’ll jump in the horse here

We don’t know if that’s a thing shadow banning

It’s like four or five messages. I can test it right now and tell you the exact number one second.