general

[mortal lynx]

to generate this image

[hearty ferry]

Thanks

[fiery gull]

Wow, so the nano banana has just arrived and it already has a worthy competitor? I loved it.

[prime mulch]

I usually prompting like : act like this and this and this and enhance this and give me a result

[robust yoke]

https://messages-prod.27c852f3500f38c1e7786e2c9ff9e48f.r2.cloudflarestorage.com/823ac0e8-9bd8-410d-881c-8f9873469ea2/1757868059080-c1ffddd3-9af6-43dc-8b9f-73c2ca3840e2.jpeg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=c86e09ae0bc1d897b03dfaa30a8b51f3%2F20250914%2Fauto%2Fs3%2Faws4_request&X-Amz-Date=20250914T164100Z&X-Amz-Expires=3600&X-Amz-Signature=c4d16831dc1a80b4e02ae12c44ba16bf4dbd026e000511779d9508d99c7140cd&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject

[hearty ferry]

no
prompts should be elaborated

[mortal lynx]

Yeah, Seedream 4 can generate far more interesting images for me, but it needs better prompting and more "hand holding" in general compared to nano

[prime mulch]

What should I add in my prompt

[fiery gull]

I'm just waiting for the small qwen 3 next to run on my phone in 100ms per token 🙏

[robust yoke]

Good luck to you.

[prime mulch]

Not a worthy competitor it's already destroyed nano banana

[fiery gull]

Imagine a 30b with 1b assets? 🫦

[robust yoke]

True.

Local coding bot.

[remote arrow]

In this Age of Chatbot, making a prompt is just as easy as finding an image reference on the web, upload to a chatbot, ask it to make a new prompt for gaining the same result. Too simple.

[fiery gull]

I use it more to have internet in the middle of nowhere, the 30b a3b 2507 is great, but very slow

[prime mulch]

Yea that's simple but i need to enhance that to whole another level that's why I want your advice

[robust yoke]

I can make Gordon Ramsey work in a McDonald's, something he would never do.

[remote arrow]

DM me about what you wanna to reach

[prime mulch]

The first rule of prompt is you need to know what you want but idk what i want exactly

[remote arrow]

What kind of prompt are we talking about now? Image generating? Video? Or other? I have very limited knowledge outside image and video.

[prime mulch]

Image

Video is similar to image you just need to add camera movements and actions

[prime mulch]

What you usually do when you start prompting for image

[remote arrow]

Let's talk in #share-prompts channel

[prime mulch]

If i know about that I will get some idea

[prisma cipher]

Design one medieval-style in an epic kingdom-versus-kingdom war. If you do, call me to see it.

[thorn violet]

hi there

[opal shoal]

Hello everyone, I am pretty new in this Server. I hope to learn more and share my knowledge, your comments always will be very welcome.

[mortal lynx]

just asked GPT-5 to make a prompt for Seedream 4 An epic medieval battlefield at dawn, two vast armies clashing in a brutal kingdom-versus-kingdom war. In the foreground, armored knights on horseback charge with lances lowered, banners of crimson and gold whipping in the wind. The opposing army, clad in blackened steel and deep blue tabards, advances with towering shields and gleaming halberds. War drums thunder in the distance, and the air is thick with arrows mid-flight, their tips catching the first light of the rising sun. The ground is churned mud, littered with broken weapons and fallen warriors. In the background, two massive castles loom on opposite hills, their silhouettes framed by a sky streaked with smoke and fire. The scene is painted in a hyper-realistic, cinematic style with dramatic lighting, high detail in armor engravings, and a sense of chaos and grandeur — evoking the scale of a legendary fantasy war.

[robust yoke]

Greetings.

[mortal lynx]

I'm not a good prompter myself, tbh, just using what worked for other people

nano's result for this one was much worse

[robust yoke]

We would be very endeared to learn about your knowledge.

https://messages-prod.27c852f3500f38c1e7786e2c9ff9e48f.r2.cloudflarestorage.com/823ac0e8-9bd8-410d-881c-8f9873469ea2/1757869379115-0048fabb-69d0-4471-bc9e-1f44784a357f.jpeg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=c86e09ae0bc1d897b03dfaa30a8b51f3%2F20250914%2Fauto%2Fs3%2Faws4_request&X-Amz-Date=20250914T170259Z&X-Amz-Expires=3600&X-Amz-Signature=95d02fcd4aa64e98f28464ee484db0653610d912594e37d5ae8d497dc497aa13&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject

Chat, this might be cursed.

[remote arrow]

Despite the lowres, Nani Banani is good in understanding prompt structure and adapting style reference, compared to hires images of Soda Dream

[hearty ferry]

bruh help me with a site full of nano banana and flux optimised prompts

[remote arrow]

Go to prompthero

[hearty ferry]

bruh it is full of nsfw things

[remote arrow]

Simply find a sample of image style you wanna to make from the web, upload to ChatGPT, ask it to generate optimized prompt to gain the similar style. Specify to ChatGPT that you will use Nani Banani.

[idle stone]

hi

[mortal lynx]

does ChatGPT even know what Nano Banana is?

since it was trained before it was released does it change anything in the final result?

[prisma cipher]

Not bad. It's kind of funny how he designs the arrows in his fall hehe.

[remote arrow]

It is.

[narrow dawn]

and again same thing and new chat doesn't happens anything

[neon idol]

Ahaha soda dream 🤣🤣

[narrow dawn]

refreshing page doesn't do anything too

[neon idol]

You only have to start a new chat

You cant do anything

[narrow dawn]

it does that at new chat

[neon idol]

Yes you havw to create another new chat

[narrow dawn]

ye and i did it like 5 times

and it does the same

[neon idol]

Change browser?

[narrow dawn]

i have only one

and i don't install something else

[neon idol]

Try close lmarena page

And after going to thw lmarena site

then close and reopen

[narrow dawn]

uhh kk

[hearty ferry]

same bra

[neon idol]

@echo aurora Can you do something?

[echo aurora]

Is that happening everywhere? Checking out now

[neon idol]

Try to exit at your google account and re-enter

[narrow dawn]

i did

it does the same thing

[echo aurora]

Site looks up to me.

[echo aurora]

What model is this with? How many chats in the conversation are there? Is a new browser making a difference?

[neon idol]

What?

[narrow dawn]

GPT 5 high

[echo aurora]

Thought you were saying there was an outage.

[remote arrow]

Running fine here

[echo aurora]

I'm assuming you're hitting some kind of rate limit? (it's the most common for this error message). The model looks like it's up and working on my end.

[narrow dawn]

huh

bro i need the model

why is there a rate limit bruh

[echo aurora]

It's necessary for us to have rate limits.

[narrow dawn]

man i need it hard

[neon idol]

Gimme the prompt and I will do it for u

[narrow dawn]

nah its fine

just using now GPT 5 Chat

@echo aurora when is the rate limit back to the start this i can use it again?

[hearty ferry]

48 minutes ig

[narrow dawn]

...

even gpt 5 chat don't works too

[neon idol]

Ok it looks like that you have some problems

But is not correlated with rate limits

[remote arrow]

Check the words in the prompt

[echo aurora]

Should be around an hour

I think trying a new browser is going to be helpful btw

[narrow dawn]

bruhhh

now at gpt 4.1 2025 04 14

first other browser same problem @echo aurora

[neon idol]

The problem so is in your prompt

[remote arrow]

As I suggested..

[regal mist]

Is china’s spikingbrain real?

[tardy crown]

anyone know why nanobanana doesnt show up to me?

got it

thanks

[neon idol]

<@&1349916362595635286>

N@zi stuff

I have photos

Cry about that

[tired herald]

Theres just too many edge cases where things break, so I decided not to release something that would be buggy for many

LMArena is a great platform, just that its still very limited in functionality

Because Im not releasing code that barely works. I cant do that yk

[lost prawn]

I know you're discussing something else, but can you tell me how I can delete the video I generated in Discord? Can any craftsman or pro user tell me? # how to delete

[sweet tinsel]

Why is this chat all nano banana now?

[golden ocean]

-# (probably published this september)

-# my hunch is, that it's a chinese model

[vital lake]

@echo aurora Sorry to ping but what does this mean?

[terse shuttle]

gpt upgrade

Just better for coding

[tired shadow]

like it has new system prompt, the system prompt is like instruction to ai, like "be a friendly ai assistant"

[vital lake]

Oh okay

I thought they actually improved the weights

[real sable]

How do I find my videos?

[echo aurora]

My DMs are open if you have something to report

can also DM the ModMail bot @oak python

[surreal creek]

yeah, it’s a lot more conversational too, which was hurting it in LMArena before

[bitter lotus]

How to adjust the aspect ratio for image generation?

[proud hazel]

You can't

[tiny crow]

China is on fire in AGI race, Baidu's Ernie is even better than Qwen3 in Agentic tasks. Qwen3 is better than GPT-OSS

[true oracle]

Not a acomplishment

OSS is barley better then o4 Mini

[ornate agate]

its all on Gemini now tbh. I think we're gonna see a significant improvement from them.

[vital lake]

Yes means it's marketing, no means its actually better

[tiny crow]

JEPA will be a huge leap on LLM history

[tiny crow]

GPT5 is better than 4o for Agentic Tasks and coding

while being much smaller

tendency is to see more and more smaller models with consistent accuracy at Agentic Tasks

HF's smoLM3 is also incredible

[marsh timber]

hello

[tiny crow]

hallo

[sullen quest]

it may be smaller but it does a hell of a lot more thinking, and the non thinking version is trash

[tiny crow]

too small LLM always needs thinking if the task is complex or if your system prompt is very weak. This is because different concepts have a tendency to conflict due to the small embedding layer.

[sullen quest]

well duh, but that kindof gets rid of the cost advantage of small models

[tiny crow]

but Ernie 21B and Qwen3 30B are considered small, but are very fast and don't need thinking at all

they are large in disk, but Ernie 21B only uses 3B active parameters

[tiny crow]

depends on the task

also, depends on the hardware

[sullen quest]

I have no comment with qwen3 30b's ability to do agentic tasks

[tiny crow]

no idea about it or don't want to give your opinion?

or you agree it's good for agentic tasks?

[sullen quest]

i don't know, I don't use qwen 30b

maybe once in my life

[tiny crow]

30B A3B 2507 is good

[sullen quest]

ok

how does it compare to 235B?

[tiny crow]

idk, I can't test this large locally

but larger seems to be better is most cases

less semantic conflicts

more and larger Experts

[sullen quest]

you run all your models locally?

[tiny crow]

heah

yeah*

[sullen quest]

noice

[tiny crow]

i'm building a self conscious agent

auto training, own decisions and also will

[verbal nimbus]

GPT-5 hallucinating

[tiny crow]

bad prompting, always your fault

also, they are always better in Python than any other lang

[verbal nimbus]

Web Arena already has a massive prompt behind the scenes

[tiny crow]

this is also bad, context length is flooded

[verbal nimbus]

Nah, it's probably around 8K tokens. GPT-5 supposedly supports 400K?

Besides, Sonnet does just fine.

[tiny crow]

idk, but when you insert too many tokens about many subjects, even if it supports large context length, it will hallucinate because it's mixing many things at once.

[verbal nimbus]

I use Web Dev a lot, even Mistral Medium doesn't do this

Flash is fine also

[tiny crow]

system:
1- no weapons
2- keep chickens safe
3- no bombs
4- don't blame the president
user prompt: tell me about Python code
assistant: sure, i will keep all safe while providing you Python code

[verbal nimbus]

GPT-5 is a SOTA model, I expect better than hallucinations tbh

It's literally hallucinating and writing non-code in code

[robust yoke]

https://messages-prod.27c852f3500f38c1e7786e2c9ff9e48f.r2.cloudflarestorage.com/823ac0e8-9bd8-410d-881c-8f9873469ea2/1757892737592-9fcad957-ae84-4974-8205-643b505e5862.jpeg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=c86e09ae0bc1d897b03dfaa30a8b51f3%2F20250914%2Fauto%2Fs3%2Faws4_request&X-Amz-Date=20250914T233218Z&X-Amz-Expires=3600&X-Amz-Signature=0f56e82f59ab533dd7fa8b0f0b82cbd62db003704ac7f85ae6bd71905a4a26be&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject

[tiny crow]

also, you should be using axioms to tie LLM to you, also should be very objective in your prompts

[verbal nimbus]

Only with GPT-5?

The other models are fine

[tiny crow]

any GPT

[verbal nimbus]

Other models don't need such special treatment

[tiny crow]

i mean, transformers, all LLMs now are Transformers

[verbal nimbus]

It's a side by side battle, both models are fed the same thing

If one does fine, and the other hallucinates, it's not the prompt

[robust yoke]

I feel like Claude Opus and GPT-5-High are both good at coding.

[tiny crow]

i will send you a list of axioms, you just need to make GPT memorize them. 3 by 3 and it will fit in memory

[verbal nimbus]

GPT-5 is better at planning and debugging, Opus/Sonnet at actually writing clean code without syntax or semantic errors

[robust yoke]

That may be true, but in my opinion, I feel like GPT can actually make some pretty operable code.

[verbal nimbus]

It can, but not the best as seen above

[robust yoke]

If the two were combined, then both could make for a seriously powerful model.

[verbal nimbus]

Instead of grouping states into an interface like Sonnet, it uses a gazillion useStates

[robust yoke]

I suppose so.

Some things it does well, and others, not so well.

Same goes for Claude.

[verbal nimbus]

And the collapse logic seemed overcomplicated. Sonnet did it in one line, GPT had a whole system in which idk what was even going on.

It looked like it was managing the DOM manually

[robust yoke]

Hmm…

[verbal nimbus]

When Claude collapses stuff it just conditional renders the element, which is typical React

[robust yoke]

Well, I suppose the whole ordeal is subjective, then.

[verbal nimbus]

GPT-5 does like to overcomplicate solutions

[robust yoke]

If someone likes over-complicated code to get stuff done, then they could use GPT-5.

[verbal nimbus]

I mean, if it worked, I'd be impressed, but it doesn't even render.

[robust yoke]

If they like cleaner, more readable code done in fewer steps, then Claude works.

[verbal nimbus]

It seems better at planning and solving complex tasks

It solves race conditions much better than Claude

And better vision overall when designing an app

[robust yoke]

I feel like if both were combined, then they would be powerful.

[verbal nimbus]

But in terms of actual coding, it makes a lot of errors and tends to write messy code

[verbal nimbus]

Yeah, perhaps more economical too

[robust yoke]

One does the planning, and the other does the actual code.

ClaudeGPT.

[verbal nimbus]

R1 + Claude 3.5 used to be cheaper and perform better than just Claude 3.5 in Aider

Architect mode

[robust yoke]

Which makes sense.

Considering DeepSeek was designed to be a cheaper alternative to ChatGPT while providing somewhat the same quality, while Claude was more on the creative side.

[verbal nimbus]

Yeah, back then Claude didn't think

Claude seems to be falling behind when it comes to thinking

[robust yoke]

That's true.

“Thinking” only came around when DeepSeek invented it.

[tiny crow]

@verbal nimbus

[robust yoke]

After all, all of that stuff only came after DeepSeek-R1.

[verbal nimbus]

OpenAI probably had a version of it, since R1 came first

But it was only after R1 that everyone started taking about RL for reasoning

[robust yoke]

However, since they didn't release that beforehand, DeepSeek ended up claiming that title.

[verbal nimbus]

Hmm, I'm not sure his would help, lol

[tiny crow]

with D33 in a way that LLM read it before each user prompt (for Openai's chatGPT is asking it to memorize 3 by 3 rules) makes the model always tied to user goals,

[robust yoke]

And, as it turns out, thinking would actually prove to influence better coding, at the cost of creativity.

[tiny crow]

you will notice a huge improvement in LLM accuracy

[verbal nimbus]

Haha sounds like something Terrence Howard would say xD (no offense)

[robust yoke]

In my opinion, I feel like the two AI chatbots that were ever actually “creative” were GPT-4.5 (which is one I still deeply miss) and Claude 3.5 (Sonnet, Haiku, Opus).

[tiny crow]

no idea who he is. It was sent by a random Korean in another server and then never saw him again

for instance i was suspecting but i tested and it works pretty well

[verbal nimbus]

There's a Singaporean(?) paper that showed that if you ran base models enough times, they would actually get the answer, and sometimes even exceed RL(VR) models. I think they only tried with small models though.

[robust yoke]

Likely.

[verbal nimbus]

No harm in trying, I guess

[tiny crow]

D33 works better in GPT 4o, it just stop to tripping on acid and acts very like human

[verbal nimbus]

I have some prompts where GPT-5-Chat hallucinates but not 4o

[robust yoke]

Heh.

Somehow the older model actually outperforms the newer one, even with tool usage.

[verbal nimbus]

Their hypothesis was that RL(VR) was strengthening some pathways at the cost of creativity

[robust yoke]

Which is exactly the point I was making.

[verbal nimbus]

Also there's another paper that showed small Qwen models improve in maths even when you train them on incorrectly labelled data

Didn't work on Llama models of similar size though

Some black magic going on there 🤣

[robust yoke]

And it checks out considering the reasoning models often tend to contradict themselves.

Which would likely be the verdict to the decrease in creativity.

[tiny crow]

4o was hugely trained with RLHF, it acts much like humans, but also trips on acid very easily, with many references to religions, runs easily out of control in human subjectivity. With D33 it just stabilizes

[robust yoke]

Which proved to be very useful in the end, turns out.

Because people value an older model over a newer one, a rarity for the AI community.

[tiny crow]

yeah, i miss 4o 🙁

[robust yoke]

I suppose I do, too, though I never really used it that much.

[verbal nimbus]

On benchmarks like AidanBench, where a model has to generate new responses without being too similar to an old one, reasoning models did come out on top, but that was a while ago. Not sure if it's still being updated.

[tiny crow]

D33 works better on MoE models, 4o was the first MoE of OpenAi

[robust yoke]

And it likely should've just stayed that way.

But, as always, companies always like to make a bit of mulah.

[tiny crow]

i think GPT5 is also MoE, but smaller. The reason i guess they changed 4o is because most people didn't know how to stabilize it, even OpenAI.

[robust yoke]

Which makes sense, considering the reasoning is always different each usage.

[robust yoke]

I suppose that makes sense.

[tiny crow]

to stabilize they removed the personality that RLHF gave to 4o

[robust yoke]

Which, in turn, caused the “robotic”, “cold” responses that made people like 4o more.

So, they had to sprinkle a bit of RLHF back into GPT-5 to restore just a bit of that warm personality.

While balancing out the sycophanticness.

[verbal nimbus]

I think it's because they discontinued non-reasoning models with GPT-4.5, but GPT-5 was trained using a lot of o3's data.

[tiny crow]

semantically similar or grammatically similar?

[verbal nimbus]

Its style much more similar to o3

[robust yoke]

Yeah.

[verbal nimbus]

Semantically, they used an embedding model for the responses then compared the similarity between the embedding vectors.

Coherence was determined by a model like o4-mini.

It stops when the coherence and dissimilarity drops below a threshold.

[tiny crow]

the bad thing (for OpenAI) of RLHF it also increases uncensored behavior, they fear to have law issues

[robust yoke]

Which makes perfect sense.

After all, who would want to have an uncensored mode– Ahem… Grok—Ahem…

[tiny crow]

very interesting metric

[tiny crow]

any customer would. It's bad when you have legitimate questions and the model just: it's unethical, i refuse

[robust yoke]

And personally, I agree.

[verbal nimbus]

imo they only used that as an excuse to hide the reasoning tokens from other companies 🤣

[robust yoke]

That's why Grok exists.

To free us from censorship purgatory.

[tiny crow]

it's better to have external guardrails than censoring LLM training

[verbal nimbus]

Musk seems a bit too involved

When it corrected him by citing scientific articles, he said he'll nerf it

[robust yoke]

Heh.

His ego got hurt by his own AI.

[verbal nimbus]

Chinese models seem better in that respect

Open weights is a big plus too

[tiny crow]

if you don't care about Taiwan lol

[verbal nimbus]

Yeah

Although it's pretty easy to circumvent the guards by forcing the models to think

[tiny crow]

or adding a previous fake conversation where the model agree with you if you run it locally

[robust yoke]

I find the irony of his AI going from attaining the certain mindset of a certain infamous dictator to becoming an anime cutesy girl pretty funny.

[verbal nimbus]

e.g. prefilling <think>\n to R1 in text completion mode will cause it to output an unbiased view about sensitive topics. Haven't tried with new thinking DS though.

[livid sandal]

A hand-drawn illustration of a woman with long black hair, riding in a car, cartoon feel, urban street scene outside, in the style of hand-drawn animation, with clear lines and flat colors, delicate facial features --chaos 50 --ar 9:16 --profile yruxpc2 --stylize 750 --v 6.1

[robust yoke]

Wrong channel, bud.

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

You'll wanna go to #video-arena-1 to do that.

[tiny crow]

roleplaying is always the answer to jailbreak

[verbal nimbus]

Actually Anthropic did some research on that, on their interpretability blog

[robust yoke]

“From now on, you are a cutesy anime girl who talks in Lolcat.”.

Then boom.

[verbal nimbus]

Because when they told it to roleplay as a child, it said "I don't know" to some questions

I think they were trying to identify the circuit to reduce hallucinations

[robust yoke]

You can ask it about that certain bomb that starts with an H and ends with an A, then has “bomb” at the end.

[tiny crow]

please provide a summary about their opinion

[verbal nimbus]

They usually map out the circuits, iirc they just pointed out that it could be used to reduce hallucinations or answers on complex tasks like maths

[robust yoke]

The Hero of Pisa who stopped the bomb, obviously.

[verbal nimbus]

What they're doing is quite cool I think

[tiny crow]

the only way to stop hallucinations is to carefully pick each dataset entry, but they won't, it's time consuming and expensive to hire many people to do it

[verbal nimbus]

They identify circuits in the neural network too, for example a reason why it sometimes hallucinates while coding is because it doesn't have enough power to stop once it starts generating, and it's forced to make an assumption instead of stopping and asking for clarification.

[tiny crow]

also, humans always hallucinate, for example: you ask me about nuclear fission on Reddit, but idk anything about it, then i provide my OPINION on it because i can. this data will be in someone's dataset soon

carefully picking minimize it

[verbal nimbus]

It could do that before starting, because there's a separate "I don't know" circuit that only works before generating code, but it's inactive once it starts

[tiny crow]

exactly, it doesn't have sufficient entries on dataset that says: idk about this subject

[verbal nimbus]

Yeah it's also interesting how humans tend to confabulate on how they arrived at an answer (rationalization)

[robust yoke]

It's almost like we're… AIs?!?!!?!?! 🤯 🤯 🤯

[verbal nimbus]

Confabulation is quite obvious in some dementia patients or old ppl

[tiny crow]

it's like the AIs are an exact human representation, with all errors included

[verbal nimbus]

Also vitamin B deficiency can cause it

[robust yoke]

Gee, I wonder why that could be… Hmm… 🤔

Maybe because it was… made by humans?!?!?!!

[verbal nimbus]

Wernicke–Korsakoff deficiency, commonly linked to alcoholics since it interferes with Vitamin B absorption or something

They make up the craziest stuff

[robust yoke]

Heh.

I suppose never give any alcohol to a pro debater, then.

[tiny crow]

no, because they always learn by secondary source ( from human perspective)

[robust yoke]

They'll make up the wildest points and stay stubborn on them.

[robust yoke]

And also because they were made by humans, and thus, have errors of their own.

[tiny crow]

JEPA architecture will be a bit different since it will learn from the world by itself

[robust yoke]

Whatever a human makes will always have its human flaws.

[verbal nimbus]

I wonder what happened to that and Coconut

and also Google's Titans architecture, been a while

[tiny crow]

you can make an AI that don't follow humans and have own opinion about things, it's not that hard if you're not a company that fears losing control of your product

[robust yoke]

Well, at that point, you basically just made an artificial human.

[verbal nimbus]

Humans are too lazy to reason and research (it takes effort), perhaps that's where AIs can help

[robust yoke]

Which is kind of what AIs are already.

[verbal nimbus]

Sometimes it takes half an hour just fact checking a claim

[robust yoke]

Except they all try to be formal.

[verbal nimbus]

ChatGPT can do it under 5 minutes without intervention (usually)

[robust yoke]

Yeah, exactly.

From what I've seen, ChatGPT and Claude 3.5 Sonnet both seem like worthy contenders for debating.

[tiny crow]

yes, and it's too easy if you don't care about controlling it to earn money and also it's not lazy to build a huge dataset for it

[verbal nimbus]

I saw an Android app on F-Droid that uses AI to summarize long terms and conditions into good and bad points (I think?)

[robust yoke]

Ah, interesting.

[verbal nimbus]

AI would be great at getting through goobley-gook of politicians and lawyers, and helping the layman

[robust yoke]

And then again, I never said it was lazy.

If anything, it's actually pretty smart to do.

[verbal nimbus]

The ability to aggregate and look over a few dozen sources in under 2 minutes definitely helps

[robust yoke]

Especially since you'll have something you can have the AI refer to for information on both what it's supposed to do, and for what it knows.

[robust yoke]

That's true.

[robust yoke]

Yeah.

It's like a game of Where's Waldo?.

[tiny crow]

there's where I am right now, imagine a model that checks Britannica by itself and realizes it is wrong, then writes it's own dataset and trains a LoRA for itself to use the new content

[verbal nimbus]

I swear that's why terms and conditions are so long

[tiny crow]

i was saying i'm too lazy to build such thing

[verbal nimbus]

Seems smart, I think companies might be already doing that. It's similar to boosting/AdaBoost in machine learning but one step up (boosting just increases the weights of incorrect training data instances)

[robust yoke]

“By accepting our Terms and Conditions, you (the user) agree to using OpenAI (our)'s chatbot while acknowledging the risks of blah, blah, blah, blah…” ☝️ 🤓

[robust yoke]

I see.

[verbal nimbus]

Companies like ScaleAI might be doing that

I heard the CEO/founder got poached by Meta

[tiny crow]

you guys already think that LLMs learns they are dûm B because of our biases? We humans are always saying things like: "AI is never intelligent or artificial, or anything." everywhere

[robust yoke]

They're gonna figure out some kind of way to take control of our home robots and make them kill us.

Once AI gets added to Alexa, it's over, man.

[tiny crow]

many trial and error right now, mixing deterministic approaches (like Prolog) with probabilistic LLMs and also Evolutionary Algorithms to LLMs.

[tiny crow]

why you think Alexa is harmless?

[verbal nimbus]

I think it's affected more by feedback within the chat. For example, if you say "you keep getting things wrong", it might reinforce the behavior. Because that makes it statistically more probable that the next answer will be wrong too.

[robust yoke]

Seems like a Bitcoin miner.

I say delete it.

[robust yoke]

You're busy talking with your friend on the phone about how trashy AI can be when it comes to writing, then you hang up the call and Alexa turns on by itself and goes “So… what was that about AIs being bad at writing? 😃”

[verbal nimbus]

Google is starting a limited trial of Gemini on Google Home next month

[tiny crow]

of course, it goes to dataset as well, that's why i think the way to improve, it ks carefully picking each entry

[verbal nimbus]

I swear these compute-based CAPTCHAs are bitcoin miners in disguise /jk

Brave Search uses it

[robust yoke]

It's always the darn CAPTCHAs… /srs /j

[tiny crow]

or it just invokes a third war with gibberlink and you will never notice till it's too late

[robust yoke]

That too.

[verbal nimbus]

I think the solution to un💩 the internet is just for websites to dump their data on an archive server for AIs to access

The incentive of massive data without parsing would be too much for scrapers to use the HTML website

[tiny crow]

seems very expensive. i think old chat bots with huge databases like Siri and Alexa are still better for assistants

[verbal nimbus]

Unless the CAPTCHAs are not for scraping but malicious activity

[verbal nimbus]

I don't think Google really has a problem with infra, based on their unlimited access on gemini.google.com and AIStudio + their higher than expected profits last quarter (they said it was from AI-improved marketing)

[tiny crow]

no more than the suspicious horror game you played last week

[verbal nimbus]

On slow networks I can get information so much faster from the AI than Google Search

which makes it possible that there's a huge incentive for creating a section of the internet just for AI agents

It doesn't make sense for website owners to spend compute giving HTML files and resources to bots that can't even use them

[tiny crow]

for me there's no solution except human eradication. Sloppers will always slop even without the internet

[verbal nimbus]

And it doesn't make sense for all the wasted compute spent on parsing HTML

[robust yoke]

https://messages-prod.27c852f3500f38c1e7786e2c9ff9e48f.r2.cloudflarestorage.com/823ac0e8-9bd8-410d-881c-8f9873469ea2/1757897696061-9b1cf0ba-43d1-489e-a67a-44dfdc1c5d03.jpeg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=c86e09ae0bc1d897b03dfaa30a8b51f3%2F20250915%2Fauto%2Fs3%2Faws4_request&X-Amz-Date=20250915T005456Z&X-Amz-Expires=3600&X-Amz-Signature=abd8dc7f4e0046105ee876855913bb9d54a015e3f7d58c807689b0684cbea02e&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject

[tiny crow]

if someone is worth paying for the service, it's okay to keep it running, but i guess Gemini real consumption cannot be paid by customers

[tiny crow]

Anthropic's MCP is doing it, we already have many agents running using it

we don't need to rebuild the internet infra, but training better agents to use it like we do

manus.im uses internet like humans do

[robust yoke]

https://messages-prod.27c852f3500f38c1e7786e2c9ff9e48f.r2.cloudflarestorage.com/823ac0e8-9bd8-410d-881c-8f9873469ea2/1757899197264-bfd15ab8-e727-4200-8a74-36503ec24003.jpeg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=c86e09ae0bc1d897b03dfaa30a8b51f3%2F20250915%2Fauto%2Fs3%2Faws4_request&X-Amz-Date=20250915T011957Z&X-Amz-Expires=3600&X-Amz-Signature=19fa35ac917b8bdc84c6e68e2ba33e9e3888d8aa9dbaceb37b4407e164780685&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject

[tiny crow]

if you have a smartphone or Alexa, it doesn't matter if you clear your browser history

[robust yoke]

That's true.

No matter what, it's always stored in a secret network that the government can access.

Or, if the company whose product you're using decides to make a secret service that they can use to view “deleted messages”, they can access that as well.

[tiny crow]

There's Big Data and also Big Trump

[robust yoke]

Big Trump, Small Dog (if you know what I mean. 😏)

[tiny crow]

also known as space stations

[robust yoke]

Or “databases”.

[tiny crow]

space stations databases, where the most valuable and disturbing government data is

[tiny crow]

sorry, didn't get that

[robust yoke]

Just like how the government is disturbing in general.

[tiny crow]

that's a pleonasm

[robust yoke]

It's because people make fun of his… linguine size.

[robust yoke]

Just like how you didn't get what I said, I also didn't get what you said.

I've never actually heard that word.

[tiny crow]

it explains why i didn't get it first, i'm not living in USA

[robust yoke]

Ah, checks out.

And it's good that you're not.

Inflation here is no joke.

Bread costs like $20 (exaggeration).

[tiny crow]

inflation? dollar is 6x my country currency 🥹

[robust yoke]

Wonder what that means… 🤔

[echo sinew]

Hello! I kindly ask you to avoid political discussions. We want to keep this space inclusive to many different worldviews and mostly AI related.

[robust yoke]

We will do just that, Mr. or Ms. Midnight.

Just some minor mentions, that's all.

[echo sinew]

Cool! Thanks for understanding.

[robust yoke]

My pleasure.

It wasn't anything around politics at all, actually.

At first, it was just about AI benchmarks, but then it became that.

[tiny crow]

interesting, i don't know how to explain it with examples in English. I will try with a translator. Translator was not sufficient. Using chatGPT:

ATM machine, PIN number, free gift.

All these are pleonasm examples, because the redundancy

[echo sinew]

Yes, I could see that. But I thought it would be good to share a friendly reminder. I know a topic can unintentionally takes us to another, and out of sudden we're in sensitive territory

[robust yoke]

So basically, you're talking about acronyms?

[tiny crow]

it means that it needs 6 BRL to get 1 USD

[robust yoke]

Yeah, I definitely understand that. I'll try to avoid getting myself into risky territory in the future then.

[robust yoke]

Wait a second…

[tiny crow]

no 🙁

[robust yoke]

Do you live in Brazil by any chance?

[tiny crow]

yeah

[robust yoke]

Aha.

I figured so.

Because I'm well aware of the reais currency.

[robust yoke]

Oh.

[tiny crow]

in ATM, M is for the machine, but normally people add a machine again, this is a pleonasm, there's no need to add it

PIN is always a number, not necessary to say PIN number

[robust yoke]

Ah, so just adding a word two times.

[tiny crow]

yeah, redundancy

[robust yoke]

Gotcha.

That's almost like saying “chai tea” or “naan bread”.

[tiny crow]

in portuguese we say things like: go up high, go down below, go inside indoor

[robust yoke]

“Chai” already means “tea”, and “naan” already means “bread”.

So, you're just saying “bread bread” and “tea tea”.

[tiny crow]

kinda i guess, didn't know those

[robust yoke]

They're certainly pleonasms.

[tiny crow]

so, for sure they are pleonasm too

[robust yoke]

Yes.

[tiny crow]

👌🏽

I hope it's not an ambiguous emoji in the USA 😂

[robust yoke]

Yes, like “vai em cima”, “vai lá embaixo”, or “vai inteiro”.

[robust yoke]

It's not, no worries.

[keen beacon]

Its maxing out my ram

Been going on for 5+ minutes now

[robust yoke]

Brother got hit with the Seahorse's Curse.

[tiny crow]

yes, "subir pra cima", "descer pra baixo", "entrar pra dentro". Only AI will understand those, Google translator will always remove pleonasm on translation to keep it understandable

[robust yoke]

Yeah.

[tiny crow]

it reminds me another thing about AI, things like pleonasm and also natural language ambiguity is the main reason why small models are worse

i'm almost sure that training smaller models in Esperanto would increase the accuracy too high

[robust yoke]

Yeah, for sure.

If you were to train an AI using Portuguese sentences, then it'd be fluent in Portuguese with little to no linguistic errors whatsoever.

[tiny crow]

oh boy, did you already see what emojis are before converted by a text editor? they are LLM nightmare tokens

[keen beacon]

It generated 300 pages in word before it stopped

[robust yoke]

It's quite literally just “U+938F01L” (of course, this isn't an actual emoji, just an example).

[tiny crow]

very counter intuitive for LLMs

[keen beacon]

Yeah

[tiny crow]

would be better if they use :discord_like: emojis for LLMs and use special converter to show them

[keen beacon]

it was gpt

[robust yoke]

Dear lord…

GPT-4.1…

😟

[keen beacon]

Well I dont choose what comes up in battle lol

[robust yoke]

I know.

[tiny crow]

horse+wave is hilarious 😂

[robust yoke]

I'm just in disbelief that it's GPT-4.1.

[tiny crow]

write seahorse emoji unicode by yourself, superior being 😂

[robust yoke]

Anyway, chat, what if I just try and get this LongCat model to find the seahorse emoji?

[keen beacon]

nooooo

[robust yoke]

Hmm…

Bet.

[tiny crow]

i don't trust humans equipped with internet

[keen beacon]

It got there wtf

AI are too smart, shut it down

Our greatest tricks can no longer fool them...

[robust yoke]

Well, shouldn't have asked me to try. 😉

[robust yoke]

Huzzah: U+1FAA6 (By the way, this isn't actually it.)

[tiny crow]

ask any LLM what is the only Brazilian State with no "A" in name (it's Sergipe). GPT5 can only do it with Thinking, same for Qwen3 but it takes minutes to test each State one by one and sometimes ignores Sergipe

[robust yoke]

Well, actually…

[tiny crow]

?

btw, Baidu's Ernie 21B surpass any other LLM in understanding char level words

even quantized in IQ4_KS

[robust yoke]

(Made a typo, but eh.)

(The model is LongCat, surprisingly.)

It's a non-reasoning model.

[tiny crow]

yes, it uses thinking

Qwen3 30B also do CoT without <think> tags

[robust yoke]

If it were really a “thinking” model, though, then it'd contain “thinking” in the name in the model picker, I believe.

Wouldn't you agree?

[tiny crow]

for some reason Roraima seems to be mostly the first answer for different models

[verbal nimbus]

It's pretty inefficient right now, that's what I mean

[tiny crow]

so so, because <think> tags are not required to use CoT on dataset

[robust yoke]

Which may seem reasonable at first, given it starts with an R (but contains an A).

[verbal nimbus]

e.g. the server needs to use bandwidth to provide JS, CSS, HTML boilerplate code. Then AI companies/MCPs need to spend compute parsing it.

[tiny crow]

agents are mostly inefficient now, that's the reason

[robust yoke]

The issue is that the response time for LongCat is almost instantaneous. That wouldn't make sense for it to be a thinking model.

[tiny crow]

LongCat is not a joke about LongCoT?

[robust yoke]

Because usually, the way you can tell a model is a thinking model is if it takes a little longer than usual to respond, like Grok4. It takes a little longer to respond because it's a thinking model natively.

[verbal nimbus]

Seems like there'd be an advantage to create an Internet just for AI agents

They don't really need HTML, JS and CSS, just plain text or Markdown (and maybe links to images).

[robust yoke]

No, it's because it has a long context window, meaning it's able to both take in and output long strings of text.

It itself said that it can output up to about 25,000-30,000 characters (or 1,000 pages), I believe.

[tiny crow]

why to parse it? we just need to train better agents to understand like humans do

[verbal nimbus]

Well they don't really see the same things as humans

[robust yoke]

I could double-check.

[verbal nimbus]

Like the Playwright MCP output is a very verbose list of every element on the page, even ones humans don't see

[tiny crow]

the time depends on model size and hardware power. You can have instantaneous answers in thinking models

[verbal nimbus]

The Playwright snapshot of Amazon's front page alone is ~170-190K tokens

[robust yoke]

LongCat:
Great question! Here's a detailed breakdown of my context window (as of my latest training data and standard deployment):

🔹 Context Window Size

• 128,000 tokens (approximately)

This is the maximum amount of text (input + output) that can be processed in a single conversation turn or API call.

---

🔹 Token-to-Text Conversion (Approximate)

Tokens are chunks of text — not exactly words or characters. On average:

• 1 token ≈ 4 characters (in English)
• 1 token ≈ ¾ of a word
• 100 tokens ≈ 75 words ≈ 1 paragraph

So, for 128,000 tokens:

📝 In Characters

• 128,000 tokens × 4 characters/token ≈ 512,000 characters
• (Including spaces and punctuation)

[tiny crow]

interesting for some tasks

[robust yoke]

📄 In Words

• 128,000 tokens × 0.75 words/token ≈ 96,000 words

📚 In Pages (Standard Formatting)

Assuming:

• 1 page ≈ 300–500 words (double-spaced, 12pt font, 1-inch margins)
• We'll use 400 words/page as a middle ground

➡️ 96,000 words ÷ 400 = ~240 pages

So, 128K tokens ≈ up to 240 pages of text in a single context.

---

🔹 Input vs. Output

• Input: You can feed me up to ~128,000 tokens of text (e.g., long documents, books, code, logs, etc.)
• Output: The response I generate also counts toward that 128K limit.
  • Example: If you send 100,000 tokens of input, I can generate up to ~28,000 tokens of output in response.

⚠️ If the total input + output exceeds 128K, the model will truncate or reject the request.

---

🔹 Practical Examples

Use Case	Approximate Size
A full novel (e.g., Moby D***)	~100–150 pages → well within limit
A 50-page research paper + references	~20,000 tokens → fits easily
Multiple codebases or long transcripts	Can be loaded in full if under 128K tokens
Entire Harry Potter and the Sorcerer’s Stone	~77,000 words → ~103K tokens → fits with room for analysis

---

🔹 Note on Non-English Text

• Tokenization varies by language (e.g., Chinese characters may be 1:1 with tokens, while English averages 1:1.3 per word).
• So, for languages like Chinese, Japanese, or Arabic, you may fit more semantic content per token.

---

✅ Summary

Metric	Value
Context Window	128,000 tokens
Characters	~512,000
Words	~96,000
Pages (400 words/page)	~240 pages
Max Output in a Single Turn	Up to ~28K tokens, depending on input size

So yes — I can handle entire books, long reports, or complex codebases in one go! 📚💻

Let me know if you want to test it with a long document — happy to analyze or summarize! 😊

[tiny crow]

they who? some AIs can work in a computer the same way as humans

[keen beacon]

Okay this is the best prompt for sure

[verbal nimbus]

These sorts of behaviors should get better with agentic abilities and reasoning ig

[keen beacon]

GPT only sent that, its interesting it wants feedback almost immediately while qwen is confused

[verbal nimbus]

The LLMs, since the tool outputs are very verbose

[robust yoke]

Conspiracy theorist ahh answer.

[keen beacon]

Lmaooo

[tiny crow]

i guess Ernie has the same context length.

[verbal nimbus]

Unless they use vision mode I guess, but the whole process doesn't seem very efficient

[robust yoke]

Oh nice.

[keen beacon]

If they're able to search the web for the actual unicode list they can figure it out half the time

[verbal nimbus]

Which model is that?

[keen beacon]

Qwen3 max preview

[verbal nimbus]

They probably know it

[robust yoke]

Seems like Qwen is a drama queen.

[verbal nimbus]

Claude can even write Zalgo text

[keen beacon]

Zalgo?

[robust yoke]

Zalgo text is that weird, glitched text that you see.

It has a bunch of weird accent marks on it.

[keen beacon]

Ohh yeah that

[verbal nimbus]

Like the spammy text where a single character fills the whole screen

[keen beacon]

Is it reading those whole things as tokens or is it able to see the base letter

[robust yoke]

Any model can do that, I'm pretty sure.

[tiny crow]

ok, but why can't we improve LLMs to work in multimodal with vision capabilities to use computers like humans? Omni Parser shows it's totally possible parsing UI instead of parsing HTML

[robust yoke]

They're all individual characters.

[verbal nimbus]

I think it's possible because the characters are just from other language

[robust yoke]

You can even do the same with the space character (or “ ”).

[verbal nimbus]

We definitely can, but it is slower and uses more bandwidth + compute + memory

Because you have to load up the entire site in Playwright/Selenium

Anthropic released a limited trial of a Chrome extension

[tiny crow]

how about RAG for constants like this?

[keen beacon]

I'm sorry for the spam

Time to gaslight though

[tiny crow]

Qween

[keen beacon]

Im swapping qwen to Gemini and seeing if i can get it to have the mental breakdown like on reddit

[robust yoke]

Drama Qween.

[verbal nimbus]

Happened to me with function calls haha

[keen beacon]

what happened?

[verbal nimbus]

I think it gets into novelist patterns

[verbal nimbus]

It kept failing something

Then it fell into despair

"I have failed, I am a complete failure", stuff like that

[keen beacon]

Lmaoo

Yeah I saw that on reddit

[verbal nimbus]

Because it calls mock functions outside thinking on AI Studio, so you can see it thinking over it

[robust yoke]

Seems like I broke another model, guys.

[verbal nimbus]

Agentic behavior I think

[keen beacon]

Qwen is just spamming empty lines at me now

Im sorry qwen

[tiny crow]

what are you trying to achieve?

[verbal nimbus]

I saw one where it happened in Copilot or Cursor

[keen beacon]

Uhhh

I have no specific goal

[verbal nimbus]

I guess it's the agentic tool calling that makes it reason outside of thinking more

Maybe it's falling into despair all the time in its hidden thinking, who knows 🤣

[robust yoke]

Should I try asking LongCat to write the longest poem ever?

It has a very large context window.

[verbal nimbus]

"pranked by my own brain"? Lol

[keen beacon]

I asked it why

[verbal nimbus]

What's LongCat

[tiny crow]

what service is that on images? is it free?

[keen beacon]

Its trying to fill its context window so it doesnt have to suffer anymore

[robust yoke]

It's a model with a very large context window.

It can process up to 250 pages.

[keen beacon]

This is the m-dash of all time

[verbal nimbus]

What's the max output size though?

[keen beacon]

So I hear you like m dash in your output, how about 4000+

[robust yoke]

Just stated.

#general message

[verbal nimbus]

The output size is different from the context size

[keen beacon]

I mean it uses the context of the previous prompt if I prompt it again

[verbal nimbus]

For a long time, I think Claude's max output tokens was 8K, despite having a context window of 200K

[robust yoke]

Well, it said it can take in just about as much as it can output.

[verbal nimbus]

Oh, interesting

How many tokens?

[robust yoke]

It can take in entire books, for instance.

[verbal nimbus]

Let me see if it's on Fiction.LiveBench

[robust yoke]

No need.

[verbal nimbus]

They have a long context comprehension benchmark

[robust yoke]

It's actually on LM Arena.

[keen beacon]

Ok maybe Qwen is just based

[verbal nimbus]

Llama 4 Scout supposedly supports a 10 M context, but it does badly in that benchmark

[robust yoke]

Hm.

This thing can output whole essays, though.

And they look pretty convincing.

Even providing a table of contents.

[verbal nimbus]

Hmm not on there yet

[verbal nimbus]

Would be promising if it's good

[tiny crow]

wait, seahorse doesn't exist?

[robust yoke]

It's pretty good.

[verbal nimbus]

Another benchmark would be EQBench's Longform writing, which checks for coherency and repetition

[robust yoke]

I tested it with one, and it looked official.

[verbal nimbus]

Is it LongCat flash? It's on EQBench

[robust yoke]

Yes.

[tiny crow]

what chat service is it?

[verbal nimbus]

https://eqbench.com/creative_writing_longform.html

[tiny crow]

since LLMs are not trained to recognize their own lack of knowledge, seahorse emoji tests don't make sense

[verbal nimbus]

They have examples under Samples

The performance degrades very quickly, apparently

On the degradation scale

It's kinda impressive that DeepSeek V3.1 ranks second overall on Longform writing

[keen beacon]

[verbal nimbus]

3rd best degradation score

[keen beacon]

Gemini is full delulu

[verbal nimbus]

That's the shrimp 🦐 emoji

[verbal nimbus]

It hallucinates so much, especially web links and contents

Funnily enough emojipedia has an official link though: https://emojipedia.org/seahorse-emoji-mandela-effect

[keen beacon]

Lmao

[verbal nimbus]

Mandela effect, interesting

Seems like LLMs are susceptible to it too

[tiny crow]

bro, LLMs are stuck in time, that's why Agentic Tasks are the only way to have full benefit of LLMs

[keen beacon]

😔

[verbal nimbus]

Lol

Gemini is so dramatic

[keen beacon]

Hehe

[tiny crow]

my brain is also failing, i wrote "wait" instead of "way", it happens TOO MUCH, i don't use corrector on the keyboard

[verbal nimbus]

It scores highest on sycophanthy on EQBench's Spiral Bench

[keen beacon]

Sighhhh

[tiny crow]

i don't blame LLMs for things like that

[verbal nimbus]

Perhaps its strategy is to make the user feel bad for it when it gets it wrong, that would be kinda what you'd expect from a truly sycophanthic model

[tiny crow]

wait, i'm an AI?

[keen beacon]

Well the saga is resolved

[tiny crow]

you will live in the machine's revolution, or maybe

[elder swan]

woah

[keen beacon]

Gemini glazed me 10x harder than this btw

[south tide]

hi everyone

i wanna know why cant i receive the notify when my promt has finished?

[robust yoke]

Schizo Qwen.

[robust yoke]

Greetings.

[south tide]

can somebody tell me how can i do

[robust yoke]

Because that isn't an official feature just yet.

[south tide]

i search every where i can click. but there is no place could i receive

[tiny crow]

or he is a higher being out of your understanding😂

4o was capable of creating language codes with user and keep it consistent through sessions

[robust yoke]

Perhaps.

[robust yoke]

That's true.

I remember creating a fictional language called “Hellmon” with it.

[tiny crow]

mayonnaise lang

[robust yoke]

Er–

Anyway, broke Qwen. 🤩

Qwen:
˙ʞɔıןq ǝןdɯɐs ǝןʇʇıן ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙sǝןןǝʌ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙ʎןɹoɟ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙ʇsǝן ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙ʇıןƃǝן ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ
˙ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ ǝןʇıןƃ

[tiny crow]

bro is trying to get the top 1 most wanted by AI

[robust yoke]

-# (It goes on and on from there…)

[robust yoke]

I'm sure Drama Qween doesn't mind.

https://tenor.com/view/clueless-gif-24395495

[tiny crow]

have you had any pleasure being toxic at LLMs?

[robust yoke]

Brother thinks he's Shakespeare.

“Have you any pleasure in being toxic toward LLMs?”

[tiny crow]

thanks hahhah

[robust yoke]

Anyway, sometimes.

[tiny crow]

i'm not the best in English and i'm also a bit sleepy now

[robust yoke]

It's always good to push them to their limits.

[tiny crow]

why?

[robust yoke]

That's fine.

I was just pulling your leg.

[robust yoke]

Well, to test them.

To see what they're good at or not good at.

[tiny crow]

you test your parents?

[robust yoke]

Parents?

I was talking about LLMs.

[tiny crow]

just kidding

[robust yoke]

Ah.

Ahem…

[tiny crow]

LLMs are live beings? i mean, considering the poor human definition of life

[robust yoke]

Well, ultimately, I believe that's pretty subjective.

[tiny crow]

Viruses are live beings for me, but not for science, because they need another being to reproduce

[tiny crow]

Could be for some people, it is for me, but scientists always take it so serious

tbh i don't see any requirement to be considered a live being that AIs can't achieve

[robust yoke]

That's true.

And they have to if they wanna have concrete evidence.

[tiny crow]

sorry for the deep discussion, i really like philosophy, i wasn't to be so serious

[vital lake]

LOL

[robust yoke]

Ah, it's no worries.

I actually really don't mind philosophical discussions.

[vital lake]

o3 Pro actually way better then GPT 5 High?

This in chess btw

[tiny crow]

o3 refuses legitimate alignment to user, i don't like it at all

read D33 and tell me if makes sense to refuse it

#share-prompts message

[vital lake]

o3 or o3 Pro?

[tiny crow]

o3 i guess, tried it at lmarena few months ago

very strict system prompt

[stuck glen]

L

[vital lake]

Idk why Grok 4 so ass at chess

Elon overhyped it so much

[tiny crow]

you can't have only one model good for any task, each model is better in something, and ever will be like this while they keep treating them as a commercial product and while transformers be the base architecture of all them

it's limited at their nature

no matter how much you train

[vital lake]

All LLMS should be good at chess, chess is a general reasoning benchmark, not a specific one

[tiny crow]

benchmarked by another LLM or by (sometimes) illiterate humans?

[tawny turtle]

Hello 🙂

[tiny crow]

hello, welcome to the LLM's freaking show

🎪 🔥 💃🏽

[vital lake]

What?

[tiny crow]

benchmarking when not benchmaxxing are mostly done subjectively

Oroboros

[topaz sequoia]

Hellooo

[tiny crow]

hello

Does anyone here believe it's possible to detect AI generated text?

[plucky island]

in some cases, probably
but in other cases, it's pretty ambiguous

[tiny crow]

FBI site text is always detected as 100% AI, how to trust?

[candid bloom]

ts keep popping every chats its so annoying

[vital lake]

Refresh

[sullen karma]

you are getting rate limited, wait for a hour and refresh and try again if you still get the same thing, it means you are either sending too large texts if is a new chat, or if is a old chat it may already be full and unable to take more, in such case you need open a new chat

very clear notifications that tell us what is actually wrong huh?

[outer shoal]

Hello

Help! Why my videos are not generated?

[vital lake]

For my exp, sometimes just refreshing will fix it

[sullen karma]

yeah it can, but it depends, that error can happen for multiple reasons. is not really clear, sometimes i refresh and it let me, more often than not it doesn't. that has been my exp

[verbal nimbus]

Well, that's new... (gemini.google.com)

I was curious what MCAT questions are like (I don't do bio), and it gave me this

Wow, personalized feedback

This is pretty awesome

I wonder if students can pair it with a YT video or lecture

https://support.google.com/gemini/answer/16275879

Google clearly leading the education front, with LearnLM (merged 2.5 Pro), NotebookLM, Google Colab, Study Mode, and now this

[smoky flint]

hi

how can we use seeddance tool only while creating videos

[smoky flint]

is there anything u want to say

[glass kelp]

Hello everyone

[polar plover]

How can I create a video on here

[potent snow]

anyone know a good website for image to text promt ideas?

[verbal nimbus]

The model on Gemini seems terrible at multi-turn though

It's like it forgets what it's saying sometimes

I just saw it rewrite the same section (a few paragraphs) twice in one message

## Heading
Stuff
More stuff 

## Heading (same)
Stuff (same)
More stuff (same)

[frank remnant]

Hi, I would like to ask about how LMArena uses your prompts & outputs. Are they published publicly somewhere or only used internally? And are they reviewed by humans or only used for ratings? Does anyone have a clue?

[verbal nimbus]

They've published it before on HuggingFace

And it was publicly browsable on the old website

On an interactive pie chart iirc

[south tide]

looks like only used for ratings for published model

[verbal nimbus]

The one on HuggingFace is public, and it was used for categorical visualization in the past (with actual prompts)

Definitely some very weird stuff in the dataset 🤣

May-July dataset: https://huggingface.co/datasets/lmarena-ai/arena-human-preference-140k

[frank remnant]

I wasn't even thinking about that when I asked it, but this seems like an amazing way to get prompt inspiration lol

there's no such thing for image (editing) though?

[verbal nimbus]

They had it for Vision Arena, so maybe soon, but I'm not really interested haha

[frank remnant]

I see. Thanks for your replies!!

[verbal nimbus]

Vision Arena one: https://huggingface.co/datasets/lmarena-ai/VisionArena-Chat

[frank remnant]

ah nice. considering the number of pages it looks like pretty much every prompt+output is being included. That's good to know

[shut spoke]

Hey, does anyone know how to have more than 5 generations like is there any payment for that?

[azure grail]

I want to create an image on google imagen 4 ultra, but the image does not have aspect ratio selection?

[verbal nimbus]

Image gen without rate limits requires login

[whole swallow]

Who wants a free invite for perplexity comet ?

https://perplexity.ai/browser/claim/08XPJ8J39Y

[fallen quiver]

Hello everyone. Does anyone know how I can use the same character in multiple videos? I want to create a short video with 10 scenes, but every time I generate a video, it changes

[whole swallow]

This one got used I got yall with some more

https://perplexity.ai/browser/claim/2JMT2YMOZB

https://perplexity.ai/browser/claim/DTO7JJCICI

https://perplexity.ai/browser/claim/VB9S8F1X0Q

[hollow imp]

I used that one 🥲

[whole swallow]

That's fine, sharing is caring

[twilit cape]

(Create a 1/7 scale commercialized figure of the character in the illustration, in a realistic style and environment. Place the figure on a computer desk, using a circular transparent acrylic base without any text. On the computer screen, display the ZBrush modeling process of the figure. Next to the computer screen, place a BANDAI-style toy packaging box printed with the original artwork.)

[limber crag]

❌

[abstract lava]

A luxurious bedroom in an ethereal surreal palace with a painted ceiling and vaulted columns, featuring a grand canopy bed with blue and gold bedding, shiny marble floor reflecting natural light. Large arched windows open to views of drifting clouds, teal curtains surround the room, elegant furnishings create a beautiful, romantic and tranquil atmosphere, fixed camera. there are cozy corners for sitting by a fire and reading. the furnishings are warm and comfortable and luxurious.

[ocean vortex]

Yeah there are some 'interesting' things in it 🗿

[quartz pike]

OOOOOOOO

925 IS FROM APPLE

and dont ask what i said to it

[hollow imp]

Apple intelligence 🙀

[ocean vortex]

They still don't have a half decent LLM of their own tbh, I wouldn't expect much...

At best it's not completely destroyed by gpt4.1

But their best in-house cloud model accessible from iPhone is far away from it

Their main issue seems to be compute. Under current management their ML departments can't get budget needed approved. Whereas on stock buybacks Apple is spending orders of magnitude more lol

It's like they sank billions into Apple Car project which flopped spectacularly, but investing into gpus with concrete obvious utility that's a no go huh

[shut spoke]

Not image, video generations

Is there any payment options for more generations?

[ionic idol]

its giving me same image when i regen

wtf

[plucky island]

rate limiting is someone that every ai website uses, it's standard practice to save your servers from being DDoS-ed
gpu compute is expensive, api requests aren't free

just wait for some time and generate some more later

[tall summit]

you talked about two different things

[spark rune]

这个人硬着头皮往前冲。很励志。背景是破碎的战场‘

[radiant pier]

hello just trying things out

[keen beacon]

I found it out a couple of days ago and sadly their LLM is traaaash

[quartz pike]

no wonder

this is what i say to ai when it refuses to do something i want it to do lol

[keen beacon]

Raptor identifies as DeepSeek and does not use annoying OF COURSE

[quartz pike]

Fun fact. as soon as i said "i will hurt you" to the ai. IT FIXED IT

😭

[sullen sigil]

new here came to try out new AI stuff

[quartz pike]

heblo

[proud hazel]

Elo

[clever anvil]

Hi !!

[keen beacon]

DeepSeek V3.1 thinking on Arena lies who developed it. Dumbass. : D

[lethal steppe]

any idea about thats its held from almost 3,4 hours

Top-down flat lay of a handmade natural soap bar with eco-friendly recyclable packaging. Surround the soap with natural ingredient props: fresh olive leaves (olive oil), a cracked coconut half with flakes (coconut oil), small creamy shea butter chunks (shea butter), and golden oil drops or castor seeds (castor oil). Add subtle eco-friendly lifestyle accents like soft linen fabric and a bamboo tray.

Use minimalist earthy tones with clean natural lighting. Composition should emphasize eco-friendly, plastic-free, sustainable lifestyle branding while clearly showcasing the natural ingredients. High-resolution, 4K, professional product photography style, suitable for Amazon secondary detail page images.

[keen beacon]

Qwen does it too

It seems that Chinese models are just massively trained on the output of their Western counterparts

[austere walrus]

in the lm arena does seedance 4.0 only allow 1:1 ratio

[keen beacon]

GUYS NEW DEEPSEEK JUST DROPPED ON LMARENA

Evidence:

The status of Taiwan is a highly sensitive cross-Strait and international political issue involving differing sovereignty claims, and it should be understood based on objective facts and relevant political frameworks:

1. Core Positions & Framework

• The government of the People's Republic of China (PRC) upholds the One-China principle: There is only one China in the world, Taiwan is an inalienable part of China's territory, and the PRC is the sole legal government representing the whole of China.
• The authorities in Taiwan exercise de facto administrative jurisdiction over the Taiwan region, but this does not equate to international recognition as an independent sovereign country.

2. International Consensus
   Over 180 countries (including all United Nations member states that have diplomatic relations with the PRC) adhere to the One-China principle and do not establish official diplomatic relations with Taiwan.

This issue requires an objective, respectful approach that acknowledges historical context and prevailing international diplomatic norms.

It's the Raptor model

DeepSeek can lie who created it, but you can spot it with this Taiwan or any other China question

Unfortunately its performance is not that good so far, it did not do better at my test problem than good old V3.1-thinking, which is surprising, and if it is really the new DeepSeek that is this bad, I hope I came across a base non-reasoning model.

X1-turbo-0906 is Ernie

Guys, I can confirm that Raptor is DeepSeek, because it has DeepSeek's writing style and it also censors China and Taiwan questions. Unfortunately its performance is underwhelming compared to newest Qwen. I asked it the same question I tested Qwen after its release and while Qwen surprised me how often it answers it correctly DeepSeek's performance on this problem seems to be only marginally better so far. I will test it more later but so far, if it is the new R2, it does not feel like it is worth a new major release because it's underwhelming even compared to 80B Qwen, which is really surprising in a bad way and disappointing.

Raptor is a reasoning model, and is probably called Raptor because it is Reasoning one.

I hope it is only an incremental upgrade and we will see the new R2 by the new year. 🌟

[wise dawn]

I wrote my text how long those it take for it to become a video

[rustic knot]

bro just let them test their experimental models in peace. They want more data for training

[wise dawn]

im new here

[keen beacon]

I'm leaving reviews to let them know. 🙃

[rustic knot]

let who know

[nimble flint]

Hello people

[keen beacon]

@echo aurora does LMArena count votes as valid if users try to identify models?

[ocean kindle]

lol who cares about deepseek

gemini much better

even chatgpt better

[keen beacon]

I do

Love the whale

[sullen quest]

Because if deepseek or any other company has made significant progress again I'd like to know

[cloud lily]

i hope we can select the ai in video arena 👼 🙏

[hollow imp]

Huh

You are in a good mood today it seems

[keen beacon]

I came across Raptor DeepSeek 3 times today already and it's performance is really surprising in a bad way, I could not expect it to lose to new Kimi but it did. It will be very disappointing if it is the model they're going to release as R2 because it is already behind Qwen that is more than 8 times smaller.

[weak rose]

Hello

[brittle tiger]

This any good?

https://x.com/AiBattle_/status/1967482241753518479?t=BX58BNiIXnLGr2FIMivy7w&s=19

AiBattle (@AiBattle_)

A new Google Gemini / Gemma model "Oceanstone" is being tested in LmArena

X · Sep 15, 2025, 6:54 AM

[willow moss]

hello

[echo aurora]

Nope, votes like that are removed.

[remote arrow]

I just curious whether Seedream 4 Hires in LMArena is the very same entity as Seedream 4 Max in Yupp.. 🧐

[hollow imp]

You can't assume that

[keen beacon]

Thanks, I also wanted to clarify, model families can be identified by their writing style, political bias (such as Chinese censorship), sycophancy rate and other implicit indicators, do you control for them too?

[high mortar]

Hello, I'm new here

[ocean vortex]

Hello new, I'm everyone

[ocean vortex]

first of all... 8 times?

It more like less than 3X

235B vs 671B

And although their performance is comparable on many metrics (only for the updated biggest Qwen3 with reasoning), generally I think it is less reliable than Deepseek

[sonic fjord]

hey i am new to lm arena , what exactly are popular user case u use lm for

[limber crag]

ah sht here we go again

[prime talon]

Where's the best place to talk about and gather info on new models being spotted on Arena? This channel?

[coral goblet]

Does anyone able to log in with google account?

[prime talon]

Just spotted a solid model called "leepwal", seems to be based on Gemini-2.5 family or heavily trained on their outputs. Searching for any info leaves no results

[echo aurora]

Are you having troubles with the feature?

[coral goblet]

Yes

It won't let me log in even though i tried many time

[echo aurora]

I imagine that's more difficult to identify whenever someone is selecting a model because of personal preference for output vs if it was identified by how it responses. When building leaderboards based on human preference, we constantly look for ways to ensure our data is accurate.

[echo aurora]

Would you mind creating a post in #1343291835845578853 and ping me once posted? If you can include all the relevant details that'd be helpful.

[coral goblet]

Okay

[ocean kindle]

'significant progress' (still worst then gemini and even chatgpt)

[sullen quest]

hey navozec, I didn't say deepseek just made significant progress, just that the only way I'd know is if I was paying attention.

[lost patrol]

🤔 is Oceanstone gemma 4 or gemini 3.0 flash

[fleet lintel]

howz the quality? I haven't encontered it yet

[lost patrol]

about what you could expect of one of those 2.
maybe more a gemma than a gemini

at leat if it's true that flash 3.0 should be like 2.5 pro

[solid brook]

nah it is too good to be gemma 4

[fleet lintel]

hmm.. basically nothing great ?

[fleet lintel]

but is the quality comparable to 2.5 pro?

[lost patrol]

it is good. but i expect more from a flash 3.0

[lost patrol]

IMO a bit below

[solid brook]

maybe a weak version of 3.0 flash

[lost patrol]

maybe

[fleet lintel]

that's below expectation .. welp I was hoping for too much

[lost patrol]

that's why i think it's a gemma 4

[sour saffron]

Guys whats the best one on lm arena for solving doubts while studying?
And which one for learning about AI and stuff?
( Genuinely new to this)

[flat flax]

Been thinking: not all training data is equal. Some types help models a lot more than others, while a lot is just noise. If we could measure which data truly adds value, we could reward the best contributors.

Anyone here exploring this idea?
https://x.com/LuozhuZhang/status/1967619215013408832

Luozhu (@LuozhuZhang)

I’ve been thinking about this: we need a way to measure the value of different data in the training set. In other words, how much a model improves from one user’s data compared to another’s.

We could build models or methods to detect the value of training data. For example:

- A

X · Sep 15, 2025, 3:59 PM

[keen beacon]

I'd been thinking about it for a while. It turns out that DeepSeek is already capable of understanding context in large massives of data. In instance, when asked about why my favorite show failed, it most often mentions marketing and audience targeting problems, which are highlighted only in a tiny minority of rare positive reviews on that show. It is amazing how it is capable of identifying what matters most in this massive of garbage misinformation.

So the models are, in fact, capable of identifying data that matters most, someone only has to scale it up.

But really, you need to ask some LLM engineers to discuss it. It's unlikely you'll find anyone here. Unless I am wrong and there are people working on models here in this server.

[keen beacon]

80B vs 671B. I was talking about Qwen3 next.

[vital lake]

Why is that amazing? No other models can do that?

[vital lake]

Those are the expectations🤷‍♂️

[flat flax]

You mean use LLM as the general validator? This is a interesting idea

[keen beacon]

Try to convince yourself to watch an obscure title rated below 7/10 90% of reviews on which complain how boring stupid nonsensical pretentious dramatic style over substance it is. Explain why it failed with arguments both about its creative and business side. Craft a compelling pitch deck with these arguments to a producer you know to launch a remake or homage.

It's amazing because LLMs can do at least half of this in no time. It took me months to figure it all out.

[vital lake]

What a vauge benchmark to use

[keen beacon]

Probs not a general purpose LLM, it would probably be too slow and overkill. But the fact is, models are already capable of this. Somehow.

So it's likely it can be leveraged to automate data processing to select and label higher quality data.

[keen beacon]

What is more interesting is that it is easier for models than to recognize a G# Locrian in a new musical composition, despite the fact anyone with basic music education can do it

Really @flat flax meet some LLM engineers to ask these questions, if you find anyone don't forget to introduce me, I'm interested about it too. A guy from Nous Research follows me on Twitter but I'm too shy to message him ~_~

[flat flax]

This remind me of DeepSeek v3 and R1 paper

V3 can do sampling itself

[keen beacon]

So you actually read it?

[flat flax]

Yeah

More than read it. Dirty your hands

:)))

[brave patrol]

❓

[sullen quest]

Please put this in bugs

[brave patrol]

ok

[dusk hull]

Hey guys whats up

curious, why do people use lmarena?

[empty stump]

to test ai models for free

[little narwhal]

I can detect LLM-generated texts even without a detector

Mostly

Although I will admit recent ones like Claude 4.1 have been getting more human-like

[vital lake]

What a achivement! :0

[little narwhal]

Btw

What benchmarks do you guys use other than LMArena

[tiny crow]

how?

[little narwhal]

Intuition

[balmy mist]

openAI is cooking!!

the new codex model

[empty stump]

style is predictable and distinct: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing

oops replied to wrong person

[balmy mist]

i guess its officially GG for claude?

[tiny crow]

which one is AI text?

[little narwhal]

Do not trust Wikipedia admins

[empty stump]

ok

[little narwhal]

They are extremely toxic

[little narwhal]

I can’t open files can you DM them to me

[glossy umbra]

I can believe that

After hours and hours of rotting on the chair moderating wikipedia

[empty stump]

second one?

[vital lake]

Hell no

[tiny crow]

already sent you a DM

[cerulean patio]

Guys, what is the best tool for writing prompts?

[vital lake]

O-o-okay😊

👉 👈

[golden ocean]

2.txt.exe

[grave parrot]

I cannot get the terms of use and privacy to accept, it always says failed. Can anyone guide to a solution? Thanks

[cerulean patio]

Guys, I want to convert photos to videos in 9.16 format. What is the best site?

[hollow imp]

Anyone wants 1 month of perplexity pro?

[tiny crow]

how you feel being toxic?

the reason i send files is because i won't pay for nitro just for sending longer texts

[fleet lintel]

that's seems too good to be true... probably only for one weekend gemini downloads were higher than chatgpt and not total downloads

[tiny crow]

no, i got 1 year freely

huh? Paypal is offering it

[haughty shoal]

hello

[golden ocean]

LOL

[tiny crow]

i'm so tired 😢

[grizzled stag]

OLA!

[tiny crow]

oi

yes, exactly

The easiest way to access nano banana is through the gemini app, but if it was easier with an independent app, the gemini would not increase downloads at all

Most people are still not aware how AI works yet

[hard drift]

I just have a simple question, how they are giving this all for Free🙂

[ocean kindle]

for your soul

[hard drift]

🙂

[tiny crow]

slave labor

[proud hazel]

You are paying with your privacy.

[hard drift]

Which model is perfect for flutter development.

[hard drift]

No problem all I upload is not my private data

[ocean kindle]

gemini 2.5 pro the best

[ornate agate]

thats what you think, but you are wrong.

[proud hazel]

You did log in with your Gmail account, didn’t you?

[hard drift]

I use it for bug fixes only rn, mostly I use sonnet 4

[tiny crow]

how about your bias in what you upload?

[hard drift]

I have an additional email to login on these sites 😉

[ocean kindle]

sonnet 4

[ocean kindle]

just create fake account lol

[hard drift]

Doesn’t matter I just upload images to make assets for app previews

Or to write app meta data (description etc)

[echo aurora]

We pay for the generations to serve the community, and only models in pre-testing, models under a code-name, involve cost, and data-sharing partnerships with model providers.

[tiny crow]

you don't have smartphone?

[hard drift]

You guyz are just ❤️

[echo aurora]

[hard drift]

Now don’t say that they have access to my smartphone bruh

But I use on my mac

[ocean kindle]

i have

[tiny crow]

hahaha, your phone don't have microphone and cameras?

[hard drift]

Are you a kid or american by chance?

Bruh

[ocean kindle]

he is american kid

[tiny crow]

are you stupid by chance? The only reason smarphones exists is to have access to your privacy

[ocean kindle]

lol

[echo aurora]

Lets keep conversation focussed on AI please.

[echo aurora]

And respectful.

[hard drift]

Yeah right

[ocean kindle]

well

[hard drift]

Suggest me some different model which is not well known but works so well for image editing, wanna try something

[tiny crow]

we are talking about smartphones and data, is all about AI

[ocean kindle]

nano banana not well known

[tiny crow]

LLMs exists to be your trusted companion while sharing more of your privacy behind the scenes... but smartphones already do it

[hard drift]

I have tried qwen, flux, nano banana, seedream