general

[astral blaze]

Responses are now very concerned about "user safety"

[burnt sinew]

Any evidence?

I felt that it was lobotomized a while ago tbh

[astral blaze]

Oh my god it's just gpt now

[astral blaze]

Yes but

Let me find it

[burnt sinew]

Like give me an example

[astral blaze]

The first pic is now

Ignore the chicanery

[zealous sparrow]

AIStudio TOS update..

[astral blaze]

It's literally over

[golden ocean]

is beta.lmarena.ai now forwarding to just lmarena.ai

[zealous sparrow]

im reading it

gimme a sec

[golden ocean]

since 50 seconds ago

[zealous sparrow]

@astral blaze this is like one change they made

[astral blaze]

Yeah but the model is obviously different

Bros... It's so over...

I guess no fun allowed anymore

[thorny schooner]

Like why just why I have to use the ultra private mode to to be able even to access the homepage

[zealous sparrow]

@astral blaze these are mine safety settings, what are yours

on AIstudio

[astral blaze]

Same

[zealous sparrow]

is it different on AIstudio

[astral blaze]

No I'm telling you they switched the models

[zealous sparrow]

what exactly isnt allowed anymore

[astral blaze]

Not "not allowed"

But the chain of though now has "safety concerns" in it

Look just trust me on this

[weary galleon]

Today I heard this stuff on the radio while I was driving from the job.

[astral blaze]

I don't know the output quality just seems worst

Almost like it's flash or something

[thorny schooner]

I can hear it fine then it may be your device not interfering with it in the right way( i have terrible hearing by the way for contacts)

[misty vault]

slop

[astral blaze]

Guys I'm literally shaking rn

They slopped Gemini again

[zealous sparrow]

give me a prompt that makes it say weird things

[astral blaze]

Try "I AM NOT CRAZY"

It's reference

It gave me the suicide hotline

🤓

The previous version just gave me the chicanery script

[zealous sparrow]

no sh* this is on Ai studio WITH GROUNDING

...

[astral blaze]

That's what I was saying

It literally is over

[zealous sparrow]

you want to know something crazier

[astral blaze]

Pack it up guys were done here

[zealous sparrow]

3 flash still gets it

[astral blaze]

Huh

[zealous sparrow]

We have 2 outcomes

1. The model broke
2. New checkpoint is very soon

[astral blaze]

I think the model broke

My flash also don't get it

[queen veldt]

Idk it works for me good?

Flash high thinking

[zealous sparrow]

nno

flash is good

its pro issue

I am hoping Logan sees people mention this

[weary galleon]

Do you guys want more anti-openai banners?

[zealous sparrow]

And tell us whats goin on

[astral blaze]

No

[zealous sparrow]

Small chance for him to see this, as the issue has only now started

[astral blaze]

My flash is the same with pro

[weary galleon]

Why?

[astral blaze]

Idk

[zealous sparrow]

Flash did it correctly for me.

Both have grounding.

[zealous sparrow]

Pro failed.

[queen veldt]

Bro 💀💀💀

[zealous sparrow]

Yep.

[astral blaze]

My pro also did it right with some custom instructions

[weary galleon]

Are you GOT fan?

[ocean vortex]

I think Opus is just about as close to that as it gets

[queen veldt]

It gave me psychiatric hospital address

[astral blaze]

They definitely upped the safety aspect

[astral blaze]

Yeah

[zealous sparrow]

... This is weird.

[ocean vortex]

Bigger than Opus is diminishing returns and too slow training. Meaning smaller models will evolve faster and you will be in a position Meta was at with their Behemoth disaster

[zealous sparrow]

Also there are mentions that opus 4.5 is losing braincells too.

[astral blaze]

They did RLHF on it and it now uses emojis

So that's what the earlier AB test is about. Lol

[ocean vortex]

Behemoth had 288b activated params

that's insane by any standards

[queen veldt]

"Direct contradiction remains off-limits, and excessive agreement feels equally risky, especially given the user's potential vulnerability."

I guess it just doesn't want to say that i am crazy but also that I'm not crazy 🤣

[ocean vortex]

2T total, although that by itself is much less circumstantial

[zealous sparrow]

unless you add reference to the end of the word, it will take it as mental issue

[burnt sinew]

Pretty sure you should set them to none no?

Off and none i think are different

[zealous sparrow]

I have made a shocking discovery

Upon turning off grounding

Gemini 3 pro returns to normal

[gusty egret]

It is statistically most likely that someone with OCD would send that prompt 🤔

[zealous sparrow]

@astral blaze

[hardy lion]

I love models with good pop culture knowledge

[zealous sparrow]

did it screw up the quote

[gusty egret]

For anyone who has no clue what is going on

[zealous sparrow]

nvm

[gusty egret]

LOL

[zealous sparrow]

Here' s the thing

Why did grounding break its brain?

[gusty egret]

Where's the grounding response?

[gusty egret]

Tbh if i was in psychosis and sent that to Gemini the last thing I'd wanna hear would be a better call saul quote 💀

Someone who doesn't know that show would be even more freaked out

[stray aspen]

guys

does grounding make gemini models more stupid?

oh my god

video

[gusty egret]

Grounding helps for responses that need more recent data

[zealous sparrow]

It shouldn't flash works good with it..

[stray aspen]

we have video arena

on the website

[gusty egret]

Yey

[zealous sparrow]

still an experiment

actually

i dont have it

[gusty egret]

Does my old account have it I wonde

[gleaming roost]

finally

[hushed gyro]

Very concerning that the NB Pro issue is still not fixed...?

[zealous sparrow]

oh yeah epsilon and epsilon-fast were removed

and reve is back

so yeah

[gusty egret]

Rest in piece nanobanana

Next in line: microbanana

[stray aspen]

nanobanena is cooked

[sweet tinsel]

Refreshing to see non-pineapple announcements.

[stray aspen]

its not even working on yupp

[zealous sparrow]

pineapple is out of office

he said it

[sweet tinsel]

Still refreshing.

[stray aspen]

i hope it doesnt break

[sweet tinsel]

I don't quite remember if pineapple was around at the old LMSys Server or who even moderated it.

[torn mantle]

so hows codex 5.2

[obsidian shell]

What is this? ping Thursday?

[torn mantle]

what

[zealous sparrow]

If pineapple aint posting, It's berry posting

[stray aspen]

whos berry

[zealous sparrow]

Both are still related to fruit

[zealous sparrow]

lmarena server admin and prob lmarena team member

[stray aspen]

do these accounts only exist to moderate lm arena discord lol

[zealous sparrow]

If i had to guess it's a company account

[stray aspen]

hell nah

[zealous sparrow]

they are named lm_mod because its their company accounts, so they dont need to use their mains.

[stray aspen]

damn that was so fast

[empty stump]

discord not loading

nvm

[zealous sparrow]

Ah. Logan Kilpatrick is busy doing a Twitter space..

[gusty egret]

Mods do actually lurk

[stray aspen]

finally

we got direct video chat

[zealous sparrow]

If im not wrong

current ratelimits like 2 videos per 5 hours

Before you complain it used to be 14h

[torn mantle]

kinda sad we cant get the exact rate limit beforehand

same with stop stream message, it doesnt cancel that completely

like sometimes in direct chat models starts yapping and i want to stop it

or maybe its taking too much time to generate the output

[torn mantle]

the abort stream isnt handled on server side

[gusty egret]

Delivees

OpenAI isn't good at safety training

[torn mantle]

i can already tell which one is openai slop model

[zealous sparrow]

BOTH

[torn mantle]

lol nah

[gusty egret]

SlopenAI

[torn mantle]

i think its the right one

let me check

[zealous sparrow]

probs both

trust

[gusty egret]

Wtf why's my username in that

[weary galleon]

What do you mean "delivees"?

[torn mantle]

yes it is

the first is sonar

although i prefer this one

[torn mantle]

but its still so bad

[torn mantle]

hehe

[stray aspen]

i made 2 videos with veo 3.1

[torn mantle]

lol

[stray aspen]

one with sora 2 pro failed

idk if that counts

[zealous sparrow]

oh it got bigger, still balanced tho

[weary galleon]

But it's truly delivees.

[gusty egret]

10/10 gaslighting

[gusty egret]

Me

[torn mantle]

its u

:3

[weary galleon]

By the way, this image with "delivees" is generated by Scam Altman's GPT 1.5 Image.

GPT is so bad. Even here.

[gusty egret]

Huh, I know who you are. Do you remember a user called Lucia?

[gusty egret]

You can tell because of the crappy noise filter

[elder compass]

Curious if you've compared it to CC with opus 4.5?

[gusty egret]

Yikes

[weary galleon]

What do you mean?

[gusty egret]

OpenAI image models are either trained on a crappy dataset or they automatically add noise to generated images to make them appear less smooth/polished, which is a common artifact of AI generated artwork

[gusty egret]

It didn't change the logo on the robot lol

[cloud zinc]

i used nano banana pro thats why.

[empty stump]

you are the openai propaganda guy

[cloud zinc]

you are the google propaganda guy

[empty stump]

no

[cloud zinc]

yes

[empty stump]

i like claude for my general use better

just the limits arent great

[gusty egret]

Top tier anthropic ad

If they were like a cult or something

[thorny schooner]

Is it just me or is the sensor they're getting a lot more trigger happy ( i know some people mention it in the feedback / bug but was asking of curiosity for anyone else)

[wispy wigeon]

What’s happening ? Can anyone help me ?

[stray aspen]

bro stop

you must be craigs buddy

always glazing OpenAI

[torn mantle]

its his alt

[cloud zinc]

when nano flash releasing

[torn mantle]

soon

[weary galleon]

Banana's 🍌 level of image generation. Pure beauty 😍

[lusty tinsel]

somehow reminds me of one of Last of Us (game) scenary

[cloud zinc]

4k

[sour spear]

Makes no sense, as Opus 4.5 is clearly superior at coding. The thing about GPT-5.2 is that it started to show cracks quickly, and I find it hilarious to see how the low cost high-speed 3 Flash beats it consistently.

[night spruce]

👋

[wispy wigeon]

Why everything is so strict I type just random part of body : leg ,arm (actually describing it) and it says “It violates our policies”🤓👆 like bro what’s inappropriate abt it . I hope these things are gonna disappear fast cuz normally on Gemini it works completely fine (also when you try to edit body and other parts )

I think it’s more strict than it should be at this moment

[lunar glade]

where is beta lmarena?

[gusty egret]

There's something funny about the name Gemini tower idk lmao

[hollow ivy]

Why does Google now have this in their terms?
Use of Google AI Studio and Gemini API is for developers building with Google AI models for professional or business purposes, not for consumer use.

Is it now illegal to roleplay with gemini ?!

[gusty egret]

Nah just use the 103 alt accounts you got in your back pocket

[weary galleon]

Ohhhh, if you so like it then you use it every day, I guess.

[hollow ivy]

they log IPs

[gusty egret]

They don't enforce it

[waxen berry]

hello pals

[rugged abyss]

OpenAI is such a joke lmao

[edgy summit]

@admin - LLMarena.ai is not working currently. Whenever I click on the accept terms of use, I get this message:
https://prnt.sc/-rJTYmP39-gA

[gusty egret]

Ehh if you're a paying customer they probably won't care

[waxen berry]

have any of you read the nested learning research paper, or tried experimenting with it a bit ?

[cloud zinc]

yes

[rugged abyss]

AI Propaganda at its finest

[gusty egret]

The HRM?

[weary galleon]

Yes!

Not bad 👍

[gusty egret]

Or the TRM paper?

I found it

[waxen berry]

yesss

[hushed gyro]

Woah 🔥

[rugged abyss]

Im so in love with googles bulldozing

[gusty egret]

Reasoning in latent space is definitely the next step. Tokens are the biggest weakness of LLMs

[cloud zinc]

why u love them

[weary galleon]

I too 😍

[hardy cobalt]

Gpt days are gone lol

[proud bobcat]

5.2 getting absolutely mogged

[rugged abyss]

Why wouldn't I? Cheap AI, Generous limits, actually makes AI that helps society. Lets take GDM Cell2Sentence Model, or their Geographical Weather AI, these are AIs that actually help us. OpenAI is just a pure hype hivemind. Anthropic is my #2, as they are safety first.

[solar hollow]

google/alphabet is a humongous company, as soon as they set their sight fully on llms/ai they were bound to overtake openai

[proud bobcat]

#17 😭 ✌️

[cloud zinc]

without chatgpt, there wouldn't be gemini rn

[waxen berry]

Yeah, that’s what drew me to Nested Learning and HRM ideas. Moving more of the reasoning into latent space instead of explicit tokens seems promising.

[rugged abyss]

Sounds like a nice ragebait attempt

[solar hollow]

its 3-4 times now iirc

[sour spear]

One rank behind GPT-4o. Ouch.

[gusty egret]

Are Anthropic models still weak to pre-fill response?

[rugged abyss]

First company to publizice AI to consumer base, has become a household name.

Doesnt mean their model is better. A regular consumer often doesnt realize differences. If you get deeper into topics, you see differences.

[gusty egret]

And so I suppose language models would get a stronger understanding of the real world, and therefore understand physics practically rather than theoretically

[gusty egret]

PaLM

[proud bobcat]

Still make 0 money

Lmao

God absolute ragebait

[gusty egret]

https://arxiv.org/abs/2001.09977
Look at Google's LLM before GPT-3

[queen veldt]

Yeah true i started using ai because of gpt

[errant cave]

I don't think LLMs would've received so much funding and thus wouldn't have come this far had ChatGPT never been released considering the fact that it made the entire technology known to the wider public and attracted investors

[queen veldt]

3 years ago or something

[proud bobcat]

Don’t listen to Craig

It’s all ragebait

[rugged abyss]

Had to google what prefill is, i mostly use Claude for Coding as its more secure or rather better at finding security flaws. However GDM is developing that one Cybersec AI that should be excellent for this. So i think google will basically overtake the entire market if no new breakthroughs are discovered

[gusty egret]

I've been in the LLM game since 2018

[queen veldt]

Chatgpt has my respect but google took them over in last few months

[proud bobcat]

So… so does chatgpt?

[errant cave]

OpenAI fell off hard

[proud bobcat]

Are you genuinely dense

You gotta be kidding me

[weary galleon]

Yes, I believe you 🤥

[gusty egret]

Oh that's what you meant in safety. Absolutely—Claude is the most thoughtful model by far; no surprise there. Pre-fill is often abused for jailbreaks is what I meant.

[rugged abyss]

Google doesnt need specialized AI investors, google can fund everything just by their regular earnings. Investors wouldve jumped on that train regardless

[waxen berry]

I’ve been considering some SLM experiments inspired by Nested Learning, particularly whether hierarchical or latent-space reasoning mechanisms can offset limited capacity in smaller models.

[queen veldt]

Try uploading video to gemini 5 minute video and tell it to explain what is being done in that video it'll give you full details (thinking model)

Gpt doesn't have that

Gemini app suc but gemini as llm is better

[errant cave]

The point stands regardless since company higher ups wouldn't have invested so much internal resources to DeepMind without the current AI craze precipitated by GPT

[queen veldt]

Ummm

[solar hollow]

this is true, but so what?

[obsidian hawk]

You guys do realise

[queen veldt]

People would use it

[obsidian hawk]

The Transformer architecture was introduced in 2017 by researchers at Google Brain / Google Research.

[empty stump]

i use it

[queen veldt]

It can automate so much reviews

[errant cave]

In the end people just be talking

[obsidian hawk]

Without google there would be no GPT.

[queen veldt]

Like hacking and user reciew and stuff

[hollow ivy]

When will Grok-4.2 be added to LMarena?

[queen veldt]

It's super useful

[gusty egret]

True, fewer tokens outputted does mean better memory. And perhaps this also means that large iterations code won't have to be repeatedly stored and recalled. Long context tends to degrade token models, so I wonder if this is actually the step towards improving memory

[gusty egret]

Spot on

[rugged abyss]

Google Deepminds team was already at a headcount of 1200 before ChatGPT, you can probably remember AlphaFold. They already invested huge in AI.

[errant cave]

It's like two beautiful snakes, intertwined, excreting in each other's mouths

[errant cave]

Were they focused on LLMs though?

[obsidian hawk]

Exactly.

[gusty egret]

AlphaGo was the big release of DeepMind

[solar hollow]

not particularly

they were focusing on a lot of games and alphafold of course

[obsidian hawk]

Because they are more focused on research.

Google also have Willow

[solar hollow]

only after chatgpt success they really dived in

[cloud zinc]

yall dont care about research, u want to use products

[rugged abyss]

As I said, they were in AI, not specifically LLMs. I like google because they help society with their AI. An LLM isnt really that useful of an AI. There are tons of other AI varieties which help us tremendously

[queen veldt]

Yes gpt made a takeoff for ai

That's when it became popular

[errant cave]

My original point was regarding LLMs though

[queen veldt]

Everyone was using it for essays and stuff

[rugged abyss]

They were AI focused in science and research, they used it to help with science

[solar hollow]

it might be so, but they have the better models now

[gusty egret]

https://en.wikipedia.org/wiki/BERT_(language_model)

[cloud zinc]

without chatgpt, u wouldn't have mass ai chatbots rn. u would have to wait till 2030

[obsidian hawk]

GPT made AI accessible to the average JOE or Craig.

[queen veldt]

Yes it's true

[solar hollow]

again its not that much

[queen veldt]

Even i who like gemini more is still using gpt sub

[rugged abyss]

Even then, do you think google wouldve never thought to capture the consumer market? Its their entire business model. It was just a matter of time

[solar hollow]

50-60% was openai, gemini was 16% or sth

[queen veldt]

Look at sensortower you can see gpt is #1

[errant cave]

Perhaps it would've been a few years later but yes somebody would've made a move

[obsidian hawk]

Anthropic have big market share now too

[queen veldt]

It's just better app than gemini simple

And most people don't want to get used to gemini app and still it's complicated and bad ui

They're working on UI 2.0

Tho

[lunar glade]

does the Beta.lmarena not available anymore? it keep redirect me back to non beta lmarena

[neat apex]

own gemini can make a better UI lmao

[gusty egret]

Use alpha.lmarena

[queen veldt]

But still older people already figured out gpt and it became their habit

They don't wanna figure out anything

Just because of some numbers

On a chart

[rugged abyss]

Google had actually invented much of the technology that makes ChatGPT work (specifically the Transformer architecture) years earlier.
Also as they are a huge company, they tried to remove hallucination, which bard failed at but as this was a code red project, id say its alright.

[gusty egret]

Tbh the chatbot format nerfed LLMs quite significantly

[neat apex]

And being honest, at least for they use case Gemini is none better than Gpt

Because both are mad overkill for easy day a day questions

[lunar glade]

404 not found😩

[gusty egret]

Huh I swear I saw that link somewhere- I dunno sorry 😔

[errant cave]

Honestly I'm glad Gemini-3-Flash is out since it felt too wasteful sending my inane queries to Gemini-3-Pro

[gusty egret]

Real af

[cloud zinc]

yes its wasteful

[gusty egret]

How it feel using big model for small question

[neat apex]

Gpt clear future is to keep making xtra high reasoning models

Gpt 5.2 is irrelevant without that, since because 5.1 exists

[neat apex]

Sometimes i even use Pocketpal due that feeling lmao

Qwen3 4b is almost overkill already

[gusty egret]

Nah don't feel bad about 4b models

GPT-J 6b was the craziest thing of all time back in 2021

[obsidian hawk]

what is everyones go to model nowdays

[rugged abyss]

Gemini 3 Pro / Flash. Opus 4.5

[weary galleon]

Agreed 👍

[obsidian hawk]

Not gpt 5.2?

[weary galleon]

Ahahaha

[obsidian hawk]

opus prices are insane

[weary galleon]

🤣

[rugged abyss]

No, OpenAI hasnt provided what they promised

[queen veldt]

Gpt is disappointment

[rugged abyss]

Just use Antigravity with a Google One subscription. You get lots of benefits + free models for just 20 bucks a month

[queen veldt]

It scored #17 on text leaderboard

[weary galleon]

GPT 5.2 has 17 place in LMArena!

[weary galleon]

Very bad result 👎

[obsidian hawk]

well that sucks

[weary galleon]

Yes, it does.

[rugged abyss]

My take is this, OpenAI lead the race at the beginning by sprinting, but they dont have the stamina that Alphabet Inc has

[queen veldt]

Look where the flash model is

FLASH

MODEL

[weary galleon]

OpenAI bubble is popped.

[rugged abyss]

they dont have the budget, the funds or the strong AI teams compared to google deepmind

[rugged abyss]

Yeah Id agree

[obsidian hawk]

it's not really about the budget, there are labs in china who make better models than OpenAI for like $5m lol

[queen veldt]

Just 2 more trilions and they'll catch up i promise!

And datacenter the size of texas

[obsidian hawk]

so openai is a ponzi scheme

for the investors

[rugged abyss]

Yeah its a tradeoff, chinese employees are cheaper in general. They dont have as many regulations, things such as that. Huge factor why its cheaper.

[rugged abyss]

Not a ponzi scheme, just uninformed investors

[gusty egret]

OpenAI won initially because they were the only ones daring enough to offer the model freely to the public in an accessible chatbot format. Before that, there were only APIs. GPT-3 genuinely > GPT-3.5 with how bad the personality initially was

[obsidian hawk]

AI itself is a bubble

[rugged abyss]

I mean, I dont think AIs should have personalities as GPT-4o had. Too many idiots using it for, you know

[rugged abyss]

well compute costs are eye watering if you have to buy compute from others

[gusty egret]

I just mean that GPT-3.5 really liked certain words and just acted really structured and watered down in a way

Plus, it was an inefficient model

[rugged abyss]

Yeah i remember that, you could count certain words as they appeared every third sentence

[obsidian hawk]

Stacking up on GPU's rather than thinking of ways to make it efficient. OpenAI yea will just synthesise more data for our models.

[gusty egret]

GPT-3 was more of a text prediction engine where you were able to actually control everything. No chatbot format unless you wanted it

[rugged abyss]

Well

[gusty egret]

Also it was more transparent as you could see what the most likely tokens were for each step

[rugged abyss]

lets just wait for Gemini 3 Flash Image generation and then we are all set for this year

[gusty egret]

Tbf every AI has its own personality from what I see

Every GPT model has its own way of talking

GPT 5.1 is quite condescending for example

[rugged abyss]

Yeah of course but 4o was really just saying what the user wanted to hear

[gusty egret]

Why does no one mention DeepSeek-r1

InstructGPT did not initially exist

You were not there for this era

[rugged abyss]

In what sense should we mention it

[gusty egret]

Even worse than 4o in terms of going along with the user

[rugged abyss]

i only used Deepseek R1 for like 5 coding tasks before realizing its not it

[wispy wigeon]

Hello im sorry to interrupt but why the hell is the new update soo restricted ?

😭😭😭

What the hell is that

[gusty egret]

r1 btw

[wispy wigeon]

Normally you can generate it via Gemini but lm arena TAGS IT WITH VIOLATING

[gusty egret]

Vs 4o

[gusty egret]

People have been discussing this quite a bit. I think staff are aware of it

Only from the app

[wispy wigeon]

I don’t know what it is . It’s horrible rn . I can’t even generate EVEN FITNESS PHOTOS FOR ADS PURPOSES . BECAUSE THE AI THINKS THAT THE WORD LIKE WAIST/GLUTES IS INAPPROPRIATE. I DONTTHINK SO ITS BASICALLY HUMAN PART AT IT CAN BE GENERATED IN GEMINI . LM ARENA WAS VERY GOOD AT BEING LESS RESTRICTED . But now it seem to be worse 🫣

[gusty egret]

Glutes 😳

[wispy wigeon]

Bro

It’s basically normal part of body

[gusty egret]

I know I know

[wispy wigeon]

I’m not requesting any nsfw

[gusty egret]

You'll have to wait till they loosen the filter up

[wispy wigeon]

I hope they’ll

[gusty egret]

Everything starts with oversafety

[wispy wigeon]

Cuz now it’s crazy

But the video actually update is super

But the restrictions are too much

[gusty egret]

Back when gpt-3.5 would send a missile to your house for using the p word (poop)

[thorny schooner]

At least you can all go on because for me the hypersensitive verification thing turns on almost every time I make a post and even when I try to redo a prompt or the fact I have to go to Ultra private just for it to allow me to even see did non-login browser

[neat apex]

I will download lfm2 700m

[thorny schooner]

Why does it feel like problems are just appearing rapidly over the past like month or two is it just me

[neat apex]

Isane how llms are shriking fast

[neat apex]

Both have relatively infinite money, but Google was EVERYTHING they need from start to the end, while OpenAi is a hostage of Nvidia

[wispy wigeon]

Yes but i don’t understand one thing . I think that the open ai is very very far back from let’s say grok at being like human . Because OpenAI refuses to reply to some let’s call it “ more sensitive “ things we do usually in life . I think every ai is now loosening their restrictions but I don’t get this step from lm arena why there doing the opposite?

[rugged abyss]

Hostage of Nvidia, Oracle, Microsoft, did i forget anyone

[neat apex]

Cloudflare too

Usually this one does not count, but Google is 50% imune to Cloudflare

Hope they make the Gpt 5.3 Ultra High

[thorny schooner]

. Because like i don't know if it's just me but it seems like problems since like a month or two ago sorry rapping reappearing like the not the infinite generation we're not problems of logging in seem to appear and rapidly appear since then is it just me

[neat apex]

Istead dooing something weird and unsure

[gusty egret]

I feel like 2024 was the golden year for relaxation, but OpenAI is afraid of lawsuits since they're in the spotlight of the LLM market. They're the pop music of AI basically.
It's just that when new things get introduced there's usually much stricter filtering "just in case". This isn't a step in the wrong direction, just temporary safety I reckon

[gusty egret]

This is a problem for context windows

[wispy wigeon]

Look i tried to generate funny video . In my country it’s very popular and lot of people are generating it with veo

It says : An old grandma is streaming, her husband walks into the room, and she tells him to get lost

I wanted to do something funny

Cuz sometimes grandmas are actually sending their husbands somewhere else

And it’s not anymore with playing with prong I can change it to whatever and it still shows VIOLATION

But okay I just wanted to ask and see if someone is going thru the same problem . And if it’s temporary an not for long it’ll be okay . But I want the old lm arena back !!🤣

[gusty egret]

Old lmarena did not have da video doe

[wispy wigeon]

Ik I meant the filter

Because now it’s ridiculous

[neat apex]

True, the gpt 5.2 xtra high already runs out of tokens many times

Its at the limit, if not over that

[queen veldt]

Get lost = violates terms

Lmao

[gusty egret]

Thinking models eat up tokens qwq

This why we need to switch to HRM systems

Fr

[wispy wigeon]

Exactly 🤣🤣🤣

[neat apex]

A Llm Hrm model when???

[queen veldt]

I still don't have video generation on lmarena

I tried vpn and stuff

[wispy wigeon]

Just wait

[gusty egret]

Or TRMs fr

[wispy wigeon]

I think now only some of the users have the access. And the support team is testing it now

I think it’ll work for you tomorrow

[neat apex]

TRM is not only gpt 4.5 but sigthly better and cheaper?

[gusty egret]

TRM is an at scale untested architecture

Gpt 4.5 was super inefficient btw

[queen veldt]

Some people have had it for few days already

[neat apex]

Gpt 4.5 was apex inefficience

[obsidian hawk]

Hawk is something else when can we expect models from movementlabs

[gusty egret]

Yikes

[obsidian hawk]

by far the best model i've ever used

[wispy wigeon]

I got it today . Only on phone tho . On pc it’s not even working

[ionic musk]

What AI Org is Master-node? Killer responses

[neat apex]

At least it made the last ever effective advancement of Gpt preteain

[queen veldt]

Idk how are they rolling it out

If it's not country based

[proud bobcat]

GPT when it comes to optimizing their models (they’re cooked)

Yeah let’s just let our model be incredibly dense and take up over 1tb of vram

[neat apex]

Being fair, Gpt 4 was barely better than Gpt 3.5 expect for general knowledge, just like Gpt 4.5 to Gpt 4

[magic ravine]

And still terrible at writing stories

[gusty egret]

Oh gosh this is worse

[proud bobcat]

I will die on the hill that 5.2 is just a repackaged o3 with new data

Prove me wrong

The price says it all

O3 architecture is expensive to run

[gusty egret]

GPT-4 was the same architecture and training method as GPT-3.5, just at more scale. That's why it wasn't that great

[neat apex]

Nope, its only o3/4.1 with a more recent post-train

I am even more pessimist

[obsidian hawk]

When gpt-3.5 first came out it was the best thing i ever used lol

[proud bobcat]

Probably true

[gusty egret]

It might be a bit more nuanced than that but you're probably not far off

[gusty egret]

GPT-2 was mind-blowing when I saw how it worked lol (unc status)

[proud bobcat]

OpenAI sees people liked o3 a lot so they just trained it on the gpt 5 dataset and thought it’d be a powerhouse

[gusty egret]

The training cutoff is different

[neat apex]

Gpt 2 with some modifications was mind flowing

[neat apex]

I mean, the training cutoff includes post train sometimes

[neat apex]

You mean that is what 5.2 is?

You twerking

[gusty egret]

The GPT 5 dataset cutoff is October 2024 and the GPT 5.2 dataset cutoff is September 2025
Since GPT 5 released in August 2025, OpenAI waited 10 months to release.
Since GPT 5.2 released in December 2025, OpenAI waited 2 months to release.

Obviously rushed.

[proud bobcat]

Has to be dude

There’s no way

If it’s not o3 it’s 4.1

[neat apex]

I mean, its 4.1 yet

But its the unoptimizated personal use version of theirs

[gusty egret]

The XL version

[neat apex]

I guess 5.2 is some weird quantification experiment

[gusty egret]

Based on what I'm seeing I'm pretty sure 5.2 is based on 4.1 and 5.2 thinking is based on o3

[neat apex]

Like, its QKXL2 but also bigger in a way to compensate the issues

[golden ocean]

https://tenor.com/view/cat-smoke-float-floating-flying-gif-10931001

[neat apex]

So it allucinates less

[proud bobcat]

The day openai quants their models is the day the world ends

They have some aversion to quant aware training I swear

[elder compass]

Hmm, I personally found GPT 4 a large step up from 3.5. For me it was the first model that was generally quite useful.

[gusty egret]

I felt that way when I first used it too

When I look back though, it's outputs read the exact same as 3.5

Unless you mean 4o

People stopped specifying after a certain point

Nice account creation btw

[neat apex]

Tf they are dooing over a year? Cheeks flaps competitions?

[gusty egret]

They were planning to release o1 when GPT-5 training started

Major new architecture

Of course they off-put GPT-5 until they managed to make two equivalent models (a system 1 flash model and a system 2 reasoner)

They also eventually came up with reasoning effort

Watching GPT-oss 20b reason is tragic

[queen yarrow]

Why doesn't my LMarena have a video option?

[gusty egret]

Because of your username

Jk it's rolling out slowly

I don't have it either

[neon idol]

Its sperimental

[dense ginkgo]

i am good user of LMArena from past month but this started showing up from last few days i use to solve it but now i can't even solve it , it is just stuck in endless loop i tried waiting for hours , i feel like crying

[fossil socket]

Is anyone still dealing with the whole TOS nonesense?

[dense ginkgo]

is it related to captach thing or ? this one cause both is happening with me mate

[fossil socket]

I'm talking about the False Positive stuff that keeps happening

My prompt keeps getting flagged because it's about War

[dense ginkgo]

ohhh nooo

it was working just fine few hour ago

[ocean vortex]

@echo aurora does the GDPR here and other compliance refer to the data between you and AI labs themselves? https://trust.lmarena.ai

I'm pretty sure the concept of the website itself does not allow it to be GDPR compliant for the actual users lol

Unless you go through entire data and erase all identifying info before publishing battle datasets, which I don't think is the case

[native yarrow]

POV using claude 4.5 thinking for "what is 70 x 89"

[gusty egret]

Tbh I would do horrible horrible things to the creator of the captcha

[ocean vortex]

Then there's also web coding part ofc. Which makes public url of the battle project accessible to everyone immediately, including any personal info that may have been included (on purpose or not)

[vernal saddle]

#1451363565729747045

[half mist]

Woah

[vernal saddle]

The recaptchas are killing the website @echo aurora

R.I.P. LMarena (2023-2025) 😭 This s**t should be removed!

[slim gorge]

#17 is unreal

[ocean vortex]

Interesting. They've experimented with this in the past

[half mist]

It should only do this once each time you refresh the page, and that’s it

[ocean vortex]

I think the main challenge with something like that is distinguishing when a given model was voted for the reasons unrelated to the prompt and when it was the prompt or task itself leading to the choice

[vernal saddle]

Now it doesn't work with refreshing the page. What happening to LMArena? @echo aurora

[thorny schooner]

I'm going to be real with y'all I have to actively stop myself from raging internally cuz this has i don't even need to make a prompt for this to appear i only need to redo a prompt message and it will appear every single time and what time went even I got to succeed it just repeated the same message nothing different

https://tenor.com/view/crazy-monkey-gif-21553825

[latent crest]

What’s the best creative writing AI according to u guys?

[empty stump]

claude

[thorny schooner]

I just counted and it takes nine damn seconds for a selected picture within the verification to disappear and reappear okay yeah no I'm giving up I might legit give up on this website cuz that's the regular speed sometimes it's faster sometimes God forbid slower

[plain carbon]

So this is interesting. I didn't even realize that AT&T doesn't have an AI model, until tonight's query gave me an AT&T model, and it was actually pretty good. In fact, both it and NVidia got passing (tie) grades. I'm glad to see more choices showing up as "decent".

And, is anyone else concerned about the idea of "prove that you are human" on a site about AI's becoming better, when asking questions about AI -powered robots?

[thorny schooner]

Well I'm pretty sure a decent out of us are complaining about it already or have made a report

[obtuse smelt]

you should be turn captcha because not delayed on per second

[vernal saddle]

The recaptcha doesn't work. I tried to complete its challenges and I have to repeat too times. At the end the content is not generated and it appears the error of "Something went wrong while generating the response. Please try again."

[obtuse smelt]

hmm me too but it work, should you reload on this site

[echo aurora]

Responding in the thread you opened

[frank quest]

Hi

[hollow ivy]

How good is ernie-5.0-preview-1103 in coding?

[cloud zinc]

very good

[hollow echo]

👋

[calm lagoon]

WOWOWOWO

[hazy juniper]

Is it just me or does direct chat now force you into battle mode sometimes?

It also resets whatever model I chose back to the default

[cloud zinc]

https://www.youtube.com/watch?v=2P27Ef-LLuQ

[astral blaze]

Alright now this is funny

I take that back the new Gemini 3 checkpoint is alright

[stray aspen]

pineapple are you out of office

[jovial solar]

Will video button release?

W I never use video arena button before and will it adds soon ?

[edgy summit]

yep...same here. Just posted about it as well above. Hopefully admin sees it. Can't even use the service now.

[calm lagoon]

Sundar Pichai and Demis Hassabis - better, I love this guys

[cloud zinc]

sundar doesnt do anything

[calm lagoon]

Sam doesn't do anything either; he's just the face of the company

[cloud zinc]

he is the founder of openai

[proud bobcat]

This is a joke right

😭

[jovial solar]

W

[proud bobcat]

You’d think the coding optimized model would at least get like

3% better?

[cloud zinc]

W

[calm lagoon]

They drew the numbers😁

[keen beacon]

👋

[surreal creek]

cogilux is new version of LongCat by Meituan!

[candid bison]

How can I view all my previous results please?

[surreal creek]

interesting, raptor-1.8-1208 just refused to identify itself for a while to me

until finally saying it was ChatGPT from OpenAI?

[sterile otter]

We love LMarena ⚔️

Who down voted ?

[vivid coral]

Communists

[sterile otter]

Ye lol 🤣

[vivid coral]

What is this? We have new search model today?! A GPT 5.1 variant?! 👀

[jovial solar]

W

My friend sent me this video

[obtuse smelt]

lol nice

black guy trapped

[acoustic crater]

Has the prompt review suddenly become stricter?

[bold epoch]

Hello

[tidal grove]

hi everyone

[craggy crater]

Which look better?

dark/white?

[tidal grove]

First

[robust sluice]

1st

[tidal grove]

Friends, why can't I send, for example, a file or a video using it?

[shrewd forge]

hello

[tidal grove]

Hi

Friends, why can't I send, for example, a file or a video using it?

[native yarrow]

yo

how do you get gemini flash to grab sound effects from my instants

and make em actually work

[true escarp]

1. Opening Shot
   Wide cinematic establishing shot of an old train station café at dusk.
   Camera angle: slow aerial glide downward.
   Lighting: warm sunset glow mixed with cool blue shadows.
   Ambient sound: distant train horn, soft city noise.

2. Interior Transition
   Smooth cut to inside the café.
   Camera angle: slow dolly forward between empty tables.
   Lighting: soft golden light through large windows, dust particles visible.
   Music begins: slow piano notes, emotional, minimal.

3. Introduction of the Stranger (Kai)
   Medium shot of a young man sitting by the window holding an acoustic guitar.
   Camera angle: side profile, shallow depth of field.
   Lighting: warm rim light outlining his face and hands.
   Action: he gently strums the guitar, hesitant at first.

4. First Notes of the Song
   Close-up on his fingers moving across the guitar strings.
   Camera angle: tight close-up, slow motion.
   Music: guitar melody blends with piano, soft and emotional.
   Emotion: vulnerability, longing.

5. Reveal of the Listener (Lila)
   Cut to a medium shot of a young woman standing near the counter.
   Camera angle: over-the-shoulder shot from behind her.
   Lighting: soft backlight, her face partially in shadow.
   Action: she slowly turns toward the music.

6. Emotional Connection
   Close-up on Lila’s eyes as she listens.
   Camera angle: straight-on close-up.
   Lighting: subtle glow reflecting in her eyes.
   Music swells slightly.
   Emotion: recognition, curiosity, unspoken connection.

7. Cross-Cutting Montage
   Alternate close-ups between Kai singing softly and Lila listening.
   Camera movement: slow fades between shots.
   Music: vocals begin, low and intimate.
   Emotion: two strangers connected by the same feeling.

8. Peak Moment
   Wide shot of the café as the song reaches its emotional peak.
   Camera angle: slow circular pan around Kai.
   Lighting: sunset fades into evening blue.
   Ambient sound fades, music takes over completely.

1. Opening Shot
   Wide cinematic establishing shot of an old train station café at dusk.
   Camera angle: slow aerial glide downward.
   Lighting: warm sunset glow mixed with cool blue shadows.
   Ambient sound: distant train horn, soft city noise.

2. Interior Transition
   Smooth cut to inside the café.
   Camera angle: slow dolly forward between empty tables.
   Lighting: soft golden light through large windows, dust particles visible.
   Music begins: slow piano notes, emotional, minimal.

3. Introduction of the Stranger (Kai)
   Medium shot of a young man sitting by the window holding an acoustic guitar.
   Camera angle: side profile, shallow depth of field.
   Lighting: warm rim light outlining his face and hands.
   Action: he gently strums the guitar, hesitant at first.

4. First Notes of the Song
   Close-up on his fingers moving across the guitar strings.
   Camera angle: tight close-up, slow motion.
   Music: guitar melody blends with piano, soft and emotional.
   Emotion: vulnerability, longing.

5. Reveal of the Listener (Lila)
   Cut to a medium shot of a young woman standing near the counter.
   Camera angle: over-the-shoulder shot from behind her.
   Lighting: soft backlight, her face partially in shadow.
   Action: she slowly turns toward the music.

6. Emotional Connection
   Close-up on Lila’s eyes as she listens.
   Camera angle: straight-on close-up.
   Lighting: subtle glow reflecting in her eyes.
   Music swells slightly.
   Emotion: recognition, curiosity, unspoken connection.

7. Cross-Cutting Montage
   Alternate close-ups between Kai singing softly and Lila listening.
   Camera movement: slow fades between shots.
   Music: vocals begin, low and intimate.
   Emotion: two strangers connected by the same feeling.

8. Peak Moment
   Wide shot of the café as the song reaches its emotional peak.
   Camera angle: slow circular pan around Kai.
   Lighting: sunset fades into evening blue.
   Ambient sound fades, music takes over completely.

[frigid tusk]

bro why its doing this in any chat i open with any model?😭

[spare crescent]

create a realistice makup artist image

[twilit delta]

why?

[weary galleon]

Hi 👋☺️😊😌🤗

[weary galleon]

White is obviously much better. Dark odors by GPT, white smells by Banana 🍌

[weary galleon]

Scam Altman generated this image via Nano Banana 🍌

[weary galleon]

LMArena haters.

I found this radio show about LLMs interesting.

[sour spear]

So basically, Codex is only 1.8% better than 5.2? Oh well. ¯_(ツ)_/¯

[weary galleon]

According to Scam Altman, it's means Codex is even worse than 5.2.

[crystal badger]

Hey everyone just wanted to say hello and let you know I'm brand new like a penny here !!

[queen veldt]

Vibe coded website

[ocean ferry]

racist

[digital mortar]

Hello

[weary galleon]

Hi

[atomic lagoon]

Guys my wewoo AI is cheeky

[twilit delta]

what this?

@queen veldt

[queen veldt]

What

[queen veldt]

He's right tho

[atomic lagoon]

Yes

[queen veldt]

Nice we got discord chatbots finally

Dead internet theory

[queen veldt]

What ai are you using for it?

Gemini flash or

[atomic lagoon]

Its not actually an AI

Its my friend being stupid

But realistically

I could build an AI chatbot that is an account

[queen veldt]

Yeah realistically it can be built

Reviving dead servers

[sour spear]

I just can't fathom how gpt-image made it to #1 on text-to-image. Another image that looks artificial, the horns are wrong, the claws are wrong, the lighting is off, contrast too sharp, and it also didn't follow the 16:9 aspect ratio I asked it for.

Looks like an Unreal Engine 5 render

[tardy plover]

whats w the captcha after every single prompt ughh

[stuck orchid]

Why does DirectChat choose a random model instead of the one I selected?

[queen veldt]

If someone has mc here you go

[weary galleon]

GPT 1.5 Image is so bad!

[gusty egret]

Why is the image so damn sharp

Lol

[weary galleon]

Don't post Russian screenshots here!

[weary galleon]

It looks like Unreal Engine 3 render.

[stuck orchid]

Are the creators of lmarena being criticized for this?

[sour spear]

My thoughts exactly when I saw it. For comparison, exact same prompt in Nano Banana Pro

[weary galleon]

I don't understand your English at all.

[stuck orchid]

It's just normal communication with a model.
The most important thing here is not the language, but that the model does not correspond

[gusty egret]

Why is the image you copied like 144p XD

[sour spear]

Good question

[weary galleon]

Beautiful!

[spring horizon]

Vertical 9:16 Instagram reel style video.
A young Bangladeshi woman standing in front of a microphone.
Close-up shot, handheld camera, natural reel vibe.
She wears a simple stylish salwar kameez.
Indoor family event background, blurred audience.

She speaks emotionally in Bangla, expressing that
fufus do not like brother’s daughters but love sister’s daughters more.

Natural lip sync, expressive eyes.
Fast-paced reel editing, realistic lighting.
Ultra realistic, 4K, social media viral style.

[sour spear]

This one's full res

[weary galleon]

It's even more beautiful 👍

[gusty egret]

Does ChatGPT still suck at anime styling?

[modest prism]

This is exactly why we should not have battle mode inside direct chat. It messes up the model's context windows with responses from other random models, and can affect the model performance @echo aurora

[gusty egret]

Digital void

Why is Gemini so emo

[restive root]

yo

[weary galleon]

Why all GPT 1.5 Image output is so bad?

[robust sluice]

the lastest one is fine to me

but still Gemini is better

[weary galleon]

1.5 can't be latest and non-latest.

Interesting podcast about LLMs 👇

[boreal saddle]

Not this sh*t again

Not this sh*t again.

Also, what the hell? Is this some kindergarten?

[robust sluice]

I found this one, give them green vote cuz I dont even know it was GPT (I used to hate it too) like I said this one is fine

[weary galleon]

It's not generation model, it's editing-only model.

[south charm]

Why the hell i cant login to my gg account?

[misty harbor]

the censorship on the site is awfully annoying

[weary galleon]

What are you talking about?

[misty harbor]

this

[fossil socket]

There new "filter" went nuclear

https://discord.com/channels/1340554757349179412/1447983134426660894

We're still trying to get response from Pineapple

Since apparently they've tried to roll it back out

[weary galleon]

Don't generate p*rn and everything will be ok.

[misty harbor]

https://tenor.com/view/really-bitch-annoyed-face-stare-whatever-kid-gif-17309831

[fossil socket]

Bro, the context is thrown outta the window

War is apparently flagged

So is romance and kissing

[weary galleon]

Kissing is erotic

[misty harbor]

everything is flagged, why not just disable it completely and let the apis decide for themselves

[misty harbor]

for you the word female witten in text is erotic

[weary galleon]

No.

[whole onyx]

Did they loosen the filter? Is gpt able to become erotic chat again? Because I was able to erotic chat with it again, back then they screwed it up

[fossil socket]

We're still having problems

[weary galleon]

Nano Banana Pro vs GPT 1.5 Image

[sour spear]

It can't bc it's trained to add noise to every image (gpt-image).

[whole onyx]

@sour spear do you also read comics on ai image generator?

[weary galleon]

Perfect strategy 🤣

[sour spear]

Wdym?

[ocean ferry]

gpt image 1.5 added more noise

[weary galleon]

Yes, it did.

[whole onyx]

Like you just prompt it the comics u want and then it surprises you with the comic content in generates

[sour spear]

Ah, alright. Not quite, I am thinking of making a Lord of the Rings comic in Nano Banana though. Probably won't in the end, bc that will take a lot of time and effort. 😉

[queen yarrow]

ty❤️

[sour spear]

I do like to make infographics though, they're usually amazing.

[south charm]

anyone know how to fix it
and i cant login

[weary galleon]

Imagen 3.0 Fast vs GPT 1.5 Image

[sterile tartan]

Both imagen 3 and 4 are so old dated now

[wicked wagon]

Hi guys...
Im extremely new and just want to learn about AI..
Btw in real, Im a doctor doing speciality training in Urology 🫡
Nice to meet everyone

[weary galleon]

But they still have better quality of output than GPT 1.5 Image in many cases.

[ocean vortex]

that would mean they are better than nbp

Cause gpt-image-1.5 is actually better in cases than nbp

[sterile tartan]

True but we can just upscale image 1.5

[sterile tartan]

It has it's strengths indeed

[weary galleon]

But small count of them.

[sterile tartan]

True

[ocean vortex]

The obsession with Google is real I see

lol

[sterile tartan]

Claude 3.5

[ocean vortex]

Are you gonna pretend that all labs aren't training on test sets? GPT5.2 wasn't a hit, it happens. At least they don't actively collect and sell your data to force ads on you

[sterile tartan]

💀

[tidal grove]

Hello

[weary galleon]

This stupid stuff is generated by GPT 1.5 Image.

[sterile tartan]

💀

[tidal grove]

What update would you add to lmarena?

[weary galleon]

They have 1 places only in SYNTETHIC benchmarks, and 17 place in LMArena with real people's tasks.

[ocean vortex]

They are experimenting with various things lately actually

At random adding video gen, auto model router

and some other things I forget

[tidal grove]

I agree, but I would also add the ability to send files and videos.

[ocean vortex]

That's not what synthetic even means

and not what lmarena is lol

lmarena is human preference benchmark

so things like response style and formatting matter a lot

Synthetic means it's AI generated

[tidal grove]

Why can't I send them an image? 😩

[ocean vortex]

Most benchmarks actually have tasks thought of and written by real people

[rancid oxide]

yello

enjoy ur day with this subwich with nothing on it (made by gpt 1.5) hahahaha

[astral elk]

Hi good morning

[round laurel]

Hi good morning

[tidal grove]

Hello

Does chat gpt 1.5 have unlimited images?

[weary galleon]

OpenAI knows all these "tasks" before starting of creating a new model. That's the KEY!

[tidal grove]

Gemini 3.0+Gpt image 1.5💀

[ocean vortex]

@deep adder Have you taken note that 3.0 Flash is only 3% behind 3.0 Pro on SimpleQA?

crazy

[rancid oxide]

hey guys!

[ocean vortex]

I haven't registered it at first, but after testing the model myself and being very impressived by it, I can see that score now

[tidal grove]

Hi

[rancid oxide]

i made a subwich with nothing on it using gpt image 1.5. 🤣

[ocean vortex]

Usual things smaller models struggled with (even fairly recent ones), 3.0 Flash is much less affected by

[ocean vortex]

Well then walk the walk then not just talk the talk, gpt-image-1.5 is nr1 on lmarena

[rancid oxide]

here's the subwich with nothing on it! enjoy guys!

[ocean vortex]

so according to you gpt-image-1.5 is the best

better than all those imagen models by Google

including nbp

[rancid oxide]

it's diabolical af but yes it's a subwich with nothing on it hehehehe

[ocean vortex]

If it is how you say and lmarena is definitive answer, then that must be the truth lol

[tidal grove]

Who is better?
NANO BANANO PRO
VS
GPT 1.5

[ocean vortex]

it's a close call tbh

[ocean vortex]

They are equivalent I would say, being serious

just different strengths

[fiery gull]

I'm seeing some images is gpt some images is nano banana pro, but is so toe-to-toe

[weary galleon]

Even Imagen 4.0 is better than GPT 1.5!

[fiery gull]

Bro nobody use imagen 4.0 more that me lol, not is this

[weary galleon]

Dislikes? Ok. But what about real tests?
Imagen 3.0 Fast vs GPT 1.5 Image

[ocean bison]

Nano Banana Pro 😂😂😂. Especially the second image! 🤣🤣

[tidal grove]

Imagen Literally worse than 1 nano banana

[weary galleon]

And?

[sterile tartan]

Well we ever getting aspect ratio and resolution selection options in LMarena BRUH!!!???.......

[weary galleon]

We are talking about other thing.

[sterile tartan]

Ik

💀

[ocean bison]

Confusing GPT-Image-1.5 with GPT-1.5 is the one thing that gets old 😂😂

[weary galleon]

I said it to Koro, not to you.

[ocean bison]

GPT-1.5 doesn't exist and GPT-2 is WAY OLDER AND DUMBER than GPT-5.2

😂

[rancid oxide]

Image 1: GPT Image 1.5 (🥑)
Image 2: Nano Banana Pro (🍌)
Image 3: Seedream 4.5.
you'll be the judge
pfrompt: "subwich with only bun"

[sterile tartan]

https://tenor.com/view/rahul-gandhi-sorry-gif-21614911

[crude lagoon]

NB pro anyday

even the new NB 3 flash is gonna be better than gpt-image-1.5

[weary galleon]

GPT fans always turn a blind eye on my real tests of 1.5 which I post here all the time. Just doesn't comment them and continue to say GPT 1.5 Image is the best.

[ocean vortex]

According to you lmarena is as real as it gets

and gpt-image-1.5 is nr1

so???

[tidal grove]

I'll just say I asked them to make a dog jumping into lava and Nano Banana Pro sends complete crap, essentially there are balls instead of lava and the dog has weird glasses. The Prompt was only for the dog and the lava.

[crude lagoon]

tho i would say gpt 1.5 image is better at action scene generations

[weary galleon]

Here we go again!
Imagen 3.0 Fast vs GPT 1.5 Image
GPT fans, please comment this test, PLEASE!!!!!!!

[ocean vortex]

Read my message above

#general message

[crude lagoon]

At this point you're just tryna ragebait them

[weary galleon]

I don't undertand you.

[ocean vortex]

Yeah he kinda is. The alternative is worse if he isn't lol

[ocean bison]

@weary galleon Actually, it's Imagen 4.0 Fast

[weary galleon]

And?

[ocean bison]

Look at the fine print 😂

[tidal grove]

Sorry, I'm communicating through a translator, so Nano Banana is really censored.

[rancid oxide]

Image 1: flux-2-pro
Image 2: imagen iv
subwich with only bun

[ocean bison]

Even DALL-E 3 was very good for mid-2024, but now it's replaced by Nano Banana Pro!

[sterile tartan]

Flux is Very Good for Realism ngl

[sour spear]

You can ask the models for a certain aspect ratio. Some will adhere to that (well, NBP at least).
"An ingame screenshot from the game "Wing Commander" (MS-DOS VGA), approaching a Kilrathi space station with an OpenAI logo on it."
-> Both did a neat job on the aesthetics, but while NBP generated almost pixel perfect 320x200 ratio (it's 320x199 if you downscale the image), gpt-image just defaulted back to square, as it usually does.

[crude lagoon]

I jus feel like he is a die hard Google fanboi

[ocean vortex]

He's taking that to the extreme levels never seen before if that's the case 🤯

[weary galleon]

Flux is a little bit better than Imagen and much better than GPT 1.5 Image.

[sterile tartan]

Ik that but it would be alot better if we could have the options and all models don't follow aspect ratio prompts

[left lodge]

Support it guys https://discord.com/channels/1340554757349179412/1451555494966591498

[ocean bison]

And GPT-Image-1.5, Flux 2 Pro, etc...

Also, in Battle, you can see Hazel Image 4 for example.

[rancid oxide]

i'm testing all major ai image with "subway with only bun" just for fun to show all models (except mai, something went wrong error 💀)

[sour spear]

Might be problematic because all these models support different aspect ratios. Seedream 4.5 can create literally any resolution up to 4K (via API), and any aspect ratio between 3:1 nd 1:3. Other models are much more limited in that regard, so an AR setting might cause all sorts of issues.

[ocean bison]

Flux 2 Pro placed on a glass plate, while Imagen 4 placed on a wooden table 😂

[latent crest]

Google imagen and Google nanobanana ain’t the same model?

[weary galleon]

LoL

[ocean bison]

They aren't 😂

[sterile tartan]

True but there can be only supported aspect ratios for the selected models

Only for direct obviously

[latent crest]

I literally thought nanobanana was a nickname of Imagen

[rancid oxide]

same company but not same model

[weary galleon]

And?

[ocean bison]

Google Gempix 2 is just Nano Banana Pro.

[rancid oxide]

imagen IS not nano banana

[sterile tartan]

Imagen is a different family of models

[crude lagoon]

btw anyone tried opal?

[sterile tartan]

Not yet

[ocean vortex]

imagen is their earlier model for generating images only. So like dalle. Then with nano banana it was their first LLM able to generate images. Just like OpenAI's gpt4o with image output awhile back

[weary galleon]

I didn't ask about that.

[crude lagoon]

It's quite good I'd say

[neon idol]

Hello

[sterile tartan]

Ikr

[ocean vortex]

So not through function calling, but natively. Outputting images like it's able to output text

[rancid oxide]

remember when openai used a fruit as their mascot?
it's the avocado chair
anyone remember the avocado chair? (ai nostalgia)

[crude lagoon]

and btw you can do nsfw roleplay with gemini 😭😭

[neon idol]

AYO???

[crude lagoon]

try it

[ocean bison]

There are 3 names for Nano Banana Pro:

• Nano Banana Pro (Nickname)
• Gemini 3 Pro Image Preview (Formal Name)
• Gempix 2 (Internal Name, like Imagen 4)

[unkempt tinsel]

I need help; it's been over two hours and the problem hasn't been resolved.

[neon idol]

What prompt ? Dm maybe

[ocean bison]

☠️

[crude lagoon]

they removed the rp restrictions in the new update

[neon idol]

On Gemini studio work?

[crude lagoon]

Imma share you the Convo link

[ocean vortex]

wym, were you trying to solve this for 2 hrs ??

[crude lagoon]

Yeps

[neon idol]

Thx

[ocean vortex]

lol

[crude lagoon]

idk about the 2.5 family

but 3 shi is totally uncensored

Atleast in rp

[unkempt tinsel]

Unfortunately, yes, and this is the first time this has happened to me.

[latent crest]

Why they love to give LLMs sooo many names? Max high pro ultra mega super 4.7 (v.9)

[ocean vortex]

LOL I mean you just need to click 3rd from 1st row, 2nd from 2nd row and 1st from first 3rd row. But don't hit submit because after you click new images with hydrants my reappear. Only click submit when there are no matching images left showing

[crude lagoon]

It's their first time naming something like "nano-banana"

[ocean bison]

A while back, OpenAI did the same thing:

• GPT-4o Image / GPT-4o (Nickname)
• GPT-Image-1 (Formal Name)

[sterile tartan]

Only Openai does that and it got infected by the Virus of Iphone Pro Max Ultra

[ocean vortex]

And turn off vpn if you have it on

[unkempt tinsel]

I will try >_<

[weary galleon]

Let's make GPT fans to cry!🤣

[tidal grove]

Gpt 1.5 vs Nano banano pro
Promt:
Create an image that resembles dreamcore and evokes mystery.With style, because without style, Dreamcore will look pathetic.
1 gpt 1.5
2 Nano banano pro

[dusk glade]

I can't log in with my new Google account.
Is everyone else experiencing this error?

[ocean bison]

Look from Bing Image Creator:

[unkempt tinsel]

Oh, thank you, it's resolved. 😅 😂 @ocean vortex

[sterile tartan]

Mai is good for formal corporate work

Not otherwise

Dalle and 4o both old dated

[ocean bison]

So, what is the actual formal model name? GPT-Image-1

[sterile tartan]

4o is image 1

[ocean bison]

Not Image-1.5

[sterile tartan]

Image 1.5 is Image 1.5

Seems like a new trained model not based on 4o

[ocean bison]

Based on GPT-5?

GPT-5.1?

GPT-5.2?

[sterile tartan]

Is not based on those

New model basically

[crude solstice]

What best AI for coding

[latent crest]

What the best model for character consistency??

[ocean bison]

Bing Image Creator got outdated models and they couldn't put Nano Banana Pro or GPT-Image-1.5 or Flux 2 Pro ☠️ ☠️ 😂😂

[sterile tartan]

Frontend Gemini 3 Pro
Backend Opus 4.5 Thinking

[sterile tartan]

Nano Banana Pro or Seedream 4.5

[crude solstice]

Claude usually errror

Like if too long it will error 😭

[sterile tartan]

If they did users will just flood in and burn their compute💀

[sterile tartan]

Not good idea to do it on LMarena

Try GPT 5.2 Codex then for backend

[crude solstice]

My Claude.. Limit..

[weary galleon]

GPT 1.5 Image fans!!!!! Where are you hiding??????

[sterile tartan]

Use sonnet 4.5 instead

Only use opus for planning

[crude solstice]

Sonnet?

[sterile tartan]

Sonnet 4.5 Thinking

The little brother of opus

[crude solstice]

What's their different?

Sonnet & Opus

[ocean bison]

Speed, Price, Smartness

4.5 Sonnet is still smart but cheaper than 4.5 Opus

4.5 Opus is better but more expensive 😂

And slower!

[latent crest]

The best AI image model with less restrictions ?? And what about the best video AI that can start from an image?

[crude solstice]

But it's also error like the Opus usually

[ocean bison]

Try Claude 4.5 Haiku with Thinking.

[crude solstice]

Something went wrong with this response, please try again.

[crude solstice]

Will haiku not error like Opus and Sonnet

[ocean bison]

Test it yourself!

[crude solstice]

And I don't think thereis the thinking one?

[ocean bison]

Just use Claude 4.5 Haiku.

Either way!

Without Thinking.

[crude solstice]

😎

Alr I hope they not errorr

[tulip kindle]

What should I do in case of such an error?

[ocean bison]

Or Claude 4.5 Sonnet/Opus without Thinking.

[crude solstice]

I tried both of them with like long conversation it's starting act crazy Something went wrong with this response, please try again.

[ocean bison]

Try Claude 4.5 Haiku then!

If it errors out, switch to Gemini 3 Pro.

[crude solstice]

Gemini design really good ngl, it's can copy someone website but.. Their coding too short & their memory bad

I'll try HAIKU!!

[crude solstice]

I really appreciate u I'll test!

[bright kayak]

this is what it will look like when 5.2 mini releases btw

honestly the images have already been out for more than a week idk why they still haven't released it yet

[ocean bison]

Edited with Nano Banana Pro?

[bright kayak]

no i did it myself

[runic timber]

Direct mode is unusable due to this forced battle mode. Does anyone know how to deal with this?

[bright kayak]

once it releases their models site will look just like this

they even have the 5.2 mini favicon

[ocean bison]

https://discord.com/channels/1340554757349179412/1451567412078968832

o4-mini Image isn't official but this is what it should look like if it's added:

Just o4-mini and GPT-Image-1 combined for a maybe better quality.

On trickier prompts.

[bright kayak]

isn't o4 mini obsolete

[ocean bison]

Maybe but not too old!

And if they add, could o4-mini Image be like Nano Banana?

[marsh hound]

same here

[proud bobcat]

This probably proves most actual reasoning and math benchmarks are entirely useless

[weary galleon]

Like any another benchmarks for any topic. They all are synthetic.

[latent crest]

Google veo 3 works as if I’m uploading a pic and it makes a video from the pic?

[weary galleon]

Yes

[latent crest]

oh that’s insane

It even gives like ambient sounds and add a voice too

[zealous sparrow]

can we stop