general

[coarse glade]

Are you a admin

[modest prism]

The API does not have those system prompts which tell the models which version they are.

[stray aspen]

no

[errant cave]

From chatgpt.com

[stray aspen]

everyone knows that

[whole wagon]

Ngl. I'm not really seeing much point to the cheaper openAI models except so openAI saves money. Because we have stronger open source alternatives that are cheaper also

[tired herald]

ngl

[stray aspen]

you have system prompts lmao

[tired herald]

uhmmm, idk what you might even be referring to hmmmm

its a very interesting hypothesis though

[ocean vortex]

It is not "better". It's just finetuned more aggressively for the style and formatting users prefer

[tired herald]

I commend you for your wisdom

[whole wagon]

Why use GPT5 mini when GLM 4.5 exists for $0.5/$2

[coarse glade]

Can u help me fix this pls

[stray aspen]

why dude

just use it

[coarse glade]

Since it worked for him

[stray aspen]

he has system prompts

[tired herald]

I have a system prompt

[stray aspen]

hes making a plugin

[errant cave]

It's an issue with OpenAI, ask them

[gentle plinth]

it should, but it needs some work for getting it to run

[ocean vortex]

because gpt5-mini-high destroys glm

[gentle plinth]

if youve never installed react it can be complicated for some

[whole wagon]

I don't think so

I tried both

[coarse glade]

Ok so now it should work for me right

[hollow imp]

What do I say to it to run?

[gentle plinth]

it cant you have to do it yourself

[coarse glade]

Ok sorry for saying you’re lying you guys should have a section for help

Like if something goes wrong

[ocean vortex]

gpt5-mini-high is essentially o4-mini-high except better in every way

[hollow imp]

You said it needs some work. What prompt do I send for it to that "work"

[modest prism]

Why do you need the model to tell you it's gpt 5 if you already know it's gpt 5 ?

[ocean vortex]

So glm has no chance...

[whole wagon]

What the hell GPT5 dropped so much in the leaderboards

[gentle plinth]

after it generated the code, just ask it for step by step instruction for how to run it on your machine, tell it if you have windows 11/mac or whatever

[whole wagon]

How did it drop that many Elos in short duration of time

[tired herald]

its garbage after all

[hollow imp]

I have android 12

[gusty helm]

ask grok 4 same q, he will say he is grok 2 or 3 😄

[gentle plinth]

not impossible, but complicated xD

[coarse glade]

Beicase a lot of websites are lying nowadays and putting gpt 4o even though the person wants gpt 5

[stray aspen]

are you kidding me

[ocean vortex]

it's still nr1

[stray aspen]

guess ill use opus

[whole wagon]

By 6 Elo lol

[hollow imp]

Like I just want it to create a website according to my customisations and instructions and it should be runnable on chrome

[whole wagon]

It's within error

[whole wagon]

To 2.5 pro kek

[coarse glade]

How did it work for this person can someone help me pls

Can someone help me by dm me

[gentle plinth]

https://v0.app/

[modest prism]

Grok 4 deserves way better. It's underrated.

[ocean vortex]

Google is always benchmaxxing for lmarena

[gentle plinth]

maybe try this then

[ocean vortex]

Grok is kinda sh'it

[hollow imp]

Paste your prompt here?

[coarse glade]

Can someone help pls

[gentle plinth]

no

just what you want

[hollow imp]

Told you

[gentle plinth]

the prompt is only for direct chat

[modest prism]

Open lmarena in incognito tab. It might help.

[tired herald]

System Prompts, I have, you do not

[torn mantle]

sorry but it deserves that spot

its not that good

[coarse glade]

Oh ok ty for helping

[hollow imp]

System prompts in lmarena how

Can you tell

[tired herald]

You will soon have System Prompts @coarse glade

Soon

[torn mantle]

im starting to think that @modest prism is close to elon or some xai staff somehow

[tired herald]

very soon

[whole wagon]

The market gives grok better odds than openAI for December

[coarse glade]

Oh are u guys going to give it to everyone

[tired herald]

yes

and im alone btw

[stray aspen]

whos funding your project

[tired herald]

no one

im doing for fun

[hollow imp]

Lmarena head

[modest prism]

I'm not, grok 4 is really bad at coding, but at solving some deep math stuff it's really good.

[hollow imp]

Mr.@tired herald Head of Department, system prompts, lmarena.ai

[whole wagon]

Isn't Gemini 3 just going to be insanely good. They literally need to add a tiny amount of perf for SOTA

They are already at the frontier with a comparatively old model

[tired herald]

Nani 😭

[torn mantle]

yea you are sus

def sus

[whole wagon]

And the 2 pro to 2.5 pro leap was immense

[stray aspen]

im confident it will be SotA

[tired herald]

I would love to release this already and send it to the LMArena team so they can improve their website

[stray aspen]

if gemini 2.5 pro still holds its ground even now

with the new models

[modest prism]

The Arc-Agi benchmark doesn't lie

[hollow imp]

2.5 pro experimental shooked the world

[torn mantle]

why? they could've finetuned even if they only have the public dataset

[whole wagon]

Grok is intelligent if you prompt it right. Issue is it takes 10 mins thinking then

[gentle plinth]

the win-rate to gemini2.5 pro also doesnt look good (was already there before the update)

[torn mantle]

idk grok 4 vibes are off

[whole wagon]

💀

[hollow imp]

Give grok 4 prompting guide

[ocean vortex]

They are not and they have some catching up to do

[coarse glade]

And also are u guys going to give system prompts for the Public soon which will be nice ty pineapple for everything

[torn mantle]

we talked about how flawed this metric is

[hollow imp]

All of these models are available for free 🤑

[whole wagon]

And it's -4 points also

Lol

[ocean vortex]

It's not really flawed just because you don't like it lol

[tired herald]

They have more important things to work on right now

[blazing bison]

Do you guys think that we gonna have gemini 3 this month?

[coarse glade]

Quick question are u an admin

[whole wagon]

Ngl these GPT5 models are good at maths though. That isn't a Gemini strong point any longer

[coarse glade]

Yeah I have a feeling

[blazing bison]

Deepthink is good

[ocean vortex]

2.5Pro is not SOTA anymore, like that's the reality

[modest prism]

Poor llama 4

[coarse glade]

And also do u guys know that Claude is good too

[tired herald]

no

[gentle plinth]

have you tested it?

[blazing bison]

But still the best 1m context option

[blazing bison]

No

[coarse glade]

Yeah that’s true

[keen beacon]

google FUMBLED

its trash now

thats the reality

[coarse glade]

Yeah 2.5 pro on ai google studio is bad now

[tired herald]

temperature is something I checked if I can do it myself, but nope, which is sad

[keen beacon]

idk wtf they did

[whole wagon]

The initial 2.5 pro release was before o3 release

[modest prism]

Their TPUs melting

[stray aspen]

its not SotA but its not bad either

[keen beacon]

but gpt-5 is amazing

[tired herald]

pdf and docs are probably gonna come soon

[coarse glade]

It isn’t up to date even with grounding with google search for ai google studio honestly

That’s why I stopped using it

[hollow imp]

How to set it on gemini web?

[blazing bison]

But it received so much updates

[whole wagon]

IMO Gemini 3 is October time. Not any time soon

[hollow imp]

Like you have Google ai ultra but no temparature setting access 🥀

[coarse glade]

https://aistudio.google.com/welcome

[keen beacon]

and im the biggest google shill

[coarse glade]

This is the website

[keen beacon]

so if im telling you it

yk its true

[solid brook]

Leaderbord does not make sense. Gpt 5 chat which is the worst model in gpt 5 because it is non reasoning is better than gpt 5 mini thinking which is a reasoning model

[whole wagon]

Gemini 3 is October, grok 5 is december

[keen beacon]

let it age

[blazing bison]

They created like 30 different endpoints for gemini 2.5 pro

[tired herald]

you need to open Advanced Settings for Top P

[hollow imp]

I SAID GEMINI WEB

Not ai studio

[coarse glade]

Did u guys see that perplexity wants to buy google chrome

[hollow imp]

😭

[modest prism]

You have to manually tell the model to search the web and fetch the websites otherwise it won't do it by itself .

[tired herald]

bot hare gemini web 😭

[keen beacon]

both mid rn

[whole wagon]

That is just refinement. The march release was already good

[blazing bison]

What

[hollow imp]

You can't use deepthink or deep research on ai studio

[coarse glade]

Yeah Ty

[keen beacon]

Grok added NSFW to their image generator

[coarse glade]

That’s nice of u

Ty so much @modest prism

[keen beacon]

Thats their only latest update 💀

[blazing bison]

Why have so many people active here today

[hollow imp]

Why

[blazing bison]

😆

[stray aspen]

elon musk made a corn generator

[tired herald]

but no, its not possible to customize on gemini web

[coarse glade]

Yeah Austin Evans did an short about it

[gentle plinth]

leaderboard update xD

(i think)

[blazing bison]

I want zenith back

[tired herald]

there was one

[keen beacon]

Same guy who thinks humans on mars is more feasible then fixing earth btw

Hes a cr@ckhead

[hollow imp]

That's what I'm saying
You pay for Google ai ultra and you can't even change the temperature
🥀

😒

[coarse glade]

https://youtu.be/pKwLS9PaeXY?si=3Ppha7opxh4o9CbR

[tired herald]

I dont pay

[whole wagon]

I heard xAI is actually building AI corn team. To refine the corn generation

[solid brook]

Man I am so excited for gemini 3. Hope they don't nerf the model

[tired herald]

I dont have ultra

[keen beacon]

so many be yappin here

[coarse glade]

Exactly

[tired herald]

indeed

[whole wagon]

I wonder if they actually using corn as training data or what

[coarse glade]

Ok bye ty for everything @echo aurora

[hollow imp]

Guyzzz

[solid brook]

Actually google ultra plan is not worth the price at all. You get like 10 prompts a day

With deepthink

[hollow imp]

And you can't even change the temparature parameter

I only use the Gemini web because of custom gems

[modest prism]

It's worth it if you are a TikToker or YouTuber thanks to high credit of Veo 3.

[hollow imp]

8 sec videos only

And I have seen how 😭 veo3 is in video arena

[tired herald]

omg why is it so hard to change the UI of LMArena

[hollow imp]

???

You can do it?

[wicked osprey]

hi im new to here，i just wondering i got limited in my gemini after few question，is it unlimited in lm arena？

[hollow imp]

No

[tired herald]

yes, thats how I do everything 😭

[hollow imp]

@tired herald

How to change the ui of lmarenn lmarenn

[wicked osprey]

so far only sometimes i got stuck in genarating page but haven't saw limited yet

[hollow imp]

I never got limited in Google ai pro free trial

[tired herald]

not that hard lol

LMArna

[wicked osprey]

it just going flash instead of pro sometimes

[tired herald]

but why tf is the Logo a collection of SVG's

even the letters

[wicked osprey]

i just feel it takes time to genarating compare gemini appwith pro😩

[tired herald]

I have zero Idea why theres a visual of the old ui behind the new ui

[burnt sinew]

when will video generation be added to the website

[echo aurora]

TBD if it will, but be sure to share feebdack in #bot-feedback

[hollow imp]

To be donest

[tired herald]

to be determined?

[indigo hazel]

i dont have this ui

[ocean vortex]

You are paying them $250...?

[hollow imp]

I said previously I'm 15

[tired herald]

im making this ui*

[hollow imp]

I do not even receive any pocket money

[burnt sinew]

can i say that lmarena should add video generation to the website in the feedback?

[steep mirage]

Hey everyone,
I keep getting this Cloudflare "Security Verification" pop-up on LMArena (see screenshot).
Any ideas how to fix this persistent verification? It's really annoying!
Thanks!

[echo aurora]

Yup!

[echo aurora]

Uh oh - are you getting the same regardless of browser type?

[ocean vortex]

Well you shouldn't pay for it even if you had the money tbh. I could subscribe today but I don't see the point in even their Pro sub...

[hollow imp]

@echo aurora 😂

[steep mirage]

Yes, I'm getting the same security verification message on both Brave and Chrome.

[tired herald]

what happens when you try to verify

[ocean vortex]

What exactly are you getting for their Pro plan than you don't already have with aistudio...? Not much at all

[echo aurora]

Is the verificiation not working or just annoying and pops up a lot?

[hollow imp]

A wise man once said money multiplicates itself. Do you agree?

[ocean vortex]

Only if money is used wisely

Their gemini subs are useless

[tired herald]

Oh, making the Model selector appear in the chat box is gonna be hell

[steep mirage]

It appears every time I vote on a generated image.

[hollow imp]

I'm not talking Abt gemini subs at all

[ocean vortex]

You may buy gemini sub if you want to support them or whatever, but then you shouldn't pretend it's a good value for the service 👀

[hollow imp]

Pls add pdf support please 😭🙏

[echo aurora]

Okay I'll flag to the team. But to confirm -> the verifcation isn't failing? Just appearing too often

[tired herald]

icons icons icons.....

atleast this works

[tired herald]

later

[ocean vortex]

Well I am..

[hollow imp]

I will do whatever you say just please add it 🙏🙏🙏

[steep mirage]

Yes, that's correct. The security verification pops up every single time, but once I authenticate, I can use it again.

[tired herald]

It's hard and tedious, ill do the easy things first

[steep mirage]

Thank you very much for trying to help me fix this issue!

[ocean vortex]

IMO the fact that they are messing with huge models (ultra for deepThink) is not extremely promising tbh. There's no obvious substantial improvement path without going to the extremes with Google for now..

I doubt Gemini3 will be substantially better than 2.5Pro. Probably more like marginal improvements (2.6?)

[tired herald]

DeepThink uses Gemini 2.5 pro

[soft kernel]

Yeah I wished they did that instead of a bot in this discord server,but I think it's about the limits and direct voting

[tired herald]

Its not a separate model

[ocean vortex]

No it actually uses the bigger model

[tired herald]

No

[ocean vortex]

2.5Pro is not mentioned anywhere for that model card

[tired herald]

Its 2.5 but multiple

[tired herald]

So what

Marketing

[soft kernel]

Ain't deep think a part of 2.5 pro?

[tired herald]

They make multiple versions of 2.5 Pro work together to all create a better response together

[ocean vortex]

So it's a different model. We already knew they had bigger model internally, and DeepThink is way to distinct from 2.5Pro to be the same model tbh

[tired herald]

Same with the GPT-5 Pro and Grok 4 Heavy

[ocean vortex]

Like it can do some tasks worse than 2.5Pro

[tired herald]

Multiple 2.5 Pro ≠ Bigger Model

Ill stop arguing

Ill go work on my stuff

[ocean vortex]

It's a DIFFERENT model. And the only one they had to use other than Pro was Ultra

[tired herald]

No

[hollow imp]

@tired herald I don't understand top p at all. What value should I put

[ocean vortex]

Yes

Stop arguing

[tired herald]

I'm not

[hollow imp]

LETS GO DEOLD VS DOMS

[ocean vortex]

Just look at their 10rpd

cap

[tired herald]

💀

[soft kernel]

Still doesn't worth 250 bucks for 10 prompts a day

[hollow imp]

DEOLD776 VS DOMS

🔥

[tired herald]

Correct

[tired herald]

Idk either

[ocean vortex]

💀

[gentle plinth]

same for me. every vote gets checked with a cloudflare captcha. i thought that it was intentional, but if it isnt, then yeah, would be better to have it less often if thats possible

[ocean vortex]

They are trying to make reverse engineering impossible. But people are still successful in doing that so it's like fighting the wind lol

[gentle plinth]

I wouldn't call that reverse engineering, just some bots who want to push certain models

If you found a way to deterministically verify a model then it's easy I guess

And a difficult to solve problem for lmarena

[ocean vortex]

No, this is more to prevent people from coding interfaces that use lmarena API for model access tbh...

[hollow imp]

A difficult problem to solve for lmarena you mean

[ocean vortex]

Rather than bots doing vote manipulation - there's not many of those if at all

[tired herald]

Its not reverse engineering. You can use their servers easily without issue....

[gentle plinth]

But the captcha arrives after voting, not after prompt

[hollow imp]

DOMS LOST

[tired herald]

That's how I found out about system prompts

[hollow imp]

🔥

[gentle plinth]

So I could still generate programmatically

[hollow imp]

WINNER IS DEOLD

🏆

[ocean vortex]

Use their API? You can't without reverse engineering

[tired herald]

No

[ocean vortex]

That's why they do it

[tired herald]

Takes 3 clicks to get around any protections on the web

Also, even if, its not called RE

[soft kernel]

I think it balances creativity and coherence

[ocean vortex]

what

[tired herald]

You are insulting every RE with your sentences

[ocean vortex]

You seem confused

[tired herald]

Im an RE you honk

RE needs something to be locked in some way

Its open

So no RE required

[ocean vortex]

I don't think you are

[tired herald]

Okay

Idc

[ocean vortex]

you wouldn't talk nonsense

but maybe a wannabe one

[tired herald]

Says the guy that has no idea what google do

[plucky otter]

Is there an option that you have always 8 sec video without the commercial at the end. Now it is random and 5 sec or 8

[tired herald]

Unfortunately no

[ocean vortex]

@tired herald Reverse engineering refers to automating things like lmarena so you could do requests without having to go through their interface, click on things and having to do it manually.

[tired herald]

WHAT

Are you insane

[ocean vortex]

It's not what you think it means

[tired herald]

Are you f*cking insane

REVERSE

REVERSE

REVERSE

REVERSE

I think bro is having a stroke

Are you okay?

[novel crater]

anyone else having this issue

[soft kernel]

I'm pretty sure that's not reverse engineering AT ALL

[tired herald]

Json escape, your message broke message handling

[plucky otter]

AE (Alternative effects) server has a free pool of VEO3 text to video generator. Tonight ofline, but most nights from 1900 to 2300. 8 sec video, without watermark free download

[novel crater]

oh ok hmm thanks

[ocean vortex]

It is. Because you need to arrive at the methods they are doing in their interface for sever requests without having to use their interface. You are reversing how it works in order to be able to do it independently of accessible interface. I don't know why this is so difficult to grasp for some people. 🤷‍♂️

[tired herald]

no

You are wrong in every sense of the word

Im sorry, but please stop interacting with me

[ocean vortex]

Enlighten us then

[tired herald]

Reverse

[ocean vortex]

Cause it looks like you have no clue

[tired herald]

Why is it reverse

Reverse

[ocean vortex]

Most likely you don't

[tired herald]

One word

Reverse

Can you explain why that word is used

[ocean vortex]

and...? Go on. You said nothing

You have no clue do you...

[tired herald]

Bruv

I asked a question

[ocean vortex]

I asked you first.

[tired herald]

Reversing something

[ocean vortex]

But you can't come up with a single sentence

without using chatgpt

[soft kernel]

Nvm bro seems confused

[tired herald]

Bruv

Dom

[ocean vortex]

🤣 🤣

[tired herald]

Can you read what an AI just said

[ocean vortex]

Bruh....

[tired herald]

Then I think

That

You

Are

Ha ing

A

Stroke

[ocean vortex]

turn on your brain

[tired herald]

This guy thinks he can code GTA 6 in 2 hours

This is the embarrassment of vibe coding

[ocean vortex]

And explain how I'm wrong

if you can't do it

without prompting chatgpt

[tired herald]

This guy is the reason why people believe in the flat earth

[tired herald]

BRO

REVERSE

[ocean vortex]

then you have no clue

[tired herald]

REVERSE

REVERSE

[ocean vortex]

LMFAO

[tired herald]

WHAT IS THE DEFINITION OF REVERSE

WHAT IS THE DEFINITION

OF

REVERSE

[ocean vortex]

are you stupid?

🗿

[tired herald]

You are pulling my leg

You cant be this unintelligent

Im being successfully Ragebaited

[ocean vortex]

You sound like a broken record having absolutely no clue what you are talking about

😭

[tired herald]

Can

You

[ocean vortex]

REVERSE

[tired herald]

Define

[ocean vortex]

REVERSE

[tired herald]

Reverse

[ocean vortex]

🤣

[tired herald]

No?

Ok

[exotic nebula]

What are you guys rambling about?

[tired herald]

Reverse engineering

[exotic nebula]

Reverse engineering?

[blazing rune]

🍿

[vocal token]

<@&1349916362595635286> See message logs. Some drama

[exotic nebula]

I see. What about it?

[tired herald]

Dom doesnt know what reverse means

[ocean vortex]

don't ask. Some silly person pretends to know it all but can't say one sentence about it without prompting chatgpt lol

[tired herald]

Agreed

[echo aurora]

Reminder:

✅ Treat others with Respect. Be kind, assume good intent from others, and keep disagreements respectful. It’s encouraged to share your disagreements, but only if it’s done in a respectful and productive way.

[vocal token]

Hey Greg!

[tired herald]

No, im proving to you that an ai can be prove you wrong without breaking a sweat

[tired herald]

Sorry 🙏

I was being ragebaited and fell for it

I promise it wont happen again

[echo aurora]

Lets just keep conversations respectful please.

[leaden palm]

reverse engineering is the process of figuring out the logic and shape of the requests; automation is separate but i can see how it would be included

[ocean vortex]

Looks more like you have no clue because you couldn't do a single proper sentence disproving anything I said lol

[tired herald]

I tried

[ocean vortex]

yeah...

EDIT: for the below... Yeah blocking is easier than coming up with a single argument that is not AI written. You can only pretend you know stuff for as long.

[tired herald]

Im blocking you for my own sanity. I cant believe how little you know

[exotic nebula]

Chill guys

Lets switch topics

Did anyone try out HRM AI?

[tired herald]

My topic is that im prob gonna release my extension today

[tired herald]

What is that

[soft kernel]

Which extension

[tired herald]

#ai-creations message

[leaden palm]

it's interesting that gpt-5-chat ranks below a few other non-reasoning models

[exotic nebula]

A new model based on the brain's system of thought.

[tired herald]

See for yourself in community creations 🙂

[tired herald]

Ill check it out after im done with what im doing

[echo aurora]

Oooo this is exciting!

[tired herald]

Hehe

Only a single bug to fix and then its onto github to release the src

[exotic nebula]

Bro stop

Stop milking a dead cow

[soft kernel]

Can you please...?

[soft kernel]

BRO

[tired herald]

. 😭

[soft kernel]

Release it asap

😭😭

[tired herald]

Yeah, theres just a little problem with it right now

I need to fix and release

Im hoping that it could be helpful for those who could use it

[vast fern]

hi everyone

i am new here

i need one help

[tired herald]

Ask away

[soft kernel]

Is there any other models for generating 3d worlds,real time,other than genie 3?

[exotic nebula]

yeah man, we all ready to help. welcome to the community

[tired herald]

Yeah, but Genie 3 tops all of them, by quite the margin actually

[vast fern]

what is this and how can i get this

i wanna be updated on the new models

[soft kernel]

Any examples regarding those models?

[tired herald]

In the battle mode, if you get lucky, one of the two random models will be toad

[gentle plinth]

I think he means he wants to get notified

[tired herald]

They all have obscure names so I dont remember, but im sure if you search genie 3 on yt, click on the channel of one of the people, youll find some more stuff

[tired herald]

And how he can get it? Or did I misread it 😭

[vast fern]

what are these notifications and how do i get these or is this like a manual thing i need to do battles and if i am lucky i will discover something new?? basically i want to know how to get this bot that will give notifications about the new model

[tired herald]

OHHHH

[soft kernel]

It's a discord server

Legit api

[gentle plinth]

These notifications about new models are on the server discord. gg/devmode

[vast fern]

Thanks man how can i join that

[tired herald]

Oh, im just stupid

I misunderstood

[echo aurora]

Sorry to say discord invite links are blocked here so will have to DM the inv link

[gentle plinth]

But can I leave the msg here?

I mean one can simply remove the space

[vast fern]

Thanks everyone, Have a great day.

[echo aurora]

Yeah you can leave that message, but if you want it to link you won't be able to is all

[gentle plinth]

OK thanks

[scenic salmon]

@echo aurora I don’t know if you saw my question yesterday, but I was wondering if you’re able to talk about the pre-seed/seed rounds at all, it wouldn’t need to include any specifics about lmarena, just how the general experience went… like finding a lead investor, pitching to angels/vcs, etc

[echo aurora]

I'm not sure I'd be able to share much info, even if I knew what that was like first-hand. I wasn't involved with those conversations so I know very little regarding that.

[scenic salmon]

Ah okay, thanks anyway

[gentle plinth]

Just have to get lucky I guess 😅 ai🎰

[tired herald]

Gacha on ai, gacha games, gacha life, gacha everything

[exotic nebula]

@tired herald https://github.com/sapientinc/HRM

[exotic nebula]

You're the ultimate gaming gambler, dude.

[keen beacon]

What is this model about? Is this some new architecture for LLMs?

Hold on, I read the github

[exotic nebula]

Its a brain inspired ai architecture which focuses on deep reasoning rather than llm. It scored some very interesting results at the ARC AGI benchmark.

[keen beacon]

Would this replace CoT then?

Just asking if you knew more

[ocean vortex]

What kind of score? 🧐

[exotic nebula]

Maybe. But unfortunately it has some flaws, you will find them somewhere in their reddit community.

[exotic nebula]

Nope, just a passerby who came across it. Interested if someone has set it up and tested it.

[keen beacon]

Okay, thanks for the info.

[keen beacon]

Would be very interesting

Esp for math and such

[exotic nebula]

40.3%. Not that impressive but good for its usecase and mostly because of its low compute.

[exotic nebula]

Exactly.

[stray aspen]

does anybody know an ai model for sound effect generation

[exotic nebula]

Lyria 2

[stray aspen]

isnt that just for music

[ocean vortex]

uh they have written a paper as well
https://arxiv.org/pdf/2506.21734

[exotic nebula]

Yeah. My bad, didnt see it through. Sound effects generation eh? Hmm, not sure about that. Did you try looking it up on the net?

[stray aspen]

yeah i found eleven labs

im gonna give it a try

[exotic nebula]

Yes! Forgot about that. Isn't it impressive tho with just 27M parameters and training on just 1000 examples?

[exotic nebula]

I heard thats good, all the best dude!

[white hatch]

Is it unsafe to use VPN while visiting chatgpt website?

[exotic nebula]

No, why do you ask?

You wont get flagged.

Just might be stripped of some features/accessibility refrained in some countries.

Have heard of account lockouts. But so far only rumors.

[ocean vortex]

It is, though it's also... too good to be true? 👀

I wonder if this would hold up as a general purpose model in real life

[exotic nebula]

My thoughts exactly. Thats why I'm searching for people who have experience with it and set it up for their own usecases.

[exotic nebula]

Yeah I guess so, this might be a huge hit in Healthcare

[neon idol]

Hello chat

How are you?

[whole wagon]

I like how opus be like this when you ask it to be your friend. But you ask 4o and it goes all in with saying you'll be the bestest friends ever and all sorts lmao

Opus 4 actually just says no basically

[solid brook]

I like this all ai models should be like this

People got so desprate for 4o when it was gone for 2 days

This will have nagetive effects

[quiet dust]

Guys, how to use the Thinking model in GPT-5, not Thinking mini?

Or is only Thinking mini available for free users?

[keen beacon]

They are still complaining about how it doesn't feel like 4o, like it's secretly gpt 5 in disguise

the whole chatgpt subreddit is on fire

[sullen quest]

GPT 5 mini high is worse than 4.1 and 5 chat, gpt nano high is worse than oss at 42'nd place in text arena

flagship model my

[keen beacon]

saddening...

[tired herald]

Sorry to say this

But the small error ballooned into a much larger set of issues so I cant release the src today

But I promise tmrw will be the day

[tired herald]

GPT 5 high

[novel crater]

the first personal fun project I made like actually was able to fully complete was done with 06-05 Gemini, I am a Gemini fan for sure

[hollow imp]

@novel crater

[novel crater]

yeah Gemini 3 would be freakin awesome

basically I made a tarkov grid style inventory, if anyone here knows that game, basically a pretty complex system and Gemini knocked it out of the park

I don't think chatgpt likes their competition at all and probably put out gpt 5 as like this big thing, because not a lot of people seem overly enthused about it

[balmy mist]

what happened to deepseek man

[marsh stratus]

Ok but mini and nano are very much not flagship models

[neon idol]

Deepseek was too strong and he was nerfed 😭🙏

[sullen quest]

They are what the average chatgpt user gets though.

[scenic salmon]

Sam has known for quite some time that GPT-5 was never going to live up to expectations, GPT-4 was this big model scale up from 3.5 and it saw huge improvements with that scale up, so people have been expecting the same from 5 since then, but it was found you can’t really scale them up any further and get meaningful improvements, so they started focusing on their reasoning/thinking models (o3/o4/etc), they tried launching another giant sized model, 4.5, but people didn’t really like it, so it was known for quite some time, whatever GPT-5 was going to be when it released was going to be more of a rebranding than anything

[sullen quest]

Didn't stop him from pretending it was the death star

Sometimes people need to realize when they shouldn't be hype men

[scenic salmon]

Gotta get those VC dollars

[novel crater]

yeah I mean it seemed like it was made out to be huge with their website and stuff, I mean they removed all the other models at first

[marsh stratus]

Gpt-5 chat is the consumer facing flagship ChatGPT model, and I think that it’s going to go down like maverick 4 in the totality of its failure

[sullen quest]

costs saving, 4o seems to cost more to use

[marsh stratus]

Nano and mini seem built for api use. Gpt-5 chat IS ChatGPT fo r most people, and it’s just so bad

[novel crater]

fair yeah thats a really big part of it I think

cost friendly is always good

especially with some of these models 👀

[whole wagon]

GPT5 without thinking is so bad

I don't know how it's even possible

It fails basic arithmetic

That even 4o could get

I'm basically permanently using thinking

Only downside is wait time between prompts

[stray aspen]

@novel craterwhat will come first gemini 3 or grok 5

[scenic salmon]

GPT-5 mini is a decent balance, it’s not as dumb as the chat model

[novel crater]

I don't have enough knowledge on that to say definitively but since grok 4 came out recently it would make more sense logically that Google would release first

[whole wagon]

Gemini 3 October, grok 5 December

That's my guesses no info

The gap between pro and regular gpt5 is quite insane. They didn't even mention pro in the livestream lol

I guess they don't really want ppl using it much. Due to capacity issues

The gap between gpt5 and GPT5 pro much larger than gap between o3 and o3 pro I found

[zinc ore]

This one's weird because mensa Norway caps out at 145 IQ

[inner gate]

Hello famalams

[whole wagon]

It doesn't?

It's just that it is rated for only up to 145

They didn't calibrate beyond that

[pure falcon]

Does anyone else think that LMArena should go back to NO style control? Now they’re also talking about “emotional control”…

Makes no sense - it defeats the purpose and mission of LMArena in the first place. The more they “control”, the more they turn into the very benchmarks that they wanted to distinguish themselves from

If LMArena is about what users like…then let the users decide what users like!

The more variables and factors you “control”…
You end up turning into a Capabilities test

And There are plenty of capabilities benchmarks out there already

[whole wagon]

Where can I find the emotional control discussion

[pure falcon]

https://news.lmarena.ai/sentiment-control/

LMArena is supposed to give insight into what other benchmarks miss

If you keep “controlling” for those unknown factors by adding style control, sentiment control, etc….the benchmark becomes worthless. No alpha in the scoring

[hollow imp]

Where is style control?

Gimme style control

[pure falcon]

You already have it! Style control is built into the default rankings

[pure falcon]

All I’m saying is, if people like emojis and bolded words or whatever, then you have to accept that, regardless of ether you think they’re dumb or meaningless

In fact, we’re seeing right now why style control is a mistake for LMArena’s clients

[whole wagon]

It's to prevent a race to the bottom where every model turns into sycophant emoji slop machine

[pure falcon]

OpenAI announced that they would make gpt-5’s personality “warmer”

Because it’s obviously been too “style-control” maxed

[pure falcon]

Take a step back and think about what LMArena is, and what its value is

[scenic salmon]

Just put that in #ai-news

Trying to work their way back up the charts

[pure falcon]

The whole value prop of LMArena, is extracting sycophantic signaling!

Again, the more you control for user preferences, the more you turn into any other standard benchmark

Making LMArena useless

[hollow imp]

I thought you're saying we can add custom instructions or at the very least a style mode like formal explanatory

[patent aspen]

The thing is: they're already controlling for a bunch of other things. Style control is controversial because it's a visible option. The other controls aren't because they aren't visible.

[pure falcon]

LMArena is supposed to captures what objective benchmarks cannot. ie, user preferences. ie, what users like. ie “sycophantic” behavior

[pure falcon]

What kinds of things is the “no style control” leaderboard controlling for?

[frigid coral]

I don't think so. LMArena is supposed to capture model's capability in general use cases. Sycophantic behavior is an undesirable side effect

[pure falcon]

Yeah my guess is OpenAI relied too much on style controlled rankings in testing, and realized that was a mistake

[pure falcon]

Wrong. There are hundreds of benchmarks that do that already

[pure falcon]

LMArena is not a capabilities test. It’s a popularity contest

[patent aspen]

Mainly integrity checks. Certain patterns of user behavior are more likely to be trying to manipulate rankings to favor certain models, and they have a relatively robust system to account for that

[frigid coral]

None as representative as LMArena

[scenic salmon]

Which doesn’t mean it’s not valuable information, it’s still incredibly valuable

[pure falcon]

Agreed! And they should keep it that way. If they keep adding stuff to control for, LMArena’s value disappears entirely

[pure falcon]

That’s different -that’s a prompting issue, not a voting / output problem

[ornate agate]

Yeah filtering scam submissions seems like it’s not a manipulation

[pure falcon]

Also, the models vary wildly by language. 50% of LMArena’s queries are not in English

[patent aspen]

I don't think style control is manipulation either. I think it's just a slightly controversial option with compelling pros and cons to consider

[frigid coral]

Where does language factor into sycophancy

[ornate agate]

I meant a fairly benign version of the word manipulation

[pure falcon]

you can have a model that is very sycophantic in English, but maybe not in other language language. Or maybe rewarded in English more than in other languages

[leaden palm]

There's some stuff you can easily "turn up and down" and stuff that you can't
I'd argue that a leaderboard that only measures the second one would be even more valuable for some uses

[frigid coral]

Do you have any evidence of that happening? I believe it's not true

[hollow imp]

@pure falcon talk to @echo aurora about all this

[ornate agate]

So filtering scam submissions seems to be a necessary baseline function of the website or it would be a joke given how popular it is. Style control isn’t, it’s a choice to have it or not. I’m not sure it’s up to date for recent models

[patent aspen]

Yeah I mean personally I'm not the biggest fan of style control, although I think it has pros and cons and is useful as an option. You get some form of model provider gaming either way

[pure falcon]

Idk what you mean by this?

[leaden palm]

Eg a system prompt can increase formatting but can't increase raw intelligence

Or it can increase sycophancy but can't induce EQ

[ornate agate]

When it was originally made I kind of liked it as a guard against llm markdown listslop. Now though I feel models like Gemini and GLM answer in this very verbose way, but the tech has improved so the lists actually have benefit/are more info dense now.

[patent aspen]

Yeah I kind of want all the models to have fancy markdown now

[whole wagon]

Removing style control would cook openAI

[leaden palm]

I wonder if they intentionally steered away from using the things style control tracks

[pure falcon]

But thats the whole question right? Is LMArena measuring intelligence? Is it SUPPOSED to be measuring intelligence?

IMO, its primary purpose is to inform model makers on user preferences. So they can improve their own models.

If an ASI model came out tomorrow but was a total a- hole that everyone hated, where would you want that model to be ranked on the leaderboard?

[ornate agate]

This is gonna happen it’s gonna answer 42 to everything really quickly unless you ask it something sufficiently interesting.

But well, such a system will be in the somewhat far future

[pure falcon]

Yes, A model that behaves like that is undesirable to model makers. Model makers want to know what users like. So if you “control” all that away (with sentiment and style control, etc), you end up with a leaderboard that is SO useless to model makers

[pure falcon]

Well, unless OpenAI is LMArenas only revenue source, then that should be the way to go

[patent aspen]

IMO a few models will eventually pull ahead by a wider margin at which point the style control problem will gradually become moot because the best models will be able to win even with extensive markdown

[whole wagon]

They can simply press a button to remove style control

Style control helped in the llama 4 situation which was favourable

[pure falcon]

Why would they do that though? They care much more about raw user preferences. There are plenty of other benchmarks (internal and external) that can measure raw capabilities for specialty / niche areas

[ornate agate]

I don’t think lmarena is just user preference. It’s a whole bag of stuff

[pure falcon]

Like what? What value, to model makers, does LMArena provide that other benchmarks don’t?

[whole wagon]

Are there really. A lot of benchmarks are saturated and/or useless

[pure falcon]

I really would like to know, genuinely

[ornate agate]

I mean I don’t care about what value it provides to model makers. I care about what value it provides to the community.

[whole wagon]

?

If they really would like to see no style control. They simply press remove style control, simple

[pure falcon]

Sure, but you could argue the exact opposite, right? If no style control is what matters more, then why is style control the default?

[patent aspen]

fwiw I also prefer no style control as the default

[pure falcon]

That’s the thing though - LMArena isn’t a “community” service project. They have enormous costs servicing top of the line models for free

They are propped up by VC funding

[ornate agate]

I also prefer no style control.

[ornate agate]

Don’t worry about that. They might not be paying for it anyway.

[pure falcon]

Who are the leaderboards most valuable to? I’d argue it’s to companies like OpenAI who use their leaderboard to determine which model to deploy. Which is exactly what they did for GPT-5

There was an alternate gpt-5 variant named “zenith” that was tested in late July

OpenAI ended up going with the “summit” variant because it did better in LMArena

If they didn’t already, OpenAI would pay a lotttt of $$$$ for that data

Which means

For LMArena

Their primary customer and purpose, is the model maker

[gritty stirrup]

Why does Claude Opus have a limit on lmarena?

[warm fulcrum]

its expensive

and every model has a limit

[gritty stirrup]

I didn't notice that other models were reaching the limit, but maybe you're right and I just didn't reach it. I didn't write that many messages and I already have a limit.

[patent aspen]

@echo aurora We've been discussing whether or not LMArena should have style control enabled by default. I know this topic has been beaten to death, although I'm curious to understand LMArena's take

[leaden palm]

that isn't really what i meant - i was saying that a leaderboard that measures the "core" of a model might be interesting to look at

"If an ASI model came out tomorrow but was a total a- hole that everyone hated, where would you want that model to be ranked on the leaderboard?" if it's better than the average model prompted to act that way, and if it can be prompted to act like whatever other model, then on a leaderboard designed to measure the core - the parts of the model that aren't superficial - it would rightfully take first place

[patent aspen]

Personally I think there are pros and cons and prefer no style control by default because I see LMArena as primarily measuring user preference

[pure falcon]

Model the core what? “Intelligence”? What does intelligence even mean? In that case, why not just feed it IQ tests, math questions, or other capabilities tests, which already exist all over the place? What value does LMArena therefore have?

Do you see what I’m getting at? The more you control for, the more you approach benchmarks that already exist. That defeats the purpose of LMArena IMO

[leaden palm]

What does intelligence even mean?
what can't be reduced by controlling for attributes of the response / forcing a specific style
why not just feed it IQ tests, math questions, or other capabilities tests
have you heard of benchmaxxing? this would theoretically be unbenchmaxxable

[patent aspen]

I think the best argument for style control is if your goal is to be the best all-in-one measure of intelligence in the areas that users care about

[ornate agate]

Intelligence is just a vibe

[echo aurora]

I believe we did a blog post about this recently...

I could be wrong

[ornate agate]

So style control is a vibe realignment from the default one, more towards a different vibe that people might want more…

[patent aspen]

I think no style control is more valuable. That's a better reflection of who is "winning" with users

[leaden sun]

a good way to test how well models follow the control instruction, isnt it, or is this the original intention of lmarena?

I think less control is better

but it depends on what purpose style control serves and what you want to test that you need style control as a normalizer

[ornate agate]

Yeah. I think it’s a bit of a fools errand to try and find some function to morph current arena votes into “intelligence” or something. It would also need to be continually updated I think.

[ornate agate]

It doesn’t affect llm responses it just affects how it’s scored in the end.

[pure falcon]

Why is communication not counted in your “intelligence” definition? Are comedians not “intelligent” in your eyes? Are entertainers not “intelligent”? Intelligence is not just about solving problems and puzzles. That’s a very narrow and limited view of intelligence. A model highly skilled in communication is, in fact, intelligent, just in a Different way than Einstein. Intentionally dismissing those capabilities (by “controlling” for it) makes no sense.

[leaden palm]

if i can't prompt another model to be "highly skilled in communication" then the model would have an advantage

there are actually a few ways to do this though and all have some problems tbh

• prompt all models to act a fixed way
• prompt one model to imitate the structure of another model
• identify the traits of each model's responses and control for each trait (implicitly assuming that each trait is easy to reproduce)

[ornate agate]

Btw do llm arena models have system prompts?

If so who sets them?

[leaden palm]

iirc it used to be that each model had a fixed (usually minimal though) system prompt, and now there's a common system prompt that says how to respond but it's different for some models

idk 100% though, i don't develop the arena

[ornate agate]

It’s not suspicious on its own. All chat apps have a (often large) system prompt

[ornate agate]

So it’s decided by lmarena what to put in it and it’s short?

[leaden palm]

i believe so

eg the o series needs something like enable markdown formatting iirc

[ornate agate]

Ah ok. That makes sense

[scenic salmon]

They added it pretty quick

[tired herald]

Damn, I should add this to my extension too

[misty vault]

But most importantly: @deep adder said it is the best by far

[tired herald]

Exactly my thoughts

[echo aurora]

Sorry I'm late to share this -> https://news.lmarena.ai/sentiment-control/

[hollow imp]

@leaden palm please some prompt engineering tips

[leaden palm]

for...?

[tired herald]

From what I checked, it seems that its either a very very simple one, none at all, or a base system prompt from the provider (gemini gave me a believable system prompt when I ehhm forc- made it give me its system prompt, all other ai's said they didnt have one/gave me mine back, which prob means mine was the only one)

Tho I cant confirm that unless I had access to the internals of LMArena

[ornate agate]

Yeah that would make the most sense. Simple as possible to stay close to the original model.

[willow grail]

IIRC cursor had big context issues 1 year ago. is this still the case, now with gpt5 high?

[fervent tangle]

why do OpenAI models suck so bad at creative writing

and why is Claude so good at it

[wintry tinsel]

Cuz open AI is the big poo poo, the prime crapola!

[rugged swan]

new here, excited to learn from you all

[fervent tangle]

🥀 fr

[wintry tinsel]

Learn how to do obscene things with Gemini?

🔥

[misty vault]

paws stop typing

[fervent tangle]

i used all openai models since 2023, even GPT5 sucks today

and the emoji blasting is corny

i hate the emoji respondings

[jade egret]

👀

[quiet dust]

What?

[fervent tangle]

google needs to release genie 3

and its game over

[jade egret]

fr

[fervent tangle]

google is actually the winner, cuz they got all the money in the world + all resources

it is actually

[jade egret]

?

when u have no money, no resources and your gonna win?

[fervent tangle]

google has the best image, chat and video models rn

wydm it isnt

[quiet dust]

Guys, does anyone know how to use the Thinking model on the phone in ChatGPT, not Thinking mini? It's just that the Thinking mini model provides the "Think longer" function.

[jade egret]

i dont see how this is gonna happen

[fervent tangle]

they literally cant lose if they do that

[jade egret]

are you payed option?

google?

[wintry tinsel]

Oh yeah baby and triple

[fervent tangle]

u can't the GPT5 chooses if it should think longer based on ur prompt

[quiet dust]

I'm on the free plan

[jade egret]

google can win

[fervent tangle]

it has already won

[jade egret]

than i dont think u can use thinking

[jade egret]

ig

😭

[patent aspen]

Google has a much, much longer runway than OAI, and it's not even close

[fervent tangle]

and in a few months they'll release Gemini 3

[jade egret]

xAI seriously

[fervent tangle]

it will crush the coding benchmarks

fr

[jade egret]

why u hating on google tho

[patent aspen]

And eventually sheer capability just wins

[jade egret]

wdym no product

[quiet dust]

There is a button. The "Think longer" button.

[fervent tangle]

they'll use it if its the best at all stuff

[patent aspen]

OpenAI is far less innovative than Google, and it's not even close

[fervent tangle]

doesn't even matter, they get their income from other stuff like google search, youtube and all google owned markets

they can throw money on AI until they get big

[jade egret]

they can just put gemini in chrome which they already started?

[patent aspen]

The wider the gulf in the capability, the less product polish matters

[fervent tangle]

u know the google cfo?

[jade egret]

if you count all the x users?

[fervent tangle]

cuz its "not woke"

[jade egret]

well than why not include ai overview and ai mode

[fervent tangle]

and u can easily jailbreak grok 4

[patent aspen]

I mean if you're going to make that argument, you have to include AI Overviews

[fervent tangle]

grok 4 is easily controlled unlike claude, openai and google models

[jade egret]

grok turned into a h*tler fan 😭

[fervent tangle]

it got no security

[jade egret]

FR

[fervent tangle]

tbh even tho claude 4.1 opus is the best at coding, creative writing. majority of people still use chatgpt

[jade egret]

true

[fervent tangle]

because normal people dont care which AI is the best

they just download chatgpt and use it

[jade egret]

still think google can win tho

[fervent tangle]

we're like 0.01% of all humans

if ur into AI and coding

the majority follows the slop

they dont care if its good or bad

[sullen quest]

honestly it might be better than some ealier versions, where it pretended it was writing a epic fantasy tale

use ai google studio if you are using 2.5 pro

[jade egret]

the gemini is different?

[sullen quest]

different instruction set

yah, its not suprising that the api version has less instructions then the non api version though

I have no clue what's going on with the front facing version since it never can act normal

[patent aspen]

I'm going to be honest. The ChatGPT web UI looks just like the Gemini web UI except with worse iconography

[keen beacon]

be me
can't choose which large language model to install
ask a friend
friend says chatgpt
download chatgpt
run chatgpt

https://tenor.com/view/giga-chad-gif-23143840

[patent aspen]

The responses?

You've got a few weeks at best

[fervent tangle]

yeah and its good enough for everything at the moment

[patent aspen]

It doesn't matter while the quality gap is small. It matters more as the gap widens

Nah they're about 1/3 to 1/2 a generation behind

Capability

Nah

Better in some areas. Worse in others. Overall parity

[keen beacon]

To my surprise, sure

[patent aspen]

But the point is that their pace isn't enough

They're roughly 1/3 to 1/2 a generation behind

Nah

Look if you want to compare bad ChatGPT responses vs bad Gemini responses, we can do that. I just don't think it would be a useful conversation

Do you just not like the bold sections?

[keen beacon]

Guys

[patent aspen]

I'm not sure what to say. Most people prefer paragraphs with bold headers

I do think verbosity is on the chopping block though

[keen beacon]

Did you know that you can compare the answers by SOTA and non-SOTA models to determine if the models you ask are right

You just ask say Qwen2.5 the same question 10 times in a row

Then Qwen3

And if the responses tend to differ, one LLM is clearly in the wrong here

[misty vault]

be me
talk with bing chat sydney

distilling from sydney fine tune before it gets shut down

[uneven lance]

'Ello guys, is there any site that generates videos using Veo3 for free?

[rare python]

For me, I don't care about the writing style as long as the LLM I'm using flawless at instructions following, consistent at long context. Gemini 2.5 Pro did very well about first 10 messages, then it broke and went back to bullet list, which I instructed to be banned in systen instructions.

[verbal nimbus]

Default ChatGPT version of GPT-5 is on there now

[rare python]

It isn't as a hype yes man as gpt 4o so it got lower elo 💀

[verbal nimbus]

Lower than GPT-4.5 too

[rare python]

This is killing me 😩 Please ship faster and better quality

maybe the thinking but the chat version is 💩

chat is the non thinking version right? It's just GPT 4.1 in cloak

and it's really bad at IF

[elder solar]

is there any api that has the 3 version of gpt 5

[rare python]

I don't care about benchmark I only care how well it follows my own system prompt

livebench rarely match my real world usage

then why did you show it as a proof of how well GPT 5 at IF 😭 💀

[elder solar]

theres not alot trustful benchmarks of ai

or prob its depend on the prompt

[rare python]

Most trustworthy benchmarks are your own

[elder solar]

maybe

who knows

[rare python]

🗿

[elder solar]

however

i still hope theres another ai model that can listen audios

that is not gemini

[inner gate]

Gemini can hear u?

[elder solar]

yeah ik

but there should be more ai models that could be it

like chatgpt

[fast halo]

only me who takes ages to load a response?

[jade egret]

?

gemini?

or direct chat

[fast halo]

direct

[lucid kite]

Good evening! I’m new here. Could you please let me know if you plan to add video generation to the site in the future?

[sinful vessel]

what happened with nano banana model?

was removed?

It doesn't appear to me anymore, before it appeared to me all the time in battle

[drifting thorn]

I guess nano banana is going to be released

I hope so

[whole wagon]

Is Google cooking smth up

Apparently they had a celebration when they saw GPT5 performance

Lol

[wind vector]

just curious, but why limit video generation / arena to discord? discord ai stuff is the worst lol

[whole wagon]

https://www.anthropic.com/research/end-subset-conversations

[hot lance]

image to video is too 💩

[hot lance]

Maybe ,the cost of data for downloading videos is very high.

[wind vector]

ahh that might be the case

[echo aurora]

welcome welcome! It's possible we add Video Arena to the site in the future, we're treating this as an experiment using Discord. #bot-feedback is where we're collecting thoughts on it.

[echo aurora]

Vid Gen is more expensive, so we do want to limit it a bit. Currently it's set to 8 generations a day.

We taking all of this feedback into account though! So it may change.

[plucky palm]

I think Discord is great, it also works as a funnel for making announcements, since you’re not collecting emails or requiring sign-ups on a website.

plus i like the community

[echo aurora]

We'll always have a Discord for that kind of stuff fortunately.

[echo aurora]

agreed & seeing our site is about gathering community preferences there should be a place to chat about it.

thinking vs non-thinking versions

[plucky palm]

exactly

[stray aspen]

any gemini 3 news

[verbal nimbus]

Lmao

Hopefully their next model is better at agentic coding

[drifting thorn]

I hope Gemini 3.0 can further extend their context window for even better agentic usage and tool use in agent. Ability of creative writing is also being anticipated by a lot of AIRP users

Gemini 2.5 Pro still holds the SOTA in long form multi-turn creative writing in actual usage but I wish they can even make it better

[wind vector]

Personally I don't see what use there is of a longer context window when they can barely utilize their full 1 million so far

They need to improve attention to the context they have, imo

Especially for creative writing, airp, or even coding in some cases. Approaching 200k tokens and you see a ton of ai'isms / quality degredation / "amnesia" creeping in

[drifting thorn]

YES

[wind vector]

Seems like a hard problem to solve, though

[drifting thorn]

I mean, they have to be train to have a consciousness of "rounds of conversation", like "this context is from the first round", "this context is from the second round" etc

[unique cave]

which model is ranked first overall for this month?

[lilac inlet]

Guys! I'm sharing my learning here, which I recently made up from my non-coding mind. I mean, I don't even know HTML.

I just discovered a way to create unlimited web apps, widgets, desktop apps, and so on for free!

We need two tools: Lmarena and Weblmarena, that's it!

Any person from a non-technical background can easily create fully customised software at no cost. With the steps below that I discovered lately!

1. Pick a pen and copy and write down about your desired app, mentioning every single detail that pops up in your head. For this, remember this rule, 5W1H (What, When, Where, Who, Why and How). If you have a better strategic model, then you can implement it here. It's totally up to you how you want to describe your app!

2. Click photos of your notebook pages and upload them to any LLM model and ask it to transcribe your images or pull the contents from them!

3. Once you have the text content, Copy it and open your lmarena and select the modal qwen coding one and paste all of your content there and also add your custom instructions how you want it to be built, for an example, the widget I made was a react js component + Typescript in node js page, same way you have to give instructions and ask your desired output in single page written code!

4. Once you have the code, Open Weblmarena and paste the code in the text box and hit enter! There you go, it will take 2-4 minutes to render your code and show you the preview of your website!

5. For Iteration and Bug error fixing, head back to your lmarena and open the Qwen coding thread where you have your previous chat. Now, you can iterate on it, fix your errors and so on. Repeat step 4 again and keep on doing it until you have the desired output!

If anyone has a better way to execute such ideas for a non-technical person, kindly share it. It would be really helpful for a newbie like me who knows nothing about tech stuff! 🤗

[frigid coral]

https://eqbench.com/spiral-bench.html

Gemini 2.5 Pro highest in sycophancy

[verbal nimbus]

Interesting, but didn't expect to see Claude below Gemini

It's even lower than 4o here

[frigid coral]

claude seems to perform really bad at mania tests

[keen beacon]

It's been only a week since I found out LMArena and I jailbroke two things already

Bruh

[fervent tangle]

calm down, ive known lmarena for a year

and they still the same

[keen beacon]

Kinda worrisome tbh

Turns out most people would rather have sycophantic delusion inducion machines than truth seeking arbiters such as GPT-5

I guess it can explains some ratings pretty well

[white hatch]

I hate the safety!

[leaden sun]

what's the definition of "safety" here?

[solid brook]

Yeah a very bad future we're heading to

[leaden sun]

without a precise definition of words like "safety, risks" in the context of AI, this is nothing but pseudo-science, isnt it?

[quiet dust]

Why don't I have a button where I can change the GPT-5 model?

I heard you'll be able to switch between Fast, Auto, and Thinking. But how?

[autumn cargo]

Why can models with lower scores end up higher? Isn't CI a symmetric thing?

[leaden sun]

if you read how they described this bench, you know it's wishy-washy science https://eqbench.com/spiral-bench.html

[keen beacon]

Aghhh, give me gemini 3

[keen beacon]

Because they want AI to act as a "bestie"

and now because of that minority, OpenAI is making GPT 5 "warmer"

[leaden sun]

hahahah

[ocean vortex]

Ohh Opus is top3 now. Seems Anthropic entered the game of user preference fine-tuning now. Opus 4.1 has bigger gains over 4.0 than Sonnet4 has over 3.5.

[reef bridge]

can't lmarena do function calls on the chat other than search

[calm trout]

Is the video generation here powered by Veo 3?

[brittle tiger]

https://x.com/IndraVahan/status/1956650892788191552?t=Q816aqHymv6huA6xls5hgw&s=19

Indra (@IndraVahan)

i think we’re looking at a breakthrough in image gen.

nano-banana is literally picking its own spots to edit. changes flow with the context + prompt. doesn’t recreate the whole thing like gpt-image-1 or step1x-edit. it's like flux-1-kontext, but way sharper on context and prompt

X · Aug 16, 2025, 9:35 AM

[copper furnace]

Hello im preatty new here, is there any limits on how many images/videos you can generate?

*pretty

[past shuttle]

gemini king!

[exotic nebula]

Yes. 8 exactly.

[exotic nebula]

Based.

[exotic nebula]

No, its powered by many video gen models. Its randomized, think of it as the exact version of battle mode.

[copper furnace]

Ok thanks

[hollow imp]

Gemini 2.5 and grok

[plucky island]

"truth seeking arbiter" and "GPT-5" in the same sentence? what?

[hollow imp]

Absolutely not

[ionic idol]

not working

[hollow imp]

Use nano banana

[full burrow]

Looking forward to seeing nano-banana on the chart.

[plucky island]

btw who made nano banana and where is it other than lmarena?

[hollow imp]

I have no clue

[drifting thorn]

I thought Google made it

[hollow imp]

Google would never put such a freaky name

And they already have veo

[scenic salmon]

They ruined it…

[stray aspen]

Lmao

[native gate]

Error: Something went wrong while generating the response. Please try again. Always when I paste some long code php

[stray aspen]

@tired herald what's the fix

[keen beacon]

I hate AI psychosis people

[neon idol]

Hello

[neon idol]

Me too

[keen beacon]

Hei

[neon idol]

Yo

I want deepseek r2

But they nerfed it 🙁

[keen beacon]

Me too... But one thing I want more is not to make GPT-5 a servile model

aka too positive and eager to please

but OpenAI is doing it

stated in a tweet

[neon idol]

Me still waiting for Gemini 3 amd r2 💀🙏

[keen beacon]

I hope R2 will wreck the whole leaderboard like with R1

in January

[neon idol]

Yesh I hope

WE WANT THE RED CHINA DRAGON ATTACK 🗣🔥

[hollow imp]

🔥

[keen beacon]

To my surprise Qwen and Deepseek answer the same question better when asked through chat, not API. Deepseek and Qwen answer my music theory question 7-8 times out of 10 and Deepseek even infers almost correct conclusions another time which makes it more like 8.5-9/10 if I'm more liberal with scoring.

In LMarena when stress tested with the same prompt 10 times they both answer 3-4/10 at best.

It seems like the workload can influence performance.

[hollow imp]

But I cannot access o3 without lmarena

[keen beacon]

o3 still gets it correctly even at LMArena

OpenAI is the most demanded AI startup on this planet either way

They know how to scale their toys

Deepseek however seems to know about it a bit less

Holy hell if Deepseek in the chat version is actually that much better than the API version then I can only imagine what is going to happen once they will rent out enough compute to run R2

It's not there (the o3/GPT 5/Gemini 2.5 level), but it is very, very close

[obsidian cargo]

I want Deepseek v4. V3 is the best model available on AI Dungeon atm but I'm a bit burned out of its limitations.

[warped totem]

hey, yesterday i lost all my chat sessions for some reason. can i retrieve it ?

[keen beacon]

Or maybe it is because LMArena proxies to Deepseek API through US servers and I am in Russia and get different entry points kek

[hollow imp]

Deepseek r2 will not be gemini 2.5 level?

[warped totem]

is gpt 5 high any better than gemini 2.5?

[sullen quest]

depends on the gemini 2.5 and depends on the gpt 5

[warped totem]

what u mean

these on llm arena

[sullen quest]

There's multiple versions of gpt5, high, chat, high nano and high mini

I used to think gemini 2.5 was slow, then gpt 5 high came around

[warped totem]

so i told "gpt 5 high"

[solid brook]

i mean you can choose the robotic personality

[solid brook]

yeah dude lol

even gpt 5 medium

[sullen quest]

depends on the task

[warped totem]

most of the time i have a feeling gemini 2.5 give better responses

and its little faster

[sullen quest]

I've heard that gemini 2.5 is better at different langauges compared to gpt 5

[ocean vortex]

They know a lot. But they are constrained by CCP forcing them to use crappy Huawei chips lol

[sullen quest]

idk I don't speak other languages so I can't confirm

[sullen quest]

its also the us goverment forcing them too

[warped totem]

yea, thats true - sometimes he makes typos

[keen beacon]

Sometimes.

One answers some of my questions better than another one. Depends on the question asked.

But they are roughly head-to-head in most things.

[ocean vortex]

Well yeah but they were able to go around that (import restrictions) to be fair

[keen beacon]

If this is true this is sad.

Teortaxes wrote about it being fake news and so on. But here in Russia we actually have this exact problem, and being a Russian he sounds like massive cope to be honest.

But really

[solid brook]

dude gpt 5 high is a lot better than the lobotimized gemini 2,5 we got

[sullen quest]

Only so much and it makes the chips more expensive

[keen beacon]

We in Russia learnt to evade this problem by importing foreign hardware and rebranding it as domestically produced one

[solid brook]

the benchmarks you see are from the original gemini 2.5 pro that was not lobotimized(the first month of release)

[keen beacon]

We call it importozameschenie

[sullen quest]

True dat

[ocean vortex]

They were still able to make it work. Now however having to use Huawei as well this is much more challenging. And there's no messing with CCP 👀

[ocean vortex]

Original 2.5Pro was a fair bit less capable than the current version. People should really stop believing these conspiracy theories.

[keen beacon]

Sure lmarena benches are from the og 2.5 too?