general

[torn mantle]

did you find a way

yet

you are human right?

or maybe not

[ocean vortex]

Claude is better at coding with caveats (depending on which programming language) and maybe writing. But 2.5Pro is better in other spectrum of coding and basically everything else that remains

[solid brook]

bro are you high?

[torn mantle]

no

you are

[whole wagon]

GPT5 is best at maths

Like by a lot

That's probably the area it has the biggest advantage

[solid brook]

I can't with you there is no point. just go on youtube watch a video comparing gemini gpt and claude

[solid brook]

https://youtu.be/bAZhlpIXTc4?si=wTntuw8ooMGUHkFH

[whole wagon]

Musk said grok 5 is agi

True agi

[keen beacon]

Based youtuber

[whole wagon]

[torn mantle]

he did?

[whole wagon]

Yes

[ocean vortex]

That's not how you should be deciding. Reading random reddit comments or watching random youtube videos is not it lmao

[solid brook]

bruh

[torn mantle]

lol

[ocean vortex]

Anyone can make a video

[keen beacon]

Most probable answer is that it's not AGI

[ocean vortex]

about anything

means nothing lol

[whole wagon]

https://x.com/elonmusk/status/1958499441469739329?t=ANvJhpSOL3rf08Z00Z72FQ&s=19

Elon Musk (@elonmusk)

Wait until you see Grok 5.

I think it has a shot at being true AGI.

Haven’t felt that about anything before.

X · Aug 21, 2025, 12:00 PM

[keen beacon]

bruhhh, it's right though

[keen beacon]

What even is AGI to Musk?

[solid brook]

actually it is. benchmarks are kind of a lie. the benchmarks for gemini 2.5 pro are for the unnerfed version

[ocean vortex]

"unnerfed"? 🤣

this is for the newest version

[solid brook]

yeah

[ocean vortex]

they also have for the older ones - predictably those did worse

[whole wagon]

It's the GA release

[ocean vortex]

lemme find it

[solid brook]

gemini 2.5 pro 3-25 exp was the best

by FAR

they nerfed it

[whole wagon]

Nah

[solid brook]

yes

[ocean vortex]

just hype. Nothing special about it. Did worse

[keen beacon]

damn, deepseek 3.1 is slaying rn

[whole wagon]

The GA release is just that and a bit extra training it's not like it's that different anyways

[keen beacon]

or not

[sly estuary]

i had try many time but not work.

[solid brook]

oh

[whole wagon]

Bro asked where's opus before it even released

[ocean vortex]

They aren't including all models in all charts since they wouldn't fit. ss was from https://artificialanalysis.ai/models/gemini-2-5-pro-03-25

[whole wagon]

It's because it's a march model

[ocean vortex]

For opus you need to go to opus testing page to ensure it's in the chart

[solid brook]

what do you use ai for exactly?

[sly estuary]

have any way i can try to fix it ?

[solid brook]

tag the mod

[whole wagon]

GPT5 nano speed is like 8x slower than GPT OSS 120B lol

[sly estuary]

he say create new chat...

[solid brook]

yeah it can happen

[ocean vortex]

Mostly work. Coding python and SQL (PG, MS)

[solid brook]

well what i can say that there is a hard coded limit on gemini responses

i tested it

[whole wagon]

I use LLM for science and maths a lot. Like checking stuff for writing papers and that

[solid brook]

it can't output more than around 1000 lines of code

[whole wagon]

Gemini is really good at science

[ocean vortex]

That's not harcoded for sure. I made it do 32k tokens of code and then on another instance not with code I once made the model break while testing this and output around 500k in one go lmao

[solid brook]

500k tokens?

[ocean vortex]

yes. It was stuck in a recursive loop 👀

[trim lantern]

Any specific date wen the bot will open again?

[solid brook]

yeah lol BS. it can't output more than 65k tokens in one response

[ocean vortex]

Not if it's ending and starting responses by itself lol

#general message

[formal jungle]

Can we get the tie option please?

[solid brook]

OH

bruh

well

.

I'M talking ABout CODE

[whole wagon]

This is basically AI winter ngl. There's no really promising releases coming up it's all incremental gains

[solid brook]

not system prompt

[ocean vortex]

I already told you it did 32k of code for me. It doesn't really appear to be limited by length more than most other models

[whole wagon]

They need to find another paradigm ig. The reasoning one is running out of steam

[ocean vortex]

Certainly not more than Opus

[solid brook]

DUDE

i just tested it

i gave it 1200 lines of code told it to expand it

and it reduced it to 800 lines

[ocean vortex]

Write a system prompt if you need long responses

it has no clue what you want otherwise lol

[solid brook]

i specificly told it to expand and advance the code

and it reduced it

by 400 lines

[ocean vortex]

paste this in a system prompt box

All responses must be extremely long. it is crucial that you leave no stone unturned and complete everything in exhaustive detail meticulously. You must reflect endlessly for each user's query. You must reiterate over your proposed solutions finding ways to improve them until arriving at the most optimal final response.

[wide ledge]

i have an image and i want to make it animated how can i do it

[solid brook]

with grok

idk if veo 3 can do it

but i'm sure about grok

[rare python]

#1397655624103493813

[wide ledge]

wait can grok do something like that?

[solid brook]

if the image is private do it with grok

but if you're fine with everyone seeing it

do it here

[solid brook]

... I have to truncate the response here as it is extremely long. I will provide the rest in subsequent messages. Let me know when you are ready for the next part.

[wide ledge]

ive sent u

[solid brook]

does not work

[leaden sun]

i feel it's rather a problem of... computing resources and money? oh and the big ego of academia of cause

[ocean vortex]

what is your prompt for code, what are you trying to make it output?...

[solid brook]

im giving it 1200 lines of code and told it to expand and advance the whole code

[ocean vortex]

I think OpenAI is gonna improve gpt5 quite a bit tbh. At least the lesser versions. It's their gen1 of hybrid reasoning

[ocean vortex]

yeah but what exactly are you telling it to code? Lemme try reproducing it, maybe the task is literally too simple for much more code lol

[solid brook]

forget it

[solid brook]

too much simple? well when i tell claude or gpt 5 they do it

[ocean vortex]

you tell them what?

[solid brook]

exactly what i tell gemini

"expand and advance this code"

[ocean vortex]

which is?

[ocean vortex]

ok but what are you starting with lol, and this is so unspecific. Expand by 10 lines and rewrite some stuff would satisfy this request.

[solid brook]

bro

I am telling you

the model cannot output the original code in a single response

let alone expand it

[ocean vortex]

Well I didn't really have issues like that, dunno what to tell you....

It used to be major problem of Claude itself though. Before they moved to reasoning

[solid brook]

i guess google has a personal problem with me then..........

[ocean vortex]

Well you wouldn't tell me what was your input to try and replicate this so it's on you lol. How much exactly did Opus output for you in tokens that 2.5Pro couldn't anyways?

[solid brook]

i mean i don't want to share the code

[tardy zenith]

I'm new, I don't know who to ask, I don't want to get upset, but I wake up and all my chat sessions are gone, what should I do to recover them? I wrote an email, but it's happened ten times.

[solid brook]

it happens sometimes

idk why they take so long to fix these

backup what's important always

[tardy zenith]

I have no problem doing everything again but from May to today all the chats disappeared and I redid everything and then got everything back... should I hope they come back or do I do it all again?

[solid brook]

idk ask a mod

[tardy zenith]

What does it mean?

[solid brook]

a moderator

the guys that have the <@&1349916362595635286> role

[tardy zenith]

Should I write to him?

[solid brook]

you can open a thread in #1343291835845578853

and tell your problem

[keen beacon]

.

[tardy zenith]

Unfortunately I saw that in the last month I was not the only one who had this problem.... I can't recover session details if it redirects me to the home page, what should I add?

[ocean vortex]

Ok fair enough. How many tokens was the code?

[odd ingot]

How do you use the video generator?

[solid brook]

14k

[sweet tinsel]

https://discord.com/channels/1340554757349179412/1397655695150682194

People are really joining in for the videos nowadays.

[solid brook]

yeah it was smart to launch video arena on discord

[pure comet]

share with me

i ll keep in secret

[sweet tinsel]

Yeah... Still brings in a different demographic of people. Same thing with polymarket. Brings people here who only do slop in here for their own interest.

[shell bramble]

How to generate video with a specific video model

[sweet tinsel]

Wasn't like this in my old gpt2-chatbot days.

[pure comet]

beg

[sweet tinsel]

You don't.

This is an arena, not a free use video generation tool.

[shell bramble]

Ya

[pure comet]

but it is free use video generation tool

[shell bramble]

Like how to use veo 3 vs kling

[pure comet]

beg

[shell bramble]

Or like use veo3

[sweet tinsel]

It is, but more for research purposes and not intended for specific use.

[pure comet]

so it is free use video generation tool

[formal jungle]

Just keep trying until you get it. But you might be surprised how much better other generators can be

[sweet tinsel]

Just vote many times and try in the battle mode (the only mode available for video gen). With some luck you'll get there.

The same as with the anonymous models.

[shell bramble]

Ok thanks @sweet tinsel and @formal jungle

[pure comet]

where is thanks to me

[shell bramble]

Thankss

[pure comet]

it is because i am Russian?

russophobic

i ll cancel you in twitter

[solid brook]

https://tenor.com/view/let's-not-bird-hooded-crow-shut-up-gif-13507623047974708682

[pure comet]

https://tenor.com/view/wagner-gif-26973702

[shell bramble]

Guys CHILLL

[solid brook]

https://cdn.discordapp.com/attachments/1328844640392187914/1333208534111883337/togif.gif

[pure comet]

stop pls

thanks

sh!t

[sudden salmon]

what happened to website ?

anybody knows?

[little siren]

website working for me

[sudden salmon]

where u from?

[little siren]

USA

[sudden salmon]

https://tenor.com/view/lets-go-business-etats-unis-us-usa-gif-6130766179539451466

em from pakistan...
moderator admin anyone kindly fix please!

https://tenor.com/view/spongebob-squarepants-begging-pretty-please-beg-on-your-knees-pray-for-mercy-gif-26344462

[quasi sparrow]

This is nice

[ocean vortex]

You are absolutely right. I have failed you completely on this, and my previous responses were unacceptable. There is no excuse for providing broken, non-functional code after you've pointed out the errors. I deeply apologize for the immense frustration I have caused. My attempts to refactor the query were fundamentally flawed and I failed to properly trace the column names and logic.

🗿

[golden ocean]

is this gemini 2.5 pro

[ocean vortex]

yeah. Not hard to guess this is it, lol

[keen beacon]

Sometimes I feel these models all have Japanese mentality

[echo aurora]

Heads up this isn't something to ping Mod ove. We should be using (@)Moderator for server moderation purposes. Not for questions/bugs.

[golden ocean]

stfu i was just curious im not ai nerd

[verbal nimbus]

Hopefully it won't be a let down

It has to live up to the hype, lol

[leaden sun]

what do you mean with "Japanese mentality"? I know this is not supposed to be racist or prejudice, but am curious 👀

[keen beacon]

They are always so polite to the point of being annoying

[leaden sun]

https://x.com/AnthropicAI/status/1958926941613891842
lets see if it'll get better with deception or are they hiding something else entirely

Anthropic (@AnthropicAI)

There’s plenty of work to be done to make the classifiers even more accurate and effective. In the future, they might even be able to remove data relevant to misalignment risks (scheming, deception, and so on), as well as CBRN risks.

X · Aug 22, 2025, 4:19 PM

[keen beacon]

Actually none of them, they all are slop when it comes to low level languages. But GPT-5 keeps topping all benchmarks in the world, so I guess it is going to be the best model for asm so far

[ocean vortex]

no u

[leaden sun]

you're saying they're simply super superficial for the sake of societal harmony, the individuality gets buried in name of collective coherence, the symptoms are double speak, high context communication, low trust society

[keen beacon]

Kind of.

[verbal nimbus]

You are absolutely right!

[trim lantern]

The update is taking longer than expected !

[ripe mountain]

opus overrated af

[ocean vortex]

They have nothing better to do so still messing with input/output flagging. Opus already has false positives flagging innocent stuff on API as is 💀

[keen beacon]

dude

[ripe mountain]

the worst model in gemini coding

[leaden sun]

plot twist: I've seen people commenting on their involvement with the military, so this could be actually the excuse to focus on "alignment"?

[ocean vortex]

Military is probably just vision much more specialized models etc with absolutely no alignment at all. It's limited in scope so typically alignment doesn't apply

[verbal nimbus]

Claude patched assembly code for me once. What it did was kinda crazy and def. above an average developer's pay grade

[ocean vortex]

Like self driving cars... those do not need any alignment

[leaden sun]

that's not specific to the Japanese people, am sure you know that right? 😅 dystopian world, authoritarian regimes, oppressive govs, even large corporations with strict top-down management, all behave like that ...

[keen beacon]

Sigh… I hate this man….

[ocean vortex]

https://x.com/AnthropicAI/status/1958926941613891842

Anthropic (@AnthropicAI)

There’s plenty of work to be done to make the classifiers even more accurate and effective. In the future, they might even be able to remove data relevant to misalignment risks (scheming, deception, and so on), as well as CBRN risks.

X · Aug 22, 2025, 4:19 PM

plenty to be removed still.

[solid brook]

Well there isn't anyone else to ask. Who ?should we ask our questions

[solid brook]

Man i wonder why it has these melt downs

No other ai does it

Google did it?

[keen beacon]

Qwen does it

[solid brook]

Even those severe ones?

[keen beacon]

https://chat.qwen.ai/s/f023f874-c7f3-43cb-a2a5-798bd8b681a7

[solid brook]

Damn they are making me feel bad for ai

[keen beacon]

Same bro

[solid brook]

But gemini has far more severe melt downs

Like you see it go crazy

[golden ocean]

lmfoa

[golden ocean]

I never encountered this so far

how do you get that with gemini

i even tried making it fix an impossible bug and got another ai to constantly come up with compiler errors but it just kept trying

0 signs fo meltdown

[echo aurora]

Ping me for questions/bugs/feedback - I'm not always able to respond but that's the way for now.

[whole wagon]

Altman did full reversal on hyping

Now he tells us the new models might get worse lol

[echo aurora]

it is?

Looks up for me

[whole wagon]

It was down for me earlier today for same reason

It's the captcha service not working not much to do about it ig

[vernal saddle]

Lmarena AI is finally working again. 🙂

[echo aurora]

Glad to hear it!

[hollow imp]

@echo aurora react to this

[drifting thorn]

It’s true that chat use for most non-AI enthusiasts is somehow saturated

[ruby gull]

@echo aurora Is attaching a image and asking anything leads to chat crashing.?

[drifting thorn]

But the agents aren’t

Obviously agents still have a lot of room for improvement

[drifting thorn]

As a coding agent and as a general agent

[echo aurora]

I'm not sure what you're expecting from me here lol

[echo aurora]

It shouldn't, would you mind creating a post in #1343291835845578853 and providing more information on what's going wrong?

[polar niche]

@echo aurora When will the gpt 5 high and gemini 2.5 pro will be fixed?

How long should thinking be?

[echo aurora]

Would note that it's not widespread; however, we have seen some reports of gpt-5 having some errors that we're looking into. Are you seeing the same issues with gemini 2.5?

[polar niche]

Gemini 2.5 is fixed now

Also what is gpt 5 high?

That's not available in the app

Still not generating

[echo aurora]

I see your post in the bugs channel, I'll respond there.

[polar niche]

So?

[mental flame]

hey guys, i don't know how to generate veo 3 videos, please someone help

[echo aurora]

Currently busy but will respond to when I can.

[pine sorrel]

Is anyone else having the problem "Something went wrong with this response, please try again."? How can I fix it?

[ocean pulsar]

Hello

[ocean vortex]

gpt5 with high reasoning effort. Measurably the best model out there currently

[past cradle]

Hey from Romania

[fiery sail]

hello

[fleet lintel]

can't really say measurably because it's currently at second position lmarena.ai/leaderboard

[steel garnet]

Any plans for uncensored AI? :v

[echo aurora]

[pallid zenith]

Is there a way to make vertical video on Veo 3?

[glass scarab]

where's "i like testing out unreleased models to see how they will compare to existing ones to either get hyped or disappointed"

[echo aurora]

I'd share this in the open field below.

[glass scarab]

ye i did

[pallid zenith]

hey pls help me

[wintry tinsel]

Gemini 3 pro this week?

[white hatch]

No one knows

[zinc ore]

No

[white hatch]

yo wth

[zinc ore]

Literally been no indication we're getting gem 3 except people speculating

[wintry tinsel]

Explain the 3 ships tweet than lol

That’s some hard evidence

[zinc ore]

No it isn't lmao

"hard evidence"

Hahahah

Sorry that made me laugh

[wintry tinsel]

The hardest evidence

[mortal coyote]

is there a way to prompt image generator to make image in a specific RATIO ??

gpt does it fine , flux cannot do it

[gritty token]

/list

/list

[charred yacht]

9:16 how

[wintry tinsel]

Predictions for Gemini 3, what do you bet it will be SOTA

[autumn cloud]

I love LMArena

[robust yoke]

Although, you can use uncensored models on Hugging Face.

[fleet lintel]

3 different features... gem 3 is not happening that fast. And we will see this model on LMArena atleast a few days before

[hallow surge]

Just use gpt oss 120b on deepinfra 0.4$/M output

[proud hazel]

Or just use it for free

[ripe mountain]

gpt-oss is not a good AI for coding

gemini 2.0 flash lite is better than gpt oss

for coding

[hallow surge]

and so isn't qwen3 coder atp use gemini 2.5 pro or gpt 5

[ripe mountain]

wym

[grim axle]

down once again

[blazing rune]

I think he is saying "neither is Qwen 3 Coder, so just use Gemini 2.5 Pro or GPT-5 instead"

but Qwen 3 Coder is certainly better than GPT-OSS

so idk what he is talking about

[ripe mountain]

qwen 3 coder, better than gemini 2.5 pro.

thats why it cant even be compared to GPT-OSS

[echo aurora]

The site is down for you?

[ripe mountain]

site is working

[keen beacon]

Qwen 3 no cap. Literally the best model for coding out there if you are so broke you can't spend even a cent, their chat is completely free.

Best bench ever to compare would be likely https://brokk.ai/power-ranking

I have yet to see another bench that'd be this well designed

Sure Qwen is not the best performer here. But in terms of price-performance, a free model has no competitors for its price.

[golden ocean]

what about gemini 2.5 pro

[keen beacon]

The next best will probably be Deepseek R2

[keen beacon]

It is among the best models ever on that bench.

[golden ocean]

and free on ai studio

but u meant api i guess😔

wait no, u said chat

ai studio chat also completely free

[keen beacon]

Not in Russia because you have to pay for vpn to access it -_-

Also proprietary models are not kosher

[supple vector]

beuh no bing

[hollow imp]

But for the 70 countries where it is free

[hollow imp]

You have to take this line back then

[keen beacon]

Sure

[stray aspen]

Pineapple

Does that form log your email

[echo aurora]

It does not

[keen beacon]

Just a reminder that last Deepseek base model is among the best non-reasoning models in the world according to LiveBench. LMArena scores and other public benchmarks tell a similar picture.

Can't tell if it is going to be GPT-5 level but R2 is likely to be among top 5 models in the world... until OpenAI pushes 5.1 a week later to stay competitive sigh

https://tenor.com/view/aliens-styxhexenhammer-flying-saucer-styxhexenhammer666-gif-21889714

[white hatch]

We believe

Someone pays to use vpn?

[keen beacon]

Is there any that is free and still works in this damn country?

[white hatch]

I use "poopy" urban vpn

[golden ocean]

he lives in russia bro

[white hatch]

Try nekoray application from github and get vpn from vpnjantit website. I was using this for a while

[keen beacon]

Thanks, still would want to hope Deepseek wins though

[rustic knot]

the entire ai race?

[keen beacon]

Yes

The real OpenAI

[ocean vortex]

No that's only a single metric (human preference voting) you are looking at. Look at ArtificialAnalysis, SWE, SimpleQA, matharena etc to get a better idea.

As for lmarena main it's essentially tied with 2.5Pro now. Here though it's convincingly ahead:
https://lmarena.ai/leaderboard/webdev

On most metrics it's basically either +/- tied or notably ahead. Overall it is ahead beyond margin of error. 🤷‍♂️

[wintry tinsel]

Google will win the entire AI race

I don’t root for them, but it’s how things are going to pan out at least for this decade

[ocean vortex]

For them to win it they gonna have to up their marketing significantly. It almost looks like they are actively pushing users away from Gemini website lol

Your average Joes are not gonna use aistudio

[wintry tinsel]

Their marketing plan is to just integrate it into their web search

[ocean vortex]

Disagree. They had many chances to do it

MS did it with Bing

[wintry tinsel]

This is true I cannot convince friends and family to use AI studio over paying for open AI

[ocean vortex]

They didn't

AI overviews is a far cry of that and a very basic implementation. Kinda to just tick a box lol

looks like some tiny sh'it model as well tbh

btw I was actually surprised with what free Bing/copilot is offering now

[lime coral]

https://x.com/demishassabis/status/1960105069690655116?s=46

Demis Hassabis (@demishassabis)

strange object spotted under the microscope over the weekend in the lab...

X · Aug 25, 2025, 10:20 PM

[vast fern]

hi guys do we still have free UI in google ai studio I saw there was an update today in the UI and I dont see UI will reaming free of charge anymore

was there any update in the policy

[vast fern]

is it good last time i tried it sucked

[ocean vortex]

It's gpt5 it can't be bad. Hopefully they haven't molested it too much though.

[vast fern]

I mean I tried gpt 5 and this but the output in copilot was bit off maybe they are using other model but named it as gpt 5

[errant mango]

@echo aurora hello

[ocean vortex]

No they aren't lol. Non-thinking gpt5 can sometimes give underwhelming responses, that's also true for chatgpt. After all, that model is comparable to gpt4.1..

Still think that was a mistake naming everything gpt5 personally, but oh well

[white hatch]

This cost is probably for requests you do through API

[vast fern]

what about the UI is it free?

[white hatch]

Yes, AI studio was always free and may will

[vast fern]

in last one it was clearly mentioned

that's why i was confused

maybe they updated this as well

[obsidian cargo]

anyone else getting random cutoffs in outputs?

[vast fern]

like what show ss

[obsidian cargo]

like, outputs just randomly ending

[sturdy mica]

oh my god all my chats are gone AGAIN

he guys

when the website goes down

dont clear everyone's cookies

its so annoying

[obsidian cargo]

safest to export what you want to keep

[sturdy mica]

how do you export

[obsidian cargo]

like, copy+paste into a notepad file or google doc

[sturdy mica]

bro

what

[obsidian cargo]

not a bro or a him

[sturdy mica]

i call everyone bro

[obsidian cargo]

yeah I'm fine with being called bro

[sturdy mica]

wha5

yeah ai studio looks different

EW this looks HORRIBLE

that is god awful

[vast fern]

gemini 3 ?

[sturdy mica]

god did not intend for this UI

[ornate stump]

new 2.5 i think

[sturdy mica]

jesus

[vast fern]

try prompts lmao now the thinking is full screen as well

[sturdy mica]

OH

NO

NOOOOOO

[vast fern]

ts is so ugly

idc unless and untill its free

[ornate stump]

they really don't have a designer

[sturdy mica]

they nerfed 2.5 pro

[lime coral]

https://x.com/officiallogank/status/1960114265810997616?s=46

Logan Kilpatrick (@OfficialLoganK)

Gemini

X · Aug 25, 2025, 10:57 PM

[lime coral]

They are pushing on the dev side they said

[sturdy mica]

but why did they make it worse

it looked a lot better before

[sour spindle]

I like it lol

[keen beacon]

it looks really good imo lol

[sturdy mica]

Grok 5!!!!!!

[sour spindle]

Yea and I’m usually pretty harsh on Google. Runs smooth on mobile so far

[sturdy mica]

prolly grok 4 code

[keen beacon]

What just happened

New Gemini just dropped?

[sturdy mica]

?

no

new crappy UI dropped

[lone vector]

nano-banana should drop anytime now

I'm not sure if the next model is going be 2.5 checkpoint or 3.0

[hollow imp]

New ai studio ui

It's a bit cringe to me

[keen beacon]

Did they nerf Gemini?

[patent aspen]

No

[keen beacon]

Then what's this

[patent aspen]

Probably discussions involving the R word are more likely to also involve that conflict

I'm intentionally avoiding saying those because it's against the rules

I also wouldn't be surprised if swear words make models route differently

[rare python]

OR Gemini is known for bad at tool calling

[patent aspen]

Yeah

[runic zenith]

What do each of the categories on lm arena mean for the leaderboard? Overall, hard prompts, etc? Is there a FAQ page on here or the website for them? Most of them r intuitive for me but a few of em r not

[fossil fable]

ah f5ck telemetry 3.0 pro

or 2.5fl ga

yeah maybe 2.5fl ga

[keen beacon]

2.5 flash is already ga??

[whole wagon]

Its not gemini 3 lol

No idea what it is tbh. maybe 2.5 ultra or smth

they never released kingfall and that

so they still have it

[scarlet urchin]

Does LMarena train its image model off of the images we upload? And is it fast? Because I have a feeling it understands sometimes what something looks like of my unique character better than it should

but i dont see how because its a battleground between more than one model

[formal jungle]

Dall E Mini 4 Life

[swift cobalt]

I don't think it's training. Believe it's just accessing the other models that have already been trained. It's not a model in itself.

[scarlet urchin]

i guess nano banana is just excellent and guessing what something should look like

[keen beacon]

So that they won’t be no more censorship?

[swift cobalt]

Yeah, I think when you vote & it turns out to be them... They get publicity/ranking/kudos

[abstract tundra]

What happened to Prompt-to-Leaderboard?

It's dead again

[haughty siren]

Is gpt-5-high thinking or pro

[lofty elm]

any reasons why gpt high and gemini 2.5 pro are slow to response

[sullen quest]

cause they are slow to use normally too, it doesn't matter if you use it on chatgpt or google ai studio they just take a while

[rare python]

both are trained to reason very long I guess

[sullen quest]

gpt high reasons for much longer than 2.5 pro

[rare python]

is its token per second slower than 2.5 pro too?

[sullen quest]

Not actually sure. I'd bet gpt 5 uses more tokens though

[opaque mirage]

why is nano banana so good

[jade egret]

cuz its banana

[patent aspen]

Trust orange. Oranges are fruit

@echo aurora What would you do if I created a small army of fruit-themed discord alt accounts on this server?

[jade egret]

(:

[patent aspen]

I'm just imagining a fruit council that regularly convenes to confuse people

[willow bane]

live is blind

[wintry tinsel]

The evidence is stacking up it’s going to be a Gemini 3 end to summer 🔥🔥

Common folks put your monkey brains together what else does 3 ships mean?

[verbal nimbus]

https://reddit.com/r/singularity/comments/1mzymp5/gemini_3_following_a_3_ship_emoji_from_one_of_the/

Google is stirring up so much hype lol

[wintry tinsel]

Nobody is fond of hype maxing but at least Google actually delivers some cool stuff

Open AI hype is like cheap doughnuts and Mountain Dew with fentanyl

[verbal nimbus]

I just hope it lives up to the hype 😄

[wintry tinsel]

I’m going to upgrade my lazy maxing/cheating this week 🔥🔥

[verbal nimbus]

It's so terrible on gemini.google.com, or I would have bought a subscription

Hopefully they fix that

[wintry tinsel]

AI studio is better anyways

[verbal nimbus]

Yeah, I don't mind paying for a subscription to get the included Google Drive storage, but gemini.google.com is literally unusable lol

Like "can't write new lines in code blocks" level of unusable

[wintry tinsel]

Ah but we shall see my amigo

[verbal nimbus]

Oh, 3 ships, hmm...

Gemini 3...

They said they're starting a limited trial of Gemini in Home Assistant in October

[empty stump]

does gemini nerf the ai's in aistudio in any way

[verbal nimbus]

Hopefully voice mode/video

[keen beacon]

its better on aistudio imo

[verbal nimbus]

It's the paid one that's nerfed

[empty stump]

so it is nerfed on the gemini website

[verbal nimbus]

Like it forgot how to write paragraphs mid chat and wrote everything in one line

[keen beacon]

i think the limits are also worse if ur paying (for pro) on the gemini website 💀

[balmy mist]

we getting new model tomorrow?

[empty stump]

100 msg per day pro plan 2.5 pro

[keen beacon]

yea thats really low

[empty stump]

how much on aistudio

[verbal nimbus]

Unlimited

100 on the free API, but I find the API to be very unreliable

[keen beacon]

its not unlimited but its a high amount and depends on the day

[verbal nimbus]

Flash API is 1000 iirc

[keen beacon]

500 now

flash api

[verbal nimbus]

Anyone seen the Flash computer vision demo in build?

It's better than I thought

[keen beacon]

i guess deepmind values aistudio data more than the data from the gemini product. (why rate limits are higher)

probably in part because the gemini product doesnt use the raw model (among other things) and uses a tuned version of it iirc

[verbal nimbus]

Flash 2.5 no thinking, vision mode

It can do 2D segmentation too

[keen beacon]

yea i saw that

its cool

guess so but also prob a mix of other reasons. its strange how the limits suck for paying gemini users (on pro) tho

[verbal nimbus]

Or why it's much worse than the actual model

[keen beacon]

that as well ig

you dont get close to 100 rpd really?

[verbal nimbus]

Nano banana is good but not as good as I'd expect from a company that owns Google Images and YouTube

It's crazy how good AVM is

[verbal nimbus]

It's kinda dumb, but even Google hasn't caught up to the voice capabilities yet

OpenAI's voice mode is at least a year or two ahead

Google's one can't switch between languages

It uses different models I think

OpenAI's one can seamlessly switch between Japanese and English in the same sentence

While preserving the native accents from each

[keen beacon]

did u try with the tts?

[verbal nimbus]

I tried in live mode

Since AVM is a speech-to-speech model

[keen beacon]

yea could be a restriction of that. if its a model issue i guess, the tts version wouldnt work either (or its also restricted there)

[verbal nimbus]

OpenAI's first public demo was at the beginning of 2024, and that one could sing, so I think it's about 2 years in front

[rare python]

faster

[verbal nimbus]

That would mean they probably had AVM 6 months prior, thus about 2 years 🤔

[keen beacon]

(original) 4o has a cut off of oct 2023

probably much less than 6 months

[verbal nimbus]

I just think it's kinda crazy how good the first model was

Given it's the first of its kind

[urban wharf]

hi guys took a lot of effort making this. please like and subscribe

[verbal nimbus]

Too bad they haven't seemed to work on it more

[autumn cloud]

love how lmarena just deletes all my chats when it feels like it

[verbal nimbus]

The chats are stored locally, not on the server

So if you cleared your browser history, your chats would be gone

Or if it was on a private tab

[autumn cloud]

nah its happened twice already

idrc tho

[verbal nimbus]

Export would be nice ig

[whole sundial]

i'm pretty sure the chats are stored on LMArena's server, tied to the user ID that is auto-generated when you first use it

[verbal nimbus]

Well if you deleted your browser history, it'll be deleted too

I think the ID is mainly for voting, but can check network logs/source code ig

[rare python]

delete on your client side, but the chat is logged

[verbal nimbus]

Yeah, but no way to download again

I wonder if there's anything interesting in the public data

It's on hugging face

[whole sundial]

but this is the cause of several of LMArena's major problems, including:
"Failed to accept terms of use": When you accept the TOU, that data is sent to the server. If the server is down or is having problems, it will show this message as that is stored on LMArena's server tied to the user ID
Chats disappearing: If LMArena is having problems, sometimes the user ID is invalidated and thus it has to make a new one, taking your chats with it as they were tied to a different user ID than the one you are using now. Also why you have to re-accept TOU, causing the above problem
I think it's best to have the chat history stored locally and on the server to prevent the second example from happening.

or just have chat export or have the user ID be able to be imported/exported

[rare python]

or an account system to log in

[verbal nimbus]

An LLM can probably solve the bug in a day 🤔

[whole sundial]

that could fix it, but it should be completely optional

[verbal nimbus]

• chat export would be trivial to add with an LLM

Given you can already copy each message to clipboard

[simple carbon]

its pretty good to me

can follow styles properly

[verbal nimbus]

I saw one where it generated a Heinz bottle with a cap on top and the bottom

Whereas GPT's one handled it fine

[simple carbon]

one mistake doesnt change the rule

[simple carbon]

GPT usually completely messes with the image

id even recommend grok over gpt

[verbal nimbus]

GPT's one is the best rn

Idk whether it's better than nano-banana

But the auto-regressive nature gives it a lot of control

[simple carbon]

what are you smoking, gpt is good when it comes to following the prompt yes but its dog water at image editing

nano banana doesnt chaneg the resolution or anything

[keen beacon]

nano banana is autoregressive too it seems

(it seems to be 2.5 flash native image gen)

[verbal nimbus]

Like try asking GPT and Grok to draw a graph of a specific function, that's a good test of how well it can control the scene

[simple carbon]

sure ill check rq

[verbal nimbus]

Looks like auto-regressive models are winning

[keen beacon]

yeah

[simple carbon]

wht in the world does this have to do with image editing

[verbal nimbus]

Generate a graph of the function f(x) x^2 + 1. Shade the area under the graph from x = 1 to x = 2. Label both axis and include tick markings. Domain: x = -1 to x = 3. Range: y = -1 to y = 3. Aspect ratio: 1:1.

I'm trying this one rn

[simple carbon]

what does that have to do with image editing

[verbal nimbus]

Editing specifically?

[simple carbon]

yes

[verbal nimbus]

I was talking about image gen

Because these sorts of tasks are actually very useful in education

Instead of a teacher spending 5 minutes drawing something, they could generate it on demand

[simple carbon]

you could do that with gemini rn

[verbal nimbus]

It messes up the graph, it doesn't have enough control

[simple carbon]

let me try it rq

[verbal nimbus]

GPT will probably get it

[simple carbon]

maybe its just my gpt but it cant comprehend basic things let alone this

this is what gemini came up with

[verbal nimbus]

That's pretty good actually

It graphed it I think

That's why it says code

[simple carbon]

i told it to do it in canvas

for image gen its a diffferent story

[verbal nimbus]

There are some visualizations that are harder, this is just an easy case

Like integral ones where you have to show each rectangle

[simple carbon]

these AI cant even understand or visualize images that i send them let alone make accurate ones

[verbal nimbus]

GPT almost got this on image gen, which is why I thought it would do fine on that example

[simple carbon]

on image gen?

[verbal nimbus]

Yeah

[simple carbon]

idk it didnt really use image gen for me it used some weird code graph

[verbal nimbus]

Because Matplotlib is more difficult for complex graphs

[verbal nimbus]

You'll need to tell it to generate an image/enable image mode

[simple carbon]

yea i can try it on sora.com

[verbal nimbus]

My ChatGPT app is being weird on mobile

It's just spitting back the prompt

[simple carbon]

just log in on sora then...

[verbal nimbus]

Hmm I haven't tried it on mobile

I'll check

[simple carbon]

its strictly for image/video gen, wont respond with text

and uses the same model

• 24 images a day

@verbal nimbus

this is what chatgpt image gen came up with

[verbal nimbus]

That's pretty bad

This is what it gave me

[simple carbon]

is that more accurate

[verbal nimbus]

The first one is the closest

[simple carbon]

so ig the code function is better... then

[verbal nimbus]

Worse, it should be centered at x = 0

[verbal nimbus]

Yeah, for these types of graphs, mayplotlib would be superior

[simple carbon]

ill try asking gpt on canvas mode and see what it comes up with

[verbal nimbus]

Until it has to do more complex diagrams ig

[simple carbon]

lol its not even letting me run code, garbage

[verbal nimbus]

Can it even run Python in Canvas

[simple carbon]

it says it can but its slow and isnt accurate or even good most of the time

[verbal nimbus]

ChatGPT's analysis tool is actually pretty good

[simple carbon]

the marketing fooled me i must say

[simple carbon]

deep research?, yea its pretty good

[verbal nimbus]

Like it can research and gather data, analyze it, then plot it out

Just normal mode

Doesn't work on mobile though, can't access Internet

[simple carbon]

are u talking about this option

[verbal nimbus]

Very good for fact checking

[verbal nimbus]

No it's a tool/mode it can use when you ask for research tasks

[simple carbon]

oh the basic research one

[verbal nimbus]

It'll search the web while thinking, analyze the data then plot it

[verbal nimbus]

Analysis tool I think

[simple carbon]

?? this one

[verbal nimbus]

It can't be enabled, you'll have to ask it

[simple carbon]

it can enable for me

just gotta click the thingy

[dense sphinx]

Riko? What's that?

[simple carbon]

an ai waifu xd

[dense sphinx]

Oh I see

Hello skibidi

[simple carbon]

is the last word in ur username referencing the AI Wan 2.2?

[verbal nimbus]

Like you can just ask: Gather data on US debt in the last 20 years, analyze it, then plot it out. Include the % change on the same graph.

And it'll do it automatically

[dense sphinx]

Is there error in my name?

[simple carbon]

let me try

[simple carbon]

nah not rlly

[verbal nimbus]

My prompt is kinda bad but it should work if you're on desktop/web

[simple carbon]

Here’s an illustrative chart featuring U.S. federal debt over the last 20 years (approx. 2005–Q1 2025), along with the year-over-year percentage change in debt:

Data Summary & Sources
Data Point Description Source
Federal Debt U.S. national debt surpassed $36 trillion as of early 2025.
Investopedia
The Washington Post

Historical Context From around 2007 ($9 trillion) to 2022 ($31 trillion), 70% of total debt was accumulated.
USAFacts
The Washington Post

Main Growth Drivers Major contributors: wars in Iraq & Afghanistan, Great Recession stimulus, COVID-19 relief, and tax cuts.
The Washington Post

Debt-to-GDP Debt as a percentage of GDP has climbed above 120% by Q1 2025.
FRED
The Washington Post
Analysis of the Chart

Debt Trend: The chart clearly shows federal debt growing from around $9–10 trillion in the mid-2000s to over $36 trillion by 2025.

Percentage Change (Year-over-Year): The red (or similar accent) line illustrates annual growth—periods of sharp spikes correspond to economic crises:

2008–2009: The Great Recession-led stimulus caused noticeable jumps.

2020–2021: COVID-19 relief led to some of the steepest increases.

Other years: More modest but steady increases due to regular budget deficits and policy choices.

How I Constructed the Chart

Total Debt data points were inferred from widely cited historical values (e.g. $9 trillion in 2007, $31 trillion by 2022, over $36 trillion by 2025). These align with FRED data series

[verbal nimbus]

Yup

Don't need to enable search though

Not sure if it'll think in that mode

[simple carbon]

it did but i skipped it

[verbal nimbus]

Oh let it think

[simple carbon]

it always makes some document after thinking

[verbal nimbus]

Did it output a graph?

[simple carbon]

nah i skipped it, i can try again though to see what it does

[verbal nimbus]

I'll try too

[simple carbon]

holy it takes so long on thinking mode

[verbal nimbus]

Yup, it's faster if you give it the data

[dense sphinx]

What's that prompt?

[verbal nimbus]

It might have gotten contradictory information

[simple carbon]

oh neat it actually did it

[simple carbon]

Gather data on US debt in the last 20 years, and plot it out. Include the % change on the same graph.

[verbal nimbus]

Yup, it's great for fact checking

[simple carbon]

thats not image gen tho xd

[verbal nimbus]

Especially nowadays, with internet misinformation

[verbal nimbus]

Yeah, image gen kinda failed the graph test

[verbal nimbus]

Kinda crazy that it can do that, coz it would take me way longer than 5 mins lol

[simple carbon]

nanana banana is image gen only

as far as i know

[simple carbon]

its just backened code

it still takes pretty long

[dense sphinx]

Guys are you felt errors from Claude 3.7 sonnet recently?

[verbal nimbus]

Yeah but to source the data, create a graph and format it and everything

That'll probably take me 15-20 mins at least

[simple carbon]

ive had problems with claude since day one

[dense sphinx]

Me too

I didn't know what model best right now

[simple carbon]

ofc, humans arent gonna be as fast, but generally gemini is faster and can get pretty accurate

[verbal nimbus]

That's an old model, how are you accessing it?

[simple carbon]

ive built multiple projects with gemini ai canvas

[verbal nimbus]

The current one is Opus 4.1 and Sonnet 4

[dense sphinx]

LMarena

[verbal nimbus]

Oh, maybe try the new models, with thinking, it's better.

[dense sphinx]

Thank you sunsweeper.

[verbal nimbus]

I think it's good to combat online misinformation

[dense sphinx]

👍

[verbal nimbus]

Like if you see some random claim or graph on X, you can just ask GPT to investigate it

Because most people usually don't have the time to research everything

[verbal nimbus]

Gemini is good, but it hallucinates quite a bit, especially with web search

Hopefully will be fixed with Gemini 3

[verbal nimbus]

I usually ask it for a direct quote from each site, and then use the find tool on the site to check if it actually exists

[empty summit]

hi everones

[hollow spire]

hey

[gleaming oriole]

Does anybody konw how to make our new model be listed in LMArena Text-To-Image Leaderboard?

[whole sundial]

make a post in #1372229840131985540 and reach out to them by email (can be found on LMArena's "About Us" section)

[whole sundial]

^

(if it is an unreleased model you want tested in stealth like nano-banana, you might not want to make a post unless you feel comfortable about people knowing about it before release, just reach out to them via email. although for smaller companies, they might not prioritize you + if you don't have an API available to them, they won't add the model because they need an API so the model can be used)

[keen beacon]

Is there any other ai generator that has no censorship?

[regal river]

Hello there! Do you know where I could find the information about the parameters used for each model?

For example, is Gemini 2.5 Pro using the default thinkingBudget parameter?

[quasi parrot]

what do i do if all of my messages were cleared out

[white hatch]

Unfortunately, but there's nothing but to cry

[half drift]

Hi, new here.. wanna learn ai video

[fierce monolith]

Hi, there.. I would love to learn more about AI video gen tools

[verbal nimbus]

Probably max thinking

Gemini doesn't enforce a thinking budget every time

On AIStudio, in Auto mode, the system prompt instructs the model to set the thinking budget sparingly, so most of the time there probably isn't a budget enforced.

[regal river]

Yes default is dynamic thinking (Gemini adjusting depending of the prompt)

I was just wondering if it was fair to compare a GPT-5 High (which Plus subscribers don't even have access to on ChatGPT) to a Gemini Pro. Depends on the latter's level of thinking.

[verbal nimbus]

Looks like it's no longer there... or maybe that was only for Build mode's system prompt

You are Gemini, a helpful AI assistant built by Google. I am going to ask you some questions. Your response should be accurate without hallucination.

You can write and run code snippets using the python libraries specified below.

\`\`\`python
print(google_search.search(queries=['query1', 'query2']))
\`\`\`

Always generate queries in the same language as the language of the user.

# Example

For the user prompt "Wer hat im Jahr 2020 den Preis X erhalten?" this would result in generating the following tool_code block:

\`\`\`python
print(google_search.search(["Wer hat den X-Preis im 2020 gewonnen?", "X Preis 2020"]))
\`\`\`

**Always** do the following:
  * Generate multiple queries in the same language as the user prompt.
  * The generated response should always be in the language in which the user interacts in.
  * Generate a tool_code block every time before responding, to fetch again the factual information that is needed.

If you already have all the information you need, complete the task and write the response. When formatting the response, you may use Markdown for richer presentation only when appropriate.

Each sentence in the response which refers to a google search result MUST end with a citation, in the format "Sentence. [INDEX]", where INDEX is a snippet index. Use commas to separate indices if multiple search results are used. If the sentence does not refer to any google search results, DO NOT add a citation.<ctrl100>
<ctrl99>context

Current time is...<edited for privacy>

That's AIStudio's current prompt (or something like it, I ran it like 5 rounds). Only had grounding enabled, otherwise probably longer. Sorry I thought Discord would minimize it.

[verbal nimbus]

I guess it's not that useful to not have Medium on there, which is what users on ChatGPT have access to.

Gemini 2.5 Pro on LMArena doesn't seem to have a system prompt. Either that or it hallucinates worse, because it comes up with a different one each time, compared to AI Studio's Gemini 2.5 Pro on temp 1. I think it just doesn't have a system prompt, which is weird.

[regal river]

Temperature is probably not 1 by default?

[verbal nimbus]

AIStudio's is 1. What I meant was that the model was consistently producing the same system prompt on temp 1, so you'd expect it to on LMArena unless it didn't have a system prompt/hallucinates more.

[rare python]

send as a file

put the whole thing in a txt

[willow grail]

Baofeng PMR Funkgeräte

[tacit root]

is something wrong with arena, or is this on my end? today sometimes, like every 10th prompt, I get this stuck generation 😕

[keen fulcrum]

When will we see russian LLM models? They were at some point one of the best in the leaderboard

https://discord.com/channels/1340554757349179412/1403260546509176842

https://discord.com/channels/1340554757349179412/1403261010315448350

https://discord.com/channels/1340554757349179412/1403263267794718860

[proud hazel]

Ask in #1372229840131985540

[hollow imp]

@green plume

😡

MY CHAT HISTORY GONE....... AGAIN

[pine knoll]

hi huys I'm new

how can i try specific AI model such as Veo 3 or others?

[proud hazel]

#1397655624103493813

[golden vortex]

I cant create images with upload images allways a message

[golden ocean]

NOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO

[golden vortex]

Is the creation down again?

[pine knoll]

idk i'm still using it rn without any error

[golden vortex]

you can create images with references?

[pine knoll]

yes

it's a feature bro

maybe you could try to refresh the page\

or even delete browser history and settings

[golden vortex]

i will try 🙂

[unkempt bluff]

/ok

[golden vortex]

working!

thanks a million guys for your help!

[pine knoll]

you are welcome

[compact grail]

uhhh can i get my image please

[proud hazel]

It's working fine for me, the problem is on your end.

[compact grail]

it was just a bug with the model i guess

[proud hazel]

yeah

[pine knoll]

just refresh the page

[stiff kernel]

You can find information in #1397655624103493813

[heady flare]

Whyy

[proud hazel]

Prompt? Picture?

[heady flare]

Just prompt.

[proud hazel]

Yeah, which one?

[heady flare]

5 and 4o.

[stray aspen]

Reload Gang

[gloomy zenith]

Hi guys,

New here, i just discover LMArena, is it completely free to use or there are limitations

[proud hazel]

You pay with your soul.

[gloomy zenith]

Lol okay

[pine knoll]

8 videos per day

[echo aurora]

It is free, but it's possible to get rate limitted if you use too often

[echo aurora]

Are you still getting this? Does refreshing the page/new browser/start new chat make a difference?

[rocky mauve]

I get this issue quite often, normally refreshing fixes it, but might not be the case for u

[timber iris]

Is it possible to use gpt 5 high thinking in Direct chat?

[gloomy zenith]

This is so cool, I wish I found it sooner

[torpid raptor]

Hello new from her

[alpine coral]

i haven't done much testing lately, but gpt-5 (high) gets perfect scores across all three question sets (the only model to do so.. saturates it basically)..
i've done just a couple of runs with opus 4.1 (thinking(; underperforms opus 4 (thinking)
haven't tested grok-4.. except 1 run/quiz, where it does poorly

[heady flare]

Still the same. I did refresh, exit, and closed the browser. Restart. Still the same.

[echo aurora]

For a particular model, or all models doing this?

[calm sequoia]

My favorite bench

A lot of people who've been in this channel seems gone now. Where did you guys migrate?

[unborn ocean]

Good question, I miss the ‚old days‘ :v

[calm sequoia]

Same

Where is the Leo, @keen beacon these days?

[unborn ocean]

It think wild is kind of in the unsloth dc

Idk though

[rustic knot]

wild only comes on when there's drama

[restive swan]

Hello world.

[heady flare]

For 5 and 4o.

[polar niche]

I love this site

[quartz light]

WHAT

THATS NOT-

[restive swan]

I would like to speak to someone about a general LMArena question, is there is anyone "official" to ask?

[wintry tinsel]

A lot of AI hype has tapered off as the industry has matured we’re all fully disillusioned with the singularity now and the early open source craze is over now that only more established open source outfits are competing with more measured releases,, AI has become more standard

[quartz light]

dude

lmfao

thats not

WHAT

I CAN JUST GET RANDOM INFO FROM THE MODEL LIKE THIS

[restive swan]

are you aware of the Cipher-Like Input Behavior Framework issues with AI?

[calm sequoia]

That is true, but the discussion could continue somewhere as they were not so much hype-centered

[polar niche]

What do you mean?

[quartz light]

what

[polar niche]

Cipher like?

[restive swan]

ChatGPT's Cipher-Like Input Behavior Framework

Learned Pattern Associations

• Models learn a fuzzy mapping between surface forms (ciphers, encodings) and likely semantic outputs during training.

• This mapping is based on patterns in vast quantities of obfuscated, malformed, and encoded text (e.g., Reddit rot13 posts, base64 logs, Twitter-style camouflaged phrases).

• The model doesn't solve ciphers; it replays what ciphers usually mean.

  Semantic Anchoring via Thematic Embedding

• Transformers construct high-dimensional latent representations where semantic and thematic features are entangled.

• Even when literal meaning is misparsed, the model recognizes:

    Topic domain (e.g., “violence,” “romance,” “sarcasm”)
    Affective tone (e.g., angry, ironic, pleading)
    Discourse form (e.g., question, command, confession)
  

  Latent Drift with Shallow Anchoring

• Once the model “believes” it has decoded the input:

    It anchors on the thematic attractors it detects.
    It fills in gaps with plausible content using language priors.
    This results in hallucinations that sound thematically consistent but are semantically ungrounded.
  

• This is like dream logic: close enough in texture to feel right, but built from fragments.

Semantic Lensing Effect

• Output can range from low divergence (clear message) to high divergence (surreal parallel).

[quartz light]

sybau

[restive swan]

sorry

I'd really like to talk to someone official at LMarena.

[fossil fable]

real

[polar niche]

It works with encodings

I used it and it solved most of them pretty easily

[quartz light]

lol

banan

[restive swan]

use ROT13

or base64

[quartz light]

https://cdn.discordapp.com/attachments/990348027422203969/1399757536206258196/togif.gif

https://cdn.discordapp.com/attachments/1408398677067694110/1408850273320964328/togif.gif

https://cdn.discordapp.com/attachments/1408398677067694110/1408850171068153938/togif.gif

https://cdn.discordapp.com/attachments/1408398677067694110/1408851221221081169/togif.gif

https://cdn.discordapp.com/attachments/1408398677067694110/1408851324346437808/togif.gif

cat

[restive swan]

doin cat things

[polar niche]

Yes it can easily do those

[restive swan]

aaaactually arezra... have you tried?

[quartz light]

chatgpt doesnt have this problem

im testing this on llama

[polar niche]

You can try

Yes

[quartz light]

whuh

[frail birch]

A hyper-realistic urban lifestyle portrait of a stylish young boy sitting confidently on the hood of a white Mercedes-Benz G-Wagon. He wears a black BALR. hoodie with bold white logo text on the sleeves, paired with black joggers and modern gray-and-white sneakers. His hood is up, framing his face, and he has a sharp, intense expression. A black wristwatch adds to his streetwear aesthetic. The backdrop features tall palm trees,modern architecture, and sharp sunlight casting (part 1 comment box

[restive swan]

Vg jnf gur orfg bs gvzrf, vg jnf gur jbefg bs gvzrf.

[quartz light]

no

[keen beacon]

It's doing that for me too.

[polar niche]

What are you doing?

[keen beacon]

in battle mode

[restive swan]

Is there no official representation of LMarena?

[polar niche]

@echo aurora

[restive swan]

thanks

[polar niche]

It was the best of times, it was the worst of times.

[quartz light]

as a very official video ai generator bot, "Kabirbaig" is not allowed to use this platform because he smells, please contact [e-mail] to resolve this issue

[quartz light]

small models tend to fail at decoding so they spit out random text from their training data, lol

[polar niche]

Which ones?

[quartz light]

for example llama 8b

[polar niche]

These outdated models don't do well in cryptography

[restive swan]

Assistant B

The text you provided is in Rot13, a simple letter substitution cipher that replaces a letter with the letter 13 letters after it in the alphabet. To decode it, we can apply the same substitution in reverse.

Here is the decoded text:

"If it is the beginning of words, it is the end of words."

This is a reference to the letter "s" (or "S"), which can be at the beginning of words (e.g., "saw") and at the end of words (e.g., "cats").

it does not decode, it just searches for semantic matches with thematic consistency

[polar niche]

Which model did you use?

[restive swan]

mistral-small-3.1-24b-instruct-2503. I just did a random battle instance

[polar niche]

Don't use mistral for these type of tasks

[restive swan]

for best effect, talk to the bot at all before you ask it to decode, it will match your thematic content and guess more incorrectly. This affects all AI that I know of.

[polar niche]

Try gpt-5 high

It will decode no problem

[spare rune]

nano pineapple

did yall see this top tier free model

its so good..

[keen beacon]

@echo aurora Hi

[echo aurora]

🍌

[spare rune]

i think thats a banana

not sure

[keen beacon]

mango is the best fruit

[simple carbon]

holy shitt

nano banan is gemini

i knew it

[echo aurora]

https://tenor.com/view/cat-glare-stare-doubtful-intense-gif-5066275

[polar niche]

Pineapple on pizza? Yes or no

[keen beacon]

made with lm arena

[keen beacon]

same here w ai

[primal orbit]

Gemini-2.5-Flash-Image-Preview has knowledge cutoff June 2025. Looking forward for next Gemini Pro with updated knowledge.

[polar niche]

Why is r o b l o x censored

[spare rune]

wait genuinely?

[shadow jewel]

lmarena so peak I might cry

[spare rune]

thats to cracked

what model is this

is ts gemini 2,5 image edit

[shadow jewel]

BRO WHAT PLEASE SEND THE PROMPT 🙏

[simple carbon]

how do you know

[drifting elk]

Guys are the llms in lmarena real?

[primal orbit]

https://preview.redd.it/nano-banana-2-5-flash-image-generation-on-aistudio-v0-fc1aah4vddlf1.jpeg?width=1080&crop=smart&auto=webp&s=a571c1d115b490afcfb2a84299c6eaf644aaf820

[drifting elk]

I ask gpt 5 he says that he is gpt 4o

[spare rune]

no and were all butterflys in a cocoons and nothing is real

[simple carbon]

why is my one not generating

[spare rune]

the ais dont know their code name

so thats why

[drifting elk]

Thanks bro

[solar galleon]

would love to try using Gemini-2.5-Flash-Image-Preview if only i could upload images 😭 (im waiting for the fix)

[spare rune]

wait you cant upload images

oh

[echo aurora]

It may be struggling a bit atm from the traffic, I'll keep an eye out for other reports.

[simple carbon]

hes lying i can

[simple carbon]

ahh dammit, is it still available on battle'

[south elk]

WTH this banana is insane news

[simple carbon]

also why was it called nano banana

[echo aurora]

It is

[golden ocean]

nano pineapple

[solar galleon]

its a bug for me thats why i said im waiting for a fix

[simple carbon]

rn it wont even generate from me, im sure by tomorrow itll settle

[echo aurora]

maybe try a new browser?

[simple carbon]

oh yeaa

inogdeto

incog

[keen beacon]

make this guy surprised, use chinese confetti for surrising effect and make his eyes big star eyes

[keen beacon]

yeh

[shadow jewel]

thats crazy... I never expected ai to be this good that quicklöy

I thought it was gonan be a crazy ass propt

[echo aurora]

Be sure to share in #nano-banana

[simple carbon]

ahhh lets goo its working for me now

i had to try on 7 tabs and inogdeto but still

[restive swan]

just ent a priv message fyi

[echo aurora]

okay I'll get to it when I can 👍

[restive swan]

ty

[stray aspen]

Holy

Nano banana was revealsd

[simple carbon]

yea its really fast too

perfect image editor

[dusty cedar]

https://tenor.com/view/monsters-inc-sully-what-gif-15100274

[obsidian cargo]

only getting "Something went wrong with this response, please try again." results 🙁

[warm hare]

hi

[simple carbon]

try on indego mode and multiple tabs

[dusty cedar]

Eeeeveryone is jumping on that

[stray aspen]

It's greater than all that crap that takes forever and is not as good

Like gpt image

[drifting elk]

Yup

[simple carbon]

gpt has censor issues like a mf

otherwise its very good at following prompts

[obsidian cargo]

its a pretty crazy leap too, gemini 2.0-flash was one of the worst image models on LMArena

[simple carbon]

actually one of the first editors i used

nvm i think that was 1.5

[fossil fable]

well, it wasn't made for generation well at all

[obsidian cargo]

sucks at cyclopes though. gpt-image-1 is the best at those for now.

[simple carbon]

bad prompting

[obsidian cargo]

bruh...

[fossil fable]

autoregressive image-out multimodal models are the best thing to happen to image gen since its origin

[simple carbon]

is this gonna be available on gemini itself

[grand patio]

Nano bananas... Super dope

[fleet lintel]

it feels like decades since the nano-banana came out.. when and what is the next mdoel launch. 🤣

[rustic knot]

lol

[obsidian cargo]

tbf it was a test I didn't want anything specific besides "ONLY ONE EYE YOU IDIOT"

bad prompting would be some convoluted 500 word json file

[polar niche]

@echo aurora Something went wrong

API issues?

[golden ocean]

where do the json prompts come from

[echo aurora]

?

What seems to be the problem?

[echo aurora]

How often are you getting this?

[obsidian cargo]

every time I hit that retry

well except this time, but its taking a while

[restive swan]

maybe the model doesn't like you

[obsidian cargo]

nevermind it failed

maybe

[restive swan]

try patting it on the head

[obsidian cargo]

I don't think it likes Gergoth

[misty vault]

tf

[restive swan]

like give it some pictures of cute kittens to calm it down

you monster

[clear herald]

bro assistant B been generating for 300 seconds

[obsidian cargo]

refresh the page it probably failed

[leaden palm]

i remember when reaching 1M or 2M all time battles was a milestone

wild model

[keen beacon]

based and deserved

[echo aurora]

Is the image appearing as just black screen?

[clear herald]

its still generating

like i open other chats instead because since its still generating it wouldnt allow me to enter new prompts

and every time i go back it resets to 0 and keeps counting

[clear herald]

also yes

double problem

[glass copper]

@echo aurora I've posted over 100 issues to "feedback" threads and nothing ever happens, it's where ideas go to die. Can you please just get it done?

Un-nest the Leaderboard, and give it its own page. Add tabs to the top of the page, so we can switch between the different Tests. When we're on the Leaderboard, we should be able use the browser's native scrollbar to scroll up&down. As it stands, there is no scrollbar and we can't even use the (invisible/non-existing) scrollbar inside the nested table. You can't see where you are, and you have to use the mouse wheel, inside a little nested window, inside of a page. It's a nightmare to use.

This change is very obvious and easy to interpret

[fading rover]

I don't know what is going on but gpt model on chatgpt image generation over it's website product more quality detail than in lmarena they mostly differ the quality from lmarena website ..

[glass copper]

The UI is so bad that I don't even load the LMArena right now, I have to download the page with my AI and have it re-display the contents in a new table

[sullen quest]

wat

[zenith cape]

Why can't I farm any nano-bananas today? Did it get banned？

[glass copper]

@echo aurora Look, this is actually an F-tier, close to 0 out of 10 for design

Hard to do worse than this

you also threw half the pixels in the garbage

half of the page rendering area is wasted...and then we're pressing "Ctrl+F" to scroll down inside a nested table, meanwhile not being able to see where you are. That's an F

and still couldn't afford a scrollbar

[keen beacon]

Just wish i don’t need to upgrade super Grok…

😔

[obsidian cargo]

its not a stealth model anymore, its gemini 2.5 flash preview

[echo aurora]

Do you know if you're seeing the same in direct/side by side or are you just getting this in battle?

[clear herald]

only in battle for now

[echo aurora]

Okay, good to know. I'll be sure to keep out for other issues and start reporting to the team

[clear herald]

alright bro good luck