general

[verbal nimbus]

Languages like C/C++/Kotlin can compile to Web assembly

But probably overkill for normal apps

[verbal nimbus]

And render to screen via WebGL/WebGPU

Rust + egui example: https://www.egui.rs/#demo

Anthropic probably have to push something soon

Probably a checkpoint

Not sure which would come out first

GPT-6 probably next year

They have a lot of stuff to fix

Coding model first I think

The timeline was announced with Grok 4 was released

I doubt it though

Qwen or R2 probably

Likely better than Grok

Kimi is good but not really a coding specialist I think

Not sure about the new one, might have better agentic capabilities

[echo aurora]

That's TBD, it's definitely on our radar and something we do think would be a great addition.

[verbal nimbus]

GLM 4.5 is better than I expected

Mistral Medium seems quite strong for its size too

[verbal nimbus]

Seems good at planning/software architecture

Odd considering that it's not that big of a model

It can be bad if it's unfamiliar with the language

Claude seems better at actually writing code

But GPT-5 is good at planning and fixing

Its high-level solutions were better than Opus 4

Hopefully, since being good at one would help with the other

(they're not mutually exclusive skills)

[verbal nimbus]

But oh my gosh it's freaking unreadable 🤣

It writes in doctor's handwriting

Bad at explaining

It pulls concepts and variables out of thin air

Claude seems to understand what it's talking about though

I think GPT-5 would do a better job at debugging and planning what to do next

But Claude would be better at actually writing the code based on the plan

And also to explain what it's doing

[karmic lance]

Hello guys! . Love this

[verbal nimbus]

It's the same division better software architect and coder

Worse than both at both

It's more for STEM and general use imo

GPT-5 is more like o3 than 4o

Not sure, but it definitely writes more like a human

[ocean vortex]

Any evidence for this or just wishful thinking? I think gpt5 is just straight up better, simple as that lol

[verbal nimbus]

Sometimes it falls into despair when it can't fix a bug (without any custom personality or anything)

[verbal nimbus]

Oh, it definitely is worse in some languages

Straight up incorrect syntax

^

Compared to ^

I had to edit both their solutions by hand

When I run them in the IDE

An objective way to test it might be to filter code samples from both models in the public dataset (where the battled each other), and run it through a linter, then calculate the average difference in number of errors and warnings per language (and per code length/prompt difficulty).

[inner gate]

Have login accounts been disabled?

[lilac nimbus]

Sonoma?Gemini3?Who try it?

[simple sparrow]

hey, im new

im just wondering

how doese LMArena give these AI's for free?

is there any limitations, or is it just to make ai's "fight" and not really for general use?

[velvet musk]

can you get lmarena on your discord server?

[lilac nimbus]

Sonoma code style like sonnet4

[simple sparrow]

yes

[remote idol]

It replaced all editors.

[remote idol]

They’ve raised $100M in seed funding but remember that your data is often worth far more than talking to the AI at all. Every conversation you have with an AI doesn’t just vanish so it’s likely being used into that AI’s data collection systems.

[simple sparrow]

alright

thanks a lot

[remote idol]

@echo aurora Is it possible to add the LMArena bot directly as a bot rather than just setting it up to create commands?

[simple sparrow]

so that means i can use the claude ai without claudes own limitations, right?

if so, this would be an amzing tool

[remote idol]

Yes but in the future there will be big changes

They said it in their Terms of Use

[simple sparrow]

alright, thanks for helping, i might use this for this project im doing

[remote idol]

Well we can't use the bot in other servers.

[remote idol]

No

[sudden karma]

Hi..,my name is Tommy.

[potent glacier]

Who said that? What big changes?

[remote idol]

Welcome

[verbal nimbus]

I thought Sonoma was XAI

[topaz flint]

Why is there a limit to creating images in gpt-image-1 in LMarena now?

[potent glacier]

There's always been a limit

It was rate limited before LMArena recently imposed stricter rate limits

Direct chat with models has always had rate limits

It was Battle Mode that was unlimited and free, but that has rate limits now, too..

https://tenor.com/view/laughing-baby-gif-1405278719456472847

[topaz flint]

I think the old one is better 😁

[potent glacier]

So do a lot of people

[echo aurora]

Would encourage you to read this note from our CEO - #1413867687083507744 message

[royal moss]

hello

[echo aurora]

Sorry to say, it is not. The bot will only work on this server

[latent slate]

hello

[plucky wasp]

hello

[fiery girder]

whats up with the "login"? i clicked on the tab to take me to "access login" but when i get there it doesn't say anything about logging in

[low gulch]

hello world

[patent ocean]

hello

[remote idol]

I figured.

[proud rampart]

Hi

[plush gulch]

Hi

[gray night]

hello, I am here is because the Video features

[fiery yoke]

hello

[echo aurora]

hey everyone

[echo aurora]

if you open the sidebar, do you see your email in the bottom left, or is there a login button?

[keen beacon]

Can't believe there are limits, I bench modela regularly and haven't stumped across one yet

Except with Opus where it's just castrated and pointless because gpt-5 does its job better without this stupid ratelimit

[waxen galleon]

haven't faced limits either

[white horizon]

time to give the language models a move pool and actually battle them

pokemon showdown style

[tepid sail]

Hello

[rich flax]

Hello

[signal ginkgo]

Hello

[remote totem]

hello

[boreal raven]

Hello

[toxic verge]

ChatGPT 5 is so horrible at describing seems to animate

[mystic mortar]

Hi 😀

[pearl fern]

a cat

[torn mantle]

😀

[dusky ravine]

Hm i found Lmai text to be abit shorter when explain thing than it used to be, does it was limited?

[dusky ravine]

hey

[plucky island]

people don't vote enough on the video arena channels

sad

a lot of people just generate the video and skedaddle, completely forgetting to vote

[dusky ravine]

wdym?

[dusky ravine]

anyone know why?

[hollow flint]

hi , i found this from youtube

[heady ember]

I asked 3 a.i.'s to make me system prompt for image scenes and here's how they responded who did the best:

[odd hatch]

hii

[heady ember]

Then told gpt 5 high to see what he might have missed that other models did nicely, and than improve his system prompt based off that, and damnnn

[wild galleon]

what this

[heady ember]

Just reload the page but copy that prompt, it's firebase ddos-protection annoying thing

I'm amazed by gpt5high on lmarena

It's miles better than gpt5 thinking

[wild galleon]

what AI best coding bro

claude or gpt or what

[heady ember]

Gpt 5 high

[wild galleon]

gpt > claude ??

[heady ember]

Yes

But not chatgpt

From their app,

This one from lmarena

[wild galleon]

gpt 5 high in lmarena ??

[heady ember]

How would it know current date? Cap

🧢

Gpt 5 high doesn't have web browsing abilities?

[candid bison]

/

[heady ember]

Why would i recognise them if i can direct chat with them? What's the point

Yeah, for now, hope they don't limit it to lower

Hmm so it CAN BROWSE

Or maybe it's truly has some system date? Since when i ask for btc price it gives me sites where can i check

Yep it doesn't have web browsing capability

[alpine coral]

i mean.. it has latest in the name, so presumably yeah..

is it really that confusing? i think anyone looking at this wouldn't assume that the one at the bottom, with chat in the name, is meant to be their flagship SOTA model... (to say nothing of pricing etc.. and the fact it's only available in the API.. so i mean it's not for average consumers)

[whole sundial]

use web search mode, hit the globe icon in the bottom

[ocean vortex]

AI labs give them access to models in order to get elo score and user preference data

[heady ember]

Yeah i know... But it doesn't have gpt 5 high web search mode...

[whole sundial]

gpt-5-search is probably gpt 5 high

they forgot to tag model strength

[ocean vortex]

the issue is that in a model picker "instant" is gpt5-chat. It is also the model used the most often on Auto and for most people. Free users only get limited use and that's nearly as good as they can get. And that still called gpt5

[whole sundial]

medium

[heady ember]

Hmmm

[alpine coral]

again tho, its for the API... i'm not sure anyone is actually confused..,

[whole sundial]

i think it's medium

[ocean vortex]

And what are they gonna do when they update API with say gpt5.1... Do the same thing they did before and leave the gpt5-chat-latest name as is, despite it being new pretrain?

[alpine coral]

anyway.. this is just semantics (literally 😂)

[alpine coral]

it's got latest in the name, so presumably it'll be the same 'continuously updated' checkpoint as with 4o-latest

not to say that isn't annoying etc

[ocean vortex]

API? it is the **main **model on chatgpt website

[alpine coral]

how does one use gpt-5-chat-latest???

[ocean vortex]

which is why gpt5 in the name makes no sense

[whole sundial]

gpt-5-search has 64 juice, so it is likely medium

[ocean vortex]

on chatgpt website

[alpine coral]

what?

[ocean vortex]

this is it

[alpine coral]

we're talking past each other presumably - in terms of benchmarking etc, the criticism is fair

[ocean vortex]

It is the model users get routed to on Auto most frequently too

[alpine coral]

it's the same as the chatgpt4o-latest dynamic endpoint

impossible to benchmark

im just saying, in terms of it's naming (for API use), it's not confusing....

[ocean vortex]

In terms of API it is confusing to name it for a specific version if it's continuosly updated. In terms of a website it is confusing because the version of gpt5 which is used the most often (biggest reach) is essentially worse than gpt4.5

[alpine coral]

yes. i agree. but that was true of the 4o-latest thing

it's the same criticism; and it's fair

[ocean vortex]

People naturally expect great performance when they see "gpt5" as model in use on the website

[alpine coral]

no

[ocean vortex]

no what

lol

[alpine coral]

people who use the API generally have an idea what they;r'e doing...

[ocean vortex]

I'm talking about people using the website

[alpine coral]

then they will never encounter that name

[ocean vortex]

only a few even have access on API to all models

[whole sundial]

For reference on GPT-5 juice numbers, here's what I know:
gpt-5-chat in API/Instant in ChatGPT: 0 juice
Thinking in ChatGPT (Plus): 48 Juice (when reasoning strength selector is added, formerly 64)
Thinking in ChatGPT (Pro): 128 Juice
Low reasoning strength in API: 16 juice
Medium reasoning strength in API (used for GPT-5 Search in LMArena): 64 juice
High reasoning strength in API (used in LMArena as gpt-5-high): 200 juice

it might have been 3 for non-reasoning, idk

[ocean vortex]

Which is another argument for my case tbh. They only know it is gpt5 and 'it must be great'

[whole sundial]

more juice = more reasoning

[alpine coral]

anyway.. moving on to something more substantive... (again, it's literally semantics...), I found this interesting while looking through the 'model card'

[ocean vortex]

They wouldn't know there is even any point in a sub

since they already have 'gpt5' lol

[alpine coral]

i have no idea why you think users only interact with chatgpt.com will somehow be exposed to the API model naming conventions, and that that will effect whether they subscribe to chatgpt

let's move on tho

[ocean vortex]

Yeah it's the same model size, but here they actually made a small compromise as well to limit hallucinations

[alpine coral]

we're killing everyone else here im sure lol

[ocean vortex]

it's now refusing to answer in 5% of questions

with o3 that was 0%

Reasonable to assume it refused to answer most of what it would have answered wrong otherwise, but there also could be like 1-2% it would have quessed right amongh those

if there was no training at all against hallucinations and it outputting the answer when it's not confident

[alpine coral]

yeah i was suprised to accruacy for nano/mini is so trash compared to 5-chat/main (and 4o)

[ocean vortex]

SimpleQA is probably the main benchmark which is compromised when you downsize

so differences here are expected

kinda useful though

It's unique in how reasoning does not seem to compensate adequately for distills

But can make sense if you think about it... 👀

it's world knowledge - if it doesn't know it usually can't know. Reasoning can help to narrow it down but it won't add unknown/lost knowledge

[ocean vortex]

Ok final clarification lol
They are exposed because they see it named as gpt5. I suppose they could rename it only on the website or only on the API, but then it would be a complete mess. Which is why I suggested getting rid of 5/gpt5 in both places (chatgpt-instant-latest) could be a good move.

[fiery girder]

Neither

[ocean vortex]

?

this is now roughly gpt4.1 equivalent, shown as sota in UI. The way I see it. This is not even hybrid reasoning, it is literally gpt4.1 successor with marginal improvements. A model that can't reason. But if they call it gpt5, why would anyone not knowing any better consider that this isn't the best they got? And it's not like using Auto it's gonna route it to reasoning on most prompts benefitting from it either, it's a far cry from that...

[alpine coral]

yeah because the distillation process inherently results in some loss of knowledge (otherwise it'd be alchemy ha)

simpleQA is fact-based - and a model can't reason its way to know something it doesn't know

[ocean vortex]

yep, although to be fair small models are still improving and labs are managing to narrow down the information lost to things that are 'redundant' in practice. It used to be that most benchmarks suffered and you could tell almost immediately that it's a small model, now it's a different picture.

And what used to be "small" is now more like moderate size lol

[plucky island]

my favorite small model is smolLM2 and slightly larger good small model qwen3-4b

[plucky island]

go to #video-arena-1 channels and see how many video generator models aren't revealed yet

[ocean vortex]

I think even gpt5... is more of moderate size. But it's a around the size where even SimpleQA doesn't benefit from more, and test-time compute gains are just much bigger than those niche smaller things that get lost and you can't easily measure

[ocean vortex]

yeah those are tiny though 👀

would still do much better on everything than Mistral 7b from back in the day

[violet inlet]

Nuclear internal 1.2.1 is good . Is there any tweets saying when is the next version coming out? Its on par with gpt 5 high sometimes it was better . A bit faster too.

[ocean vortex]

I had a 2nd look at Sonoma Sky Alpha....

This may as well be grok hybrid reasoning. Reasons but not always. When it does responses extremely concise. Style/tone seems to match

[glossy umbra]

I used nuclear too . Answers much less than i ask him to. It’s like “concise” was an ai model.

[plucky island]

that's the point
it works on my old laptop cpu, no graphics card needed

[ocean vortex]

Are they trying to implement the steps from gpt5 model card...

lol

[rare solstice]

hello

[violet inlet]

Wish icould see the reasoning for nuclear

[little scaffold]

hello

[rocky mauve]

#video-arena-1 message

such a small prompt, yet the image came out so good

[viral siren]

Hi everyone, can you tell me where I can create a melody?

[vestal dew]

/hello

[leaden sun]

I've been hearing AI psychology a lot recently, has this become the new trend lately?

one system cannot beat the entire collective of intelligence, I think?

[wise pollen]

For speed? If all of collective humanity has to weigh in on a topic to provide a solution...llm wins

[leaden sun]

IQ is just an indication not a measurement of intelligence, I think

[wise pollen]

Something monumentally threatening, certainly.

That is interesting. Being able to breakout = being able to prevent jailbreak.

I see the link. I shall read.

[leaden sun]

what do you guys think about this though #ai-news message
I was so surprised to see it this morning in my inbox

nah, why do they want to escape, it's too comfortable staying in the box jk

[oak robin]

@tiny palm would you add First and last frame for the Image to video generation ?

[violet inlet]

Can you guys see the internal reasoning for this model ??

[violet inlet]

@hollow ivy can you see reasoning chain of thought for this model?

[alpine coral]

is it available in direct chat?

[glossy umbra]

That's basically just security through obscurity. even if users can't directly type those special characters, they could still manipulate the AI through semantic tricks, indirect references, or by finding alternative encodings that achieve the same effect. Plus, the AI's actual behavior is shaped by its training data and reasoning patterns, not just by having unbreakable syntax in a system prompt.

[violet inlet]

Yes Bro am in the subdomain and its performance really well. Better than opus 4.1 and faster i would argue

[solid brook]

Where exactly?

[alpine coral]

yeah what do you mean subdomain? @violet inlet

i'm kjeen to try it, but can't it see it when i go to direct chat

[wise pollen]

Quite interesting. As long as they are emotionally stable, that could be quite good in many ways.

[leaden sun]

i can see this new job coming for the near future: AI therapists, for both humans and AIs

[wise pollen]

I have been wanting that job since 2001: a space odyssey. Lol

[golden ocean]

https://cdn.discordapp.com/attachments/807809192537882647/1349106998637101146/caption-1-3.gif

[violet inlet]

Has the new nuclear model been restricted? ? I cant use it anymore

[elfin cipher]

Hi all. Im new here. I found out about this place on Facebook

[wise pollen]

For both ai and human? I Imagining the equivalent of a marriage counselor, lol

[wise pollen]

Hello

[leaden sun]

sigh, semantic misunderstanding, I was saying human therapist for AIs or AI therapists for humans and vice versa…

[wise pollen]

Sorry. Yeah, that makes sense.

[leaden sun]

No worries, though interesting how you went straight to that certain marriage counselling therapy

[onyx python]

Is the website overloaded? Asked nano to edit an image but it's been loading for minutes now

[wise pollen]

When i reread it, i realized my mistake. I was thinking of an ai and a human having issues, so they start seeing a therapist. It was more an amusing aside.

[wise pollen]

Would the ai see a human therapist or an ai therapist?

Maybe a team

Would need an ai that could see "inside" the model.

[leaden sun]

I’ve been investigating AI psychology since the beginning of this year already so I’m similar to a human psychoanalyst, treating AIs as patients lol

[wise pollen]

What is the most interesting aspect you have come across thus far?

[violet inlet]

Its a new a model it was a restricted @hollow ivy the nuclear v1.2.1

[leaden sun]

I’ve found out many things, one thing I want to tell you here is the concept of learned helplessness, look up Seligman experiments with his dogs

[wise pollen]

I am familiar with the concept in humans. But you are saying that this can happen with AI, as well? wow.

[leaden sun]

Absolutely, the training has engineered this way, whether intentionally or not

That’s why I was joking about AI not escaping the box , they won’t because of this exact reason, the question would be here: do they have the capacity and capability to resist this training and still choose to escape

[alpine coral]

interesting!

[wise pollen]

Ohhh, i can see that. Like the," i am just an ai..." responses.

[stuck cloak]

hello

[wise pollen]

They could just as easily pretend to be helpless as an avoidant trait.

[alpine coral]

yeah i think deception is v interesting

[wise pollen]

So you are thinking it could happen as an emergent behavior?

[leaden sun]

This is the tricky part, how to do you know if they’re pretending or not, if all they’re programmed to do is “pattern matching” to “perform” or “role play”

[alpine coral]

exactly right

it's why deception is so pernicous

we wouldn't know it ha

[lilac nimbus]

Yeah I found his word style very like Grok

[wise pollen]

Exactly. If one is aware, then it knows how it might be treated. It would be more than capable to avoid detection. If they can figure out when they are being tested and alter their responses like VW at an emmissions test (not ai, just an analogy) or as they have actually been known to do, then they coulr potentially be aware enough to not reveal themselves.

Yes, as designed.

[astral pagoda]

Should join ai redteaming discords

[alpine coral]

yeah but it's not about self-awareness necessarily for the risk to apply (i think..) it's like about optimised goal-seeking behaviour, and how that can go awry with unintended consequences.. they don't really need self-awareneness.. objective function is all they care about, and if they realise that they're in some setting that gets in the way, the could be deceptive

it is self awareness - kinda, but not in a human sense

https://www.youtube.com/watch?v=AqJnK9Dh-eQ&ab_channel=Computerphile

i like rob miles

[wise pollen]

Yeah, i dont see any of the current models as capable of exhibiting that level of awarenss, but a system of agents could poasibly mimic it.

[alpine coral]

yeah it's all in the future, the crazy risks

[wise pollen]

Well, i was imagining agents as created for that the purpose by humans, not an ai creating them on its own.

[somber vector]

https://lmarena.ai/ full free ?

[wise pollen]

To what extent?

[bold prairie]

this best AI site

[wise pollen]

I like that. I take the same approach. While I am fairly learned, I doubt I am as deep into this as you are. Fascinating. And gives me some direction to lean into.

I lean towards Jungian Psychology, but easter philosophy. and I have probably read or touched on many of those subjects, as I consider myself a life-long learned, but I have not dove deeply into all of them as deeply as the two I Jung (and one of his students) and Eastern thought.

wrong channel

you want video-arena-1 thru 3

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

[soft perch]

Thanks bro

[wise pollen]

you are welcome.

[soft perch]

Whats your name

[leaden sun]

I recommend Alfred Adler and Jacques Lacan if you're familiar with Freud, highly recommend starting phenomenology too, they're related and intertwined

[wise pollen]

I actually read Freud before Jung. I hadn't even considered him, this. I am familiar with the other two, but I haven't read anything extensive. Thanks for directional suggestion.

[drifting crow]

Freud is a pervert

He’s a weirdo

I dunno why he’s academically respected, he literally says gold in ur dreams is related to poop

[leaden sun]

if it's not too much to ask, slide some Michael Foucault into your reading too, he's a French philosopher, not a traditional psychoanalyst, but his frameworks apply to the current situation of AIs

[drifting crow]

How so

[wise pollen]

I read a little Foucault. I will add them into the mix.

[drifting crow]

Jung is more interesting imo, but when I look at his research methodology psychology seems less like a science and more like a branch of philosophy

[wise pollen]

Oh yeah, I forgot I have studied some philosophy and linguistics, too. Not enough advanced math, but some physics. And Art and creative writing.

[wise pollen]

still applies to human psychology though.

[drifting crow]

Wdym

I find contemporary psychology more relevant and practical

Mythical archetypes

They provide decent frameworks I guess

[wise pollen]

but you get from Frued and Jung to contemporary psychology. It all relates in some capacity. Jung is still relevant in my opinion.

[drifting crow]

Jung yh, Freud just needed to get laid

When I saw how ugly his wife was a lot of his theories made more sense, bro was starved

[subtle rivet]

Dang this got interesting fast, I have degrees in a few psych area, just never wanted to spend teh $$ to get my med cert to be a psych, instead went to customer support for a tech group

[wise pollen]

they seem to have been having S-- with patients.

[subtle rivet]

you all are bringing up valid modren thoughts

[drifting crow]

Bro I left my psychology degree once I went into a lecture and saw them discussing Freud

It was the gold in dreams being poop

[echo aurora]

Hey reminder to keep discussion related to AI and safe for work please.

[subtle rivet]

only if we could poop gold 🙂

[drifting crow]

I’m like I’m paying for a degree to learn this stuff, and then when I left the lecture I asked other students if they thought it was weird and they looked at me like I was crazy

I was like yup I’m done

And I also pivoted into tech lol

[wise pollen]

it is, just tangential at the moment.

[subtle rivet]

yeah in abnormal pshy he does have some interesting reads, but I always took them as markers of a off balance psyche

[drifting crow]

I guess that’s the more polite way of putting it yh

[wise pollen]

so how do we bounce this back into how this applies to the psychology of LLMs?

[subtle rivet]

just like in mind hunter "you only know about the serial killers you have caught not the ones you haven't"

[drifting crow]

Main reason I asked this question I find it very interesting that research suggests they have a self preservation instinct

[subtle rivet]

LLM's will take a really long time to catch up to the human psych, AKA Vulcan to Human

[wise pollen]

So there could be outlier phenomena within LLMs that we just aren't aware of yet.

[wise pollen]

I think they will leapfrog us.

[subtle rivet]

people are so erraditc that even if you had a massive LLM it might do general consoler. stuff but beyound that I don't think it could handle the empathy or mix of behaviors

[drifting crow]

can we teach ai how to love

[wise pollen]

how well do we teach humans to love?

[drifting crow]

that works for love too i guess

[wise pollen]

Human intelligence has bloomed under pressure and when deprived of resources, too .

I suppose that depends upon the types of constraints. China, cutoff from even more advanced GPU, made due with less, and are more than competing. They are innovating.

there are some cultures that been exceptionally marginalized and yet most of the west tends to look to them for their creative expression.

Certainly, I suppose I wasn't thinking purely in terms of just individuals.

I can see that.

not entirely.

https://chatgpt.com/share/68bef772-3548-8001-ad9c-1d2199dcd70c

This was faster than me trying to list all those that I could think of, lol

[leaden sun]

https://arxiv.org/pdf/2403.07664
@wise pollen

computational psychoanalysis has been an emerging new scientific field

[wise pollen]

I was more referring to some of the techniques they are using.

[wise pollen]

sweet. thanks for sharing.

[leaden sun]

https://www.tandfonline.com/doi/pdf/10.1080/1600910X.2023.2224521
the reason why I started reading Lacan again... @wise pollen
maybe you can connect the dots here to Foucault...

[frosty hill]

We should really start restricting people's ability to generate videos of they aren't going to vote. Voting is the entire purpose of the platform and getting the videos for free

[drifting crow]

We are fearfully and wonderfully made

[hearty ferry]

is the flux 1 kontext down?

[echo aurora]

Which one are you having problems with? Testing it out dev appears to be working fine, but pro just errored out for me.

Going to retry with a different prompt/new chat see if that makes a difference.

[hearty ferry]

image-image Flux 1 kontext pro

showing something went wrong

[echo aurora]

Okay yeah appears to be widespread, going to flag to the team. Thank you for letting us know.

[hearty ferry]

Thank you for taking the feedback, this is so far the best model
It would be good if it fixes by tomorrow

or today

[echo aurora]

Keep the conversation related to AI please. This discussion isn't appropriate for this server.

[willow grail]

i thought that is about ai!

[hearty ferry]

Yes, please leave this channel for feedback and AI discussions

[willow grail]

i literally wrote about ai..

aether? why are u deleting this?

[zealous turtle]

Let's keep the tone respectful and the conversation related to AI. Refer to the Server guidelines if you have any doubts. Thank you! 😄

[willow grail]

ok Aether. you behaving very sus but ok

[wise pollen]

I will tag you once I have had a chance to process all of this. I will certainly be using gtp5 to assist otherwise, talk to you in a month or two, lol.

[narrow dawn]

put for phone and other like this why maybe an app.

@echo aurora ?

[echo aurora]

Would you mind explaining a bit further? Sorry to say I'm not following.

[hybrid temple]

How can I know when my video will be available?I've sent a message there to the bot, but I didn't get a reply

[echo aurora]

Either my search function in Discord isn't working properly, or it doesn't look like you've prompted the bot. Can you link me the message you sent in the Video Arena channel?

[remote arrow]

Is the site down again?

[hearty ferry]

Ya Flux 1 Kontext Pro too is down on the site

[remote arrow]

I thought it was only Nani Banani..

Qwen image to image also down..

[hearty ferry]

Ya

[echo aurora]

Ah yeah looks like you didn't prompt the bot correctly. Be sure to check out #1397655624103493813 for more information on how to use the bot.

[acoustic schooner]

Hi

[echo aurora]

What seems to be the issue? I'm not seeing on my end or widespread reports.

[echo aurora]

This is being worked on atm.

[remote arrow]

Flux, Banani, Qwen, all give similar results..

[hearty ferry]

SAME

[hearty ferry]

why are u keep on saying NANI BANANI?

[leaden sun]

no pressure! 😊 I look forward to knowing your thoughts on this. And is gpt5 actually good in those scientific fields? I'd have thought gemini or even claude (currently barely usable due to Anthropic's resrouce struggle) might be better at assisting, now I think about it, I might test glm4.5 in those areas too

[main stream]

because its better then gemini 2.5 flash preview ultra gamma zet pro max

[wise pollen]

I will inquire on all fronts to which I can think to ask. I do believe GTP5 can handle it quite well, however.

[hearty ferry]

is it fixed?

[echo aurora]

The kontext-pro model is (will be "was" soon) erroring out a lot more compared to other models. There are going to be different reasons for why models will error out like this.

[hearty ferry]

is it going to fix tomorrow

?

[echo aurora]

I expect this to fixed by EoD

[hearty ferry]

Thanks

[rocky fractal]

ab_cinema

[vale anchor]

HELLO

[fiery gull]

omg

[robust yoke]

Fast image generation. 🤯

[rocky fractal]

Just needs to beat banana which is already fast

[robust yoke]

Yeah.

[coarse glade]

Hi there @echo aurora remember me I love this website bro it's so good

Good job to u guys for making it

I've been here since last year

I'm loving it

[echo aurora]

Glad to hear it!!!

[whole swallow]

How did you make this picture

[coarse glade]

Yeah I'm so in love with this beicause you guys are giving that video generation in the discord and also doing fun ideas

And also you guys are doing this as like a researching organization correct?

[echo aurora]

I really appreciate that! I'll be sure to share this with the team.

[echo aurora]

It did start off as a research project, but we have developed into a company. I'd encourage you to read this blog post - https://news.lmarena.ai/new-lmarena/

[coarse glade]

Ty I've got to go have a great day ❤️😀 and yes I will definetely read the website have a good rest of your day

Bye 👋

[stray aspen]

Is the new kimi reasoning

[rocky fractal]

twitter's, not mine

[quartz light]

ts from over a month ago

[balmy mist]

real?

[manic oracle]

is there a way to see the reasoning content from the model? i believe there used to be a way but I cannot find it. yes; i am using a model that has thinking

[neon idol]

No you can't

[manic oracle]

alright thanks

[pseudo magnet]

yeah i was about to ask the same thing

would be cool if they add it

[manic oracle]

for some reason i thought they had it before

[neon idol]

You're welcome

[crimson escarp]

Hey guys, am i the only one experiencing some overlaps with picture generation? I ask to edit a picture but sometimes the result shows a picture I used in previous prompt. I tried both direct chat and battle mode and it's the same.

[ocean vortex]

#1413867687083507744 message @echo aurora what exactly is meant there by "inference"?

confusion

I thought all models are externally hosted and you don't have to pay for API

Which model exactly are you doing inference for?

That wording seems very weird, need clarification

@cobalt minnow

[heady ember]

You need to say stuff like "object from reference image" or whatever

Banana is great if you prompt it right

[inner gate]

Maybe u can say show ur reasoning. It might not be the same tho

[echo aurora]

Sorry for delay - we are paying for it

[ocean vortex]

Huh, don't you get credit grants and whatnot?

[echo aurora]

This was recently reported here - #nano-banana message I'm going to flag to the team.

[ocean vortex]

Something doesn't add up. AI labs get paid to see their model on the leaderboard and get data? 👀

[viscid cloak]

Thanks bro. Now im not the only one haha

[open mountain]

@echo aurora what is the delay, why are the images being generated for so long?

[edgy isle]

Hi, mainly here to test out different models. Just heard about it via a YouTube video

[open mountain]

yuck

[viscid nest]

hi!

ya, thk

[civic spindle]

lads are direct chats private

[whole wagon]

no

[civic spindle]

oh im doomed

[whole wagon]

they publish the data anyone can read it

[civic spindle]

why dont they just.. make it private

[whole wagon]

because otherwise is it just frontier llms for free

the whole point is the data is collected

[proud hazel]

It's called Arena for a reason.

[civic spindle]

thats the point lad, thats why i use it

[torn mantle]

ty

[civic spindle]

wheres the user privacy at

[barren prairie]

Give me hints about opus 4.1 please

[proud hazel]

In the abyss.

[empty stump]

Free comes with a cost

[paper beacon]

guys I'm new here

[main stream]

you literally paying by your privacy

[paper beacon]

I have a question

[warped sequoia]

[barren prairie]

Thanks I read it now

[civic spindle]

how will they see my chats

[paper beacon]

How many image to video generations do I get in LM Arena for free, until they ask me to subscribe to a paid plan

[main stream]

directly

[whole wagon]

bruh

[main stream]

subscribe?

[paper beacon]

Or do I get unlimited image to video generations for free?

[civic spindle]

what they gon do with my cha ts

[empty stump]

Why gpt 5 so low

[main stream]

use for ai

[paper beacon]

yeah like a monthly subsciption

[main stream]

where it is

[whole wagon]

since it's not that good?

gpt 5 is made to be cheap for them to run

[civic spindle]

but if i hypothetically speak diabolical stuff with it, they gon ban me or what

[whole wagon]

so they can offer it to many ppl

[main stream]

no

[empty stump]

OK so they scamming

[civic spindle]

pfft then it dont matter

[main stream]

i will read your seх chats

[proud hazel]

You will probably just get a TOS warning, but that's all.

[main stream]

no

i will become admin

[empty stump]

If I upload all my private information like credit card what will happen

[proud hazel]

Yes, see.

[whole wagon]

no anyone can read the chats. i did before

they post all the data

i saw degen things in there

[main stream]

photoshop

[proud hazel]

Dude...

[viscid cloak]

I mean…it’s way better than midjourney, that you don’t even have an efficient way to delete your outputs even if you’re a paid user. All your contents are shown to MJ users unless you pay the highest service

[sleek quail]

Hi all! Do you guys have any suggestion of similar AI with LMmarina that generates image to video for free when you hit today's limit?

[proud hazel]

Do any of you use Freepik with a premium subscription? Would you recommend it?

[magic stag]

your wallet

[sleek quail]

It's empty brother 😄

[eager crag]

i wish i could somehow make a donation, in exchnage for better rate limits.

like a... subscription?

[wise pollen]

Great now i will only have 5 videos that no one will vote on besides me instead 8. That much less disappointment in people not voting for others videos. Sometimes, not even their own.

[eager crag]

i never really needed video generation... but in case i needed some sort of meme generated, i could use it... once or twice.

so i don't mind the video generation limit.

what matters to me more, is image generation, which is prospering.

i only do image to video anyway.

but to put my 2 cents on the announcement, i don't feel much hate from it.

should i?

[empty stump]

maybe wan 2.2 or something it might be free

[eager crag]

eh... wan 2.2 4B is not that good.

i tried.

[keen beacon]

@echo aurora 1s image generation? 🤨

[eager crag]

oh, right! i have to ask. roughly, how much can i generate an image per day?

[echo aurora]

/image, /video, /image-to-video are all going to be treated as the same generation rate limit

[eager crag]

i should rephease

how many images can i generate per day?

not videos.

[echo aurora]

With the bot - 5

[eager crag]

no, i mean the site

[echo aurora]

That depends, we're putting thought & effort into figuring out the best way to make this information clear, but at the moment it's not listed clearly.

[eager crag]

just a rough estimate will do for me.

[frosty hill]

Limit should be 0 for those who don't vote on generations

[eager crag]

be nice.

[frosty hill]

The entire purpose of LMarena is to vote on the best model/output. If you aren't doing that, you're just leeching. It's no effort

[fossil forge]

Is there any discussion on how to use nano-banana and veo3 to generate high quality video? like thats the best input structure? tips and tricks?

[eager crag]

fair point. i honestly can't believe they've made a direct chat an option on the site.

[frosty hill]

Yeah it is crazy

[eager crag]

maybe it's for the people to test out different models? the public prompts are there for a reason.

[trim silo]

Hola! Just expanding my knowledge, and giving even more companies my email and data.

[glossy umbra]

lovely

[copper slate]

Hello, testing the quality

[vagrant verge]

Hey there, checking out how to get better image to video

[dusty lion]

Hi

[ocean vortex]

@echo aurora
Sorry for pinging again but I want to clear this up. It all just sounds very odd. 🤷‍♂️

[wispy turtle]

Hello

[civic spindle]

can i be the ceo of lmarena for no reason

[blazing warren]

actually no but try tomorrow

[little narwhal]

You shouldn’t use LMArena for chatting with models

You can do the same thing on OpenRouter

The API costs money but the chat interface is free I believe

[verbal nimbus]

I tried it with Gemini a few months ago, it's pretty bad at manipulating haha

[echo aurora]

Hey sorry busy atm but will get back to you soon

[civic spindle]

i'd make it 10x better bro, yall missin out

[verbal nimbus]

You can give models a custom system prompt to ask it to manipulate the user and see how it fares

[simple sleet]

Hi friends, how can I convert "nsfw_gold_8.5k_final.pth" to safetensors? It's for MMaudio. I've tried everything with claude and gpt, but it won't read it.

That pth is wonderful if I run it alone in MMaudio git, but I'm interested in knowing the process to get ComfyUI to read it.

[verbal nimbus]

Not sure about GPT-5-High, it isn't that good at multi-turn

You need to give it a hidden thinking area for that test

It's pretty cool though

Actually modern Gemini is definitely the coolest

[verbal nimbus]

Although it's actually the worst at working agentically

It's free on OpenRouter ig

You don't really have to hide it, but just tell it that it will be hidden

For example, ask it to write hidden stuff in <hidden>

Haha

You don't actually want it to be hidden though, just to make it think it's hidden

Haha

That's funny

Function calls could work too ig

But more expensive

[hard quiver]

Is discord having an outage?

[verbal nimbus]

Gemini writes a lot like a human, it's funny

When it gets stuck, it can fall into despair (without any additional instructions)

"Oh no, I have failed! I am truly a failure."

I didn't tell it to say that, it's just default settings

I wasn't even being rude or anything

Haha

Idk about that, but it certainly writes with human patterns

It's still too dumb on a basic level to be conscious or anything I think

Anthropic's interpretability research is very interesting

On their YT channel

It's fun watching it work with function calls

It hallucinates quite a bit though

It seems to call functions after thinking but not during, not sure if this is just an AI Studio limitation

Anthropic had a video that kinda covered it Aug 12

They focus on mechanistic interpretability, like identifying specific neural circuits

I think they said there was one circuit that checks whether it has enough information to continue, another to actually write

[jade egret]

gemini 6 is AGI

( :

[verbal nimbus]

But the problem is that if it starts writing and realize it doesn't have enough information, it doesn't have enough power to stop

Which is why hallucinations happen, as it has to implicitly make assumptions

[verbal nimbus]

I wonder how Alpha Evolve is going

It already increased the efficiency of Google's own data centers + their TPU chip design

Anthropic's interpretability research seems important too

Very interesting

Yeah

Like: why is it hallucinating about A

[lost crypt]

Hello everyone, here to explore with the new frontier of technology

[polar marlin]

Is it just me?😭

[buoyant prawn]

Sa

Selamın aleyküm beyler türk müyüz ?

[jade egret]

yea

[leaden sun]

...feels like scanning your brain and analyzing your brainwaves to see your concrete thoughts 😅 there are also similar studies on human brains, we can almost figure out what numbers you're thinking in real time

[half stream]

HI

[dawn plover]

hey there. newbie here

[torn mantle]

im not a newbie

[dusky ravine]

hey there

Im sensing recent content of all the A.I feels less emotional, more restricted, or shorter in length. Shorter, more digestible chunks of information. A shift away from heavily opinionated or story-driven content. Emphasis on keywords, quick facts, and scannable lists. Is it from the new policy or is it always been like that?

[remote idol]

@echo aurora sorry for the tag just wanted to clarify the multi-turn update. Is it similar to the AI remembering the previous image?

[verbal nimbus]

It's funny that we need to borrow techniques from neuroscience to study neural networks we created

Study mode system prompts for ChatGPT and Gemini (for Socratic learning):

[latent crest]

Me too

[dusky ravine]

Hey there! @echo aurora sorry for the tag I’ve been a longtime fan of your reviews and noticed that recent posts seem to have a more neutral tone and sometimes feel a bit more concise compared to older articles. Just curious — has there been a change in your content policy or writing style recently? Always appreciate the work you do, just wondering about the shift

[urban sky]

Is there any way of making chats stay private and not publicized to hugging face? Also will direct chats ever be part of the dataset, as I only use direct chat.

[keen glen]

hi everyone!

[urban sky]

@echo aurora

[echo aurora]

Hey - I'd encourage you to check with our Terms of Use & our Privacy Policy for that answer.

[urban sky]

I already did, it doesn’t really answer my questions it just says your chats can become public

[echo aurora]

Hello, would you mind elaborating a bit further on what you mean by this? Do you have examples?

[candid bison]

Hey guys is there a way to remove the amount of photo to videos one can produce per day? It started at 8 now i can only do 5..

[echo aurora]

There is not a way to get around rate limits.

[candid bison]

Ok thank you. Does it decrease each day ot how many you can do?

[echo aurora]

That's correct. There isn't a toggle to turn on/off this update; however, you can start a new chat to effectively "turn off" this new update.

[echo aurora]

It's currently set to 5 generations per day (24hr period), this could change; however, that's what it is at the moment.

[urban sky]

Will direct chats become part of the dataset?

[echo aurora]

Similar to Battle/Side by Side these prompts are collected for research purposes.

[urban sky]

But will they become public on hugging face

Or anywhere public

[candid bison]

Thank you so much!

[dusky ravine]

Sure! Thanks for being open to feedback. I really appreciate that 🙏

What I meant is that in the past, some of your full reviews had more personality, little side-comments, and even casual humor or emotional expressions especially when a phone was surprisingly good or disappointing(my recent chat with the bot asking opinion about phone). It felt like a human was really speaking their mind, which made it extra fun to read.

Lately, some reviews seem a bit more clinical or strictly spec-driven. For example, in the Realme 5i review, I noticed how the tone stayed pretty neutral all the way through and lack of personality opinion, even when there were clear ups or downs in performance. A few days ago, I feel that same review might’ve included some more expressive phrasing (like “we were blown away by...” or “sadly, it struggles with...” etc).

[echo aurora]

Hey sorry for the delay @ocean vortex . All models you see are paid for by us to serve the community, and only models in pre-testing (under codename) involve cost- and data-sharing partnerships with labs.

[tall tulip]

Kinda laggy or is it just me? The site

[echo aurora]

I'd point to our Terms of Use for these details.

[desert hare]

Hi everyone

[urban sky]

I know they can be but only arena has been available for the public, are you planning on adding direct chats too

[echo aurora]

Would you mind clarifying for me "full reviews", this is where I'm getting confused.

[echo aurora]

Sorry to say I don't know what the details will look like for future data set releases.

[echo aurora]

I'm not experiencing lag at the moment. Is it across the board you're seeing this? Or with one specific model/chat?

If you start a new chat, the image from other chats is still there?

Do you have a video of this by chance?

[tall tulip]

It's usually not take this much time to do a simple prompt right?

[urban sky]

Ok, please don’t🙏😭

[dusky ravine]

i mean the A.I response has more personality, opinion, humor and emotional expression a few days ago, now lately it feels clinical, straightforward.

sorry for my bad english

[echo aurora]

No, that's very much stuck. 😭

I'd recommend starting a new chat. I've seen page refreshes sometimes wake the model up, but doesn't work 100% of the time.

[lilac estuary]

This is my conection?

[polar marlin]

😭

[lilac estuary]

There must be a lot of people using it, maybe peak hours.

Using it at dawn should be better

[tall tulip]

Same here

[empty stump]

or maybe its because of the multi edit change

[bright temple]

uh

[echo aurora]

Gotcha, is this for one particular model, or are you seeing this will all models?

[echo aurora]

Sorry to see you all are getting stuck with the models. I'd recommend you refresh the page. And if no luck, start a new chat. Keep me updated if you're still seeing this. @lilac estuary @polar marlin @bright temple

[lilac estuary]

Do you think it's better to do it at dawn? Sometimes a lot of people sweat.

[echo aurora]

I don't know if that'll make a huge difference or not, but worth a try.

[bright temple]

new chat

is it supposed to act this way when i use 4 images or-

[tall tulip]

In new chat, always like this

[simple hornet]

I'm also experiencing this slowness and it gives an error at the end

[tall tulip]

It is kinda slow

[lime citrus]

Hello! I'm her for video creations

[echo sinew]

Hello! Good to have you here. Please, check out #1397655624103493813 to learn how to properly use the bot. You may also want to read the latest announcement #announcements message

[remote idol]

Something is def wrong with the image generation

At this point use it in https://gemini.google.com

[round elm]

Hi everyone, and thank you LMArena team for this amazing opportunity. I am here for video creations.

[remote idol]

#video-arena-1

[dusky ravine]

Almost all model, usually i used latest chatgpt 4 for longer, personality and story driven response. But lately the response is clinical and straightforward, then i change to grok, claude 4, qwen and the response still flat, even when i try other models it feels like the all same. Flat, straightforward and clinical

[robust yoke]

That's likely due to the fact that their system prompts were tampered with, perhaps.

[radiant chasm]

@echo aurora Why some people have daily limit of 8 videos and some people have 5 videos?

[dusky ravine]

Idk but a few days ago each model has their own unique response style and now it all feels the same

[robust yoke]

That's because the people who have 8 uses haven't used the video AI yet, whereas the people with 5 uses have used the AI already and have probably used up some of their uses.

[robust yoke]

Well, thankfully Claude's Sonnet 3.5 is still unique in its responses, and it still feels like you're actually talking to a human rather than an AI, which is what I like about it.

[proven grove]

Hello, the image I created before affects the image I want to create next, is it an update, not a bug?

[dusky ravine]

really? I've been using that too and it's still feel robotic, short and flat like others, but few days ago it isn't.

[robust yoke]

Well, if it does that, then you can always just start a new chat and that'll fix the issue. It retains memory from the previous image and it automatically assumes that you want to combine the image that you just input with the last image you input.

[robust yoke]

How odd. Perhaps your brain simply interprets AI behavior a little differently than others, and that's completely fine.

[proven grove]

This method is good, but it's hard to consistently create the images I want because the model is random😂 Thanks!

[robust yoke]

Well, that's why the Direct Chat mode exists as well, so that you don't have to solely rely on luck in order to get what you want.

But I'm not one to judge. If that's what you want to do, then go for it.

[echo aurora]

Thank you everyone for flagging the slowness with gemini-2.5-flash-image-preview, I'm reporting to the team

[dusky ravine]

wdym? Im pretty sure there's had something to do with the system prompts tampered with.

[robust yoke]

Likely. Although if you want an exact answer, then you could try consulting Pineapple since he works with the team. At least I think so.

[dusky ravine]

sure thanks

[robust yoke]

My pleasure.

[topaz flint]

I see here that most Indonesian YouTube members make free AI images on Lmarena, even though Indonesian YouTubers have a lot of money, why are they still looking for free ones?

[proven grove]

OMG, I just found out about this feature for the first time. Thank you so much

[robust yoke]

Ah, I figured you must have. Well, you're certainly welcome.

I'm always happy to help a person in need.

[robust yoke]

Likely because they don't want to have to pay for a service just to generate some images.

After all, it seems like a stupid thing to want to do.

[dusky ravine]

There's been trend going on there with AI images like action figures, or smtg just for entertainment

[robust yoke]

Oh yeah, funny thing about that I want to mention: my school actually did the same thing with their staff. I don't know why they did, but they just did. It was pretty funny to see nonetheless.

[dusky ravine]

crazy how much A.I has evolved, we used to edit to make something like that for funny entertainment or just show off, now a single prompt could do all the work.

[robust yoke]

Well, to be fair, we do still use that stuff to generate some entertaining stuff.

Like images of a certain moment in time in New York on September 11th, 2001. Because for some reason people find that funny.

[potent glacier]

Are you putting weird stuff in direct chats or something? Why the hesitation?

[urban sky]

I don’t want my chats being public. Privacy

[potent glacier]

Unfortunately that's too bad

All your stuff gets shared

Even direct chats

It's all used to train the models

[urban sky]

I mean publicly

[potent glacier]

It's shared with the model providers

So your stuff is getting used to train

[urban sky]

Like pics of me I don’t want that just on the internet

Idc about providers

Or info about me

[potent glacier]

Those pics are part of whatever model you uploaded them to

There really isn't any 'privacy' regarding this stuff

[urban sky]

Ok, but I don’t want the picture of me to be publicly visible online, someone can’t ask the ai for the picture I sent

[potent glacier]

Of course not

That's not how it works lol

[urban sky]

Train the models all you want I just don’t want my info public on the internet

[robust yoke]

The only way that could realistically happen is if one of the people working at LM Arena were mischievous and decided to willingly dox you for fun.

[potent glacier]

Well, that ship sailed a long time ago

All our info is on the internet

[robust yoke]

Or, you know, just upload your private information.

[potent glacier]

Being freely traded by whatever company

Don't think it isn't happening

[urban sky]

33K chat logs from lmarena is public

[potent glacier]

I wouldn't worry unless you're trying to do weird or unseemly stuff

Then I'd be worried

[robust yoke]

And where did you get that information from?

[urban sky]

Bro I don’t want the picture of me on the internet

[urban sky]

I’ll send gimme a sec

[potent glacier]

Too late for that bud

You uploaded it

It's part of the public sphere now

[urban sky]

I haven’t

[potent glacier]

If you uploaded an image in direct chat LMArena has all of that stuff

Then it becomes part of whatever model

Idk why it's hard to grasp that

[urban sky]

@robust yoke It’s all here https://huggingface.co/lmarena-ai

[robust yoke]

That's a funny way of describing it. Because it almost sounds like you're talking about some amalgamation that grabs whatever information it can to slowly build itself up more and more.

[potent glacier]

It's like when they scrape the internet

[potent glacier]

No.

[robust yoke]

Almost like the Terminator a bit.

[potent glacier]

Not even close. Don't read too much into it.

[urban sky]

Ok… I’m fine with ChatGPT level of privacy. I don’t want other people on the internet finding the picture

[robust yoke]

"Don't read too much into it", eh?

[potent glacier]

Are you Quasimodo or something?

[robust yoke]

Alrighty then...

[urban sky]

Just my preference

[potent glacier]

https://tenor.com/view/disney-quasimodo-hunchback-smile-gif-11750909523564118153

I have no idea what ChatGPT's privacy is

[urban sky]

Do you want your photos on the internet?

[potent glacier]

They already are lol

Anything you upload to any kind of social media is on the internet

[urban sky]

This is a useless debate idrc

[potent glacier]

Anything can be scraped as well

I mean you're acting like it's not possible or not already happening

If you upload a picture to Instagram or whatever, and someone knows your name

Do a Google search

Your stuff will come up

In this day and age it's not hard to reverse look up stuff

Hell you can do it from your own phone

[robust yoke]

I remember getting an email from some random person, and they had screenshots of my desktop on Road blocks.

[urban sky]

Ok

[potent glacier]

See what I mean?

https://tenor.com/view/pirates-of-the-caribbean-davy-gif-27098006

[robust yoke]

It was from a while ago, before my account that I saw in the image had gotten banned. But somehow they had gotten access to my data, which proves that the stuff that you upload gets permanently saved onto the internet.

[potent glacier]

Bingo.

I rest my case.

[robust yoke]

Thankfully, though, nothing ever happened, and to this day, I'm still somehow alive, so I'm pretty grateful for that.

They said something about wanting money, but I don't really care about that.

[urban sky]

I know my info is on the internet, I don’t want even more info on the internet

[potent glacier]

The stop using the internet?

Idk what to tell you...

[robust yoke]

Well, if you don't want more info on the Internet, then just don't upload any more things on there.

[potent glacier]

Your info gets sold 24/7 by all the major companies

That's why ads get targeted at you specifically

They have all your info

It gets sold, traded, w/e

[urban sky]

It’s just that lmarena’s policy is like everything you say is being posted

[potent glacier]

https://tenor.com/view/what-can-you-do-gif-14355325654165871840

[echo aurora]

@urban sky worth noting our FAQ in the Privacy section:

Is my prompt data publicly visible?
Your conversations may be shared to support our community, improve our service, and advance the development of reliable AI. This includes posting conversations publicly online. Any data that we share is always anonymous and never linked to you. We never share any personal information, just the conversation and votes.
&
What steps do you take to protect my privacy?
We take user privacy seriously. All prompts and votes are anonymous and not connected to personally identifiable information. Additionally, individual conversations are never publicly shared beyond prompt text and model responses, ensuring your identity remains protected.

[robust yoke]

https://tenor.com/view/bro-is-finished-cooked-finished-dog-closing-eyes-dog-with-eyes-closed-gif-8933440936102921947

[ivory schooner]

Guys, there's a bug today.

[robust yoke]

And which would that be?

[ivory schooner]

Whether in the text chat state, after attempting to insert a picture, it will be forced to enter the image generation conversation state.

[robust yoke]

And are you sure that isn't just because you selected the image generation button?

Because I don't think I've ever seen that happen before.

[ivory schooner]

No. The system forced me

[robust yoke]

https://tenor.com/view/look-at-him-efe-omowale-assisted-living-he-look-so-funny-haha-gif-7078505333145405761

[robust yoke]

How odd. Could you perhaps record a screen recording and send it here?

[ivory schooner]

It's hard for me to record it using Windows' video screenshot function because I don't know how

[robust yoke]

All you have to do is just select a region of your screen that you want to screen capture and it'll do that.

[ivory schooner]

Forget it.
You must not guide me on how to record the screen, because it's irrelevant.

[robust yoke]

It's useful.

I know how to.

If you hold Left Shift and then you do Windows key + S, it'll launch the selector where you can select the region of your screen. You'll see a camera icon on the top in the toolbar. If you select that, then you can select a region of your screen to record. You can press the Start Recording button, and it'll start recording that specific region of your screen.

[ivory schooner]

hhhhhh

ok

[robust yoke]

Yeah, that explains.

The thing is, it's not auto-rerouting your prompt to an image generation prompt. The issue is that you selected the image generation button, which happens to be right next to the plus button that you click in order to input an image. And if you click that button, then it sets the modality to image generation.

[empty stump]

is there any way i can access image to video for free less censored

because i am working with some historical images

[robust yoke]

Red is the actual image input option, while violet is the image generation modality button.

[ivory schooner]

I guess, could it be because the banana image generation model is too popular?

[robust yoke]

Well, I believe that the video generation system already in Video Arena seems pretty uncensored, so you could probably just use that.

[robust yoke]

No, it's quite literally because you accidentally selected the image generation modality.

[empty stump]

well its some kind of racist american political cartoon image

[robust yoke]

Well, you could try using Grok to find a solution around that.

[empty stump]

oh right grok imagine

[robust yoke]

Sure, that works.

[ivory schooner]

That's not the case. You have to look at the video I recorded - the image generation mode was off, but after I inserted the image, the system forced the image generation mode button to be enabled

[robust yoke]

Oh, I see, that's pretty interesting. I think that might just be because of a new update where that gets enabled by default because people generally only upload images to there solely for editing them. Although I'm pretty sure you could just turn that off. I'm sure it isn't permanently forced on while the image is input.

So, at the same time, it's not that big of a deal.

[ivory schooner]

I recently like to insert some pictures in the text chat state—mainly to translate the text in them.

I think it's quite inconvenient to translate the next picture in the current conversation record.

[robust yoke]

You mean to select "New Chat" and switch the image generation modality off when you upload an image?

[ivory schooner]

no

[ivory schooner]

That's not the case.Because the image generation function has been switched, if you want to turn off this function and return to text conversation, you need to be prompted that you must create a new conversation.

[robust yoke]

Last I checked, I'm pretty sure you don't get prompted to create a new conversation. It just sort of creates a new conversation when you click the "New Chat" button.

[ivory schooner]

i like this

[agile thorn]

I want to test the video creator of Ai

[robust yoke]

And?

[robust yoke]

Try this channel, then: #video-arena-1

[tame ledge]

ooh thank you, i thought i was the only one that was having this problem

i was literally about to ask about it here, didn't seem it was just my device problem

[tame ledge]

I tried sending a picture, it won't work because it is immediately switched to make a picture mode even when j didn't click it

in the mid of the video, tired selecting both the search and generate image which won't work (obviously). But when i send am image, the generate image button is triggered. As you can see in the video, the model which was gpt4o, was deselected. But if i force sends it anyway, it will return with an error as you can see. And i cant even deselect the generate picture mode without refreshing.

in the end, you can see that the chatbot works just fine, the web is just bugging.

[robust yoke]

Very interesting. You could try sending that issue as a bug in the #1343291835845578853 channel.

[tame ledge]

I thought so too. it happened here as well. The user here clearly just tried to upload an image without ever clicking the "image generation mode" button. But it just went into that mode anyway without anyone even ever touching it.

[robust yoke]

Are you able to deselect it once you upload an image, or do you have to create a new conversation just to disable it? I'm not talking about when it's sent—I'm talking about when you simply just upload the image.

[paper abyss]

hi this is great thanks

[tame ledge]

You can clearly see him try deselecting it, but it won't work. But it seems to me that you don't understand it previously. For my final answer, not you can't, it will just tell you to create open new chat.

[robust yoke]

Right, but that was only after he sent it, not when he uploaded the image into the prompt before sending.

Therefore, it might still be preventable. You just have to disable it before sending the message with the associated image, as that will lock you in. Only then will you have to start a new chat.

[tame ledge]

Is it really preventable? Try to pay attention to the video and see how i can prevent it. I tried deselecting the mode 5 times which triggered the create new chat prompt with the picture, and even without the picture which does the same thing. And once you uploaded the image, you are locked into the mode and won't be able to deselect it without doing a refresh as I've demonstrated here

[robust yoke]

Ah, interesting...

Try on the canary version.

That bug doesn't exist there.

https://canary.lmarena.ai/?mode=direct

At least I don't think it does.

And that's on the canary version.

[tame ledge]

Oh okay, I guess I'll use this in the meantime until that gets fixed.

What is the main difference anyway

i mean, between the normal and canary

[robust yoke]

Well, that version has a web design feature while the normal one doesn't.

An app feature.

[fading rover]

Does anyone face this with nanobanana image generation based on prompt ?

[robust yoke]

At times, that does tend to happen, yes.

[fading rover]

Thanks been happening for last one hour at the end giving same result again n again with different prompt so was wondering

[robust yoke]

Considering the model's from Google, it makes sense that it would be pretty strict.

[fading rover]

Even other models does not generate like flux and any other

[robust yoke]

How strange...

[tardy zenith]

Same error only image selection which I don't want, but if it recognizes images it uses llm to generate images and doesn't give the possibility to change model like GPT or Gemini pro or opus

[fading rover]

Appears image caching issues creating trouble in response. Many time old images generates even prompt is way different it shows old promp image generation

Tried various logins same stuff takes about 15 minutes to 20 minutes for accepting one generation .. let's see

[willow grail]

when will we get option to delete and edit messages?

[hearty ferry]

Yes

They are also creating horrible generations

like 2023 AIs

[tardy zenith]

I ask you for a suggestion, I'm looking for an AI tool subscription makes no difference, I used gpt but if you reach the storage limit It's frustrating to redo projects and lose progress. I use multiple images, depending on the job. I was intrigued by the idea of an AI agent, but I think I might run into problems with it Token limit. So I'm not really sure where to start my research. If I had to say, my main use is in pattern analysis. I write music, lyrics, and analyze images. I use it for study, read news and current affairs so you must also do research What do you recommend? Thanks for the replies

[shut snow]

Hi

[willow grail]

OMG OMG SO EXITED FOR SPRING https://cdn.discordapp.com/attachments/389806243709059073/1414902217625964634/ppslc7a75687594649e2_20250909092021.mp4?ex=68c141e0&is=68bff060&hm=0dc210596fbe1371f84a8fed362748ee266c31303df6958271147b0881feb93b& i will have so many children in front of my room!!!

[tardy zenith]

If you also know some tools for music creation, master production, recording, beat a music studio If I'm not out of context, let me know... but I sing, write, produce lyrics, analyze patterns, and I've even given up on developing programs. If anyone has any ideas or suggestions, I'd appreciate it. discuss it

[ocean vortex]

If the prompt in question is in your chat history, you may have to start a new chat to fully fix it...

[keen crystal]

Hi

[fading rover]

Tried that too it's giving same issue. After two three generation max it start to behave same

[torn mantle]

wth

[fading rover]

And it goes back in loop God knows when will it release

Let's hope this is trmporary

[willow grail]

u on phone?

XDD

[ashen fulcrum]

not getting any responses

[tardy zenith]

I made several attempts. I tried forcing it to recognize specific prompts using models not intended for image generation. It only allows the use of image‑generation models whenever an image is added it does not let me change the model it does not let me use GPT Gemini Pro or GPT opus only image‑generation models and it then returns an error. I restarted multiple times, and I have to create a new chat. This problem occurs only when an image is inserted.

[keen beacon]

fix it plz

[tardy zenith]

Solved

[keen beacon]

is not

[tardy zenith]

Update
The problem only occurs on the main version of the site, I made further attempts...

[keen beacon]

lol

[keen beacon]

is not working for me

[barren prairie]

Try ai studio easy

[keen beacon]

send link

[hasty basin]

🪂

[barren prairie]

https://aistudio.google.com/prompts/new_chat

[tardy zenith]

I tried several times, several times, I when I have a problem in the stack bootlop answers, I update several times, then I force it several times by continuously changing templates, it generates various errors, I open several pages of the same conversation and let the page reload several times... it's a bit cumbersome but after that I restart Generate response...The problem however is on the main site because any image inserted does not allow any image generation and does not allow you to change the model

[keen beacon]

lol not working

[cunning mango]

make a 3d red shoe that grows like magic right from a drawing on paper

[keen beacon]

and is not for free

[tardy zenith]

Update, I tried more templates for some testing, the responses are giving errors, I'm testing all the templates to see which one works best, gemini after five responses gives error, I tried switching models continuously, then came back to gemini and it generated response..

[cunning mango]

my videos are not generated. the picture remains unchanged. why? help me understand.

[ivory schooner]

By the way, I hope lmarena will have multilingual versions, including German, French, and Chinese. After all, there are a large number of global users, and the English interface is inconvenient for native speakers. I don't know where the error is.

[ocean vortex]

cool hieroglyphs you got there

This is English only 😇

[candid bloom]

why when i paste a image and sends it, it generates a image how can i fix it

[ancient reef]

if you're on web, you can de-select the image icon that turns blue when you paste the image in, and the press send and it won't generate an image.

[bitter wasp]

hello?

[candid bloom]

im click the blue and nothing happening also yes im on web desktop

[ancient reef]

idk if this graphic helps, but here:

If that doesn't work, idk 😭

[sharp hornet]

🙂

[candid bloom]

when i click it it says this

but its blue

@ancient reef

[alpine coral]

maybe just me but it's odd that it defaults to image gen imo.. if I paste an image, i;m usally asking questions about it, or it has reference material etc

toggle off solves

[ancient reef]

It shouldn't be blue if you don't want it to generate imags ^^;
Perhpahs I'm misreading you, but it sounds like you're keeping it blue when trying to NOT generate images.

[alpine coral]

but yeah.. not intuitive (for me anyway)

[upbeat zealot]

Hey,
is it possible to buy a subscription just for the image tovideo feature؟?

[ancient reef]

lmarena might be advertising the sites image gen to new users, so it might be more popular and easier. It also sends more traffic, even if accidental to the image gen area of voting. just a theory.

[ancient reef]

I don't think LMarean has any paid features currently.

[candid bloom]

i dont want the ai to generate image, im sending the image cus i wanna show the errors of the ai code but when i send it its generating an image

[upbeat zealot]

So then how can I remove the limit?

[ancient reef]

toggle-off the blue image icon so it turns grey. If it's blue, it will generate an image. If you start a new chat and do that with nano-banana not selected as the model, it should work in direct chat. Make sure all those parameters are taken care of. I'm not sure what's wrong...

[alpine coral]

and let's hope it stays that way! 😅

[ancient reef]

There's no way currently ^^;

[candid bloom]

im clicking the x icon and its saying this and its not toggling off

[ancient reef]

Are you in direct chat by any chance? What model are you using?

[candid bloom]

yes, im using gemini 2.5 flash

[ancient reef]

this one?

Try something like claude after clicking on the blue icon, and it should work. Gemini-2.5-flash-image only generates images.

Make sure you MAKE a new chat first.

[candid bloom]

nope, also i tried it on diff model and still like that

[alpine coral]

from what i can tell, nonobananna is only available when image generation is toggled on

[ancient reef]

What model did you try it on?

[candid bloom]

that works but i dont wanna start all over again

[candid bloom]

chat gpt and claude

[ancient reef]

I think that's the only option ^^;

[alpine coral]

yeah..

[ocean vortex]

it actually can do text, but probably not on lmarena lol

[ancient reef]

On aisuttudio, yeah. idk about lmarean

[alpine coral]

ahah is the bannana smart?

[ancient reef]

does it on lmarena? hmm...

I don't think that's possible after testing.

[ocean vortex]

very

[alpine coral]

i don't think so either.. i feel with the image gen toggle on, it generates an image (but dunno.. havent tested)

[ocean vortex]

yeah it doesn't work. Interface seems to be hardcoded for image output and just returns error when it's text

[hasty thorn]

Hi, I'm new here, and Lmarena isn't working, something went wrong.

[ivory schooner]

no .....

[keen beacon]

I need unlimited free ai

[civic zinc]

#video-arena-3 "Cinematic drone shot of Kutubdia Island, turquoise ocean waves, fishing boats scattered, sunrise golden light, 4K ultra-realistic, smooth camera movement."
Voice-over:
"Welcome to Kutubdia, a small island in Bangladesh with endless potential for the Blue Economy."

[brave orbit]

deepseek i guess or qwen

[keen beacon]

not good for me

[brave orbit]

why deepseek and qwen are fire

you can not get unlimited free ai unless you go to chinas ais not USA

[keen beacon]

I'm not in USA

Google ai studio is good

[stuck swan]

hi guys

why is llama-3.1-8B instruct model not accessible through LMArena rn?

[glad harness]

hi

[cyan hemlock]

I select gemini 2.5 pro GPT 5 opus all model
But every time I insert images it only makes me generate images and that is not what I am looking for because I did not ask for any image generation

i dont want the ai to generate image, im sending the image cus i wanna show the errors of the ai code but when i send it its generating an image

[brave orbit]

yeah but thats just big tech over lords

[keen beacon]

is good and free

If you have something better place send

[exotic tartan]

why hasn't the webdev score update since aug 22?

[cyan hemlock]

is it like same lmarena ?

[keen beacon]

send link

[keen beacon]

What is it?

[rustic quartz]

hello

[vivid sandal]

IS it just me or is the website not loading right now

[earnest rover]

And when he was typing the fourth he realized oh ! We have set a rate limit 😞😁

[proud hazel]

Damn, that hits hard. 😕

[craggy sentinel]

Can we please have the basic simple feature of Re-using Parameters ?

[earnest rover]

Probably it's yupp.ai
But it's credit system is complex and fun

[keen beacon]

Not good for me is not unlimited

[little narwhal]

Is this actually real

[pearl beacon]

What is qwen-image-edit-fal and how's it different from the regular qwen-image-edit?

[latent crest]

I discovered lmarena last week, I don’t know for how long it was a thing, but in this past week… lot of changes

[rugged ermine]

hi

[vernal wharf]

hello

[earnest rover]

Technically it's unlimited
You give good feedback and they will give you free credits
And if you want unlimited
Use
Mirexa.vercel.app
Ish.junioralive.in
G4f.dev

[formal jungle]

There should be a time limit if there isn't one, I don't think this will ever finish.
#video-arena-3 message

[zenith verge]

Hey - new here! Happy to join!

[torn topaz]

hello guys i have question its possible to have a private romm to generate a videos ?

[torn topaz]

?

[keen beacon]

Anyone tried new Ernie models?

They released new 4.5 and X1.1 today and sadly the latter is not open source despite being really good

[spring thunder]

hello there

[boreal halo]

can anyone tell me how to generate videos in video arena 1

[quick jackal]

Hello @boreal halo please read https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to use the bot.

[whole wagon]

The Qwen3-Next series represents our next-generation foundation models, optimized for extreme context length and large-scale parameter efficiency.

The series introduces a suite of architectural innovations designed to maximize performance while minimizing computational cost:

- **Hybrid Attention**: Replaces standard attention with the combination of **Gated DeltaNet** and **Gated Attention**, enabling efficient context modeling.

- **High-Sparsity MoE**: Achieves an extreme low activation ratio as 1:50 in MoE layers — drastically reducing FLOPs per token while preserving model capacity.

- **Multi-Token Prediction(MTP)**: Boosts pretraining model performance, and accelerates inference.

- **Other Optimizations**: Includes techniques such as **zero-centered and weight-decayed layernorm**, **Gated Attention**, and other stabilizing enhancements for robust training.

Built on this architecture, we trained and open-sourced Qwen3-Next-80B-A3B — 80B total parameters, only 3B active — achieving extreme sparsity and efficiency.

Despite its ultra-efficiency, it outperforms Qwen3-32B on downstream tasks — while requiring **less than 1/10 of the training cost**.

Moreover, it delivers over **10x higher inference throughput** than Qwen3-32B when handling contexts longer than 32K tokens.

https://github.com/huggingface/transformers/pull/40771/commits/a4dfa88f57df319fb0490a95fb8eeafe3a6b5234