general

[quartz pike]

but i was asking for like an ACTUAL ai image model that exists on lmarena

im stoopit

and lazy

so me no do dat

[wild sapphire]

this is going wild now

only limit is imagination

[mental scaffold]

Hello, here to check out the videos

[wild sapphire]

yeah and the wild imagination worldwide

[quartz pike]

Chat is ts tuff?

Duh its ai generated. this is lmarena 😭

correct. but personally i find the seccond one a bit funny

Bro just casually lets go

and f*cking dies 😭

lol

btw can u vpte

i wanna see what the models are

that made it

[fervent pebble]

Hello. I'm here because i'm excited at how fast Ai Image generation is evolving and looking forward to seeing how it develops.

[quartz pike]

😭

[alpine coral]

nice one - that does suggest some kind of RAG/internet access almost surely

is nightride a strong model otherwise? like aside from up to date info

@echo aurora i had scroll through so many images and greetings before finding discussion about models in the arena (which is what i liked about the server in the past).. perhaps there's a way to separate out like general chatter from discussion / speculation about anon models (the juicy stuff ha)

[neon idol]

Sorry for the image in italian just use google lens for translate

[solid brook]

intresting

ask for knowledge cutoff this time

[neon idol]

Ok

[ocean vortex]

Huh it's still R1 (newest ver)

[ocean vortex]

But ig it's old and people are bored wanting smth new lol

Deepseek V3.1 was a non-release for the most part. You can now have no reasoning or reasoning that almost matches R1 within the same model. Ok great, moving on...

[rustic knot]

lmao rekt, what i do know is Qwen 3 has overtaken R1 in math, other fields idrk

[ocean vortex]

Actually looked at this and dunno anymore. I disagree with this 😠

I suppose qwen3 performs, but my IRL experience was somewhat differing from this

[rustic knot]

Qwen 3 is better at math and competitive programming

that's why its AA score is better

but ngl they did r1 dirty by that aime 2025

[ocean vortex]

Qwen3 seemed to me like it's much easier to break and not nearly as reliable as R1

[rustic knot]

another thing to note is that it seems like ds keeps trying to use grpo while Qwen sswitched to gspo

[worthy sleet]

I need nano banana and I can use it for free from google aistudio. I can also do it from lmarena directchat. Is it preferable for lmarena doing it there because they can use those prompt or is it more of an expense for them?

[ocean vortex]

R1 is just reliable though and shines in several metrics on unseen data (new benchmarks), whereas Qwen can quickly fall apart and seems more benchmaxxed tbh

Also smaller

[rustic knot]

which new benchmarks?

[ocean vortex]

uh there was creative writing one

[rustic knot]

this is also a new bench lol

[ocean vortex]

that became a thing after that model was released

[rustic knot]

kimi does the best.on that and it doesn't even use test-time compute

[ocean vortex]

things like this: https://www.reddit.com/r/LocalLLaMA/comments/1ieooqe/deepseek_r1_takes_1_overall_on_a_creative_short/

[rustic knot]

that seems like from a long time ago

[ocean vortex]

Like it really seems that model had good fine-tuning. It is performing on things they didn't explicity train for

[ocean vortex]

Yeah but still kinda relevant

[rustic knot]

like kimi overtook everyone on the eq creative bench

it's basically the best writer, even gpt 6 (underscore) agrees

[ocean vortex]

As it should, it's the biggest of those 3

[rustic knot]

there seems to be a tradeoff b/w what certain RL algorithms are good for, GRPO might be better for writing where the rewards are uncertain while GSPO is better when there is only 1 right answer

[ocean vortex]

They can't do RL training (reasoning) though it seems 🗿

kimi has a lot of potential for it

[rustic knot]

have u seen the GSPO paper?

[keen beacon]

gspo is just better overall

vs grpo

[rustic knot]

in grpo, there is a higher variance for answers which might be better for writing purposes. Trentk used it to turn a small Qwen model in mechahitler

more creative yk

[keen beacon]

not too sure about that but in my experience gspo is just better

[rustic knot]

cuz ur probably using it in technical cases where there is only 1 right answer for the most part

@ocean vortex you gone buddy? xD

[leaden sun]

Anthropic is losing the game it seems, they're not racing to AGI, but to establish as a coding product company?

[white hatch]

will there be added support for uploading files other than images?

[ocean vortex]

Nah, why?

No opinion on this. I'm more interested in how it turns out in practice. But haven't looked into technical details of this enough to comment tbh

Plus I would say it's a bad idea to generalize like this. I think there's more than 1 way to train a good performing model

[keen beacon]

Sure.

Until you remember that GPT-5 is the best model overall, that is, most likely to give a good answer to ANY prompt.

Regarding if Qwen is benchmaxxed or not - we quite literally do not have appropriate benchmarks to figure it out

In psychometrics for example, tests are created to distinguish between different broad abilities, which are measured with different subtests

Let's say that we did an exploratory factor analysis on a LLM and figured out that it has a strength in broad factor that is responsible for creative writing, and each other thing this broad ability accounts for

So it's going to be great at benchmarks that measure this broad ability

And if a model does somehow bad at all benchmarks at the same time except those it was measured at, it is an evidence of benchmaxxxing

But we do not have extensive psychometrically valid benchmarks for LLM so far, so...

[summer cove]

hi

[leaden sun]

Those psychometric are just theoretical frameworks that don’t necessarily reflect the reality and can be easily gamed, those top consultant firms are using psychometric for their assessments and people know very well how to game the system, same will be for LLMs

[keen beacon]

As a person with some background in quantitative psychology, nope.

[leaden sun]

Ok, that speaks volume about your background

[tawny kite]

HEY EVERYONE, I AM NEW HERE,

[sullen fern]

hi

[languid crescent]

mah chat history is gone 😭

[fervent tangle]

GUYS

GOOGLE RELEASED NANO BANANA

ITS GEMINI 2.5 FLASH

on google ai studio

[keen beacon]

OMG NANO BANANA

[sullen fern]

I have a question

[fervent tangle]

finally imma use nano banana for free

[sullen fern]

How can i change the aspect ratio on lmarena ?

[fervent tangle]

on google ai studio

[fervent tangle]

u cant

[keen beacon]

its also available on lmarena direct chat btw if u run into ur limits there

[sullen fern]

damn i wanted to make a horizontal image

[fervent tangle]

wydm

[keen beacon]

aistudio has generation limits for nano banana

[fervent tangle]

on google ai studio its better, u can choose it right away and edit stuff with it

[fervent tangle]

really

[keen beacon]

yea u have X amount of requests

[fervent tangle]

10 amount?

[keen beacon]

im not sure about the specifics, they dont document it and it changes

[fervent tangle]

idk i been using it for a while now

[fervent tangle]

it also has limit on lmarena

[keen beacon]

yea im just saying if u do reach the high limit

[fervent tangle]

i can just change my google account

[keen beacon]

🤷 im just saying it is available on lmarena direct chat if you want. no login required there to get extra use

[fervent tangle]

yea ig

imma try it if i reach the limit

also did the quality drop?

after they released it?

or am i tripping

[keen beacon]

personally i dont see much of a difference/if at all compared to when it was anon but i barely used it

some people have said it was nerfed though

[fervent tangle]

yea it used to be much clearer

[south vigil]

What do you gets think is better at coding; GPT 5 high or claude opus 4.1/sonnet4

[white hatch]

gpt for architecture, claude for coding

[fervent tangle]

4.1 opus for sure, but it costs alot

try 4 sonnet if u dont want alot of costs

[south vigil]

claude opus 4.1 was unable to help yesterday fix a simple api concurrency limit issue, had to take over, and do it the old fashioned way. i use opus 4.1 often and it's decent but was thinking about switching to gpt5 high though because it's a better reasoning model.

[fervent tangle]

did gpt5 fix it for u?

[south vigil]

didn't try, should've, i just fixed it myself, i'll see if gpt5 can try optimise it today

because claude opus 4.1 completely failed

[fervent tangle]

google ofc, easily

[drifting crow]

Nvidia

[sonic bear]

golden hour backlight,soft rim light on hair,clothes & subjects,lens flare sunlight,create shadow on ground

[topaz bay]

wen it comes to text model only, then XAI,openai and Anthropic have a big chance

Only problem is that the west only does close sourced

[wet pewter]

Google farming your data through androweed

[topaz bay]

I hope china keeps doing good open source, so that it's a fair market

[topaz bay]

ai doesnt need data from people anymore, it self trains and does a feedback loop, only picture and video needs new data, that's why youtube is so good at it

[fervent tangle]

google already has everyone's data lmfao

thats how they made their AI models good

[quartz pike]

tuff or nah. vote ples

yall just asking. are the devs planning to add video arena to the website?

[stiff blaze]

Make him run like super mario 3

[fervent tangle]

brodie i cant vote its fowarded

[quartz pike]

click the thing at the bottom

to go to le channel

[neon musk]

How generate image to video on this LMARENA to ratio 9:16 or portrait??

[quartz pike]

so you can vote

[fervent tangle]

nvm i did

the right one is veo3

but it doesnt have audio..

[quartz pike]

oh?

[fervent tangle]

tbh veo 3 looks good

[quartz pike]

dis one is even better but idk wtf is going on with the left one

The ai on the left crashed the f out

[exotic tartan]

why do i see only video-arena-1? aren't there supposed to be more?

[grand valve]

hello friends....how to use google flast 2.5 i.e. nano banana

[solid brook]

gemini app lmarena google ai studio

[keen beacon]

https://www.lesswrong.com/posts/aFW63qvHxDxg3J8ks/nobody-is-doing-ai-benchmarking-right

[fervent tangle]

google ai studio

[keen beacon]

Another interesting writeup from these guys:
https://www.lesswrong.com/posts/SJARqiCTcqyGPSKJw/anthropic-is-going-all-in-on-ability-without-intelligence

In short, while Claude models seem to be great at coding, they severely lack in GENERAL ability

[sacred wharf]

hi cant wait to see whats posssible here

[keen beacon]

They are great at coding - but when it comes to anything but coding, they are suddenly so behind the frontier it is even difficult to consider them a frontier model anymore

I've written before about METR's time horizon benchmark. While I consider it a valuable benchmark, it doesn't measure exactly what it's trying to. In order to only measure a model's time horizon, a benchmark would need to only vary the task length. Instead, the short tasks tend to be easy and not specialized-knowledge-dependent (i.e. doing a web search), whereas the long ones tend to require far greater specialized knowledge and intelligence/problem-solving (i.e. ML coding tasks). So it winds up measuring an amalgamation of time horizon, coding ability, ML knowledge, problem-solving, etc. Very roughly speaking, it's a decent benchmark of (partly narrow) abilities useful for AI automation of AI progress.

[quartz pike]

Hey anyone here know code?

Cause just asking. Wich assistant did better since i dont understand code:

Assistant a:

Assistant B:

[keen beacon]

Bruh 💀

[quartz pike]

?

im a dummy

me no know

or how to put code somewhere and start it

[astral kayak]

hello

[keen beacon]

Did you know that you could just GPT-5 about it...

[echo aurora]

This is a bit tricky as creating new channels doesn't always lead to community members using them for the topic of the channel, we're currently experiencing this in the other existing channels. What I've seen other communities do, and what I'm encouraging everyone here to do, is if you'd like to have a conversation stay on a specific topic -> create a thread for it. It won't be great for new people to join in on that conversation (as it'll get pushed up and off the general text area), but it does establish a more "private" place for members to discuss something specific.

[quartz pike]

Ik i can just ask gpt 5. its a great coding model and at anything. But... I wanna see how other ais compare

oh wait im a dumbfuc

i just realised what you meant

ima just shut tf up 😭

[keen beacon]

Bro 💀

[quartz pike]

Chill english aint my first language 😭

im greek

Ελλαδα!!!!

-# that says greece in greek

[feral python]

aaaaaaaaaaaaaa

[tribal peak]

Hello

[torn mantle]

AAAAAAAAAAAAAAAAAAAAAH

WAAAAAAAAAAAAAH

😱

[feral python]

NAAAAAAAAAAAAAAAAAA

[feral python]

Hi

[keen beacon]

Neither is mine

[quartz pike]

cool

[unborn spoke]

Hello this Forum is so cool

[echo aurora]

welcome welcome 👋

[zealous sparrow]

MAI-1 is the stupidest model name i heard

[humble sonnet]

What is mai-1 ?

[zealous sparrow]

Medium Aritifical Inteligence?? I dont know the short say.

[keen beacon]

Lmaoo

[zealous sparrow]

Nevermind its probably just Microsoft Artifical Inteligence..

[rustic knot]

oai stream rn (it's incredibly bad)

[ember whale]

hi

[quartz pike]

Yall how good is mai

[keen beacon]

No clue

[quartz pike]

i expect it to be trash

since well... Microsoft

microsoft's models suck ass

[zealous sparrow]

13 on text arena thats what we know rn

I will test webdev

prob goin to be so bad

[quartz pike]

hmmm

how do you test webdev?

js asking

[zealous sparrow]

personally use codepen

for html

too lazy to open vsc

[quartz pike]

oh

me stupid

what is codepen

and how do i use it

[zealous sparrow]

online html editor

[quartz pike]

And can you pls compare it to gpt 5 chat?

[zealous sparrow]

it has no chance

[quartz pike]

yes

but its worth it

lol

[zealous sparrow]

it takes a while to generate a simple html game

[hollow imp]

Gpt 5 high vs this new model

[zealous sparrow]

100-0

[quartz pike]

hell no 😭 it stands no chance

Maybe againt gpt 5 chat or nano

[hollow imp]

So the new model is bs

I'm not even gonna try it

[quartz pike]

but high there is no chance

[hollow imp]

Then

[quartz pike]

we arent sure

But since its microsoft

its safe to assume its dogsh1t

[zealous sparrow]

im doing the high vs mai comparison
Topic: Single html shooter

simple

[hollow imp]

Bro

They were using o1

[zealous sparrow]

GPT 5 high is already done thinking and mai is still thinking

[hollow imp]

And now they merged with github and released this sh it

[quartz pike]

bruh

no wonder it was ass

o1 is respectfully. DOGSH1T

[zealous sparrow]

its obvious who is winning.

[quartz pike]

left lol

[hollow imp]

I'm saying before they released this

[quartz pike]

mai1 thinks its einstein

[hollow imp]

They were using o1

[quartz pike]

thinking that much

[civic flame]

mai-1 has been in development since early 2024 btw

☠️

[zealous sparrow]

unless mai servers died

[quartz pike]

DAMN

ALR ITS LOOSING

EARLY 2024

THATS A 1 YEAR OLD MODEL

SONNET 3.5 RELEASED 1 YEAR AGO

[zealous sparrow]

btw

[hollow imp]

🤮

[zealous sparrow]

mai was designed to be a gpt 4 competitor

[civic flame]

4 turbo*

[glacial mulch]

lmao what is mai

[civic flame]

lol

[keen beacon]

mid ai

[balmy mist]

so is openai speech on seasame now?

[zealous sparrow]

prob short for Microsoft Artificial Inteligence

[keen beacon]

#announcements message

ah

you meant that

[whole swallow]

This is on web lmarena or normal one?

[zealous sparrow]

lmarena.ai

[whole swallow]

Aight

[zealous sparrow]

yeah mai breaks if you tell it to webdev

its a generating loophole

[civic flame]

literally agi

[zealous sparrow]

been like 5 min

if it ever finishes and the game is worse than GPT-5s

i mean i told it to make a cafe site

it finished...

i am not going to lie

the site aint half bad

It found images.

[supple vector]

Lmarena free API when????

day one of asking for LMArena free ai api

[zealous sparrow]

my thoughts on mai

the model is slow but it makes decent sites

if LMArena errors at you during gen you are in for a lot of waiting

LMarena redid its captcha?

[cyan zodiac]

what could it possibly be writing

[prime mulch]

What is the result bro

[zealous sparrow]

cant quite host it so ill show in a few images

theres also a contact us section

the generation takes so long tho

im pretty sure the next prompt i gave it which was add JS is already generatin 10mins

i did say a lot of js

it errored 🙁

[misty vault]

@deep adder mai-1-preview is sydney from mcdonalds

[zealous sparrow]

mai-1 must have a very small context window

will error if you ask for too much

[golden ocean]

Omg its reaL

[supple vector]

@echo aurora 🍍 Mr pinapple I beg

[keen beacon]

Guys

I just joined

What's up with the hype

[keen beacon]

More like middling artificial intelligence hahaha lmao kill me pls

[echo aurora]

Could happen, can't say though I'm aware of these plans

[keen beacon]

what

[wintry tinsel]

No context

[zealous sparrow]

MAI is not that bad but also not that good

context window is prob small

will error if you are demanding

takes long to gen

[talkin webdev]

[unborn ocean]

is currently labelled as "exp", so might just be the not fully finished post training

[magic stag]

Who's the one with nailoong enojis?

[magic stag]

#announcements message

Saw this reaction in announcements

[keen beacon]

What's up with oAI streaming?

[ocean vortex]

gpt4o got updated lmao

They refuse to let that name go

[keen beacon]

why did they ever upgrade 4o when they have 5?

now we have to retest all these stupid benchmarks again

[ocean vortex]

oh wait I can't read

it's a previous model

it's now called gpt-realtime huh

[quartz pike]

mai*

[ocean vortex]

4o the text model was not actually updated ever since before gpt5 release

And now they renamed gpt4o-realtime into gpt-realtime, which makes sense actually

[vast fern]

what new features?

[ocean vortex]

Marginal update looks like. They just polished it and renamed I think

As for gpt4o text model... Current version of that is gpt5-chat. If they haven't renamed, it would be that.

[quartz pike]

any updates on mai?*

[ocean vortex]

it already cannibalized gpt4.1 so they held on to that name too long as is tbh

[vast fern]

it sucks

[ocean vortex]

"mai" ??

[keen beacon]

microsoft ai 1 😂

[quartz pike]

oh lol

[ocean vortex]

wtf 🗿

[ornate agate]

I thought that was a r1 finetune? is this one actually their own model now?

[keen beacon]

yes 🤣 🤣

[ocean vortex]

What is is exactly though... is it like a huge model or renamed Phi... 🧐

[keen beacon]

theinformation had an article about in may 2024

lol

[ocean vortex]

Lack of any metrics at all is not very inspiring

[ornate agate]

"MAI-1-preview is an in-house mixture-of-experts model, pre-trained and post-trained on ~15,000 NVIDIA H100 GPUs. " :\

it better be really good if it used 10x the GPUs of DeepSeek

[ocean vortex]

It probably isn't if they can't sell it to public convincingly lol

[ocean vortex]

First prompt impression it's thinking for ages and can't decode things nowhere near as well as gpt5-mini

[ornate agate]

I asked it an aime question and its still "generating..."

[ornate agate]

theres literally no way anyone's gonna click on random audios posted with no words

[ocean vortex]

go into voice channel

[echo aurora]

I do

[ornate agate]

If I was a mod here I would run discord in a VM with all the wierd stuff people spam onto this channel 😄

[ocean vortex]

correct decoding:

sh'it decoding:

[ornate agate]

it crashed on trying to do an aime question

lets try again

[zinc ore]

Bruh 14 day ago poll

[ocean vortex]

how did it go?

[ornate agate]

generating...

[solid brook]

bruh 2 of them really chose elon musk....

[ornate agate]

whats the text string you used as text? I wonder if a local model can do it

[ocean vortex]

PzA1MSBuYWh0IHJlbGxhbXMgdHViIDAyMSBuYWh0IHJlZ2dpYiBzaSBlcmF1cXMgZXNvaHcgcmVnZXRuaSB0c2VsbGFtcyBlaHQgc2kgdGFoVw==

[wooden salmon]

Oh thx

[ocean vortex]

MAI doesn't seem to actually be horrendous, but I really do not think it's gonna challenge the current best ones. Maybe og R1 level.

[neon idol]

I am one of these sigma 🗿🔥

[gray junco]

hello

[hardy lion]

leaderboard has og r1 at 1395, mai-1-preview at 1402 and r1-0528 at 1417, so it's pretty close to between them

[ornate agate]

gpt-oss can decode this.

[unborn ocean]

decoding is a really boring test honestly

only openai ai is good at it

does not really reflect much

[ornate agate]

it seems that other AIs fail the base64 decode itself, for some reason

[ocean vortex]

Boring test is a good test. You can't overfit or contaminate for this

There are endless possible prompts

[unborn ocean]

you can obviously rl your model on this as a task in post training
openai might be doing (based on their performance in the area)

very easy to implement

not overfit in the traditional sense, yet still optimised and not reflective of overall reasoning power

[ocean vortex]

This one is one of the tougher ones. But in general other models can do well too

It's reversed and then encoded. So next level. But gpt5-mini can do it reliably so 👀

[unborn ocean]

in my experience (which admittedly is from a couple of months, if not halve a year ago) openai is just stronger than expected on this

[ocean vortex]

Not really. This kinda converts to many other reasoning tasks too. But they didn't specifically do this for sure as with tools it will solve it in seconds

[unborn ocean]

the same on coding competitions, which does not mean the overfit in the traditional sense here as well, it just means they have trained (rl'ed) a lot on the format

[unborn ocean]

a year ago every lab was doing the same thing for basic number calculations (e.g. 1231*2333), so why should they not do that?

[ocean vortex]

That's how you should be doing it. This means model is able to generalise and you have improved performance.

[ocean vortex]

why should they is the better question. There's no such benchmark that they quoted which would test it...

[sonic tendon]

i don't think that that's true? qwen and DS can do it fine, as well as 2.5 pro and flash

[unborn ocean]

yes, but using the same structured environment they rl'ed in will result in the model looking better than they are in similar tasks (that they should actually learn, like in the coding comp example the actual coding vs the short-ish coding competitions they are rl'ing on)

[ocean vortex]

They just didn't. It makes no sense to do it for them. Target something specifically you can't even promote or sell to the public? Makes no sense at all.

Unless they did smth similar to improve the reasoning performance in general

but then it just kills your entire point

lol

[unborn ocean]

the idea is not that it provides benefit to rl on this, but that it is an easy way to implement rlvr + curriculum learning to get better perf in other areas

[ornate agate]

actually training specifically on this is exactly the sort of thing I imagine OpenAI to do.

[unborn ocean]

that is obviously the point here, duh

[sonic tendon]

this seems fairly trivial to test

[ocean vortex]

They don't do random things they can't sell if it doesn't improve measurable performance

[unborn ocean]

yes would also be interested, the last time i did this was in the later o1 days (so i think o3-mini just released)

and openai was king back then

and i felt like they did use it for training

[ocean vortex]

You are essentially saying "yeah it's better but only because they focused on making it better". Or like "yeah it is better on this thing but only because they focused on this specific thing they didn't advertise anywhere at all and there's no benefit at all". None of these statements make any sense whatsover tbh

Metrics do though. No one has time to dig through this nonense of what people are doing lol

They would degrade what actually matters to them if they were training so randomly

It's just one of the things that highlighs where OpenAI is clearly ahead atm

[whole wagon]

Microsoft made their own LLM Top 10 anime betrayals

[sonic tendon]

i really don't think that openAI is particularly ahead in this

[sonic tendon]

they made phi

and an R1 finetune I think?

[ocean vortex]

They kinda are measurably ahead. Other models can do it, but most of the time accuracy is visibly worse

[whole wagon]

This is the first major llm release out of Microsoft

[unborn ocean]

i dont think you are getting my point:
they did not focus on this to sell the 10 nerds who try it out on their model, they probably used it in their rl training process to enhance other areas, why?:

• it is really easy to implement (no humans needed)
• one could suspect that the models would benefit from more concise and flawless reasoning in other areas aswell (bc one flaw results in broken output and the token usage indicates efficiency)
• it is really easy to implement curriculum learning on top of this (scaling the difficulty of the decoding problem)
• no real overfit

[sonic tendon]

https://huggingface.co/microsoft/MAI-DS-R1

[sonic tendon]

maybe

[ocean vortex]

Well then if they used it to make the model better (unlikely, but ok, let's assume that) and arrive at the current SOTA to improve it general, what is the problem?

[whole wagon]

https://microsoft.ai/news/two-new-in-house-models/

[ocean vortex]

Well almost everyone is doing math or cyphers with reasoning

for training

[unborn ocean]

i don't have a problem with them doing it, my point is just that using the very environment they (potentially) used for training as a benchmark seems like a really dumb idea

[whole wagon]

I actually knew about this Microsoft ai for ages. They operated in secret as a redundancy for if openAI ever betrayed Microsoft

[ocean vortex]

And decoding is fundamentally just math

So yeah... 🤷‍♂️

Someone just does it better than others lol

[unborn ocean]

yeah they heavily scaled up, they always had some interesting people, but now they are all moving into the proprietary team

[whole wagon]

Will be interesting to see how fast Microsoft climbs to the frontier

Everyone's first LLM is bad

[sonic tendon]

people seem to be sloptimizing for LMArena

[unborn ocean]

amazon, lol

[ocean vortex]

wdym. RL training for reasoning pretty much evolves around math. That's how you can see deterministic results and eval reasoning traces

Like that's not news lol

[rustic knot]

the entirety of AI is math

[sonic tendon]

mistral is still a surprise to me

i always liked them for their terse responses

but then

[unborn ocean]

same, still mentally ignoring that they are n.2 without style control

can't explain why

[ocean vortex]

This is still math if you write the prompt correctly. Nothing to do with puzzles when we are assessing just decoding. If you do base62 there's gonna be a TON of math

[sonic tendon]

i feel like DS might be the first to top Google on the non style control leaderboard

since 2.5

[ocean vortex]

Every model knows how to do it

[unborn ocean]

yeah, if huawei gets their sh*t together (big if)

[ocean vortex]

but not every model can do the math part with good enough precision

disagree. It's simply testing how well a model can reason and for how long while still being able to complete the task and not go offtrack

If it gives up half-way through and hallucinates an answer - you now know the limits of it.

tools render it invalid test lol

But that's also why it is unlikely for AI labs to focus their training on it

It is not. ML models are not humans

Ever since reasoning became a thing, this is the way to test that (test-time compute) tbh

If it's allowed to reason as long as it needs to and it had good RL training, eventually it will arrive at the answer with high enough confidence

And generally... I don't think it should be surprise to anyone that OpenAI is leading on fine-tuning at this stage to be completely honest. Their model is not the biggest and nor do they have the most compute

As I've said, it makes no sense at all to do this UNLESS it improves performance for them elsewhere they can show. So it's kinda irrelevant?

Testing this is meant to show what it can do what is applicable to IRL tasks. I think this does it well. Regardless of if they found it to be the case by training for it or if they didn't train for it.

It is positively not possible for them to train for it "just because" if they hadn't found it improving performance on the things they can showcase and sell (published metrics). Like this is just ridiculous and totally not how things are done lol

[tall summit]

LMAO

[ocean vortex]

No I gave you the arguments lol

I find it unbelievable what mental gymnastics you are doing to turn some thing a model does well at into a negative. Coming up with some wild theories 🗿

Don't be THIS biased 👀

[whole wagon]

openAI does a bit of chess training I think. There's no way GPT5 is this good at it otherwise

[ocean vortex]

Dunno about chess... 🤔

maybe

lol what

You probably realise yourself this is nonsense

At least I would hope so

just stop it lol

example 236 from their RL gym ???

came up with the prompt myself, reversed it myself, encoded it myself. Have more prompts than just this one

well you are just spitting non-sense

🗿

My goal is to eval models regardless what lab was behind it. Believe it or not in my testing gpt-oss didn't do great overall at all

Even though I did have fairly high expectations

You think wrong...? Or maybe didn't read attentively

[whole wagon]

Gpt oss is not well rounded. Great in some things bad in others

[ocean vortex]

yeah it falls apart fairly quick

o4-mini considerably better model, like no comparison better, imo

[verbal nimbus]

It performed well on my prompt, but I rarely encounter the model.

[whole wagon]

o4 mini different price class. Compare to GPT5 nano

[ocean vortex]

they were selling the biggest OSS variant as the model equivalent to o4-mini though. Their goal was indeed comparable performance I believe

It wasn't meant to compete with nano

[whole wagon]

The small 20b model is just straight up bad. Like qwen is better lol

[ocean vortex]

Oh I mean the 120b one

[whole wagon]

Yeah ik

[ocean vortex]

The smaller ones are naturally worse still

[whole wagon]

I'm just saying the small one is basically useless lol

Qwen 30b a3b outperforms

[ocean vortex]

#general message

Other than that we have no way of knowing what exactly they have done. You can say this about everything

but at a certain point you just need to accept the facts

[ocean vortex]

yeah I can believe that. Personally I started with the biggest most capable one and lost interest fairly quickly lol

[whole wagon]

There are fine tunes that remove it iirc

[verbal nimbus]

OSS 120B is interesting, only 5B active parameters

[ornate agate]

those guys also usually degrade the model significantly, I prefer to just use another model really.

[wintry tinsel]

It operates using math although the raw data it is learning from is not necessarily math so data science is not purely math

[rancid knot]

Hello, I want to learn how to use Ai properly and change style of image and video

[errant condor]

Guys any way to generate 9:16 videos from an image?

[empty briar]

Hello, I am here to learn about the new AI tool with LMArena

[golden ocean]

true

[maiden fox]

helli, I love just you

[white hatch]

Any benchmarks where mini models are better than their big brother?

[keen beacon]

High-level: yes, this happens. Smaller models can beat their bigger sibling when:

They’re task-specialized (e.g., a 7B code-tuned model edging out a general 70B on coding benchmarks like HumanEval/MBPP).
The benchmark rewards concise instruction-following (some 7–8B instruct models have topped their family’s larger base models on MT-Bench–style evaluations).
There are tight context/latency budgets (short prompts, low-temperature, or limited tokens) where smaller models are less prone to over-elaboration.
The pipeline uses strong retrieval/tools, making the generator size less decisive.
Evaluation variance is high and differences are within noise; smaller models can win specific subsets even if they lose on average.

got it to google

[minor schooner]

Hi all fellow AI enabled self expression utilisers..! Here to compare different models abilities to bring hard sci-fi animations with a few very exact requirements for details expected to be present, and keep the consistency in check.

[errant rover]

Guys what's the best AI for upscaling right now ?

[bleak jewel]

Hi everyone

[rocky frigate]

Hey I'm new to here tell me how to add a prompt and generate a video. How to type the prompt?

[barren prairie]

Hi

[echo aurora]

More info can be found in #1397655624103493813 that should help!

[sonic tendon]

wait, when did this happen?

i don't think it's even on OR yet

no idea, I haven't seen it till now

[jovial wolf]

when i try https://web.lmarena.ai/
it gives me like ksx files that are a big headache to run and are buggy, rather than the quick .html files im used to getting from chatgpt's site. does anyone know more about this?

like the sandbox doesnt work, ever. maybe ill just try a different browser?

[urban mulch]

New here

[hardy lion]

my guess would be that it happened on or around 2025-08-15 🤣

[echo aurora]

You'll want to review #1397655624103493813 for more information on how to use Video Arena

[median oasis]

Hello all, Her to attempt to 'level up' my content creation skills for non profit association member engagement

[sonic tendon]

I think that's just when they finished it internally?

https://x.com/JustinLin610/status/1960692051185688646?t=qS_urFTZ8QuWJifepTXGfQ&s=19

Junyang Lin (@JustinLin610)

Qwen

X · Aug 27, 2025, 1:13 PM

this seems like it could've been when it hit the arena

if so it wouldn't be uncommon

[echo sinew]

Hi! You'll want to check #1397655624103493813 for more information on how to use our tools for generating content.

[keen beacon]

Is it even on chat.qwen.ai already?

[rustic knot]

god bless Qwen, they will overtake deepseek in fundamental ai research and development

[ornate sleet]

New here 🧡

[keen beacon]

Insane cope

[rustic knot]

ok, what do u think then

[keen beacon]

Idk, it just did far worse against Deepseek each time I compared them both

To be honest we do not even have any good benchmarks to start with

[jovial pelican]

I'm curious to see and be a part of

[lofty elm]

nano-banana not showing in direct chat?

[keen beacon]

Its Gemini 2.5 Flash Image now. Also titled as nano-banana on LMArena because people just loved this name more

[lofty elm]

i cant see it in direct chat options?

[keen beacon]

Click on the image generation option

[lofty elm]

oh thats it im seeing it now, i just woke up

thankss

[keen beacon]

okay, this Qwen seems to be a bit better than previous versions

still far away from frontier

seems like they did something to reduce hallucinations this time

[livid venture]

Hello, LmArena

[lofty elm]

thanks its working

[keen beacon]

by the way, it's also in aistudio.google.com now since they released it

[errant rover]

yo

[verbal nimbus]

Q1 2026? 💀

The new AI Studio is bad, can't even scroll up

[rare python]

https://ai.studio/banana

[coarse charm]

hi

[fiery wraith]

hi

[muted delta]

hi

[gloomy badge]

ya

[rain rune]

Hey

[keen beacon]

Did anyone know that AIs trained with reinforcement learning lose the ability to be creative?

https://gwern.net/doc/reinforcement-learning/preference-learning/mode-collapse/

Mode collapse harms esthetics: outputs start to sound the same, like “AI slop” or “ChatGPTese”, or look somehow similar, like “the Midjourney look.” This cripples creative uses like creative writing. And this damage can manifest in strange ways, like models refusing to write non-rhyming poetry & subtly steering non-rhyming inputs towards rhyming, or being unable to generate random numbers.

Basically, you just train a model to give precise and correct answers - and it starts to create art with the same precision and correctness as if it was a reinforcement learning challenge. It loses the ability to synthesize from weakly related and totally unrelated ideas - it loses the ability to be creative.

[steep cypress]

helo!

[novel umbra]

hello

[gusty gull]

Hi just dicovered lm arena on discord 🙂

[echo aurora]

hey everyone

[exotic nebula]

yo congrats on your promotion to admin 🥳

[molten fern]

hello

[exotic nebula]

hii

[echo aurora]

https://tenor.com/view/huh-dog-hug-dog-question-mark-dog-question-mark-gif-14144426802562365039

haven't I always had admin

[exotic nebula]

https://tenor.com/view/bruh-meme-gif-26978290

bro i thought you were just a mod 😭

[slow galleon]

hello

[limber acorn]

hello

[tired lichen]

hi there

[stoic raft]

hi.. im newbie.. how generate video here?

[willow steppe]

hello

[quartz light]

bruh

where is the coding model

[keen beacon]

hello

[granite igloo]

Hello I have the reached the limit for nanobanana, it's telling me to wait another 50 mins to generate so I have a question for that, How many Images can I generate before hitting the limit? I have generated around 15-20 images before hitting the limit and once the 50 minutes pass will I be able to again generate 15-20 images or will I be limited to a few images only? TIA

[hollow horizon]

Hello

[limber crow]

the model of video generate is random?

I can't choose the model

[glossy tinsel]

Hello

[desert abyss]

Yes, it's random here in Discord.

[desert abyss]

You can go to the battle mode to choose the model: https://lmarena.ai/?arena

[limber crow]

so only the picture mode can choose?

🤔

[lilac pivot]

Hello. I new to this arena.

[proven reef]

hi

[echo aurora]

[echo aurora]

If you check out #1397655624103493813 you'll find what you're looking for.

[echo aurora]

to the battle mode to choose the model:
slight correction - selecting Side by Side & Direct is what will let you select specific models https://lmarena.ai/?arena=&mode=side-by-side

[vague owl]

hi everyone

[lone notch]

helo

[keen fulcrum]

they released Grok Code Fast 1

[quartz light]

its ass

[keen fulcrum]

I was expecting more as well

[quartz light]

wait

[keen fulcrum]

Its ok for its price

[quartz light]

is that the only thing??

[keen fulcrum]

Yes

[quartz light]

seriously????????

ive been waiting for over a month

for that sh?

@deep adder grok fell off

[quartz light]

no

its many times worse than gpt 5 nano

[astral pagoda]

Isn’t it Sonic ?

[quartz light]

idk

yall check this out

https://websim.com/@rat/ai-game-generator

its goofy but

i guess camera rotation kinda works

ill fix in like 10 hrs

btw it uses nanobanana

and yes it is super inefficient but uhhh

yes

[empty salmon]

Google aistudio remove aspect ratio 9:16 for veo ?

[latent pagoda]

hello

[empty salmon]

anyone have same issue?

[keen beacon]

Insanely difficult task for LLMs:

What emotions does the following piece of music evoke? Why? Conduct a comprehensive music theory analysis.

Kick = 1/4 * 32
Snare = 1/16~, 1/32~, 1/16, 1/8, 1/16~, 1/16, 1/16., 1/32~ * 64
Closed hat = 1/16 * 128
Open hat = 1/8 * 64
Bass = (B1 1/4 * 6, G#1 1/4 * 2) * 4
Celli pizzicato = ((B3 1/8 D#4 1/8 F#4 1/8) * 21), B3 1/8
Organ = (D#5min 1/8, 1/8~, D#5min 1/8, F#5maj 1/8, C#5maj 1/8, 1/8~, C#5maj 1/8, F5dim 1/8, B4maj 1/8, 1/8~, B4maj 1/8, F#5Maj 1/8, A#4min 1/8, 1/8~, A#4min 1/8, C#5maj 1/8) * 4
Choir = (B4maj 1/4.., F5dim/B4 1/4.., F5Dim/B4 1/8, C#5maj/G#4 1/2, C#5maj 1/2) * 2
Trumpet = (B3+G#4 1/4.., F4+G#4 1/8., C#4+G#4 1/8., C#4+G#4+B4 1/8., C#4+B4 1/4.., D#4+B4 1/4.., D#4+B4 1/8, B3+G#4 1/4.., F4+G#4 1/8., C#4+G#4 1/8., C#4+G#4+B4 1/8., C#4+B4 1/4.., D#4+A#4 1/4.., D#4+A#4 1/8) * 2

Legend:
C#5maj/G#4 1/2..

C - note label
5 - octave label
maj - chord quality
/G#4 - note in bass (for inversions)
1/2 - note duration

Other:
. - dotted note
~ - pause
B3+G#4 - notes sounding together

Here it is:

Qwen 2507 Thinking thinks that the tempo is

Implied Moderate to Slow (Dotted half notes in choir, long bass notes). Likely 70-90 BPM – slow enough for emotional weight, fast enough for rhythmic drive.
Which is absolutely ridiculous. Other models, however, determine it correctly at 120-160 BPM.

But if I tell that it is 70 BPM in the prompt, they all pretend that the tempo is "slow and ceremonial march", which is again ridiculous

That Qwen is absolutely horrible at this. It invents things I never included in this track and never ever studied:

Organ/Choir/Trumpet Progression: This is a direct adaptation of Pachelbel's Canon progression (I-V-vi-iii-IV-I-IV-V), but transposed to B minor with minor-mode alterations

Picardy Third (B4maj): The tonic chord resolving as major (B-D#-F#) instead of minor (B-D-F#) is the emotional core. It transforms expected sorrow into luminous hope—like sunlight breaking through clouds. This is the primary source of "bittersweet" tension.

Okay, maybe I accidentally rediscovered Pachebel's Canon, but there are literally no Picardy third in the end here, all chords here are strictly diatonic to F# major/B Lydian, the mode is definitely major so it can't be a Picardy third. It completely made it up.

New Qwen-Max which is at LMArena right now is much better. But it is still horrible, making up things and hearing modes that are simply not present in the track

Deepseek V3.1 thinking in the chat is much better. It makes something up sometimes, but at least it is able to clearly identify the tonal center at B most of the time, and call it a major mode. Which is very close, because it is actually B Lydian, a major mode of F# Major scale.

[keen beacon]

The worst offenses I can accuse Deepseek in are "chromaticism" and "modal interchange" (which are bs because all chords are strictly diatonic here), but not goddamn Picardy thirds

Qwen on other hand completely slops everything up. Also the first model that calculated the tempo to be 2 times slower than it actually is.

[zinc ore]

Crazy grok is on the list but not oai

[keen beacon]

GPT 5 is the only model that correctly identifies B Lydian. It succeeds around 2-3 attempts out of 10, but it is still better than Qwen and Deepseek that never managed to do it.

Also the only model that keep insisting on F# or B Major and disregards other modes, which is almost correct

[ocean vortex]

grok code only there because it's a new model lol

[whole wagon]

Who is top

[ocean vortex]

Now I see why I haven't heard much about it. This is probably the first coding model I see that doesn't have any coding metrics neither in their press release nor even in model card https://data.x.ai/2025-08-26-grok-code-fast-1-model-card.pdf

https://x.ai/news/grok-code-fast-1

Last week, we quietly released grok-code-fast-1 under the codename sonic. During this stealth phase, our team carefully monitored community channels and deployed multiple new model checkpoints to address feedback.

So basically 'sonic' was no good and they buried the results. 🗿

Their blogpost absolutely screams "AI written" lmao

grok-code-fast-1 was crafted to shine in the tasks developers face every day, striking a compelling balance between performance and cost. Its strength lies in delivering strong performance in a economical, compact form factor, making it a versatile choice for tackling common coding tasks quickly and cost-effectively.

[wild galleon]

How if i want use API of Lmarena.ai to my openweb ui ??

[unborn ocean]

Can’t. There is no api, scraping etc. is obviously not allowed.

But you might want to look at this: https://github.com/cheahjs/free-llm-api-resources.

If you are looking for resources in general.

[keen fulcrum]

It does in the model card

[ocean vortex]

where

[zinc ore]

Where's the coding benchmarks (outside of cyber security)

[keen fulcrum]

If they were that proud of their coding model we would have seen promotions by Musk

[ocean vortex]

biology and chemistry I wouldn't classify as coding. Cybersecurity cybench perhaps, with a stretch... Though I wouldn't expect to be looking at this and nothing else for a coding model

EDIT: Ok they did low-key mention SWE in their news article. That's something I suppose

[pine kraken]

https://x.com/noob_contrarian/status/1961335098327929081

Hope you all are having a good day. I just wanted to take a moment and tell you all about a research study I’m conducting for an app that focuses on AI Brainrot Detox with some new features and benefits for people who are struggling. I would really, really appreciate it if you all could just take out a few minutes from your day and participate in this.
I’m reiterating that this is completely anonymous so please don’t feel uncomfortable.

Ash (@noob_contrarian)

working on recovering from AI overload - quick poll to see who else needs help: [https://t.co/dcy1UF0iOG]

(anonymous, of course; we’re all in this together)

X · Aug 29, 2025, 7:48 AM

[hollow imp]

Bruh

[rocky mauve]

🤖

[fleet narwhal]

hello..just wanted to learn more about AI,that is why joined..have a good day to all

[tender trellis]

Wha5h is this

[vapid sapphire]

hi

[leaden sun]

thanks for sharing, finally something substantial to back what we all feel about claude for a long time now

[gaunt void]

Hello

[tall summit]

lesswrong

[cedar estuary]

hi everyone!

[tall summit]

lesswrong moment, disregarding it immediately

[leaden sun]

ok? i should check their credibility then 😅

[rich field]

hi

[leaden sun]

hey nice to see you here Eric!

[rich field]

🙂

what is it all about?

[leaden sun]

well...talking about benchmarking mostly 😆 feel free to look around and play with image/video generation

[neon idol]

Bruh what the hell of AI are you?

[rose vessel]

hello all !

[fleet lintel]

I am lookoing for some hard (text base) questions that we expect that LLMs should be able to solve but currently cant

[tall summit]

some of HLE

[leaden sun]

Hi, I'm Sydney. How can I help you today?

[daring schooner]

hello

[neat quarry]

Hello!

how i make videos here?

[dusky tangle]

Hello! 🙂

[wicked thicket]

hello

[neat quarry]

I sent my prompt, but the video wasn't made T-T

How long does it take to make?

Damn, I think I've already used up my daily credits even though I haven't done anything lol

[wanton knot]

Hello!

[stark sand]

they are enjoying burger in a happy and smiling mood

[grand pewter]

Hii

[latent crest]

Good morning guys , can nano banana do nsfw content or soft nsfw content ?

[merry geode]

Helloo

[midnight mesa]

i love lmarena

[tough onyx]

heloo

[golden ocean]

-# (Claude Opus 4.1 thinking is included in the second option)

[quartz pike]

yo yall

on lmarena

is there a limit to how many messages i can send

or image requests to an ai?

-# personally i think yes. because it would cost a SH1T ton of money for lmarena to run o3 for free to anyone. and 4.1 opus.

-# but im not sure so pls fact check me

[lilac inlet]

I have been using LMarena for quite a long time, and I have experienced all the top-notch models, especially Opus 4.1, which is available for limited use. Once you exhaust your limit, you have to wait for some time for it to run again!

[quartz pike]

oh lol

what is the limit?

[quartz pike]

And me too. but personally i come from websim lol

its a really great platform

[lilac inlet]

I think it depends on your token. If you're using it to generate a long code, then it won't even run thrice!

[quartz pike]

damn

so its a token limit not a run limit.

Got it.

[willow grail]

hm

banana sucks at

convert style of image into japanese anime style. this shows a dinosaur anime. we see a ceratosaurus.

someone help me ?

[crisp spear]

Is this your prompt?

[quiet kestrel]

How to make video

Pls tell me

Hello?

[willow grail]

ya

this is the best so far which it could make

but its not really anime

[keen beacon]

Is there any Ai generator that does nsfw content? Or I meant say everything in general literally

[willow grail]

convert picture of dinosaur so it looks like manga death note and similar manga.

[keen beacon]

Stable diffusion

[chilly parrot]

hi

[viscid sparrow]

hi

[quiet kestrel]

How to see my videos

[rocky wedge]

Hi

[midnight mesa]

is it possible to nano banana make videos

[lofty compass]

a 3d wireframe of a battle ship

#share-prompts a 3d wireframe of a battle ship

/video a 3d wireframe of a battle ship

[timber iris]

Does GPT 5 HIGH in imarena have thinking?

[frank elm]

I am exploring the new Features

[neon idol]

Yes

[echo sinew]

Hello! You'll want to check #1397655624103493813 to learn how to prompt the bot and generate videos in #video-arena-1 #video-arena-2 #video-arena-3

[timber iris]

Great, i guess it have auto thinking, sadly we can't see what are they thinking

[cinder oar]

yoh great to get a chance for different ai.

[keen beacon]

Yes, it's literally the best thinking model of OpenAI and on the market right now.

Sad that we can't really see the reasoning process.

[willow grail]

why am i alive. its suffering. nothing else.

[hybrid yacht]

Hi all, complete newbie to this Ai thing,
just here to learn

[willow grail]

nice

how did you find lmarena

[hybrid yacht]

Been watching Jack Vs Ai videos

[willow grail]

oh thats a new youtuber

[hollow ether]

Hello there! It was good to be here and learn more about AI

[boreal rampart]

hello! I'm here for make a content video AI!

[stray dock]

broooo wtf just booted up LMArena and all my previous prompts are GONE

WTF

[willow grail]

omg jack

[eager crag]

is it... down?

[exotic gust]

seems like it

nope

[eager crag]

no?

[exotic gust]

back to normal

[eager crag]

oh yeah it's back up cool

[exotic gust]

prolly some cloudflare issue

[eager crag]

or it's me who did too many prompts

[rich compass]

fix your @$$hole site😎

[eager crag]

hey, not nice.

[rich compass]

what not nice?

im trying to fix my script

AND SITE

JUST DOWN

[dense sphinx]

Mine not working too

[eager crag]

me too, but it isn't necessary to insult the developers.

[dense sphinx]

Is it bug again?

[eager crag]

they're gonna fix it like they always did. patience is a virtue.

[opal hamlet]

WTF？

[rich compass]

badass site

dont worry

your ass soon gone

[dense sphinx]

JN Pavel durov dangerous.

[rich compass]

nah

IM NOT IN THE DANGER SKYLER

I AM THE DANGE

R

https://tenor.com/view/dayni-gif-16797284359068260185

[dense sphinx]

Can it be fixed?

[eager crag]

it can be, just an outage.

i'm guessing at least.

[rich compass]

use free grok 4 for now

grok com

idk

5 requests

[sonic flax]

Hello friends, I get the message "No models found," how can I fix this?

[tight oriole]

🙀 💀 ✌️

[rich compass]

nice nick

kitler

[tight oriole]

😽

[eager crag]

it might be an outage. we and other usual users don't know yet.

[sonic flax]

tank you

It's back, friends.

[dense sphinx]

@rich compass where?

On LM?

[rich compass]

nah

grok dot com

[merry wren]

hi

im having an issue

there are no more models on the model selector

[night geode]

Hi

[merry wren]

hello

[night geode]

I am completely new to this platform, how can we generate videos in the website?

[merry wren]

video arena i think

[night geode]

Alr

Thanks

[static stream]

hello

[merry wren]

you're welcome

[merry wren]

hi

[lean coral]

hey

is there any way i can generate more thn 8 vid in 4 hr

[echo aurora]

Are you still seeing this?

[echo aurora]

There is not

[rugged brook]

is sonnet 4 thinking better then gemini 2.5 in coeding

[lean coral]

@echo aurora there is limitation for generating image?

[rugged brook]

ye

[lean coral]

ah

i was looking for any ai that close to veo 3 that going to give vid with sound with any limitation

[merry wren]

refreshed, resolved now

[merry wren]

thank you for asking

[quartz pike]

Tuff or nah pls vote i wanna see what sh1tty ai made the one on the right.

[sterile pulsar]

Hello!

[quartz pike]

hello

[lean coral]

hey

[rose lintel]

hi

im new here

[upper venture]

It makes sense. Veo 3 is one of the best video GenAI rn and seed dance isn't made for those type of gaming videos after all, you can see it generates yeah but as the name suggests and given the company who made it (Bytedance, owner of TikTok and others) that isn't the main goal, it's more about dances, IRL videos, etc not games even tho it can generate xD.
Btw what an impressive result from Veo 3 even with the audio

[lean coral]

can i generate ai vid in lmarena compare model

[upper venture]

#video-arena-1 /video

ig, I didn't use or test it yet, I just like seeing others' prompts and gens lol

[willow grail]

wtf is happening with users

so many japanese nicknames

[unborn ocean]

the main point is that it is the lite version: good visuals, but too small to actually understand the query or physics

i acutally called that it would be the lite model

very easy to notice

[upper venture]

That's also very true

It is a capable model ngl

Not as capable as Veo 3 tho. But yes, this result doesn't show all the power that the Seed team created with their GenAI models

[unborn ocean]

no name like yours though :v

[willow grail]

o.o

[unborn ocean]

nah, again the name change, lol

like the 10th

[willow grail]

yeah

i played catch with wild crows

itsg fun lol

they need to train tho

[dense geyser]

hello

[willow grail]

i will teach them how to catch food in air

[willow grail]

like there is a pedestrian bridge above a river. and crows live there

so u go up there and throw the food down and they try to catch it

[unborn ocean]

sounds pretty cool, the crows in my city are really brutal though

[willow grail]

if its baby season then u just need to feed em and they stop attacking

[unborn ocean]

like they kill (weak) pigeons and stuff like that

[willow grail]

weak pidgeons sounds like virus etc?

[unborn ocean]

yeah idk, usually it is just old looking ones (and then they team up etc.)

[willow grail]

like normal carrion crow?

[unborn ocean]

was probably just confusing em with ravens (because we dont really have any of them)

thinking about it

i was also surprised

[willow grail]

ravens usually are parents and children.
they dont do big groups like crows

[unborn ocean]

well i did see ravens / crows teaming up on single pigeons

[willow grail]

lol

wher ulive

was it winter?

achso

[unborn ocean]

maybe i was drunk

[willow grail]

no idea

[unborn ocean]

but saw it multiple times, so idk

[willow grail]

king jon un would perhaps also attack very weak birds

[junior forge]

Hi, im abdoulaye from fench nice to meet all!

[willow grail]

whats going on so many new people

https://tenor.com/view/ooh-ooo-cat-shocked-funny-gif-14366308

[echo aurora]

hi everyone lots of new folks, welcome welcome! Don't hesitate to ping me if you have questions or problems with the site

[misty vault]

bring back alpha.lmarena.ai so that i can continue enjoying lmarena while it is down for everybody else

[dusty niche]

Guess the modle

[clear trellis]

Hey im pretty much new and i just heard of lm arena like 10 mins ago.

My question is, is it 100% free ? Can i test and generate images and make comparison fully free or is it limited by credits ?

[modern flume]

how to generate video

[dusty niche]

yep 100% free, just go to the website and start testing

[modern flume]

how to generate video

[remote lagoon]

even has an extra btn that links to webdev arena!

[dusty niche]

go into video arena an type /video then add your prompt

video arena is chat here not a website

[echo aurora]

Lol good to know, we shall consider

[echo aurora]

Check out #1397655624103493813 for more info on how to use this!

[ripe orbit]

hello

[echo aurora]

welcome! Yes, it is free. There are limits in that the amount of prompts per day/hour by model is in place. But when you hit that limit and you can wait it out kind of thing.

[upbeat idol]

hi

[rustic knot]

oh this model literally said it was everyone

[quartz pike]

WHAT

YOU ARE HERE

WWWWW

[austere kiln]

Had the same issue on a smartphone. On pc, however, there is a long history available. Using the same wifi network on both devices.

[tall owl]

hello

[teal mantle]

what is the gpt 5 pro quota of chatgpt pro?

team is 15 per month

[weak timber]

Hello

[torpid roost]

hi

[blissful carbon]

hi

[echo aurora]

hello 👋

[nocturne fiber]

happy to be here

[teal mantle]

hello

[grizzled plaza]

hello world

[vagrant idol]

Hello

[median ginkgo]

Anyone here knows how i can do the image to video thing

[verbal nimbus]

#1397655624103493813

[naive sedge]

#video-arena-1

The do /image

Then u will see the bot Givin u the command

Then*

[echo aurora]

Yeah check #1397655624103493813 for details, you'll want to use /video

not /image

/image creates images, /video creates videos, /image-to-videos makes videos with a reference image

[leaden palm]

hi guys - this place grown into something new.

remember when there was just one lmsys discord, when the icon was the same as gradio's or vicuna's, when there were just a few million votes? these days with all the new image, video, and text models, lm arena has really became a phenomenon.

thing is that i've also became a person who isn't as interested in these new video models, these hyperactive chats, these things that come with scale. i haven't even been using lm arena or moderating the chat that much lately. as such i'm leaving this server and leaving the job to the other moderators here.

it was great being with all of you. see you around, hope to discuss some more things then 👋

[fathom plover]

hi

[echo aurora]

Thank you so so much for everything @leaden palm

I really appreciate all of the help you've put into this community! You shall be missed ❤️

[sterile saddle]

helo to everyone

[proud hazel]

Hey @echo aurora, how do you prefer to be contacted about “private” matters? Is it okay to just send you a DM?

[echo aurora]

Can you send a DM to the @oak python bot? That's a good way.

[inner gate]

What’s up pals

[umbral veldt]

can't wait to try video gen!

[sweet imp]

Yo YO

[quartz pike]

Chat i think gpt image 1 is cooking

[proud hazel]

I'm curious for the result

[quartz pike]

it crashed

i cri

me depresso

me die

[proud hazel]

🫡

[verbal nimbus]

Direct mode never works for me

[quartz pike]

it aint direct

its compare

or wait

is me stupid

yeah

compare

[magic rock]

How to generate video with audio?

[quartz pike]

L u c k.

pray to lmarena discord

to give you veo3

[keen beacon]

#1397655624103493813

ah

with audio

it is random, so no

[neon idol]

Or just use clickbait video in YouTube tutorial for get Free Veo3

[quartz pike]

yes

[raw arrow]

hello

[echo aurora]

Not all video models have audio support, and since it's random which models you get, it's also random if your video gets sound or not.

[sour wyvern]

Hello

[dense quail]

Hello I am here to learn more on how to generate good videos from prompts

[golden ocean]

true

[echo aurora]

Be sure to check out #1397655624103493813 for more information on how to use the bot.

[hollow geyser]

hello, video generation

[ocean vortex]

so in short, veo3 > seedance, lol

[upper venture]

As it stands rn, in short: Veo 3 > anything else.

And listen, I'm not a big fan of Google or Gemini.

[ocean vortex]

it does seem that is consensus yeah. Although I'll admit I'm not huge fan of video models in general to use them extensively. Don't see a terrible amount of use for them other than just playing around for fun, personally 👀

[upper venture]

Same

But from all the texts here and even that Minecraft first pov generation, I see why people talk so great about it

Veo 3 is really capable (I mean, expected from the owners of YouTube and the ones who have the largest amount of GPUs out there)

[proud hazel]

Kling 2.1 Master, Wan 2.2 and Seedance 1.0 Pro all generate higher-quality videos. However, Veo 3 has audio.

[upper venture]

Yes but look at the rank. They're all 100 points behind Veo 3

That's a big thing

[proud hazel]

Yes, just because of the audio. If they all had audio like Veo 3, Veo would be at a disadvantage.

[upper venture]

Even Veo without audio is still above them, don't get it wrong

[ocean vortex]

They don't really have many GPUs. They use TPUs instead. But yeah they have abundance of compute relative to everyone else...

[upper venture]

Lm arena separates the Veo models from sound and non sound. Yes, the difference is big too. The ones with sound have way more points. But the ones without sound are still scoring better than the competition. Don't get it wrong. Veo 3 is the most capable video model.

[upper venture]

Yeah, exactly 💯

[proud hazel]

In most cases, I actually find the results from other providers to be better than those from Veo itself. But taste is subjective.

[upper venture]

Yes and the majority agrees that Veo 3 is better. It is trained on way more data, follows prompts closely and also comes from Google and all its processing power.

That said, I prefer Wan, it's on the bottom of the rank but c'mon. Unlimited free generation is better than limited generation from Hailuo, Kling, Seedance pro or even Veo.

[proud hazel]

Unlimited Wan?

[upper venture]

Yes! Qwen/Alibaba provides Wan for free. Unlimited. Like there's guardrails I believe (you can't be making 100 videos per minute with a macro or bot). But aside from that, it's truly unlimited! How many videos you want to make you just go there and make.

At least for the time being

[proud hazel]

480p or 720p?

[upper venture]

I'm not sure about that. Honestly as I agreed with Dom. I don't quite use GenAI as much as I use LLMs

I prefer it but I don't use it lol

[ocean vortex]

it's subjective until we measure it:

[proud hazel]

Could you possibly generate a sample video in Qwen and post it here?

[ocean vortex]

If majority of people agree on their "subjective" opinions, that kinda becomes not subjective anymore. The findings 👀

With bias obviously ruled out since it's a blind testing/voting

[upper venture]

LM Arena (this server) ranking as of last update

[upper venture]

I mean it won't be the best but I'll try it

[ocean vortex]

https://wan.video 🧐

[upper venture]

Yes, that one

[ocean vortex]

this seems "slow unlimited"

How long does this typically take?

[upper venture]

Yes. It is slow

A few mins. The get member is not needed even for free members you get I believe 50 coins per day where you can get 5 fast generations per day.

People needs to understand that those video models are extremely expensive to run. Being provided unlimited generation per day even if it takes half an hour or an hour each. It's still free. Remember that it'll be using wan 2.2 their best model too.