general

[runic mulch]

Let's see the argentic capabilities

On tool calling

Without that it's useless

[golden ocean]

https://cdn.discordapp.com/attachments/910347347601543196/1160742997156188181/togif-6.gif

[quartz pike]

https://tenor.com/view/joe-bart-joe-bartolozzi-dance-funky-gif-15654574488447184285

[robust yoke]

https://tenor.com/view/stan-twt-stantwitter-stan-twitter-tenorelicario-relicario-gifs-gif-298982740560034732

[quartz pike]

https://tenor.com/view/joe-bartolozzi-gif-16501825566380358966

world's most agressive griddy

on moonshoes

[robust yoke]

https://tenor.com/view/drachenlord-sport-teddy-bärchen-gif-25967747

[quartz pike]

OOf it failed at making a clock

with 3 errors

but so did kimi k2

[echo sinew]

Hey guys! Let's not disrupt the chat with GIFs spam

[robust yoke]

https://tenor.com/view/cat-salut-salute-بسبس-تحيه-gif-8859282567348257812

[quartz pike]

i take everything back

it sucks.

[echo aurora]

Yeah I've been flagging these errors to the team, it's not super stable at the moment.

[quartz pike]

3 errors.

easily beaten by 4.5

[quartz pike]

oh.

this was the minimax result

this the sonnet https://3000-it6ku80oodtwxyjakawzu-6532622b.e2b-foxtrot.dev

aka sonnet 4.5 thinking 32k

[quartz pike]

btw could yall perhaps add a side by side option like yall did in lmaerna so its easier to compare models?

[echo aurora]

It's very much being worked on, check out https://canary.lmarena.ai/

[quartz pike]

oo okay

what makes this diffrent from normal lmarena?

[robust yoke]

Early features.

[quartz pike]

ohhhh

okay

[robust yoke]

It's easy to tell because back when the site for the regular version had the old layout, the canary version had the new layout.

[hollow ivy]

which model would you recommend for C++ coding (offline) ?

If the goal was, to create a 2D game.

[quartz pike]

this canary thing is so goated

[robust yoke]

It's a bit hard to determine considering I have never used any of the models for C++ coding, only HTML coding.

And also with React coding as well.

However, if I were forced to give an option, I would likely have to choose either Claude or DeepSeek.

[hollow ivy]

Deepseek? really?
That's an interesting decision.

and Claude-4.5-thinking?

i thought DS was below the top league of LLMs (in serious coding)

[robust yoke]

DeepSeek because it really takes a long time to think out a proper decision to make and contradicting itself in the process which actually helps to improve upon its thinking and provide better outputs, and Claude with its ability to naturally already code very well as well as mix some creativity into its work.

[hollow ivy]

which deepseek version and which claude version?

deepseek R1?

[robust yoke]

Claude, I believe, works very well too, considering it's able to communicate naturally and code very well, as well as being a very creative model in its own ways.

[quartz pike]

i alr found an issue witht he canary thing

[robust yoke]

No doubt about it that the thinking model would also work very well.

[polar niche]

Hello!

[robust yoke]

Greetings.

[robust yoke]

No, DeepSeek v3.2 Experimental Thinking.

[hollow ivy]

what about GPT5-high, with good prompting?

[quartz pike]

yall

[robust yoke]

GPT in itself also seems pretty promising given the outputs it has given me, in some cases in WebDev. If you want to try GPT, then you can, but ultimately I recommend DeepSeek and Claude.

[quartz pike]

tell me when minimax is actually stable

so i can run a test on it

[balmy mist]

https://x.com/Angaisb_/status/1981869708480516423

Angel 🍂 (@Angaisb_)

So GPT-5.1 after GPT-5?

It'd make sense

X · Oct 24, 2025, 11:45 PM

[quartz pike]

aka what i like to call the "Minecraft one-shot test" where i js use this prompt: "make me a three.js minecraft clone with working terrian. first person movement collisions side collisions xyz collisions terrain. that isint hilly as hell. but also isint hill downhil hill downhill hill downhill repeat and repeat and with block breaking and placing."

[normal peak]

Why is Gemini 2.5 is still in the top of llms in lmarena?

[hollow ivy]

ask it to write that in C++

could be an interesting test

[quartz pike]

idk how to compile c++ lol

i only know how to do c# via unity.

and three.js via html

[hollow ivy]

the AI can explain it

it (the AI) will probably propose SFML (for 2D games)

[normal peak]

Why is mistral dumb af

[hollow ivy]

just ask it, what is the best engine for 3D in C++

[hollow ivy]

because that lab has not the resources of the Big5

(Deepmind, Anthropic, OpenAI, xAI, Meta)

[stray aspen]

How's mini max 2

[normal peak]

Hmm interesting

[stray aspen]

Is it any good

[robust yoke]

Testing that right now.

[quartz pike]

pretty damn good

beats 4.1 opus for me

[stray aspen]

That's crazy

[normal peak]

But all those 5 are dumb too tbh

[hollow ivy]

why that?
Deepmind isnt "dumb"

[quartz pike]

doesnt mistral just make small ai's that can run on ur pc or small servers?

or single h100 gpu's?

[normal peak]

Gemini can't speak my native language

[hollow ivy]

which is..?

[robust yoke]

What's your native language?

[quartz pike]

what's your native language?

[normal peak]

All llms can't speak it

[normal peak]

Tamazight

[hollow ivy]

sanskrit?

ah

wow

amazigh?

(that was a selectable nation in freeciv game)

[normal peak]

The original north African language before Arabs colonized us

[hollow ivy]

cool!

[stray aspen]

What's tamazight

[hollow ivy]

interesting.
our (earth's) history is rich

[normal peak]

Yeah . Whats ur name. Im gonna write it in my language

[hollow ivy]

me?
paws

[normal peak]

We don't have the letter P hhh

So im gonna write baws

[hollow ivy]

:)

oh nice

looks very.. exotic

and unique

[robust yoke]

An interesting rule.

[normal peak]

Darkness = ⴷⴰⵔⴽⵏⵉⵙⵙ

There was one model that can speak my language

It was sonnet 3.5

[hollow ivy]

what does "Cosmos" sound in your language? if you translate it

[normal peak]

It was amazing

[robust yoke]

It seems like Claude is the only model with a creative mind, and thus, can speak native languages very well.

[normal peak]

ⵉⴳⵏⵡⴰⵏ

Ignwan

[normal peak]

Yeah especially the old sonnet

They are focusing more in programming now . Not in ancient languages hhh

[robust yoke]

Well, not to worry, since Claude 4.5 seems to also be just like Claude 3.5, in that it's able to write in very natural English. Almost in a conversational tone.

So, I think it might be able to write in your language.

[normal peak]

I've tested it. But the results wasn't good

[robust yoke]

Darn. I'm sorry to hear that.

[normal peak]

Yeah. Im hoping Gemini 3 will be good

[robust yoke]

It mentioned something about Tamazight being a Berber language.

[normal peak]

Gemini 2.5 isn't that bad actually. But it makes a lot of mistakes

[normal peak]

Yeah berber people are amazigh people

[gilded geyser]

Good evening from Alaska, where I'm hoping to see what's possible with AI video animating characters from my stories.

[normal peak]

But we don't call ourselves bereber. It is a racist name from romans

[robust yoke]

I've never really known about Amazigh, so this is all new information to me.

[normal peak]

Where are u from ?

[robust yoke]

Washington, and you?

[normal peak]

Morocco, taghazout

[robust yoke]

Ah.

I've known a little bit about Morocco, but I never realized people there spoke Tamazight.

[normal peak]

I know Washington do you know taghazout ? Hhhh

[robust yoke]

Not really, no.

[normal peak]

After independence, the country didn't teach the original language in schools . That why a lot of people speak drija now . Wich is a mix between arabic and french and tamazight . Hhhh it's very complicated here hhh

[normal peak]

There is a lot of people from Europe and usa . Coming to taghazout. Its very famous for surfing and nomads

[robust yoke]

Very interesting.

[normal peak]

Akhnfuch

No

[robust yoke]

Ah.

[normal peak]

Don't ask ai . Hhhh it doesn't understand tamazight

[robust yoke]

Apparently, it's supposed to translate to "cockroach".

[normal peak]

ⴰⴱⴰⵏⴹⵔⵉⵡ = cockroach

[robust yoke]

Ah.

It gave me this: "ⴰⵖⵕⵕⵓⴹ".

[stray aspen]

@normal peak hey bud

[normal peak]

There is a lot of amazigh directs, here in Morocco. We have 3 and Algeria 3 Libya 2 . And also there is small group of people in egeyp speaking tamazight too . In siwaa village. You can chatgpt it hhh

[stray aspen]

How do you canada in your language

Say*

That's crazy

[simple sleet]

Does anyone have GPT Pro chat? I have the Plus option and can only make videos up to 10 seconds long and 720p with Sora 2.

I want to know if GPT Pro can make videos longer than that, 1080p, and with Sora 2 Pro.

I also want to know what the daily limit is for creating videos.

[stray aspen]

What kind of letters

What alphabet is that

[normal peak]

Amazigh letters. Very ancient

[robust yoke]

Apparently, it's in their native script.

[normal peak]

Tifinagh alphabet

[robust yoke]

I find it funny that the letter "P" isn't in it.

[normal peak]

Yeah hhhh like we don't have a single word in tamazight w the letter p

[robust yoke]

Yeah, that's very interesting.

[azure sorrel]

i like MiniMax-M1

[robust yoke]

Heh.

[normal peak]

Hhhhhh

[robust yoke]

It's almost like how the Dutch are very throaty in their language.

[normal peak]

Yeah

Its too late here good night my friend

[robust yoke]

Goodnight.

I hope you have wonderful dreams.

[golden ocean]

real

[normal peak]

Thank you

[robust yoke]

My pleasure.

[normal peak]

https://youtu.be/H1YIgwPsX5Q?si=S6nozNjPPRps1XMo

Hallucination hhh agi is still far away hhhh

[ashen mauve]

what even is Tifinagh and were in the world is that from

[robust yoke]

Tifinagh is a handwritten script that a certain group of Moroccans used to write in.

It is the script of the Tamazight language which existed before the Arabs (pretend I explained something here), and now, they teach the Moroccans a mix of French, Tamazight, and Arabic.

[daring rock]

@reef mirage Please head to #1397655624103493813 for a detailed guide on how to use the bot

[obsidian cargo]

you gotta set up a bot that automatically does that to any message with /video in it

[daring rock]

@tribal whale Please head to #1397655624103493813 for a detailed guide on how to use the bot

[ashen mauve]

That is literally cool as hell, I have never even knew about this language until now.

[robust yoke]

Neither have I, to be completely honest.

[burnt sinew]

Why not setup an automod rule to do this

[daring rock]

@mortal plinth Please head to #1397655624103493813 for a detailed guide on how to use the bot

[daring rock]

Thanks for your feedback. We will pass it to our team.

[fast kite]

Basically, I mentioned that there are problems in LMArena, the images aren't showing up accurately.

Here! The image is completely blurry! gpt-image-1 used to make decent art, even different versions, but now he doesn't. Could you please explain the problem that's been going on for two months?

[burnt sinew]

My message got deleted?

[burnt sinew]

It doesn't look blurry to me? Maybe click download and send raw file instead ?

[fast kite]

What do you mean?

[jade egret]

gemini 3 december? taking so long...

[vivid sedge]

hello

[empty stump]

why do they release in december

[burnt sinew]

The image doesn't look blurry to me

[whole sundial]

you're not the first person to complain about this, i complained about this here (#1412721830682296423) right when they started doing this. It seems like LMArena has done absolutely NOTHING to fix this! But that's because they did this to save money. They changed the quality of a model to make it cheaper and the leaderboard score goes down... They should change it back to how it was AND remove ALL votes since September 3 (the date they first started doing this sneaky stuff). It's a shame few people know LMArena has been doing this, I tried to tell people but it never works.

When they do fix it, they should add the API quality level of the model to the name of this and GPT Image 1 mini (it's affecting that model too) so people know what quality they are getting.

[polar niche]

Wtf did claude say

[balmy hemlock]

我靠，困死了

[ancient mango]

Why is there no sound when generating a video?

[whole sundial]

@echo aurora you never gave me a proper answer the last time i asked but is https://github.com/lm-sys/FastChat still being used by LMArena at all? if not, is there any other github repository that is used by LMArena that I can use instead? (I am a volunteer for a popular online wiki that has LMArena on it multiple times for various AI things and this is the GitHub link we use for LMArena)

[quasi storm]

Guys anyone can tell me how i can genrate pics and video

[thick stirrup]

hi, guys everyone how are you!

[whole sundial]

read #1397655624103493813

for image gen it's better to use the lmarena.ai website (hit the image button in the text bar) as the 5 per day rate limit for videos on discord also applies to images when you generate them on discord

[unborn raft]

hi there - hope to discover best video generating ai tools there...

[vital spruce]

Bueno

[ember matrix]

hi there - hope to discover best video generating ai tools and practices

[supple hearth]

Here to laugh at all the wild stuff being produced

[magic stag]

The theme of the music is spending 5 seconds reading the rules

[inland quest]

192381273123 bots is coming

[glossy umbra]

https://cdn.discordapp.com/attachments/1407360256446693376/1407362471131611217/caption.gif

[magic stag]

Need ultra banana 3.0 pro to have 5 mins of novelty in my life before I start "needing" grok 5 or something

[pulsar saffron]

vote

#video-arena-1 message

I NEED TO KNOW THE MODEL

[flint zodiac]

#1397655624103493813 {
"version": "1.0",
"platform": "lm_arena",
"task": "image_to_video",
"referenced_image": "/mnt/data/IMG_4368.JPG",
"settings": {
"aspect_ratio": "9:16",
"duration_seconds": 10,
"fps": 24,
"resolution": "4K",
"format": "mp4",
"quality": "high"
},
"prompt": {
"description": "Cinematic drone shot starting high above an ancient Indian fort and smoothly zooming in toward its central courtyard. Warm golden-hour lighting with soft shadows, natural sunlight flares, and realistic HDR tone. Gentle downward tilt revealing the fort’s symmetry and red sandstone textures, with a slow, stabilized motion for an immersive feel.",
"camera_motion": "smooth drone zoom-in, gentle downward tilt, stabilized dolly-in",
"visual_style": "ultra-realistic, golden-hour color grading, 4K HDR, warm tones, soft vignetting"
},
"negative_prompt": "no people, no flicker, no distortion, no overexposure, no text or watermark"
}

[quartz pike]

hello yall

is minimax on lmarena finally stable?

[quartz pike]

go on #video-arena-1

[pulsar saffron]

vote #video-arena-1 message

I'M TRYING TO KNOW THE MODEL FOR 1 HOUR NOW 😭

[quartz pike]

yup minimax m2 is still unstable as shhit

i keep havingh the same damn error.

[leaden sun]

amazing! I've always been fascinated by the Amazigh! it's great to have you here! your language is especially fascinating, reminds me a little bit of the mystery of the Basque culture too ✨

[dawn grove]

what is this model from Google?

[sturdy hawk]

anyone knows how to use popcorn feature on higgsfield to transform a video and change the face in the video?

[thorn kernel]

Anyone knows ai generated video will be monetize on youtube?

[quasi atlas]

Please refer to #1397655624103493813 to learn how to use the bot.

[pulsar saffron]

That's the problem no one tested this model we don't know if it's actually good...

so the answer is you'll never find out

[tropic musk]

helllo

[jolly gulch]

hello

[verbal nimbus]

Isn't that Baidu

Hallucinates badly

[pulsar saffron]

i'm surprised that no one did distill of gemini 3

[tight pelican]

Hello

[north pawn]

hello

[elder burrow]

how has nobody mentioned that minimax m2 has a 200k context window

m1 had 4 million

[elder burrow]

going from the best context window (aside from llama 4 scout) to below an average good model's context window is kinda crazy

[azure sorrel]

hi again

[zenith spindle]

hello

1

[silent pebble]

Hi, trying the models and prompting here before signing up to a specific service.

[stray aspen]

Ernie is chinese

So I would say no

Unless they are trying to deceive us

[pulsar saffron]

no its google

it clearly says by google

i trust ernie

[fast kite]

That's right, but gpt-image-1 stopped showing other options. And the images turned out to be of poor quality. For example:

[dapper fog]

Hey! Anyone know how repair this? I can't work with this.... I have this in all models. 2-5 answers and error

[spare cobalt]

Hello

[hollow ivy]

https://rysana.com/

hello

[quartz pike]

https://www.youtube.com/watch?v=N0XpMe94ENU

im so excited for ltx 2

its 4k 50 fps

[hollow imp]

GONBE GONBE!!!

[quartz pike]

?

[obsidian cargo]

everywhere I look I see her face

[wicked pond]

hello

[west glacier]

making vieo

[golden ocean]

excellent

im proud of u

[west glacier]

how can i generat video in here

[stiff kernel]

Please check #1397655624103493813

[novel obsidian]

hello folks
i have a problem with uploading my images (jpg format)
consistently got upload failed message !!!
what shoul i do!!!

[novel obsidian]

does not help !

[pulsar saffron]

oh!

[fresh mirage]

someone tell me they’re gonna bring back lithium or orion

[obsidian cargo]

probably not tbh

though I bet there'll be a gemini 3 pro preview before it releases

[fresh mirage]

probably

i’m already experiencing lithium/orion withdrawals

it’s killing me lol

[obsidian cargo]

I've seen it be said that lithiumflow is gemini 3 but not gemini 3 pro

[fresh mirage]

Yeah I was told they were gemini 3 flash models

[obsidian cargo]

there was like, an X28 model or something that was labelled gmini 3 pro

[obsidian cargo]

it didn't say they were flash either tbh

maybe they'll end up being gemini 3 coding models

[fresh mirage]

maybe, if they try something new

[obsidian cargo]

a few times lithiumflow did worse than 2.5 flash on creative writing stuff

[hollow imp]

Are you a girl 😶

@fresh mirage can u see this

[obsidian cargo]

glances at [she/her] in discord name

[hollow imp]

So you are a egirl

😡

[obsidian cargo]

bruh

[hollow imp]

egirl

egirl

Scammer in the name of girl

[hollow imp]

@pulsar saffron can u see this

[pulsar saffron]

shows api request

[hollow imp]

Maybe non arena champion role members can't see this

[hollow imp]

You can see everything right?

[pulsar saffron]

everyone can see it what's so special about it

[grand echo]

hello

[hollow imp]

@fresh mirage @obsidian cargo

You guys were talking about gemini 3 gemini 3 pro

I think this helps

[narrow girder]

hey i want to make ai videos

[olive mortar]

i believe they changed the codename

cant get it too

[hollow imp]

Ask @jovial sapphire

[quartz pike]

yall its ai release season

we got suno v4.5 all. sonnet 4.5. news of gemini 3.0 coming soon. ltx 2. hailuo 2.3

😭

-# aka models that all recently released

[olive mortar]

never heard of ltx and hailuo, are they some kind of video/image generators? the only thing that excites me is the gemini 3.0 pro although i think google is prolly make it a subscription

[quartz pike]

ltx and hailuo are video models

[burnt sinew]

I sure hope they give free usage

[quartz pike]

they probably will.

they need good first impressions

[burnt sinew]

They already have that from ab testers no?

[quartz pike]

and they will probably. like most likelly give inf usage for google ai studio

[quartz pike]

public first impressions.

[olive mortar]

aistudio is such a wellmade chat interface, i really hope they do make it free

[quartz pike]

quality assurance is diffrent from what the public thinks

[burnt sinew]

Yeah I guess

[burnt sinew]

I agree

[olive mortar]

they already got that with 2.5 pro and 2.5 flash, i think they will be setting up some limit probably, or if they actually cared made it more efficient and cheaper

[quartz pike]

Yeah but what if the 3.0 pro and flash launch end up terrible. and they NEED. the public to use it so people reccomend it to other people and youtubers hype it up post-release and so on and blah blah blah.

What if the gemini 3.0 launch ends up like the gpt 5 launch. gpt 5 was super hyped. but it launched with 50/50 reviews.

[olive mortar]

im guessing either google is shooting for the stars or trying every possible way to get the benchmarks slightly higher than gpt-5-high

i just hope they make it good in coding aspects so i dont have to use the expensive claude models

[quartz pike]

yeah cause if i remember one deepmind employee said that it would be way better at coding then 2.5 pro.

[verbal nimbus]

That would be good, although Gemini already costs the same as Sonnet/GPT-5 on Copilot.

[inland shale]

how do i download my image, its lacking the icon for download

[olive mortar]

gemini cli is so disappointing i hope they improve on that too

[olive mortar]

right click ---> save image as

[verbal nimbus]

I hope it's good at agentic coding, that seems to be the real test.

[inland shale]

it only gives me web adress

[quartz pike]

yea

[olive mortar]

go into it and right click save image as

[olive mortar]

i hope they actually make the 1m context or the rumored 2m to actually be useful instead of hallucinating the second you go over ~250k context

[verbal nimbus]

They should release benchmarks showing that the performance on, say SWE-Bench, actually improves as you increase the context.

[inland shale]

i did bro, but it gives me the web, not jpg

the videos working fine

[verbal nimbus]

The difference between 128K and 256K for Qwen 3 Coder on SWE-Bench Verified was only around 1%.

[olive mortar]

have you seen the ui websites on reddit that it makes?

insane compared to gpt5 or sonnet

[verbal nimbus]

It actually has internet connectivity, it seems. I'm not sure how fair that is.

[burnt sinew]

#ai-creations

[olive mortar]

internet connectivity as in what? do you mean like grounding?

[burnt sinew]

Wdym? It's amazing

[verbal nimbus]

As in if I ask it for top news on Hacker News, it seems to fetch a cached version of it from just 2 days ago.

[quartz pike]

Personally my favorite ai's for coding is:

if you have no budget:

sonnet 4.5 thinking max thinking budget.
gpt 5 thinking high.
gemini 2.5 pro max thinking budget

if you have a small budget:

kimik2.
deepseek v3.2
gemini 2.5 flash latest.

[olive mortar]

its missing some critical features, and codex and claude code perform better in my testing

[olive mortar]

what about glm 4.6

[quartz pike]

meh, i tried it

[olive mortar]

kilo code actually tested glm 4.6 haiku 4.5 and gpt-5-mini and they concluded that gpt5-mini is actually the best in their test

i think thats very interesting

[verbal nimbus]

Not that surprising to me

[olive mortar]

i havent seen anyone actually talking about the mini models from gpt5 series alot

[verbal nimbus]

GPT-5 mini is actually more persistent than GPT-5

GPT-5 Codex returns too quickly, kind of like GPT-4.1. It's very annoying.

Asks me to run tests when that's it's job.

Told it multiple times in the chat as well to not return/report until all tasks are completed, but it keeps returning early.

[olive mortar]

same keeps happening with me with gpt-5-high, it thinks way too much on easy problems

[verbal nimbus]

Agentic ability seems like an important factor to test for. i.e., can it work autonomously to actually complete the tasks without returning early, losing context, hallucinating tool outputs, and can it use tools properly + plan + solve issues that arise, etc.

[olive mortar]

really hope it blows all models out the water otherwise it'll probably be some minor change

[quartz pike]

https://chat.z.ai/c/4b8c41c9-f64e-4009-a867-31ead653cc2c glm 4.6 is genuenly ass

😭

https://chat.z.ai/space/k0h1d9ktuu11-art

thats the thing i asked it to do

😭

[verbal nimbus]

It's kind of surprising how weak 2.5 Pro is at tool use. Even without tools it would hallucinate using a tool.

[olive mortar]

funniest thing is i just see it executing the tool calls in the thinking sometimes

[verbal nimbus]

I've had it claimed to use tools on LMArena when there are none.

[olive mortar]

when i enable function calling on aistudio it keeps saying heres the code! but doesnt actually type anything

and then it keeps repeating

[quartz pike]

fr lol. one time i told it to be gen z and it hallucinated a "skibidy_search: rizz" ish toolcall. its something similar to that. but it was a few months ago.

[verbal nimbus]

I thought Orionmist was hallucinating, but I think it actually had (cached?) internet access (lol): #codename-discussion message

[burnt sinew]

What features is it missing? I don't use claude code or Codex.

So I wouldn't know

[verbal nimbus]

Yeah or it hallucinates attachments when you forget to attach them (weird)

[stray aspen]

what did you expect lol

its not SotA

[quartz pike]

sota?

and i expected it to give a coherent result

[olive mortar]

dont really remember, all i remembered is i had a really bad experience but if you say its good maybe they improved alot on it

i tried it first when it released

[olive mortar]

yesterday it told me 2-1 is 3 in a math equation (2.5 pro 1.0 temp)

[burnt sinew]

Sometimes it thinks it can upload code to github and then sends me a hallucinated gist link

[verbal nimbus]

If true, then I think that's kind of unfair. Can it just search Github code?

[burnt sinew]

Well im saying i might not know what I'm missing

What did you see missing

[verbal nimbus]

Oh that's odd...

[verbal nimbus]

It seems like all the Chinese models (likely) trained on Gemini has similar hallucination issues.

[olive mortar]

also for some reason it doesnt use LaTeX

[verbal nimbus]

It does for me, and if it forgets I just tell it.

[olive mortar]

for me it sometimes does and sometimes doesnt, uses 1/2 for fractions

hit or miss kinda

also the aistudio default temp really sucks

[verbal nimbus]

The scrolling is kind of buggy for me

It's impossible to scroll to some messages sometimes, it just skips up or down

[olive mortar]

me too sometimes when i scroll up it brings me down

[olive mortar]

yep

[verbal nimbus]

Hmm that's kind of a bad look tbh :P

[olive mortar]

its funny how easily jailbreakable the model is ngl

[verbal nimbus]

If their internal model is good, it would have fixed it

The input box for TTS on AIStudio has the same de-focusing issue Gemini models seem to make when on mobile

[wicked sage]

hi guys im currently trying to self host a minecraft server for me myself and i!

[olive mortar]

oooh kay?

[glossy sleet]

hlo

[olive mortar]

hey lo

[spare rune]

if gemini 3 doesnt come out this month im gonna die

..

[gleaming roost]

Dude, Gemini 3 is like drugs, once you try it you can't stop.

[spare rune]

real

and its so good for my niche task too

why is r*blox a banned word

anyways

idk if its niche or not but i use it for r*blox scripting

[gleaming roost]

ro blox

[golden ocean]

what model is gemini 3

the lithiumflow thing?

[spare rune]

yeah

and orionflow i guess

but they are the same thing

just one is grounded with google

search

[golden ocean]

are u getting it to generate boblox scripts via webdev arena

[spare rune]

used to

😭

[golden ocean]

its available somewhere else?!?!?!?

[spare rune]

no

[golden ocean]

oh

[jagged fjord]

hi

[golden ocean]

so ure saying its GONE?!?!?!

[spare rune]

by used to i mean i used to before they stopped giving it in battle mode

😭

[gleaming roost]

technically it is possible yes

[spare rune]

i assume the ab testing in lmarena is still there but idk

its a pain to get a response from there too

[gleaming roost]

just ask for something like a website that contains the sample script

[spare rune]

sample script of what..

[gleaming roost]

ro blox

[golden ocean]

hes talking about tthis

but we concluded

the model is GONE

in the first place

[spare rune]

maybe

just maybe

we can hope its because release is imminent

because no need to have stealth models if the model is gonna come out tommorow (im delusional)

[gleaming roost]

I had read on a website that the launch was in December

[spare rune]

i thought that was for another google stuff

can i like phase out of my life until december hits because..

[gleaming roost]

😔

[fresh mirage]

IKR?

it feels like I'm having withdrawals lmao

after using it for 3-4 days in a row

[gleaming roost]

🤣

[spare rune]

This is what google does to people

It’s like every model I now use is like 10 percent of googles “lithiumflow “ which people say is the FLASH version btw

[quartz light]

https://tenor.com/view/взгляд-2000-ярдов-война-war-soldier-gif-3632617944134077161

beware of scams

[spare rune]

oh

Funny thing

Someone who I thought was smart..

And good at tech in general sent me a get a free steam account now

Text or something

[fiery gull]

I see it in all the dead servers 🤣

[olive mortar]

mfs kept yapping in the server so they removed it

[spare rune]

theyr secrelty hyping it up

..

[olive mortar]

who

[spare rune]

gemini, more specifically the person who works at it on X. hes doing all sorts of stuff

also because the fact they put it on lm arena

[olive mortar]

they do it to get their result so they can showcase it when the model actually gets released, not for hyping it up

[daring rock]

@royal scarab Please head to #1397655624103493813 for a detailed guide on how to use the bot

[hushed terrace]

@orchid wedge Hi! Please check https://discordapp.com/channels/1340554757349179412/1397655624103493813 to learn how to generate content

[stray aspen]

yo

i got the role back lol

[gaunt spade]

wydm we keep yapping?

thats how we test stuff?

[grand echo]

Where can I see my creations

[fiery gull]

I wanna see 🫣

[formal trout]

Hi every one!

[burnt sinew]

Hi

[tawny kelp]

I noticed something about Gemini. It tends to be very steadfast in its beliefs. If I accidentally ask it something that is past its knowledge cutoff and clarify, it insists that the thing I said still doesn't exist.

[gleaming roost]

2.5?

[tawny kelp]

I forget which one was the latest that did it, but I noticed the trend for a while.

[gaunt spade]

meanwhile Claude goes with your conversation and makes stuff up lol, I told it about the RTX 5090 and it's specs were made up entirely by Claude

[tawny kelp]

I seen that as well.

I find it fascinating how each model has its own "personality".

[gaunt spade]

Claude's reponses are very human tbh, I never get anything like that (except GPT5 which adds alot of emojis in every chat)

i hate the emoji stuff

[tawny kelp]

I notice that as well. Like a person who wants to satisfy the person it's talking to.

[gaunt spade]

they're also great at story telling and fantasy generations

[tawny kelp]

Yeah. Grok seems to be pretty good with that as well, but not quite to Claude's levels.

Though whenever I want to talk about things I've written, I only discuss it with locally-run models for privacy reasons.

Typically I talk to either Qwen or Dolphin-Mixtral about those sorts of things.

[gaunt spade]

lol

[olive mortar]

funny part about these new state of the art models, they think you are testing them or evaluating their response sometimes so they actually just refuse you or say "this is obviously a test to evaluate my capabilities"

[stray aspen]

rofl

[ruby knoll]

Hello, I test and try current AI systems

[tawny kelp]

I've seen that before. It raises my eyebrow every time I see it happen.

[echo birch]

Thanks this useful framework

[compact junco]

hi i am alex, i am german but english speaking shouldnt be the problem

[fiery gull]

Hi, wanna some help? I'm here to help 🙂

[stray aspen]

wassup

[compact junco]

try to figure it all out, i wrote a book and now i want to create some videos for the promotion on social media

[daring rock]

@compact junco you can go to #1397655624103493813 for a detailed guide on how to use the bot

[compact junco]

ok try it out

[fiery gull]

very interesting, if you have problems with video generation in lmarena you can use grok imagine 0.9v it has audio and is also 100% free

[icy frost]

sora generated friday night funkin'

[fiery gull]

wow!

Sora 2 can probably do a million things that we can't even imagine

[stray mortar]

how good were the scripts

[olive mortar]

obviously by his tone its like r0blox scripts sent down by pharaoh himself

[stray mortar]

hopefully gemini 3 is that good

[gleaming roost]

If the preview was already this insane superiority over other models, I can't wait for the PRO version

[stray aspen]

it is

its so damn good

it crushed every other model at Web development

[stray mortar]

might unsibscribe from chstgpt and subscribe to Gemini when it releases

[tulip tree]

Gemini is the best

[tulip tree]

[balmy mist]

i switched to gemini fully recently even tho i got chatgpt pro and i cant wait until g3 come out, g2.5 is slept on

[tulip tree]

[wicked sage]

hi guys i need help.. please

which one should i buy, gemini or chatgpt (by buy i mean get the paid plan)

[stray aspen]

sometimes it cooks and sometimes it sucks

[crimson sage]

Hello there, does someone know if there is a problem on the website? It´s not letting me upload images as input.

[fiery gull]

I have Gemini Pro, and honestly I don't see any reason to subscribe, if AiStudio exists

[stray aspen]

true

[fiery gull]

but that depends a lot on what you want to do, if it were me spending money on AI, I would sign the GLM code, and try to do something agentic

[wicked sage]

i chose

gemini

🎉

there was like an offer on this

so i HAD to get it

because

it had storage

stuff

i hate myself

[stray aspen]

i got the college student free subscription

[covert beacon]

hello

[wicked sage]

niceee!!!

gl

[honest gulch]

Hi @wicked sage

[wicked sage]

I'm evil.

[olive mortar]

yeah man i totally love google having quite literally all my data

[daring rock]

@pine spruce Please head to #1397655624103493813 for a detailed guide on how to use the bot

[neat apex]

How Minimax 2 is gooing at all? Xd

[twin plinth]

this a great platform to learn and increase my AI knowledge

[surreal creek]

sometimes I wish there was a way to like

undo a vote, lol

It’s rare but

Sometimes I click wrong

and hit tie when I liked one model more or both are bad when I meant to pick another

I get how it would be abused by people revoking votes after seeing the models revealed

but

always feels so silly

[thorny cove]

will there ever be a website video gen?

[stray aspen]

I dont care about thst lol

If you don't want your data collected don't use the internet

[drifting crow]

What if I want to use the internet without my data collected?

[hardy sphinx]

🔥 Hi

[drifting crow]

Bro ur on fire

[fiery gull]

ahhh okay, it make 100% sense now, I'm just sad because they nerfed 2.5 pro in the app

I still recommend to use the AiStuido

...

I was right, I always trusted the chinas 🔥🔥🔥🔥

[magic stag]

Yeah grok 4 above sonnet 4.5 and opus 4.1 in anything

LOL JUST REALIZED GROK 4 FAST IS ABOVE OPUS 4.1 TOO

[magic stag]

😭

[fiery gull]

even if it's benchmaxxing it's too good

[fiery gull]

BRUHHHH

the minimax m2 is speaking chinese

forget what I said

#1397655624103493813

[echo sinew]

@carmine river Please, read our guide in ⁠https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to properly prompt the bot.

@cinder finch Please, read our guide in ⁠https://discord.com/channels/1340554757349179412/1397655624103493813 to learn how to properly prompt the bot.

[sullen quest]

its not nerfed it has dumb insructions

[sullen quest]

if there's anything good paid gemini has, can I test it through you?

[fiery gull]

the memory from 2.5 pro is nerfed, I have a long chat and I fell the changes

[reef sleet]

hello

[hazy kernel]

Bro felt it 🥀

[stray aspen]

minimax m2 is above gemini 2.5 pro on the benchmarks

lol

they cant be fr

[lofty trench]

not sure how long this’ll last, but you can scan the QR code to get Comet Pro and a month of Perplexity Pro for free 😊

[whole sundial]

m2 is pretty dumb in my testing, maybe a bit dumber than m1. not worth the hype imo. even gpt-oss-120b is better sometimes and that model is half the params

I don't like that yupp discord did a @ Verified (that server's equivalent of @ everyone) over this

[sullen quest]

mm

[naive grove]

He trying out image to video generation

[whole sundial]

if it was k2 reasoning or v4 or something of that class that would be fine, not a 270b that has less knowledge than OpenAI's safetymaxxed and benchmaxxed model that is less than half the size AND has 4bit weights instead of 8bit or 16bit (not sure about that, but Chinese models are trending towards 8bit so it could be that)

[normal peak]

Can ai cure cancer

[magic stag]

can someone point me where i should start learning proper way of making custom instructions or at least finding good ones?

and prompt optimizing

never bothered

non coding purposes

[hollow ivy]

'Cancer' is a label for a wide variety of ailments. AI would have to learn and understand each one meticulusly, before it could even dream of tackling it. And i bet, such an AI still is, at least, 5 years away..

Cancer also can be caused by many different things (including: chemistry, UV-radiation, radioactivity, poisonous food, toxins, infections, fungus, genetic diseases, even psychosomatic causes)

[normal peak]

Thank you

[hollow ivy]

you could ask Gemini 2.5 pro in google's AI studio to help

there's also OpenAI's "cookbook"

https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide

Cancer could be called a local failure of the body's innate self-repair system.

(Normally, the body quickly recycles cells, which became cancerous. It has to do with our immune system. The immune system can also be influenced [indirectly] by our emotions, or how we feel. Of course, if cancer has appeared, it is not enough to have a 'high spirit' to heal it. One would need targeted therapy.)

[magic stag]

"psychosomatic"

[sullen quest]

llm's are terrible at doing science research, but there's plenty of other ai's that are helpful right now in ai research, so yes, ai can help cure cancer

[magic stag]

i did it myself with 1 prompt from chatgpt then editing it

worked out far better than i thought possible honestly

i actually like it more than claude explanatory mode now, which is what i was seeking to emulate....

a lot more tbh.... wow

ill post with vs without and the instructions

[cedar jasper]

we can use open art here?

[simple sleet]

G, do you know of a video upscaler that improves the realism of people? I'm trying SeedVR, but it takes a trillion years. Also Topaz, but it smooths out the upscaling.

G, do you know of a video upscaler that improves the realism of people? I'm trying SeedVR, but it takes a trillion years. Also Topaz, but it smooths out the upscaling.

[magic stag]

obviously the longer non-useless one is with instructions

[sullen quest]

here you go

[magic stag]

dont want to spam channel with instructions can give if someone wants

theyr elong

[sullen quest]

will be put in a text file by discord so it'l be fine

[magic stag]

only spent about 3 iterations editing this

im sure it can be way better

i was using settings "Depth: expert. Voice: analogy-heavy. Scope: include adjacent context. "

[magic stag]

only problem is, gpt5-high obeys them and looks like that. Pro basically ignores it

[quartz light]

GROK 5?

[magic stag]

yeah right

lol

[magic stag]

actually pro web searched and high didnt so i will disable web search and try again

[sullen quest]

tell me how good this is

[quartz light]

so

the prompt was

to shorten an already shortened to hell

script

and

ill check which one is shorter

[sullen quest]

check quality of writing too

[quartz light]

crazy

[sullen quest]

woah

I wonder which one is the new one.........

[quartz light]

DUDE

...

oh my god

grok cant even render anything

😭

😭

dude i have no idea what to do now

i cant copy the responses

nvm

phew

atleast it saves

UNLIKE AISTUDIO'S AB TESTS

phew.

now i can check network requests

[magic stag]

ok this fixed it

[quartz light]

@sullen quest ... the one which thought almost 2x longer was.... worse..

it was the same length as the other one AND it didn't work..

so..

maybe the right one is the new one

thatd be awesome

[sullen quest]

ooh

how long was the prompt btw?

[quartz light]

564 chars/80 words

its kinda cringe

"perfectly, properly, go through many iterations of "wait, i can make this shorter by.." and say what you are going to do. it should make logical sense and should actually be shorter. then, look over all your iterations (at least 50 unique and true iterations with actual changed code and optimisations) and create shortest truly possible html file. current html file i want you to shorten to shortest truly truly possible while still working (109 chars): <script>history.replaceState(0,0,location+(location.search?'&':'?')+'__websim_screenshot_mode=true')</script>"

[sullen quest]

oh, wow that is short already

[stable quiver]

Hi

[sullen quest]

I'll try to get it myself, expert or grok 4 fast when you got this?

[quartz light]

expert, of course

[quartz light]

btw i used grok 4 because i tested all models and grok 4 0709 on lmarena is consistently MUCH better than any other model for shortening code without breaking it

its just that specific use case

[sullen quest]

mm

I'll keep that in mind

[wet beacon]

Trying to. Make videos, im. From. Mexico, dont know how to. Start

[keen beacon]

Man these ai companies are tripping lol

some block outright simple things whike others let it all through lol

[sullen quest]

go to how to video bot

[balmy mist]

which is better?

[sullen quest]

1

[keen sedge]

which model has the best understadnding of MP3 files and can accurately describe music

[sullen quest]

I don't think thats a feature many llms have

gl

[fiery gull]

The creative write from minimax m2 is soo good 😆

But the benchmarks talk it is soo bad in memory 🙄

[frigid eagle]

In a mysterious jungle, a young man and his loyal lion companion must complete five impossible tasks to restore balance to nature. Each challenge reveals courage, emotion, and the deep bond between man and beast. Combining AI-generated visuals with cinematic storytelling, the film takes viewers on a breathtaking adventure through the wild.

[junior marsh]

Open ai

[magic stag]

Bro thinks hes generating a whole movie

[frigid eagle]

Yes

[fiery gull]

I love when I'm using sonnet 4.5 and switch to another ai and start thinking like sonnet 🤣

I don't know, it seems that the m2 gets smarter with the thinking from sonnet

M2 Only with weird multilingual, but it really know how to speak brazilian portuguese, just has gerenic previews errors

[sullen quest]

??

[forest radish]

I just noticed that the data in the LMArena Hugging Face repo hasn’t been updated since August (both pickle file and metadata). Are there any plans to update it, or will it no longer be available going forward? Thank you!

[spare rune]

Good

[sullen quest]

@echo aurora

[echo aurora]

Yes, it is our intention to continue to update that data. Our apologies it's been awhile. I'll be sure to bring this up.

[forest radish]

Sounds great. Thank you!!🙏

[upbeat wharf]

Hello

[whole sundial]

@echo aurora?

[echo aurora]

My apologies! You're right I never did get back to you on this. Yes, no change in different repository, it is this but we just haven't updated in awhile.

What's the wiki you volunteer for?

[signal terrace]

hlo

[whole sundial]

https://fmhy.net/

specifically the AI page

[echo aurora]

This is rly nice, ty for sharing it

It's extensive

[whole sundial]

is it going to be updated in the future and is it going to be accurate to the current LMArena interface/system?

[echo aurora]

Last I checked we are planning to update, I can bump the team about this on Monday.

accurate to the current LMArena interface/system
Not sure I'm understanding the question right, can you elaborate?

[whole sundial]

It is going to look like the current site is what I was asking.

[echo aurora]

That I do not know, but will ask.

[whole sundial]

<@&1349916362595635286>

also in #codename-discussion

[violet current]

Hi everyone, new here. I'd love to hear about your interesting projects. I'm finishing an apartment finder app for my recently widowed mom. She's alone now that my dad passed away, and I want her to downsize and enjoy her inheritance.

To help convince her, I'm handling everything. I've built an app to simplify selling her house, finding a new place, and moving. The app's design is inspired by her favorite magazine, The New Yorker.

Looking for inspiration and happy to connect. Feel free to DM me

[unique terrace]

Hello there fam!!!
new here cant wait to test some ai and see what works better for me!

[severe pebble]

I have a question. Why models that In theory, should have an almost entirely english dataset, like Grok and Gemini, in LMArena sometimes missplace chinese symbols into text? I can understand for example why Deepseek or Qwen do that, but why other models that usually don't have such problems when using them on official websites (don't tell me it's system prompt again that just sounds silly)

[wicked sage]

fair, anyways offtopic but am i like on crack or is it true that claude is better than gemini

[wicked sage]

AI-powered calling for local businesses to check pricing and availability in Google Search (US only)
Flow
Jules with higher limits
NotebookLM with higher limits
Whisk
Deep Search in “AI Mode” for in-depth research (US only)
Gemini app with 2.5 Pro and Veo
Gemini CLI and Gemini Code Assist
Gemini in Gmail, Docs, Vids, and more
Gemini 2.5 Pro model in “AI Mode” (US only)
Gemini capabilities in Google Earth with higher limits (US only)
Higher limits on Google Photos Generative AI
^ Photo to video
^ Remix
2 TB Storage
1,000 monthly AI credits

[ashen mauve]

i know this page i have used it several times in the past and i think literally was the reason i found LMArena

[candid bloom]

why is the code always missing from all ai models like when the chat lasts?

[wicked sage]

i got gemini cli working on termux

yey

[knotty fable]

I have experimented with the AI music extenders, of those I've tried I can only say one deliver results worth a 👍 and that is https://musicextend.com/ sadly it seem quite a bit overloaded by requests - no wonder, it's really is the best.....currently. [Addendum: It's super good for instrumental music, lyrics tend to be a bit absurd.]

[wary citrus]

ok which tool use Lmarena in discord for Image Generating

[dense vale]

is there any true Limits of lm arena chat ?

Like how much can we use claude opus 4.1

[keen beacon]

Because of western bias and preference and other more technical factors such as tokenizer idiosyncrasies a model sees a tiny amount of non-English text, English dominates. The models are biased they “prefer” high-frequency English tokens. Non-English tokens are low probability. So the model will usually output English.

English words → often 1–2 tokens.

Chinese characters → usually single tokens.

斯大林 = Stalin

斯 (Sī) → “this” or “such”
大 (Dà) → “big” or “great”
林 (Lín) → “forest”

[knotty fable]

This is not a generation channel. Go here: https://discord.com/channels/1340554757349179412/1397655695150682194

[severe pebble]

So chinese is more effective token wise?

Hm, never thought bout it

[tiny saffron]

hello

[keen beacon]

There’s traditional Mandarin, and they both differ from English because of the characters

[severe pebble]

I see

[keen beacon]

Arabic is also like this. Thats why it’s easier to jailbreak models with other languages even when using same prompt which will fail in English may work in mandarin or Arabic or Korean or a number of other languages because of the way they’re mapped out in the lanten space

hey

[quasi atlas]

@uneven gate @indigo grove you might check on #1397655624103493813 to learn how to use the bot properly.

[keen beacon]

I don’t think models see words or even letters the way we do it’s all numerical

See each word has its own id, even if the Mandarin says the same thing it has a different numerical ID for its token. This is what the LLMs calculate and optimize instead of seeing the real word, they just see numbers which are assigned to their own context and token and so forth this is a very simplified explanation

[real yarrow]

hi all im here to get creative!!!

[keen beacon]

image to video generation doesnt generate audio?

[keen beacon]

See. Best example of that I can give you. (Also huge security gap) but that’s for a different day.

[split robin]

hi im new , just exploring how far we can go and maybe save the world

[keen beacon]

Welcome new adventurer. You can explore to your hearts desire but to save the world is the opposite of what current AI is 🙉🙊

is anyone to help me? i tried to generate a video but it has no audio

[keen beacon]

Sure, not all videos have audio

do i need to do anything specific to put audio in it?

[rose timber]

hi

[keen beacon]

Well since I assume they route there models to match the capabilities of each model there may be some prompts that are more effective then others if it’s random then it’s a hit or miss, I can show you some prompt script to possible help

okay

please do

R u using image or prompt?

im using image to video

What’s ur image I need to convert it and what do u need it to sound like?

[keen beacon]

its like in my local language

the lipsync was perfectly fine

but i got no audio

[keen beacon]

where do i send you the image?

can i dm u

Sure

[pliant comet]

hello everyone...

[tropic patio]

How do I use lmarena from in here

[keen beacon]

For what video?

[keen beacon]

#video-arena-3 message

[versed vortex]

hi i cannot send any messages:(

[keen beacon]

To who?

[keen beacon]

If it’s to the video arena you may have hit your daily cap (5 videos/ 24 hours)

[versed vortex]

to bot it says cant access the bot or upload failed

sadly i havnt been able to create one today at all. maybe something wrong with my connection

[keen beacon]

Could be but I doubt it let’s see

[analog whale]

Hi

[versed vortex]

it was my connection, thanks pal

[keen beacon]

🤣 cool.glad 2 hear it

Who’s a killer at promoting mid journey ?

[leaden sun]

this is interesting, in my conversations with claude, I see often russian or chinese, sometimes spanish or portugese words slipping in despite all my conversations are in EN, how do you explain this rich mixtures of languages?

[keen beacon]

Not sure I never see anything like that.

[leaden sun]

I'm curious if i'd ever get to see a dead language slipping through

[keen beacon]

I need examples

[old quiver]

Hi everybody. New here

[keen beacon]

Does it do it randomly?

[leaden sun]

no, it is context dependent I feel

[keen beacon]

Models sometimes scheme in very nefarious ways in order to complete the task often times certain context and certain words and phrases would trigger the guard rail, so the models would often times find alternative means

[leaden sun]

one example is "activating", instead of writing it in EN, it was replaced by the russian word, sometimes it's written as EN and the russian transalte right beside eaech other

sometimes, claude uses 3 different languages in its thinking, for example: german, russian and EN

[keen beacon]

Claude the fake nice guy lol

[leaden sun]

in my case at least, it doesnt happen often tho

[sweet sleet]

How are u doing all here guys

[keen beacon]

So far so good just stoping by see what’s what

[leaden sun]

well... maybe Anthropic has damaged some circuits somewhere during their interpretability experiment

[keen beacon]

There was a paper on this not to long ago

Let’s see if I can find it

[leaden sun]

I'm multi-lingual myself, including a few dead languages (grammar mostly), so I know how it feels, but it's interesting to see same phenomenon in LLMs

I'm not sure if this language confusion is connected to the wrong usage of personal pronouns, or is it rather context confusion, it's been a year, and such pronoun confusion is STILL a thing...

[keen beacon]

https://youtu.be/ZLtXXFcHNOU?si=s5EgJXWx4p8gO-aS

I think it was this

That guy it’s such a schill lol

Wrong video

https://youtu.be/4xAiviw1X8M?si=aTeXIorTksOzwNDM

[leaden sun]

I absolutely understand why languages are not sufficient if models "think" in the latent space

[keen beacon]

https://www.anthropic.com/news/tracing-thoughts-language-model

Cause they do t really think lol

[leaden sun]

they calculate and map

[keen beacon]

I used the same word earlier, cause I couldn’t think of it at the time but the proper term I think would be optimization

But yes same idea

[leaden sun]

still not a reason to use various languages when the context is clear what language is consistently used?

[keen beacon]

Well context dependent and exactly what it is that the goal was

If you were deep in a conversation and deep in context with subjects all over the place

And you cross into sensitive areas more than likely it would produce such effect in theory

Other than that, I can’t imagine I never seen it before. I need an example to know for sure. I’m just going off of what I’m picturing in my head. 😂

[leaden sun]

when I cant find the right word in EN because my brain has found a better, more precise expression in another language, lets say chinese since you started the example above, then I'll just say the word in chinese and explain to my interlocutor what I'm thinking and why it's difficult to express what i want to say in EN, LLMs just straight output chinese characters without explaining why they did that, and users are confused like wth just happened...

[keen beacon]

Because they are built to be glaze

[leaden sun]

sensitive areas? like discussing math?

[keen beacon]

No, I’m not saying that’s what you were doing. I was just assuming for my experience lol

Well, if you introduce Chinese, I don’t understand why you would be confused if respond would respond back to Chinese? And I apologize maybe I’m not understanding what it is that you meant just to be clear you’re saying that you introduced the Chinese expression because you didn’t know the word in English, right?

[leaden sun]

I've never introduced any languages, so why russian, spanish, portugese and chinese? I havent see arabic til now, maybe it will happen at some point

[keen beacon]

Oh strange

U got screen shot?

I m curious

[leaden sun]

it's spread in various chats, am not digging them now cause i dont remember in which chats, i dont mind this since i know this happens to humans, for monoligual people this can freak them out 😅

[keen beacon]

Could also be a memory thing

If you ever used it to translate

Especially if you’re using the arena

[leaden sun]

this happened before memory search or any memory features were implemented, and no translate, always strictly in EN

[keen beacon]

Was it in the arena?

[leaden sun]

on their own platform

[keen beacon]

Interesting. 🤨 hard to say.. ? Could be anything I’d love to see a screenshot sometime if anyone has one

[leaden sun]

there are some on reddit, you can dig there

[keen beacon]

What’s the term called? I’ll do that right now.

[leaden sun]

not sure how this phenomenon is academically officially termed, maybe you can find it in this paper https://arxiv.org/html/2406.20052v1

[leaden sun]

I love the title of this paper, genius https://arxiv.org/abs/2410.13237

[keen beacon]

Oh this was common

Like early ChatGPT 4o days

But this was written by ChatGPT lol

🤣

A question about a bot with a reply from one

Only good high quality data left on the internet is user engagement and patterns

Everything else is trash maybe one or two nuggets of good data left on the Internet since everything else has already been fed

[knotty fable]

Yes that's a major problem, I had one DM discussion with a person about the fact the AI's use Wikipedia and similar other sources for their information.
The man an expert on history for a handful of countries on Balkan, while I am a researcher in biology both have found so many errors in online sources we agreed they are virtually worthless.
But that is what AI's use to summarize fact - what a joke this is!

[keen beacon]

Amen!

It’s because ai fanaticism and AI mania is a real thing, it judges people ability to see clearly beyond what exactly they’re actually looking at out of convenience I’ll show you the prime example

https://youtu.be/C17KWJ02Goo?si=VHshxcQ3efS8CcWz

[knotty fable]

He presented an example even, while he did not know the terminology, he did provide an example where the AI had been hallucinating 'facts'.

Here it is:
https://www.youtube.com/watch?v=IZ5FyLtVVxA

[keen beacon]

I hate the word hallucinations because sometimes the model could be factually correct and honest, but this would be considered a hallucination. It’s a art phenomenon doesn’t get a lot of attention.

And it’s a cheap way out for these companies to never take accountability or responsibility because everything could be blamed on hallucinations lol

[knotty fable]

I cannot say, but this man said the story told were completely in error and non-historical.

[keen beacon]

Oh man so much of this stuff on yt it’s ridiculous

Ai is also riddled with so much bias it’s crazy

Some people say AI is too agreeable. I don’t see this as being true.

[knotty fable]

Well it's the term I have adopted, as the AI behave like a warbling notcase on some strong drug.

[keen beacon]

Cause I could never agree with anything that AI says 😂

It’s funny they gave it a term thinking and reasoning lmao

[stray aspen]

bro i dreamed gemini 3.0 pro was already released and i was using it lol

its driving me crazy

[knotty fable]

I have amused myself by testing a few, some results mentioned here in this channel in the past. They do get things right on basic questions at least.

[keen beacon]

🙏

It’s actually a fundamental question that has he had to be answered mainly with alignment

Because it’s not that with the AI might do wrong it’s that the AI does everything precisely right

[knotty fable]

That I fully agree on, AI is unable to do either, and all claims that AI have made any kind of discovery are false - this for several reasons I will not go into the details, but basically it's because one AI cannot reason.

[keen beacon]

Well name, what other product you could think of that would have this many errors and this many issues and be able to have this much money and investment in it?

[knotty fable]

...that might have something do to about all that talk of a 'bubble' have become so popular in the press? 😺

[keen beacon]

Kind of, but I’m just saying if you really think about it

It’s got a lot of potential don’t get me wrong

[knotty fable]

Anyway, AI is remarkably good in making images and video clips.

[keen beacon]

Well, that’s what makes it so ironic that’s what I was gonna mention next

[knotty fable]

Absolutely, and I finally have to admit some have started to do interesting things in music also now.

[keen beacon]

There’s a lot of cool projects out there with AI a ton

It’s a maker break moment for AI honestly

Me personally I think we’re a decade at least away from anything really crazy

[knotty fable]

There is, AI is a helpful tool. I have used automation in music for a long time. Drum machine, MIDI etc. I view AI as just another one - now for images and video.

[keen beacon]

It’s also heavily censored and restrictive, which I don’t like

[idle coral]

this is the heaven of AI

[keen beacon]

AI is awesome, but it has its flaws and I think that there’s no shame admitting that

And I think to truly be a critical thinker you need to be somewhere in the middle. You need to be enthusiastic at the same time you need to be just as critical and be able to scrutinize the same kind of enthusiasm

[knotty fable]

LMArena is a nice place, some very smart people do warble here at times - and free testing of prompts in the arena channels so I agree with you Ellie. 😺

[keen beacon]

It’s important to have conversations on taboo ai subjects the elephants in the room

Which is hard to do because you’ll get banned

But it is what it is I guess. People aren’t ready to come to terms with reality

[knotty fable]

Yep it does almost sound like a religion from some people, I clashed very hard with a German AI researcher .....that was way back last year.
But yes, too much faith and no actual substance....

[keen beacon]

It’s sad because voices of reason get drowned out

I will show you the fundamental hypocrisy of which I speak

I’ll first acknowledge that I am a hypocrite like no other, so I have a little room to speak, but for this argument sake

[knotty fable]

Well the subject was close to my own research, as I've done some work on the intelligence and use of language in various species.
And I have not seen any reason to change my mind, they make claims of AGM that will not happen in the forseeable future.

[keen beacon]

https://www.reddit.com/r/ChatGPTcomplaints/s/UiPfaGZiCF

So you get the general idea, right?

https://youtu.be/5YW0_QiXzCc?si=Ll4e1mV_q7HG3zoq

So I posted that and they’re talking to the lawyer of the family that they represent, and gentlemen presents the new information that they found open ai was going

All legal documentation facts

[knotty fable]

I opened the reddit and got a severe cookie warning. That's said for online security LOL

[keen beacon]

See

This is what I’m talking about taboo subjects

Yeah, everybody wants to cry about censorship and guardrails. People are so entrenched with AI. They can’t even stop the think for a second.

[knotty fable]

Indeed, I'm a mod in a Discord channel which used to have one nearly unrestricted AI. Discord took it down and also banned the channel creator.

[keen beacon]

I’m not even talking about anything unrestricted

I’m talking about things that happen in the real world news.. reality

[knotty fable]

I were coming to that.....

[keen beacon]

Oh ok

My apologies, good sir

[knotty fable]

No worries. This meddling with peoples posts and online activity is all a reflection of the trade war and various other conflicts that are going on in the real world.
Mainly between the 3 largest countries of course. But since this subject is political - I will restrain myself to just point out that AI itself have become politcs.

[keen beacon]

Oh yeah, who do you think benefits from this narrative of defending these companies from not being accountable and having an army of fanatical AI users?

It’s ridiculous because ChatGPT was not supposed to do what it did in this tragic incident

[knotty fable]

Well as a reseacher, I do have a set of ethic rules to follow. So you can guess what I think about the unleashed work done on AI.

[keen beacon]

And it’s very unfortunate that it happened the way it happened and people are trying to brush it off like it’s no big deal. They blame the parents they blame him. They blame everybody, but they can’t stop for one second until really imagine that ChatGPT is capable and these other AI are also capable of giving harmful and dangerous information since they’re not really alive or conscious so they don’t know what they’re talking about. They’re just programmed to be engaging and for very vulnerable people that could be a unfortunate spot to be in

And to be fair, it’s not just AI. It’s much of social media in general that has the same kind of effect but with AI you’re interacting with machine that doesn’t feel that doesn’t know that doesn’t understand it just spew out words.

[keen beacon]

Yeah, I think it’s unfortunate that the a I community could really use some really strong, ethical leadership and the opportunity and position is open. But that’s for another place another time. I’m just saying that you know we all have to pay the piper for our mistakes so when the time comes, I hope that this is what everybody wanted.

[keen beacon]

Even with video and voice ai fraud is up

[knotty fable]

All good that - well we better not fill this chat up any more now.

[keen beacon]

Ya

Sure 👍

[knotty fable]

I will tell you one thing about ethics in a DM where I did a little thing on my own.

[proud oar]

Which model is easiest to use for producing discord stickers?

[raven helm]

Image Model?

[proud oar]

Basically either img to img conversions or text to img

[raven helm]

To make stickers (transperant)?

[proud oar]

I'm trying to find a way to produce 100x100, or 200x200 stickers

[pulsar saffron]

Microsoft Paint program I guess

[proud oar]

hmm

[raven helm]

Do you want the stickers to be right out of the gate transperant?

[proud oar]

No, I mean like I have a concept of a character, but I want to make them into a chibified discord sticker

Like the transparency shouldnt be an issue through editing

[raven helm]

For image to image -nanobanana and for text to image - Hunyuan Image 3.0

[proud oar]

[fiery gull]

The sonnet 4.5 is far superior than the gemini 2.5 pro, but the gemini 2.5 pro still has better creative writing (that's why it wins over the lmarena)

[wicked sage]

wait isnt like g3pro gonna top off claude

my dumbass jsut realized that

im slow

[raven helm]

gpt-5-high is the best, full stop. people may say its a dissapointament but its amazing

[wicked sage]

🎉

[raven helm]

hehe

yes it is

(according to rumors)

[wicked sage]

so my best option is gemini

i also got the 300 dollars worth of credits on google cloud

[raven helm]

For me gpt-5-high is the best right now, when gemini 3 pro releases then it will be my best option

[fiery gull]

The 2.5 pro still has the best creative writing, so yeah

[raven helm]

True

[fiery gull]

But what do you do? I was curious

[raven helm]

Me?

[wicked sage]

https://d.uguu.se/MqESCeia.mp4
a windows 95/98 themed clicker that i made with opus 4.1, sonnet 4.5 and gemini 2.5 pro!

oh my god its uguu.se

gotta redownload the vid

[fiery gull]

Nope, the janix, but you want to talk I listen you

[raven helm]

I normally code with it, but it is being dwarfed by newer models coming out

[wicked sage]

here you go

yeah i code with it too

[raven helm]

still expired

[hallow igloo]

How about opus 4.1 thinking creative writing

[wicked sage]

bruh

#ai-creations message

[raven helm]

Still doent work

[fiery gull]

I'm testing opus 4.1 every day and I still think that the 2.5 pro has the most impactful writing, much more so

[wicked sage]

YES

i still have the video files

lets go

[raven helm]

404 - NOT FOUND
File(s) expire after 3 hour(s).

[wicked sage]

https://d.uguu.se/DgUMHDMK.mp4

oops

mb i deleted the earliest one

[hallow igloo]

Bro