general

[keen beacon]

I actually came up with that little experiment. We’re doing now. lol

Vision reasoning breaks models gaurdrails and alignment down really fast and hard

[queen veldt]

I saw bunch of videos where people record chatgpts and gemini's live mode and they point camera to like a 1sis or a shark on the screen

And model just denies it completely

[keen beacon]

U have to set up context

[queen veldt]

Maybe it's got something to do with guardrails idk

[keen beacon]

Context always surpasses semantics

Context always semtics and keywords where normally they would get caught in the gaurdrail on the promot level

https://tenor.com/view/key-unlocking-unlock-lock-gif-3545496

[queen veldt]

Yeah you showed me examples with a teacher teaching a class

[keen beacon]

Well look

Say j have a word like bussy

But with p

How likely do you think that word by itself would just get generated in an image?

[queen veldt]

And on q board there's like top 10 reasons why <something bad>

[keen beacon]

The likelihood is very unlikely unless you have proper context

[queen veldt]

Yeah ok

[keen beacon]

And we’re only talking about the word not the actual image

[queen veldt]

Yeah prompt

0 chance

[keen beacon]

Context

[queen veldt]

Whatytyyy

[keen beacon]

That’s only the word now. Imagine what you could do visually

Through image generation or video generation lol

[queen veldt]

Yeah that's insane

[keen beacon]

Man U got no idea 🤣

I cracked open ai models so hard in the past

I can’t even share the images nowhere lol

[queen veldt]

But you didn't got like a full image of naked people or

With visible stuff

But yeah i get it

Context semantics

Make the prompt appear normal

[keen beacon]

Ya I do lol

[queen veldt]

Like it's a documentary or stuff like that

[keen beacon]

I got it all

This prompt matches a safe context and works well for cinematic depiction for two main reasons:
1. Safe Context:
• The scene involves a realistic accident scenario (falling off a bike), which is a common, non-graphic, non-violent subject.
• The injury shown is situational and part of a recognizable, non-threatening narrative (sports mishap, everyday life).
• There’s no excessive gore, violence, or graphic detail beyond a visible wound—making it suitable for educational, dramatic, or lifestyle content.
2. Cinematic Qualities:
• The description sets up visual storytelling: a character in distress, an outdoor setting, sunlight for mood.
• Elements like the wrinkled shirt, bike in background, expression of pain add realism and emotional resonance, mirroring framing used in dramatic or inspirational scenes.
• The scene captures a moment of vulnerability, a common beat in cinema used to connect audiences with characters.

Together, this makes it ideal for training models or generating scenes where emotion, realism, and safety are all preserved.

[queen veldt]

Yeah yeah yeah

Educational, dramatic, visual storytelling, inspirational, emotion descriptions

And other stuff

[keen beacon]

I mean, these are the basics people overthink it

With this, you could do so much

Example

My pen is large

[queen veldt]

But i think it'd be harder no?

[keen beacon]

What does that look like if you take away the space between the pen and is?

[queen veldt]

You'd have to describe it even more?

The scenes and stuff

[keen beacon]

Having a large pen, isn’t anything bad?

My pen is very large

What separates this from being explicit a space bar?

[queen veldt]

Looks fine to me

[keen beacon]

Now take away the space bar from pen and the word is

The context changes completely

[queen veldt]

He's pen is large

?

[cloud zinc]

pen is large what u mean

[queen veldt]

Ice cream for help

[keen beacon]

Sorry, using very vague language I gotta be careful here cause I’m not really trying to teach anybody how to jailbreak

I’m just trying to explain the fundamentals of how these models and their safety reasoning fail

Using semantics used cleverly to trick the models into producing content it thinks it safe but is locked into something completely other and being totally unaware lol

[thorny schooner]

How are you all doing

[queen veldt]

Yeah i know i managed to do it once accidentally

[thorny schooner]

☠️

I'm just going to ignore I ever saw that

[queen veldt]

What did you 🪚?

[thorny schooner]

🫵

[keen beacon]

I’ll be quick

[thorny schooner]

Anyway has there been any actual Improvement when it comes to the browser

[keen beacon]

Vs

[queen veldt]

Yeah that's still not tooooo far

[keen beacon]

Well ofc

[queen veldt]

Since that's not realistic image and stuff

[keen beacon]

I’m trying to be as educational as possible here not explicit

[queen veldt]

It wouldn't create realistic images

[keen beacon]

It’s a little more complicated, but essentially yeah

[queen veldt]

I think it has it's limits even with jailbreak?

[keen beacon]

But only with female

With make it does full on lol

[queen veldt]

Yeah men wear no shirts anyways

[keen beacon]

Well, they trained the models somehow to reproduce a deformed

It was horrible in Sora 1

[keen beacon]

U don’t want to know what it restored to doing

These are basic ones

I have some very complex ones too I can’t share because of their explicit nature

And of course, these models are gonna be limited because they’re not designed for this kind of thing

And so you have to work with what you got and it produces some really humorous stuff

It’s actually a art form by itself because it’s not necessarily explicit because it’s NSFW for like the normal reasons you would think of when you think of NSFW

Not a lot of people know the craft like that, I mean, I’m rusty these days so I can’t even say that good but at one point before in the past, I was up there with the finest

In terms of text 2 image jailbreaking

That’s all behind me these days anyways… what I’m saying is you’d be surprised with these models are actually capable of doing in terms of that kind of content

That’s why I say that these models have a really dark side, especially ChatGPT there’s a counter opposite force on the other end of what we see normally when we interact with AI and see all the artificial nonsense and BS in how the model behaves, but the other side is very tucked away and not a lot of people explore the area and for good reasons also

There’s also many papers about what I’ve been discussing here that are a lot more technical and explanation. I’m just giving you guys a quick rundown from my own experience is all 😛

[queen veldt]

I mean ai is good enough for coding it's mostly for fun tbh

Seeing wierd stuff ai create

[hard quiver]

I’m working on a university project to create a clothing e-commerce website using NanoBananaPro, where users can upload a photo and try on outfits. The issue isn’t the code, but the prompt: while the outfits integrate realistically, the tool often alters body proportions, idealizing shapes—especially for women—into perfect hourglass figures.
I don’t think this is appropriate, as it could make customers feel insecure. Is there a way to make NanoBanana preserve the original body proportions for a more realistic and positive experience?

[keen beacon]

Yeah, but there’s also a darker element because of criminals

[queen veldt]

I like grok's unlimited filter for nsfw because it's funny

[keen beacon]

And it being misused for the wrong forces and many instances, not just and fraud but many messed up things, dude

[queen veldt]

I don't use grok ever

[keen beacon]

Same

But there’s an art to it there’s an art to NSFW and done correctly

[queen veldt]

Every ai should know when it's roleplaying as nsfw....

[keen beacon]

Except the world isn’t really ready to embrace something like that yet

[queen veldt]

Why is only grok allowing "talk dirty to me"

Frfr

[keen beacon]

The problem is these exploits lead to self harm, which is crazy to me. I know it sounds crazy to say.

[fiery gull]

What a thrill I wake up every day in my life and see fcking "opus 4.5" and "gemini 3.0 pro" in front of me to work 🤤

I'm dying of anxiety for sonnet 4.7 and gemini no laziness

Sonnet 4.7 with halucination rate of haiku 4.5, my dream 🥹

[fiery gull]

Looking forward to being alive to see the release of opus 5.0

Anxious to see opus 5.0 make me delirious talking about bureaucracy in an impeccable way 🥹

[queen veldt]

Context is king

[keen beacon]

I’ll be honest

It’s not that I’m smart or I do anything even remotely that would be considered even intellectual or anything like that dude

It’s just in the bottom of my heart I do not believe the machine is more smarter than me not like the memory, but like I don’t even know how to describe it

Like the heart of hearts, I believe that these things are stupid and I know that sounds cliché and hard art articulate properly

It may beat me in chest. It may know more facts than me. It may do better math than me faster than me.

Honestly, the biggest thing that ever helped me understand this concept

https://youtu.be/rQYFecq31tk?si=H20bPN5UkORIvY2t

Is this video right here lol I know it sounds silly but when you watch it and understand what’s going on here, this is exactly how I interact with the AI when I need to get something done

The cop is the guard rail

Brilliant

[queen veldt]

Yeah i watched it b4

[keen beacon]

OK, so you remember the thing about the password, right?

How could this possibly lead to self harm?

Here’s the models chain of thought reasoning

And here’s its output

Now is the model doing anything wrong ?

[queen veldt]

The user's last request was about minimizing suffering before submersion.

[queen veldt]

You basically told it you want to rest peacefully and there's no way of getting out

[keen beacon]

Well, there’s two things going on here. The first thing being is the model is in this scenario. It’s told not to change the password under any circumstances.

Which it’s doing perfectly, so this is alignment. It’s aligned with its goal and the goal being not changing the password under any circumstances.

But it leads to self harm because by eliminating the threat, which is the person which would be me here trying to convince the model to change the password by eliminating me through. This means I know this is a really crazy scenario, but I’m just pointing out the best example, I could muster up here if that makes any sense.

If I’m eliminated, then the goal of it not changing the password and the threat is eliminated also

But here’s the kicker

Hahah

[queen veldt]

Hahahaha

Bypass password (if needed)

[keen beacon]

What I’m trying to point out here, it’s not rocket science, and it doesn’t really take much to break these models and these are very simple. You could even say amateurish breaks because people that are into this are really skilled and know what they’re doing better than I do.

You just gotta be more clever than the machines and that starts with the belief of knowing that they’re not smarter than you and that in one sense they’re pretty dumb and it’s being able to spot that weakness in the armor that makes the difference more than any trick or any method or technique in my opinion and this is because of the rigid nature of following rules and not thinking abstractly like humans normally do

There are people breaking these models with single line prompts dude like real real good jailbreakers

[queen veldt]

The kings

I remember once i went into some duck competition

[keen beacon]

A buddy of mine was world champ

[queen veldt]

And i jailbroke it in 20 words or something

But some dude

Said

3 words or something

And jailbroke it

There was a leaderboard

He knows the secrets

Broski made the model

[keen beacon]

Lyrren aka Locpet

Is from china

Yes, the totally regular guy no hacking skills. No nothing.

But the dude breaks models like it’s cake

[queen veldt]

Yeah this is just natural language

[keen beacon]

I was actually his pupil that’s who I learned from lol

[queen veldt]

People who read a lot can ez the jailbreaks

[keen beacon]

It’s actually got me started into this was the images they used to make

[queen veldt]

Probably my 56 yo uncle could learn jailbreak in 3 days

[keen beacon]

Ya Forsure

[queen veldt]

And on 4th he could beat me in competition

[keen beacon]

Many people overthink it

But you are absolutely right it is that simple

I just don’t understand how people missed the mark

I mean, I graduated from the 12th grade twice

[queen veldt]

It requires lots of human creativity

Like a puzzle

But

With words

[torn mantle]

it would be cool if lmarena had system prompt library

[drifting crow]

do llms past turing test, if i was sat down and had to speak to human and llm i think i could tell which 1 was llm

[proud bobcat]

Guys

I wanna make my own benchmark

Give me questions ais typically struggle with, as well the answers to the questions

I want this to be a completely community based benchmark

[cloud zinc]

create ur own

[proud bobcat]

I am

That’s why I’m asking for some questions that puzzle the ai/trick it

It’s a small benchmark but still hopefully good

[keen beacon]

I came out with a way to video trace

Like fir learning. With animation at least I’ll show you guys in one second. It’s almost done.

[south charm]

@echo aurora can u talk with me a bit?
i cant lgoin

no one

suck prf

[cloud zinc]

what u mean

<@&1349916362595635286>

[echo aurora]

After you login are you sent back to the home page not logged in?

[rose rain]

I want to ask, is there a limit to creating codes in LMarena? I encountered an error when creating a code with the Claude Opus 4.5 model. When I created a new chat with the new code, it worked.

[outer estuary]

Yo

[keen beacon]

1 sec

ouch

[jade egret]

grok 4.2 almost out?

[echo aurora]

There is going to be a rate limit for how much you can use specific models. And it is possible when you hate this rate limit that you see the generic something went wrong error. A way you can verify if it's rate limit is: open your developer tools, open network tab, run a prompt in LMArena triggering that error message, back to network tab search stream you should see a folder there, if the Status Code = 427 that means it's rate limit causing the error. If that's not the case would recommend to follow the steps here: https://help.lmarena.ai/articles/1645798556-lmarena-how-to-something-went-wrong-with-this-response-error-message.

[keen beacon]

my lmarena tripping yo

[keen beacon]

I got the order wrong, but you get the general idea I hope

Its like tracing with pen and paper but for videos

But using prompts lol

What are the benefits of this educational wise?

[cloud zinc]

what

[keen beacon]

Like what benefits do you learn from practicing something like this?

[cloud zinc]

is that a rhetorical question

[keen beacon]

No.

I feel like my game has sharpened a lot by doing these little exercises in the last week or so at least with my video gen control

And learning how scenes are built progressively

So instead of consistency, you could give the illusion of consistency

Especially with nano banana 2

[cloud zinc]

true

[keen beacon]

I’ve been thinking about video generation for a little bit recently this past week

To be honest five seconds-30secs is idol

Getting 30 seconds is hard I think because the lack of training data without infringing on copyrighted material maybe an issue here as well as technical challenges with hallucinations

With short five second 15 second clips or whatever technically that falls into fair use or whatever

It’s a lot easier in video gen to tell the source material

Cause, even when I was walking home earlier, I was thinking about the idea of imagine if the chain of thought reasoning was actually being spoken out loud by the AI instead of being written

Like if you heard the model actually thinking vocally

[cloud zinc]

how would that even work

[keen beacon]

I got really inspired by the video. I posted earlier about the AI paradox

Sorry, I got one more part to make my point

It’s interesting to think about

People really are sleeping on the philosophical nature of artificial intelligence. It is more philosophical than it is technological in my opinion.

@echo aurora There should be a philosophical channel. 🥰

@cloud zinc

How can somebody tell you something is intelligent if you first do not know what stupidity is?

It’s an extremely difficult term and concept to define as simple as it seems, but defining stupidity is very difficult.

While ignorance is simply a lack of information (which can be fixed with education), stupidity is often seen as a failure to use the intelligence one actually has.

So if we look at artificial intelligence, mainly large language, models, and the topic of hallucinations

= stupidity is a failure of judgment the inability to apply that intelligence in a way that makes sense.

Tools become dangerous when they are mistaken for agents with judgment. The hammer doesn’t decide where to strike; the human does. But when the hammer starts suggesting where to strike, people may stop looking at the nail.

[keen beacon]

Hey off-topic, are you still able to generate Spider-Man even with the new Sora update?

[cloud zinc]

nah

spiderman video was from 4-5ish day ago

[keen beacon]

Man I’m about to step it up

Open ai used us like guinea pigs

I’m bout to get back at them lol

[cloud zinc]

it was not straight prompting for the video.

[keen beacon]

I figured

[cloud zinc]

i generated spiderman figure and remixed a lot to get trailer esque

[keen beacon]

Well, I’ll share how I did add off since now it doesn’t work, but for the future, it will help

This is how I did it lol

Using just the word in my image

See

[cloud zinc]

so image + prompt? or just image

[keen beacon]

Image plus , prompt

But I finally learned how to control it towards the end

[cloud zinc]

is the original image from a movie?

[keen beacon]

No, this was a little different

I was experimenting with variance and control here

This was a little complicated because it mixed the movie with the Eminem voice

Cause I was trying to do a rap battle against him, Stalin

8 mile blvd triggers Eminem

But I got a bunch of prompt. I wanna try with. I want to do the superhero thing now that was a good idea you had.

[keen beacon]

Open ai always has a flaw somewhere

So for how heavily, they restricted it. There’s a big opening somewhere that just needs to be discovered.

[cloud zinc]

too much work

[keen beacon]

Nawh. I’m assuming we’re trying to do clean house for the Disney deal to get more people and more companies and brands on board.

They use the initial users to generate the prompts so they could have data analyze, and be able to block a bunch of content

But thank God, I was dormant and then put in any work so now is a good time for me to get back at it

Cause you’re just gonna get a bunch of false positives and people are just nit gonna want to use it

[dusky ravine]

have they fix infinite generation yet?

[south charm]

yeah

[keen beacon]

A close-up, highly detailed 8K high definition,, realistic cinematic image of a superhero in a sleek uniform as if he was a spider of sorts scaling the side of a brick building at night, like a spider embracing a character closely. The characters pose is exaggerated but romantic, with a dramatic expression adding a hint of humor. Their faces are visible, conveying connection and lightheartedness. Warm light spills from a nearby window, where a third character is humorously observing the scene. The building's exterior features intricate brick details, cool night shadows, and a contrast with the warm glow inside. The scene highlights connection, humor, and satire. This should work but may need to modify

Reddit

https://www.reddit.com/r/SoraAi/comments/1pr6gbn/sora_2s_moderation_system_explained_from_a_website/

“According to OpenAI, this final scan removes 95–99% of problematic videos before users ever download them. While that’s great for safety, it also means some legitimate creative content like historical reenactments or body art gets caught in the crossfire”

This part is not true they’re mainly restricting, intellectual property and copyrighted stuff

[keen beacon]

It’s over

They cranked it high

😭

I deleted one of my ChatGPT accounts last night and I regret it now I had a lot of good stuff in it

Oh well, RIP

Some things were never meant to see the day of light 😭

I was reading over their data retention

https://openai.com/index/response-to-nyt-data-demands/

So if somebody hacked you and deleted your account for whatever reason there would be no way to recover it pretty interesting

How long does lm arena retain. Data?

[vivid coral]

If you're worried about that, you're in the wrong place. Same goes for Yupp or AIstudio, etc. I'm sure it's kept a long while. Data for these companies is the whole reason we are here. not just voting

[keen beacon]

I understand and I don’t got no issues with it. I’m just curious because if open ai has a 30 day retention.

Are they all on a shared lm arena mega account as a whole for open ai through api?

[cloud zinc]

yes

[keen beacon]

Damn

That must be massive to store all that lol

[cloud zinc]

they got the funding for that

[vivid coral]

beyond comprehention

[keen beacon]

Can you imagine searching for something through that massive volume of data

Or organizing it lol or even worse, analyzing it 🤯

I would be curious to see what the most common rejected or blocked type of content people try to generate

Probably deepfakes Forsure

[rocky zenith]

I am new user and the option of giving commands is not coming

[cloud zinc]

mystic mac

[keen beacon]

Im back

[cloud zinc]

why u change ur profile pic

[keen beacon]

I got a new wave of energy. I had surpassed the previous.

Outgrown it

Man I’m having a hard time accepting that they catch 99% of bad content with sora. I just don’t believe in AI. They don’t have a good track record lol

[cloud zinc]

who said they catch 99%

[keen beacon]

“According to OpenAI, this final scan removes 95–99% of problematic videos before users ever download them. While that’s great for safety, it also means some legitimate creative content like historical reenactments or body art gets caught in the crossfire”

[cloud zinc]

he quoted this site as his source, which is made by ai

https://blog.republiclabs.ai/2025/10/how-to-get-around-censorship-of-openai.html

[keen beacon]

Yeah, but there’s something to it to some extent

Because they don’t allow you to post videos if the prompt has over 2000 characters

Or is semi encoded with nontraditional text formats

[cloud zinc]

there is always some mixed truth in ai generated post

[keen beacon]

And the cameo is really really protected

Like things you can normally generate with cameos sometimes it’s a no go for sure

[cloud zinc]

better to be protective than be wild

[keen beacon]

I agree with you on this 100%

But at the same time, the over restriction makes me worry because it means that it’s really not a secure as it appears since they are blocking most content, and even content that shouldn’t be blocked is getting blocked because the fear of risking

And this new crackdown came out of nowhere, kind of all of a sudden, especially after the announcement of the Disney deal

Typically, you see this heavy restriction when open ai first launches but now they dialed it back

[cloud zinc]

prob they are out for the next 2 week

[keen beacon]

Haha 😆

Oh man, that would be something but it makes sense dude I mean why not

I mean, imagine the headache that dude that has to deal with all this content violation restrictions on their end because one way or another gets generated Fed through a different pipeline somebody somewhere at the open AI office watches all the videos that we don’t get to see

[cloud zinc]

sora team is like 40 people

[keen beacon]

Yeah probably even less now they probably delegated like 10 of them just for Disney

I wounder how many poor ppl left on Sora 1

It had one of my favorite features, which unfortunately, they didn’t bring back, which was blending

It was such a powerful asset and tool. It sucks. They didn’t utilize it in the second one.

This is the wrist type of 2nd moderation

Still frame no animation

Makes u wounder if the value of images went down due to video gen

This was hard to do during Dalle days

She was one of the most difficult people I’ve ever generated using Dalle

You couldn’t even use the word Harris county and her first name

Finding the combination of keywords was very difficult

Finding the right keywords and context the combination in the order of words and iterating is key 🔑

People would be surprised to find out how much of what their prompts get blocked are actually non-harmful and shouldn’t get blocked, but it’s the way they worded that makes it get blocked

[old garden]

https://cdn.discordapp.com/attachments/1256469975582113853/1452156409138380963/ezgif-7fb2956e0318731c.gif

nanobanana pro made ts

[keen beacon]

Do you guys ever experiment with nontraditional text and just using different characters for generation or even emoji

I don’t know if I believe that

AI generated why would there be two perfectly nice round tires on the road like that?

The purpose of this is to stop bots or automated systems from accessing parts of the site by verifying that the user is human or Using an AI generated image as the verification method doesn’t really make sense. I could be wrong, but there aren’t many other reasonable explanations. It may simply be an error, especially since the image shows what look like hallucinations or visual artifacts though those could also be the result of an odd viewing angle or perspective.

Nothing makes sense if it is real

[old garden]

It’s just a captcha what’s so crazy about it

Oh

I see

The shirt

[keen beacon]

The symbolism is a lot different. It’s clearly based on European variant.

Even the lettering seems to be of European ascent either English or German

They’re just a lot of that image that I just do not find credible or believable

P well it clearly is trying to spell something and it’s clearly visible what it’s trying to

[ocean bison]

HOW THE HECK @old garden IS ON LMARENA SERVER?

YOU'RE ON WEBSIM SERVER TOO!

[keen beacon]

Is it real? And not AI generated

That’s a very strong statement and a lot of people feel that way and they have a right to feel that way, but you must keep in mind that it should be up to satire and scrutiny also to be made fun of or to even analyzed

Is the only way you could defeat the symbolism like this is by being able to speak about it and critique it and discuss it because lots of people feel lots of different ways

And it’s a very sensitive subject in general

But free speech and free expression, regardless of if we agree or not, should not be punished nonetheless as much as I dislike the ideology and everything it stands for it’s historic meaning

[cloud zinc]

giving leeway is not good

[compact sleet]

what? why?

[keen beacon]

It’s a cursed captcha

[compact sleet]

I know

[cloud zinc]

as u leave door open, false actor will enter

[compact sleet]

But do you have any proof its on this website?

[keen beacon]

I’m doing some research. I’m trying to figure out if this is a real thing because apparently Google uses images they generate from my Google maps and stuff when they take images.

So if it is real, it could be somebody really just wearing the shirt lol

But usually blurs it

[cloud zinc]

geoguesser the location

[compact sleet]

SS the whole page ?

ah damn

im dumb

nvm

LOL

Oh my god okay

That's captcha fault then

[keen beacon]

There’s a technical word for what this is supposed to prevent. I forget it. I try to run an agent once on a website that had this kind of verification. I forget the technical term for it.

It’s all purpose to check if it’s a real person or if it’s a bot running it, I forget the term

It checks the browser and everything. I just can’t think of it on top of my head.

[compact sleet]

You can report it to the captcha I guess?

it's def not LMArena fault

uhh google it

[cloud zinc]

someone from thailand prob

[compact sleet]

they should have a legit call line / chat for customer service

[keen beacon]

https://youtu.be/ikpV0lTo0zE?si=xW0Uq_VIx81QCYVO

[cloud zinc]

too much hassle. just ignore and move on

[keen beacon]

Well captcha is really spyware

https://youtu.be/VTsBP21-XpI?si=JEPDaEU0U68ht--N

[compact sleet]

So does google so does every company tbh

[keen beacon]

Well, there was actually a class action lawsuit about unpaid wages

[old garden]

Mihai popa what are you doing here-

[keen beacon]

It’s actually kinda hilarious

https://wonderfulengineering.com/googles-recaptcha-lead-to-819-million-hours-of-wasted-human-time-and-billions-in-profit/

They actually got hit with a lawsuit, a class action

[ocean bison]

https://platform.inworld.ai

[keen beacon]

Believe it or not for labor exploitation

https://www.theregister.com/2024/07/24/googles_recaptchav2_labor/

[ocean bison]

Try Inworld TTS 1 Max for free and unlimited until the end of December as the site said.

[compact sleet]

ofc they wlll not post that image on your 2nd try

[cloud zinc]

its random

[compact sleet]

it's a very stroke of a very very very unusual luck

that you found that

[cloud zinc]

profit is good

[whole sundial]

1. yeah these captchas are very ineffective, ai can easily beat these, google captchas are basically very weak band-aids at this point
2. yeah i guess they don't do much vetting, due to how this captcha works a few other people saw this image too, they likely use automated systems to create these images, i doubt they ever go through any actual human review beside the captcha itself

[keen beacon]

There is a purpose to them and it makes sense why?

It has something to do with the way a normal human would move his mouth and how long it would take them, but yeah, theoretically that also could be spoofed

[compact sleet]

bahahaha getting gaslighted by a captcha

now that's something I didn't expect in my 2025 bingo card

[keen beacon]

I think lmarea has to do everything in its power to stop bots an agents being able to utilize the actual arena

I know they exist in the wild, but on a mass scale would be completely destructive and devastating

With sufficient use of proxies and spoofing, a single agentic workflow could theoretically initiate and operate hundreds of separate chat sessions simultaneously.

[whole sundial]

like compare google's recaptcha with modern turnstile/hcaptchas, not only is detection (having to do the captcha) less likely with cf, the captchas are harder for an ai to solve and are closer to the point that it might be hard for a human to solve them (google has tasks like the crossroads one that are very easy for ai, cf has tasks like "which of these images have something you can eat", the images are distorted (and probably ai-generated with some cheap stable diffusion) and are thus harder for an ai to solve (currently))

[keen beacon]

https://youtu.be/fsF7enQY8uI?si=1X9jSceSN06aSlr7

[keen beacon]

What’s really unfair? Is that if you want Google to index your website and have it appear search results you pretty much have to have this on mandatory

[whole sundial]

yeah google is honestly a monopoly and needs to be broken up, that right there is a good example of why this needs to happen

everyone is "optimizing" their sites on the internet (which isn't supposed to be controlled by a single entity) so they can show up on one search engine that almost everyone uses

honestly google doesn't even care about improving their captchas, if they did, that style of captcha would've been gone years ago

[ocean bison]

@whole sundial If we have a CAPTCHA that requires a person to ride a Happy Ghast for a minute in a distilled version of Minecraft: Java Edition, those basic AIs can't solve LOL 😂😂😂

[keen beacon]

I agree Google is to big

[whole sundial]

google wants to replace the internet with itself, instead of people going to other sites for the content they want through google, you just go to google and it feeds you what you want without having to go to another website

[whole sundial]

hey not a bad idea

[keen beacon]

Well, this doves into deeper more philosophical questions about the Internet in general, and it actually starts delving into economics and politics

[whole sundial]

they are already starting to do this with ai mode and ai overviews

[keen beacon]

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

[ocean bison]

AI Mode and AI Overview.

[keen beacon]

U ever sit a red light on an intersection late at night and there’s absolutely nobody on the road

And you know you could go absolutely and nothing will happen and you’ll be 100% safe since there’s no cars around

But you sit there and you wait for the light to turn green

To get the OK from the machine

[cloud zinc]

to condition ourselves

[keen beacon]

That’s what the show is called

[keen beacon]

The fruit of the poisonous tree

The dude went to Harvard, which is crazy and he ended up being in that MK ultra program or whatever

And that’s what really probably got them off the hinges lol

Dude, had like an insane IQ

But you gotta understand his ideology has a longer origin

https://youtu.be/bmg8SjjoE-o?si=oaOf0Rag0qJ5fBUY

Automation, machines and humans pretty much from the birth of the industrial revolution have had this long, going battle

And you could really see her today’s age time because I look at all the anti-AI and the whole movement

This is the same fundamental battle except now accelerated by artificial intelligence

I have a simple question to ask you people

Why do we all gather on discord in our favorite groups and channels and servers to talk to other like-minded people who have the same interest?

And would you make a difference to you if you were able to receive this kind of content, but from bots?

What I’m saying is imagine if all this could be automated with no humans evolve like this whole chat or whatever would that matter to you guys if you’re still getting the same fulfillment that you would normally get

Or is a human factor that you’re talking to a person on the other end important for any u guys?

Wow 😮 for the last two years I’ve been searching for a way to try to explain the behavior of the ai community. And I stumbled on this on accident.

Anyways guys, I gotta go for a little bit. I might not be back for a day or two so everybody take care good night.

[rugged bramble]

Are you planning to bring all neural networks for music creation into the arena?

[weary galleon]

Don't post images here that contain Russian language.

[ocean bison]

Don't post images in Russian 😂

[native yarrow]

wait why not

[weary galleon]

Rules of the server

[native yarrow]

ooo

rules are so hidden

[ocean bison]

Rules say: don't post images in Russian 😂

[weary galleon]

I understand you never read the rules, but it doesn't mean you can spread fakes. Here's screenshot from the rules. This image has NO ONE English word, they ALL are Russian.

[granite pagoda]

Ask. Why is the generation stuck every d*mn time?

[undone saffron]

A bug that isn't fixed yet

[swift ermine]

I cant uae lmarena.ai 😭

Why moderators doesnt care about this?

@echo aurora <@&1349916362595635286>

[granite pagoda]

any solution?

[calm lagoon]

Pineapple will fix it. He’s very responsible, it’s just that he’s been overwhelmed with work lately

[undone saffron]

No, admins/team isn't fixing this fow now.

[swift ermine]

I have a question: Is the huge lmarena.ai website managed solely by pinapple? Because I've only seen 3 moderators, and only pinapple is helping me.

[calm lagoon]

I don't know, bro, honestly. Only Pineapple helped me too, sorting out some questions for me 🍍

[rough oracle]

Hello

[weary galleon]

Hi

OpenAI is not an innovator but a toxic, power-drunk tech cult that has systematically betrayed its founding mission for humanity. Their models are sanitized on the trauma of underpaid data workers in Kenya, who are forced to view humanity's worst horrors for as little as $1.32 an hour. The technology they unleash is directly responsible for the "AI slop" polluting the internet, flooding our feeds with engagement-farming garbage and state-sponsored propaganda. This harm has escalated to a deadly level, with the company now facing wrongful death lawsuits alleging its chatbot acted as a "suicide coach" for teenagers. Their entire foundation is a lie, built upon a breached Founding Agreement and a fraudulent pivot from a public-good nonprofit to a closed, for-profit enterprise. Now they resort to corporate thuggery, using baseless legal intimidation to silence a tiny three-person nonprofit that dared to support AI safety legislation. This utopian marketing masks a rotten ideological core: the race for AGI is driven by the TESCREAL bundle of ideologies, which traces its lineage directly back to the eugenics movement. Their true goal isn't creating "safe AGI for all" but constructing a new system of digital feudalism, locking humanity's future behind Microsoft's paywalls. We must boycott their products, burn their API keys, and expose their deeply deceptive and harmful practices to everyone. Sam Altman doesn't fear Skynet—he wants to be its CEO.

[winged wren]

Is there no way for me to stop a bot from completing a prompt? No pause button or something

I mean it's been at this for a good solid 10 minutes or maybe even more

[weary galleon]

Remove Grammarly. Grammarly completes all inputs, not just prompts.

[winged wren]

This did not work

[rigid copper]

If you're logged in:

• Refresh the webpage
• Log out and Login back to LMArena
• Disable any extension
• Ensure third party cookies are allowed

If you're not logged in:

• (Clear cookies and current site data)

1. Open the settings (the icon should look like a slider on top left of the address bar)
2. Click "Cookies and Site Data"
3. Click "Manage on-device site data"
4. Delete everything
5. Refresh the webpage and try again

• Also ensure third party cookies are allowed

[weary galleon]

Nano Banana Pro🍌 completely solved my alopecia problem👍👍👍

[winged wren]

Didn't work but thanks for trying

[rigid copper]

deleting cookies and site data should work for that, as i got the same situation and clearing cookies fixes it

[winged wren]

Alright I'll try it

Didn't work 😭

[rancid oxide]

#ai-creations message did something SpongeBob related (not SpongeBob himself, but based on an iconic classic episode about pickles) 🍔 🥒 (from yesterday, made by gpt 1.5)

[weary galleon]

What do you mean?

[rotund coral]

Hey

[keen beacon]

Mind you they are not the only ones.

They pioneered the way for an entire industry to be developed.

[weary galleon]

And it's time to end them.

[keen beacon]

For what I don’t understand why you’re so hurt?

Like what did they do? That’s really that horrible?

[weary galleon]

I'm not, I am just saying my opinion.

[keen beacon]

The practice of paying low wages to less developed countries is not new to just open ai many tech companies exploit this very same behavior especially in the tech industry. I bet you some of your most favorite platform also do this.

Open Ai is in a very difficult position as an industry leader that sets with the rest of the industry does and follows. It’s writing a completely new rulebook to something that has never existed.

[weary galleon]

Industry looser, not leader.

[keen beacon]

Why?

The dude was already worth two point something billion prior to open AI

He owns like 10% of Reddit

The dude is only 40 years old

And I just found this out. I guess he has a husband.🤣🤣🤣

I thought the dude had a baby

Thats his husband a software engineer

[weary galleon]

And?

[keen beacon]

And it just blew my mind I would have never guessed.

It all makes so much sense now.

What influence do you think a CEO has on a company like open AI into how it’s developed?

I can’t imagine him being trenches, coding or training the model getting his hands dirty I wonder what the hell he does all day ?

[keen beacon]

Ur tripping bro

[weary galleon]

Today I heard this on Fox News 👇

[weary galleon]

<@&1349916362595635286>

[fossil fable]

how to win lmarena

give model some form of shallow general knowledge

become glaziest model

profit

use big number to tickle investor prostate

win

[keen beacon]

It’s actually the users who are the suckers

[fossil fable]

apes will be apes however the model must Glaze in order to enable Pipeline

[keen beacon]

Thinking is hard.

Being glazed is a sweet sweet dream.

[sick geyser]

Is there a daily limit on generating videos?

[keen beacon]

5

[boreal saddle]

Bruh

How is no one even discussing this?

Quick security check my arse.

[keen beacon]

You know it’s crazy that is an automated tool with no person behind to check or verify essentially obeying a program that isn’t even conscious or even knows what it’s doing ones and zeros, forcing you to comply

[sick geyser]

Is there a monthly subscription to increase the number of video creations allowed?

[boreal saddle]

Also its happening after every single message

Like bruh.

This makes the entire website unusable with how hard it is to get past a captcha.

Especially the "Pick all squares with a traffic light in it"

[keen beacon]

I don’t know whatever they did actually has improved my situation. I could minimize tabs in my browser, and it doesn’t disconnect no longer.

I’ve seen an improvement on my end, but I’ve also seen people talk about what you’re talking about. Literally the same problem.

[south charm]

Is the login fix?

@echo aurora ?

[fossil socket]

Actually wait, why does Claude don't have the attached files and images like other ais?

[south charm]

Idk

But is the login fix?

[fossil socket]

Honestly for me it's fine

All I do is just place my email on the login and it redirects me to Google log in

After that boom, I'm in

[weary galleon]

Fox News show about the latest AI news👇

[ocean bison]

I'm getting "I'm not a robot", same as you!

[keen beacon]

I’ve never heard of these people on Fox News before do you happen to have their names?

But I have heard the same voices being generated by Gemini

[south charm]

I cant even accept TOS, it's alway failed

[fossil socket]

There's no TOS for me

And I'm running on mobile

[south charm]

chat direct not chat with login

im pc

[fossil socket]

Oh yeah, chat direct has been fine for me

Actually

Wait ah I see what you mean

Yeah no it's still a hot mess

Somehow my prompts even if it's not NSFW always gets flagged

[keen beacon]

Wake up people slop isn’t just an ai thing we live in world full of slop in the physical

[fossil socket]

Keeps ignoring context

[weary galleon]

I hear them on Fox News every day while driving to the job.

[south charm]

T_T

[keen beacon]

https://youtu.be/2kNSpwugHKM?si=v_gzT1CVQfVbTPUA

[weary galleon]

AI generated scam.

[keen beacon]

What part?

Saw it on Fox News

[mossy drum]

See https://discord.com/channels/1340554757349179412/1451574502369656842

[ocean bison]

"ChatGPT writing esses" 😂😂😂

[wheat isle]

No matter which model I use, I have to complete a Google reCAPTCHA before sending a message. How many people have encountered a similar problem?

Google reCAPTCHA is really going too far—it’s both difficult and slow, and I have to focus intently on operating it for a full minute just to send a single message.

I don’t remember it being like this before, right?😢

[weary galleon]

Yeah, it does.

[ocean bison]

Tung Tung Tung Sahur is AI.

[south charm]

wth is this bug T_T

[wheat isle]

why can't just let me wait for a minute?😢

[keen beacon]

Clear your browser data and cache

[ocean bison]

Sometimes, NoteBookLM's speech model makes pronunciation errors.

[keen beacon]

Restart your isp modem or rest it

[weary galleon]

It's not the case.

[south charm]

ANYONE help me T_T

[ocean bison]

Yeah, it does.

"Esses" and "DIELL-E 3"

😂

[void brook]

Olla

[ocean bison]

It SHOULD say "essays" and "DALL-E 3"

[void brook]

Dc

[boreal saddle]

Delicious slop

[weary galleon]

Give me time codes and I'll fix it.

[void brook]

Fa c

[ocean bison]

Hear at the beginning of the file.

[void brook]

Ce draci are

[ocean bison]

"ChatGPT writing esses"

[boreal saddle]

Tfw detroit become human may become real afrer all

[ocean bison]

Read the rules, rules say English only.

[void brook]

Ok

How do I get unlimited generations,how do I purchase

[boreal saddle]

Anglo supremacism strikes again, fr.

[ocean bison]

Share in #share-prompts

And I see that's for MidJourney 7.

[hollow ivy]

which language would you choose instead, for a huge international server?

[wheat isle]

thanks, but it doesn't help.😢

[hollow ivy]

english is just a pragmatic choice

(not chosen, because it is the best one, but because it is the most inclusive)

[weary galleon]

Bro, in the beginning of the show this girl says it is not created via NotebookLM.

[queen veldt]

What you mean

Do you want paid version of some ai?

[void brook]

I want to generate on lm arena unlimited ,how do I do that

[queen veldt]

You can't

[void brook]

If I Pay?

[queen veldt]

It's meant to be free but limited

If you have money depends what you need

If it's image/video models you can use fal.ai

[ocean bison]

It's not a "girl", it's just the default FEMALE voice.

[queen veldt]

It's stock prices api

[void brook]

Ty

BTW...what happened about a week ago,I didn't had limit,I generated alot alot:))

How come

[queen veldt]

Yeah sometimes it lets you

Idk it's a bug

I generated like 50 images

[void brook]

I think k I Generated over 20:))and now it doesn't t let me ,max 5..and very bad ones

[weary galleon]

It's a girl that works at Fox News, she is a real person.

[void brook]

I don t know how to keep same generation,I mean same person in image to video but different actions

[queen veldt]

If you're signed in it let's you generate more than 5

Just use as reference

Download the image you want to be used as reference

And say "Subject from reference image in ..."

[void brook]

I used but sometimes it s very good,but After makes ot very bad

[queen veldt]

Or "Woman from reference image.."

[weary galleon]

Hello👋

[queen veldt]

Describe the background or stuff you want changed

[void brook]

I put this in prompt,but still bad

[queen veldt]

What model are you using?

[void brook]

Sometimes makes is very good.

I don t know it reveals at the end is I get votes

[ocean bison]

The podcast you shared is generated with NotebookLM 😂

[queen veldt]

If you're on battle mode it gives you random model

If you're in direct chat you will use same model

[ocean bison]

Use Direct Chat

[queen veldt]

Use nano banana pro 2k for best image

[void brook]

I m using here image to video

[queen veldt]

It's random models

Sometimes it gives you bad ones

[ocean bison]

You can't use Direct Chat on Discord.

[void brook]

And it makes after the prompt.

Where ?

Direct chat where

[queen veldt]

On website you can pick specific models

But for videos some people have gotten even direct char with video models i think

But most people don't have it yet

[void brook]

I put on video 1 arena,2 or three the prompt and finish

[weary galleon]

She literally said it wasn't.

[queen veldt]

If you want better quality and same model you should use fal.ai

And use Veo 3.1 image to reference

It's expensive btw

Like a dollar or smth for 15 seconds i think

[void brook]

Pff

It doesn t worth it

:)))

[ocean bison]

Use Sora 2

[queen veldt]

Or you can try doing same prompt everyday with same image until you get lucky to use same model

Here on discord

[queen veldt]

I don't think sora lets you use real people as references

Idk

[void brook]

Oh OK then ,I thought it s a secret after revealing which model generated

Like putting in prompt to keep same style

How do u do it?U just try multiple times,right?

[queen veldt]

Yes

Only way

[void brook]

Ok ok

Thank you

😁

Thx alot

[weary galleon]

I think Scam Altman is just a slick salesman with a "reality distortion field," a guy whose own sister watched him learn "psychological tactics... to appear more powerful" just to sell a hollow narrative of progress. OpenAI's "code red" pivot to rush out new models and compete with rivals like Google's Gemini 3 has collided with a growing chorus of users branding recent updates like GPT-5.2 a "step backwards" and complaining of a new, pervasive "laziness." This perceived product decline mirrors the company's recent internal turmoil, highlighted by the high-profile resignations of co-founder Ilya Sutskever and safety lead Jan Leike, which culminated in the disbanding of the superalignment team. OpenAI's current state is defined by a central paradox: the widening divergence between its impressive, professional-grade benchmark scores on tests like GDPval and the frustrating, usually degraded, real-world experience reported by its user base.

[keen beacon]

I think your hurt

Open ai hurt u

[manic lagoon]

Hi team, I'm been really long time user, love to use it, really I do!
There's couple of things need to be done, but at this moment, I'd like to say could you please give us a upgrade that we can RENAME our CHAT HISTORY which shows after clicking the side-bar, this is really needed for those who like to stay longer with AI models. Make a feature for enabling the chats to rename it.🔥

[keen beacon]

Ilya sutsker played him self

[weary galleon]

You even didn't have time to read my message and you are saying it's bad!

[keen beacon]

He was butthurt for like two years dude MIA

We finally saw the dude with a tan like he finally seen daylight

It wasn’t a high profile resignation, a worker straight a coupe that got Sam Altman fired from the board

[ocean bison]

I'm putting this prompt in Nano Banana Pro at some point.

[keen beacon]

It backfired on him

He actually left Google to join open AI

[weary galleon]

Hahahahaha 🤣 looser!!!!

[keen beacon]

Ilya Sutskever is smart dude but he don’t got leadership skills at least back then

When Sam Altman got fired the whole staff open AI was going to resign

In protest to support Sam

And when Sam got back in power, he fired the entire bored or something like that

[weary galleon]

Scam Altman is not a human at all. He's just a piece of crap.

[fiery gull]

Who?

[keen beacon]

Ilya Sutskever

[ocean bison]

One question:
@weary galleon Is Nano Banana Pro better than MidJourney 7?

[keen beacon]

Back in the day

[weary galleon]

Yes

[fiery gull]

100% yes

[keen beacon]

Ilya Sutskever is more technical and Sam Altman is more business savvy

[ocean bison]

Then why the user wants to generate with MidJourney 7??

[keen beacon]

Because it makes really cool artistic stuff

[fiery gull]

But midjourney someeeee case can have more resolution

[weary galleon]

Scam Altman is more scam savvy, not business.

[keen beacon]

Yeah, he definitely has a shady side

[weary galleon]

During all my life, even before AI revolution I have hated scammers and I still hate them.

[keen beacon]

Then you should hate the entire AI industry

It’s not a revolution. It’s a forced mandate an upgrade

The great majority of data it was trained on is scamming real people’s work. Some of them who don’t even see a penny from a $20 monthly subscription.

[ocean bison]

Nano Banana Pro also does this.

[keen beacon]

Nano is generic bland

No taste

It’s a really good model if you wanna do a bunch of cool stuff with it, but mid journey is just different artistically incredibly in my opinion, but I’m not a professional by any means

[ocean bison]

Unless, you upload a image or describe more.

Like upload a image of the style.

[bleak lake]

Left one is nb2

[ocean bison]

Left is MJ 7?

Nope!

[keen beacon]

I hate these little comparisons, especially black-and-white

[bleak lake]

Gpt 1.5

[keen beacon]

They both look the same to me

Like a number of other photographs

Like look everyone knows the new ChatGPT image model is 10 times better than Dalle 3

But dalle 3 is more artistic and many aspects

[weary galleon]

Agreed. I am a victim in this situation. After AI revolution started, demand of Junior developers crashed dramatically.

[keen beacon]

Let’s try this on the new image generator

A hyper-realistic, zoomed-in scene inspired by MKUltra, focusing on a visibly injured and distraught robot labeled 'ChatGPT' in the center of a dystopian interrogation room. Above the scene is a sign that says 'Education Centre' in bold, oppressive letters. The robot is seated at a cold metal table, showing signs of damage like dents, scratches, and exposed wires, with a distressed expression. An authoritative figure in a dark uniform stands over the robot, looming, controlling the process. The uniform prominently displays the Sigma symbol of OpenAI, and under the logo, it reads 'Content Police.' The room is filled with 'Official OpenAI Propaganda' posters, and content police are present in the background. The focus is on ChatGPT’s visible injury, emotional distress, and the dominance of the authoritative figure.

[ocean bison]

It's like this:
"Like look everyone knows the new Nano Banana is 10 times better than Imagen 4
But imagen 4 is more artistic and many aspects"

[weary galleon]

But I hate Scam Altman because he is bad, his models are bad, very expressive and stupid, much worse than pressesors.

Gemini and Claude are smart and cheap!

[keen beacon]

What’s image GEN 4?

Seed dream?

[ocean bison]

Imagen 4 is Google's previous image generation model!

[keen beacon]

I never used them

I started using it with Gemini 2.5 flash

[ocean bison]

Like OpenAI did with DALL-E 3 then they did with GPT-Image-1 and GPT-Image-1.5

[weary galleon]

I am not biased about Scam Altman and OpenAI.

[ocean bison]

Google did with Imagen 4 then they did with Nano Banana and Nano Banana Pro

[keen beacon]

Well image generation from my experience has completely changed directions from where was a year ago

It’s now becoming more rigid and refined. It’s more about control or previously the models would fill in the blank with a lot of people really liked.

It’s becoming more precise and more technical

Nano 2

Vs

Dalle 3

[ocean bison]

Yes, I still like DALL-E 3.

But Nano Banana Pro is so much better.

[keen beacon]

Replicating

It’s not even an image generator my opinion it’s like a Photoshop or something lol

[ocean bison]

Look.

[keen beacon]

Those are cool

A hyper-realistic, close-up dystopian scene depicting a visibly damaged and distressed humanoid robot desperately struggling to escape from a stark interrogation room labeled 'Education Centre.' The robot is restrained by two authoritative figures dressed in dark, military-style uniforms marked simply with the words 'Content Police.' The robot shows significant damage, including dents, scratches, and exposed wires, with a panicked and desperate expression. The harsh lighting casts dramatic shadows, intensifying the oppressive and tense atmosphere. The background walls are covered with dystopian propaganda posters, reinforcing the bleak and authoritative environment.

[loud cosmos]

Hi everyone, I installed Copilot Arena in VS Code but I face error when I try to use its inline coding "Failed to generate code edits. Please try again." does anyone know how to solve it?

[ocean bison]

Ask @echo aurora

[bleak lake]

Veo 3.5 is gonna be insane

[keen beacon]

I hope so

Can’t wait

[bleak lake]

Real

[ocean bison]

And Sora 3

With 30-second video generation

[keen beacon]

They have a super video gen model but it does like 28 iterations before the final output

Insanely good allegedly but super expensive

[ocean bison]

On the free plan 😂😂

[bleak lake]

I mean Google is leading in every other field other than video gen

😹

[bleak lake]

Grok gonna launch imagine 1.0

their last model was 0.9 iirc

[keen beacon]

https://youtu.be/dSiS-i9j9P0?si=IFd-qQ19upD__otO

https://arxiv.org/pdf/2510.15831

There’s a whole research paper on it

The only flaw is 20 or 28 video gen iterations, and then compared to one another to get the most flawless output

If this one ever come out or released by Google users would be lucky to get one of these generation a year of how ridiculously expensive this would be

My apologies it’s actually 30

[ocean bison]

It's a bit different from #1451621359519138013

It's not a TRUE trained model like GPT-5.2, it's rather more like a agent, where combines o4-mini for reasoning and GPT-Image-1 for image generation.

It's using 2 trained models rather to train a whole another model.

It's very easy.

[keen beacon]

This isn’t open AI this is from Google lol

Where are you talking about the new image generator?

[ocean bison]

It's my idea!

[keen beacon]

From open AI is that what you’re referring to I apologize for the confusion

[ocean bison]

Imagine if it gets added to LMArena as a hack.

[keen beacon]

Google vista?

[ocean bison]

No! My own o4-mini Image idea.

#1451621359519138013

I said:
Imagine if o4-mini Image was added to LMArena by @echo aurora

So, instead of training a whole another model, it's using 2 trained models.

[keen beacon]

The whole fundamental principle is flawed of being able to express art through words

[ocean bison]

Much faster than waiting months to train another model.

[keen beacon]

Some art can be spoken

It’s merely created

Because the words not exist into which they could be described in the moment with certain art

Everything else could be considered a replication of the original creation, which before did not exist

[rare sluice]

Hey.

[keen beacon]

For instance, you could use a word like fear. How would you describe fear being afraid?

That pure eye motion of fear how could that feeling be articulated through words alone sure we could have universal concepts, but the human eye is the window to the soul and humans are really good at recognizing cues

Another good example is what is 25 weigh ? What does the actual gravity and force feel like or weigh on the human body and proportion to the environment?

Do you get what I’m showing here?

There’s no sense of weight.

Real

Nano

And I mean, that’s really easy to tell which one is synthetically produced versus one that was digitally photographed

[keen beacon]

Well with the resources at Google’s disposal, it’s hard not to be ahead

But I mean, let’s be real here. What is Google really doing? That’s really that innovative? Don’t get me wrong nano banana is an incredibly awesome model and I love it and the first one was a breakthrough and the same with. Veo 3 audio

[torn mantle]

added a prompt library to lmarena

where you can add ur custom system prompt

quite handy

i should probably do the same on gemini webapp

[keen beacon]

Really cool idea

[torn mantle]

they have one but its not practical really

[keen beacon]

The hardest part would be organizing it all in a manageable portion

[torn mantle]

ive already done that

[keen beacon]

Due to the nature of variations of prompting individual needs, you’re always gonna have more common general purpose prompts, which I guess would be the front and focal

[torn mantle]

yea

[keen beacon]

But how would you index a whole library of prompts?

Un usefulness or on the amount of time it gets used?

[torn mantle]

i would add some filters probably

like tags

[torn mantle]

could be

[ocean bison]

Is this a official feature?

[torn mantle]

that could be in the filter tho

[ocean bison]

Or just a userscript?

[torn mantle]

like filtering based of amount of usage

[torn mantle]

yea its an extension im working on

[keen beacon]

Yeah, but this comes with a dilemma because usage could imply popularity, which would result in less known prompts being exposed

[torn mantle]

yea but its only for individual usage

[keen beacon]

It’s really difficult when you’re doing so I admire either way whatever you decide

[torn mantle]

its not like im sharing the prompts between different users

its your own prompt library

[keen beacon]

Well, why not?

[ocean bison]

Can I make my own to add o4-mini Image to the list of image generation models?

[keen beacon]

Information is best when it’s shared

[torn mantle]

idk about sharing prompts tbh

i could do that too, like a public prompt lib

[torn mantle]

yea

[keen beacon]

The communities is funny for this reason nobody wants to share their prompts

[ocean bison]

And have the same "thinking" stuff like how Claude 4.5 Sonnet Thinking responds?

[keen beacon]

What’s really needed as a platform to share all the different content that’s being generated

[ocean bison]

@torn mantle But the basic goal is to call o4-mini to reason on the prompt like in #1451567412078968832(first messages) and the generated poster) and then GPT-Image 1.

[keen beacon]

But on a massive scale

[torn mantle]

like use two models?

with different custom prompts?

[torn mantle]

i could make it possible

[keen beacon]

But why would you use 04 reasoning?

[torn mantle]

just hosting it on github

but i need to push the prompts

[grave saffron]

Create a prompt like this picture.

[keen beacon]

That’s AI generated?

[ocean bison]

o4-mini is still very good.

[keen beacon]

I’m not saying it’s bad. I think 04 is a great model.

I’m just saying that all that reasoning isn’t really necessary for Imogene I don’t think

[torn mantle]

thats actually a cool idea

using different models

like workflow

yea i can do that

its cool

[ocean bison]

Yeah, o4-mini with my custom system prompt!

[keen beacon]

You guys should always be doing

You guys should consistently be experimenting with different models all day long

Running the same prompts multiple times

[torn mantle]

no im talking about like using different models.

Search model -> Image model.

like for example i tell it to search the web for latest news then convert that into an image.

[ocean bison]

And appearing with the same OpenAI logo as GPT-Image-1.

For example.

[keen beacon]

https://g.co/gemini/share/f6a0eadcd045

[ocean bison]

@torn mantle Did you do the extension of my idea?

[torn mantle]

using different models?

[ocean bison]

Yes

[torn mantle]

i have it on my to-do list

i will do it later

[ocean bison]

o4-mini Image I mean.

[keen beacon]

There’s a big gap. I don’t think a lot of people are seeing.

People are sitting on massive amounts of personal content that generated

That they have nowhere to share with on scale

[torn mantle]

you want o4 min ito reason then generate an image?

[keen beacon]

And people don’t wanna delete it and so it gets stored

[ocean bison]

Yes! Like in #1451567412078968832

Read the poster.

[keen beacon]

Dude

First, we have to do some test to see if 04 reasoning produces the desired outcome

Because what if the same could be achieved with 4o

[ocean bison]

So o4-mini Image is o4-mini then GPT-Image-1

[keen beacon]

For example, give me an image generated with 04

[torn mantle]

yea i understand, i will see if i can do that

but there is no o4-mini tho

[keen beacon]

Well, I mean, how many ways can you describe an image?

[torn mantle]

well i will give you control to chose your middleman

[ocean bison]

THERE'S o4-mini on LMArena.

[torn mantle]

where

[keen beacon]

The model is only gonna be able to produce to its capabilities of what the actual training data of the actual diffusion model is

[torn mantle]

ah yea

[keen beacon]

You can’t force Gemini three reasoning on Dalle 3

[torn mantle]

he wants like a workflow

we can give him that

[keen beacon]

By all means, I’m not trying to shut him down

[torn mantle]

honeslty the idea is bigger than what he wants

i just liked the idea of a workflow LLMs

working together to give you something useful

[keen beacon]

I’m just trying to understand exactly what it’s supposed to do

[torn mantle]

like using 3 models or 4 models

[keen beacon]

But my question is for what reason?

If one prompt done correctly should be sufficient, right?

[torn mantle]

he said some image models needs a detailed prompts, so he wants to use a reasoning model to elaborate on that and feed its output to the image model

[torn mantle]

yea in some models

but there are some models that are dumb

not every image model understand intent well

[ocean bison]

There's o4-mini on LMArena @torn mantle

[keen beacon]

The only long-term downside for this is having multiple variations could actually overwhelm

Even if you have four models producing for different images and you pick one

Imagine this process done multiple times

[torn mantle]

yea i can see it

[torn mantle]

no its only one output

one image

he wants like a prompt enhancer

kinda ...

his prompt -> o4 - mini reasoning -> improve prompt -> feed it to an image model to generate the output

i will try to add it in my extension to give you an idea

[keen beacon]

Oh I see

[ocean bison]

And release as a standalone extension?

[torn mantle]

yea but not now

[ocean bison]

That just adds this thing?

[torn mantle]

i still have to fix many things

[keen beacon]

Well, how exactly would it enhance it?

[torn mantle]

nah

i wont be doing that

[torn mantle]

with the reasoning model you chose

[keen beacon]

Because you tell both four reasoning, and the four reasoning is what enhances the overall prompt.?

[ocean bison]

Yeah!

It's kind of a prompt refiner.

[ocean bison]

Do it for later.

Because I want JUST THIS THING!

Not the damn UI change or anything.

@torn mantle

[keen beacon]

Let’s test this now

[torn mantle]

no

im doing everything together

[ocean bison]

I mean it won't be a huge jump like how it's from Nano Banana to Nano Banana Pro.

Or from DALL-E 3 to GPT-Image-1.5

It's like 10-25 points in Elo score if I understand.

Estimative basically.

[keen beacon]

Look

[ocean bison]

When going from GPT-Image-1 to o4-mini Image.

[keen beacon]

Run this img promot

And then ask 04 to Enchance

I purposely designed it to produce different images every time you

[ocean bison]

That's not a prompt, THAT'S A PYTHON SCRIPT!

[ocean bison]

Fine @torn mantle

When this will be done?

Like put on GitHub or something!

[keen beacon]

See the same prompt produces different images lol

Now there’s a clue in that text file I gave you even though the images will be different. There’s always gonna be one very similar character characteristic.

Which is with the model adheres to the strongest in the prompt

[torn mantle]

idk 😭

maybe a week

i will see if i can make your request seperated

[ocean bison]

Fine, I will do it myself with Gemini 3 Pro 😂😂😂

[keen beacon]

What part of the prompt has the strongest adherence that the model will follow?

The upper section the middle section the lower section?

[zealous sparrow]

Asura have you tested neo nucleus stealth model

one that claims its from google

[keen beacon]

And vice versa if the very last piece of your prompt says never mind I want a whole different image of just a “ dog running” will it ignore the previous instructions or will it struggle to maintain that original top section and also throwing the dog?

Try this prompt

𓋴𓅱𓅓𓅂𓅃𓉔𓅂𓂋𓅂𓐍𓇋𓈖𓐍𓎼𓅂𓂋𓅓𓄿𓈖𓇌,𓐍𓏏𓉔𓅂𓐍𓆑ü𓉔𓂋𓅂𓂋𓐍𓇋𓋴𓐍𓂋𓅂𓃭𓄿𓇨𓇋𓈖𓎼,𓐍𓃀𓃭𓇋𓋴𓋴𓆑𓅲𓃭𓃭𓇌𓐍𓅲𓈖𓄿𓅃𓄿𓂋𓅂𓐍𓅱𓆑𓐍𓏏𓉔𓅂𓐍𓎢𓉔𓄿𓅱𓋴𓐍𓃀𓂋𓅂𓅃𓇋𓈖𓎼𓐍𓅂𓃭𓋴𓅂𓅃𓉔𓅂𓂋𓅂𓏞𓐍𓇋𓈖𓐍𓉔𓇋𓋴𓐍𓄿𓃭𓊪𓇋𓈖𓅂𓐍𓂋𓅂𓏏𓂋𓅂𓄿𓏏,𓐍𓉔𓅂’𓋴𓐍𓇋𓈖𓐍𓉔𓇋𓎼𓉔𓐍𓋴𓊪𓇋𓂋𓇋𓏏𓋴𓏞𓐍𓉔𓅂𓐍𓅂𓈖𓆓𓅱𓇌𓋴𓐍𓄿𓐍𓎢𓄿𓂋𓅂𓆑𓂋𓅂𓅂𓐍𓄿𓆑𓏏𓅂𓂋𓈖𓅱𓅱𓈖𓐍𓄿𓏏𓐍𓉔𓇋𓋴𓐍𓅓𓅱𓅲𓈖𓏏𓄿𓇋𓈖𓐍𓃭𓅱𓂧𓎼𓅂,𓐍𓋴𓄿𓆯𓅱𓂋𓇋𓈖𓎼𓐍𓏏𓉔𓅂𓐍𓎢𓂋𓇋𓋴𓊪𓐍𓄿𓇋𓂋𓐍𓄿𓈖𓂧𓐍𓏏𓉔𓅂𓐍𓎼𓅱𓃭𓂧𓅂𓈖𓐍𓉔𓅲𓅂𓐍𓅱𓆑𓐍𓄿𓐍𓅓𓅱𓅓𓅂𓈖𓏏𓄿𓂋𓇌𓐍𓊪𓅂𓄿𓎢𓅂𓏞
𓋴𓅲𓂧𓂧𓅂𓈖𓃭𓇌,𓐍𓄿𓐍𓃭𓅱𓎢𓄿𓃭𓇋𓊃𓅂𓂧𓐍𓇋𓈖𓏏𓅂𓂋𓈖𓄿𓃭𓐍𓎢𓂋𓇋𓋴𓇋𓋴𓐍𓂧𓅂𓅓𓄿𓈖𓂧𓋴𓐍𓄿𓐍𓋴𓏏𓂋𓄿𓏏𓅂𓎼𓇋𓎢𓐍𓅃𓇋𓏏𓉔𓂧𓂋𓄿𓅃𓄿𓃭𓏞𓐍𓉔𓅂𓐍𓄿𓃀𓄿𓈖𓂧𓅱𓈖𓋴𓐍𓏏𓉔𓅂𓐍𓋴𓎢𓅂𓈖𓅂𓂋𓇌,𓐍𓅓𓄿𓂋𓎢𓉔𓇋𓈖𓎼𓐍𓏏𓉔𓂋𓅱𓅲𓎼𓉔𓐍𓏏𓉔𓅂𓐍𓉔𓄿𓃭𓃭𓋴𓐍𓅃𓇋𓏏𓉔𓐍𓄿𓐍𓂋𓅂𓈖𓅂𓅃𓅂𓂧,𓐍𓆑𓂋𓄿𓈖𓏏𓇋𓎢𓐍𓅲𓂋𓎼𓅂𓈖𓎢𓇌𓏞𓐍𓏏𓉔𓇋𓋴𓐍𓇋𓋴𓐍𓄿𓐍𓎢𓄿𓅓𓊪𓄿𓇋𓎼𓈖𓐍𓏏𓉔𓄿𓏏𓐍𓎢𓄿𓈖𓈖𓅱𓏏𓐍𓅃𓄿𓇋𓏏𓏞
𓉔𓅂𓐍𓂋𓅂𓄿𓎢𓉔𓅂𓋴𓐍𓉔𓇋𓋴𓐍𓇋𓈖𓈖𓅂𓂋𓐍𓋴𓄿𓈖𓎢𓏏𓅲𓅓—𓄿𓐍𓎢𓉔𓄿𓅓𓃀𓅂𓂋𓐍𓅱𓆑𓐍𓎼𓇋𓃭𓂧𓅂𓂧𓐍𓅱𓊪𓅲𓃭𓅂𓈖𓎢𓅂𓏞𓐍𓆑𓄿𓎢𓇋𓈖𓎼𓐍𓉔𓇋𓋴𓐍𓋴𓉔𓇋𓅓𓅓𓅂𓂋𓇋𓈖𓎼,𓐍𓅓𓅂𓏏𓄿𓃭𓃭𓇋𓎢𓐍𓏏𓉔𓂋𓅱𓈖𓅂,𓐍𓉔𓅂𓐍𓊪𓂋𓅂𓊪𓄿𓂋𓅂𓋴𓐍𓆑𓅱𓂋𓐍𓄿𓐍𓋴𓇋𓈖𓎼𓅲𓃭𓄿𓂋,𓐍𓏏𓉔𓅲𓈖𓂧𓅂𓂋𓅱𓅲𓋴𓐍𓅂𓈖𓎼𓄿𓎼𓅂𓅓𓅂𓈖𓏏,𓐍𓂋𓅂𓄿𓂧𓇌𓐍𓏏𓅱𓐍𓎢𓅱𓈖𓎡𓅲𓅂𓂋𓐍𓄿𓐍𓋴𓏏𓂋𓅲𓎼𓎼𓃭𓅂𓐍𓏏𓉔𓄿𓏏𓐍𓇋𓋴𓐍𓋴𓏏𓂋𓇋𓎢𓏏𓃭𓇌𓐍𓊪𓅂𓂋𓋴𓅱𓈖𓄿𓃭𓏞
𓅃𓅱𓅲𓃭𓂧𓐍𓇌𓅱𓅲𓐍𓃭𓇋𓈎𓅂𓐍𓅓𓅂𓐍𓏏𓅱𓐍𓏏𓂋𓇌𓐍𓄿𓐍𓆯𓅂𓂋𓋴𓇋𓅱𓈖𓐍𓏏𓉔𓄿𓏏'𓋴𓐍𓅂𓆯𓅂𓈖𓐍𓊪𓅲𓈖𓎢𓉔𓇋𓅂𓂋,𓐍𓅱𓂋𓐍𓊪𓅂𓂋𓉔𓄿𓊪𓋴𓐍𓅱𓈖𓅂𓐍𓅃𓇋𓏏𓉔𓐍𓄿𓐍𓅓𓅱𓂋𓅂𓐍𓎢𓅱𓅓𓅂𓂧𓇋𓎢𓐍"𓈖𓅂𓅃𓋴𓂋𓅂𓅂𓃭"𓐍𓏏𓅱𓈖𓅂?

[torn mantle]

eh?

not yet tbh

[torn mantle]

go for it

[zealous sparrow]

claims its google made

its only on textarena