general

[keen beacon]

I need them to be serious

[blazing bison]

I feel that gemini is being nerfed

But maybe it's just me

[versed totem]

im noticing the same

[keen fulcrum]

@echo aurora Can you make the bot show us the result of the votes?

[echo aurora]

It’s possible we make that change. Be sure to note the #bot-feedback channel though for specific bot feedback so it’s easier for us to track.

[stray aspen]

avez vous une date de sortie estimee pour le video arena

[stray aspen]

im noticing that too

[tight silo]

think it's a sign of a new model?

[stray aspen]

bro

why is romlox word banned here

[digital umbra]

people spammed it in the video arena

[stray aspen]

or maybe grok is just better at roblocks coding

[grizzled bobcat]

is unlimited time?

My question

[wheat onyx]

https://x.com/chetaslua/status/1949905375546708242?s=19

Chetaslua (@chetaslua)

GPT - 5 attempt on minecraft

ZENITH TO BE EXACT

This is impressive this is magic .

@tszzl @aidan_mclau @apples_jimmy

OPEN AI has literally made something insane 🙏

X · Jul 28, 2025, 6:50 PM

[unborn ocean]

sharing this for the 3x time because it seems relevant

very few knew zhipu

[digital umbra]

zhipu have used a few different names iirc

[gentle plinth]

https://fixvx.com/chetaslua/status/1949905375546708242

Chetaslua (@chetaslua)

GPT - 5 attempt on minecraft

ZENITH TO BE EXACT

This is impressive this is magic .

@tszzl @aidan_mclau @apples_jimmy

OPEN AI has literally made something insane 🙏 https://t.co/iMl1LnK3A2

▶ Play video

vxTwitter / fixvx · Jul 28, 2025, 6:50 PM

With embed

[wintry locust]

i know all of these am i cooked

[ocean vortex]

LOL

[wintry locust]

youre missing minimax and baichuan and iflytek

[ocean vortex]

Yandex

[void tusk]

is GLM-4.5 an improved version of gpt4.5?

[wintry locust]

that's russian...

i guess lg is korean

ok fair enough

[ocean vortex]

Yeah but they are also irrelevant

[digital umbra]

completely unrelated

[gentle plinth]

It's a different model from a different company. But it's open source

[wintry locust]

AI Sweden

[gentle plinth]

https://huggingface.co/zai-org/GLM-4.5

[void tusk]

oh i see

thanks

[stray aspen]

why do all the chinese models have a similar reasoning process

[grizzled bobcat]

Is free?

[stray aspen]

yes

[grizzled bobcat]

I see it

[daring rover]

is glm openai?

oh

it's a random company

[keen beacon]

Yes, chinese

[daring rover]

baited

i thuoght it was openai's OS model for a split sec

[keen beacon]

GPT-5 will come early next month probably

so soon we'll have some new models

[digital umbra]

i'm just really curious to see what the open source model will be called

[keen beacon]

GPT-5os

lol

open source

they will probably continue using some weird names

perhaps

Or maybe it's 3.5 finally open-sourced. I'd laugh

[digital umbra]

if they include GPT or "o" in the name it's going to be so confusing lol

i'm guessing it's going to be a new model architecture rather than based on any of their proprietary models

[keen beacon]

I really hope that it will be useful and not be just some random tiny model

[digital umbra]

well, considering they delayed it when kimi k2 released...

i guess it will have a few hundred billion parameters, or a dense model equivalent

if they release a 50b moe it's going to be so underwhelming lol

[keen beacon]

Yeah, google and others will overtake them easily in that case

Gemma 3 27b is real good for it's size for example

a bit aged already

[reef pawn]

Gemini 3 pro when?

[blazing bison]

I bet August

[reef pawn]

Same month as GPT 5, right?

[blazing bison]

I think gpt 5 will be released this week

[reef pawn]

Oh okay

Can't wait

[warm fulcrum]

how are some people using gpt-5 before its released

[blazing bison]

They aren't

[reef pawn]

WHO IS USING GPT 5?

[digital umbra]

people speculating that openai is rushing their model before the eu ai act goes into effect, if that was the case i would think google would also be rushing something

[warm fulcrum]

https://x.com/chetaslua/status/1949905375546708242?s=19

Chetaslua (@chetaslua)

GPT - 5 attempt on minecraft

ZENITH TO BE EXACT

This is impressive this is magic .

@tszzl @aidan_mclau @apples_jimmy

OPEN AI has literally made something insane 🙏

X · Jul 28, 2025, 6:50 PM

[blazing bison]

Actually there is people that get access weeks before release, but these people generally don't talk about it

[blazing bison]

We can't be sure that zenith is gpt 5

I think it is but...

[warm fulcrum]

well ye how are people able to use that model?

[blazing bison]

It was available on the arena in the weekend but already got removed

[reef pawn]

Can you tell me it's rank in LM arena?

[blazing bison]

The rank is not public yet

[warm fulcrum]

unfortunate

[reef pawn]

Oh okay

[warm fulcrum]

how does lmarena even have ahold of these models?

[keen beacon]

OpenAI provides the pre-release models to LMArena

[reef pawn]

Labs give them early access for testing

[digital umbra]

they get a lot of useful feedback for putting models here

[blazing bison]

They hack ceos accounts and proxy it

[warm fulcrum]

wowie

lets hope openai new model lives up to the hype

[blazing bison]

Its good

But it's not agi good

[reef pawn]

AGI is buzzword

[blazing bison]

Its like 25% improvement from o3

[reef pawn]

Nice

[blazing bison]

And 25% is a lot

[keen beacon]

We ain't near AGI at all unless some type of new architecture goes live

with no hallucinations and consistent answers

[digital umbra]

zenith will probably be a great model, let's hope it won't be too expensive or behind a router that gives you garbage most of the time

[keen beacon]

and learns from mistakes independently

[keen talon]

can someone tell me what are the limits for claude 4 opus?

[blazing bison]

I think it's gonna be a router that gives you garbage if your prompt is not good enough

Because it was already like this on the arena

[reef pawn]

I think we at least decade away from true AGI but Scam Altman and Melon Musk keep milking that word

[warm fulcrum]

why is gemini 2.5 pro rated #1 on all tasks

theres no way it actually is that good

[blazing bison]

Because people vote for it

[keen beacon]

Yeah, they should be honest and not lie to people

[reef pawn]

Gemini is my fav model

[keen beacon]

It's an all-around excellent model

in many areas

[warm fulcrum]

everytime i ask it to code it just blabs a lot

it adds more comments than code

[reef pawn]

You use Google AI Studio?

[keen beacon]

Ah, I just use it for random specific questions. Querying the knowledge base, ethics, morals

[blazing bison]

I agree, but we are not a decade away for people that use llms getting 2x leverage

[keen beacon]

And of course in Finnish

[warm fulcrum]

yes

the actual gemini website doesn't even want to code

it just says it isn't capable

[keen beacon]

Lol uralic languages are probably in the 0.005 percent of votes/prompts

[reef pawn]

Yes, Not trying to precise or anything the point was it's not anywhere near yet.

[blazing bison]

Try ai studio

[warm fulcrum]

i alr do

[blazing bison]

I mean, we don't need agi to make useful things

[reef pawn]

It works fine with me, the thing is you have to give specific prompts to the model and break bigger task in small chunks for better output

[reef pawn]

True

[keen beacon]

Well medicine would be real useful for AGI to know

[blazing bison]

If gpt 5 is not a big leap, I'm sad

The bubble will burst

[torn mantle]

grok 4

[surreal creek]

Gemini in first place 30 points ahead of o3 on the coding leaderboard lol

[stray aspen]

which claude

[keen fulcrum]

I think Arena should reconsider the evaluation process and include pregenerated results for prompts

That way a prompt can be evaluated from multiple users

[meager harbor]

no the actual is good, but your idea is good that you also add the possibility to rate pregenerated prompts

[unborn ocean]

you know chinese labs are afraid of repercussions if prompting for "a taipei vacation" is already considered an inappropriate topic

[meager harbor]

so any gpt 5 whispers ?

[sonic tendon]

glm 4.5 is surprisingly good

[echo aurora]

We're aware of issues related to non-text models struggling at the moment.

[quiet moss]

Ok

[echo aurora]

All fixed

[whole sundial]

guys I think they might have distilled glm 4.5 off of gemini, I just had a response start with "Of course!"

[quiet moss]

just because it said Of course means its trained off of Gemini?

[whole sundial]

ok then tell me another model that starts with "Of course!" all the time

seems to happen when reasoning is off

[wintry tinsel]

And writing too Claude is always the bomb

[whole sundial]

yeah it starts with "Of course!" just like Gemini

at least with reasoning off

must of post-trained it off of gemini conversations, at least partially

but this shouldn't be a surprise, Chinese companies distill off of US models all the time

[stray aspen]

glm 4.5 no think is gemini

[whole sundial]

I feel like the "Of course!" is a watermark put in by Google

I'm not saying it is Gemini, I was just saying that they distilled Gemini into the model

and it has long response, kimi gets straight to the point

that might be better for some people though, but this means glm 4.5 is going to have more slop

[stray aspen]

i love the glm UI

[whole sundial]

thinking glm 4.5 does not have the "Of course!" stuff, i think it only does that for non-thinking due to likely gemini distillation. As they can't distill their reasoning traces anymore, it won't do it in reasoning mode because it's distilled off of a different model

[leaden palm]

no dark mode tho

and a short default max tokens

[whole sundial]

i was using this site https://huggingface.co/spaces/zai-org/GLM-4.5-Space to try it out, disable thinking and lower the temperature and you'll see what I mean

it identifies itself as being by Zhipu like it should, but the "Of course!" threw me off a bit

[blazing bison]

Even if it have a little of gemini data, it's not a problem if the model is good

But for me it's no good

[sturdy mica]

why is it responding with a slutty highschool girl system prompt wtf???

whats wrong with that website

☹️

[whole sundial]

it seems to be fine without thinking, maybe it messes up with thinking?

or when multiple people are using it at the same time?

[sturdy mica]

prolly this.... i wonder whos using it for weird fetish roleplay.....

[keen beacon]

my bad

[leaden palm]

what ai mode suggestions do you guys have

#1 doesn't really make sense to me and #3 isn't really relevant but #2 is definitely personalized

[gusty night]

Hello !!

[harsh flume]

What do you guys think it's capping AIs from performing well in frontier math benchs?

it doesnt seem like it would be an unsurmisable problem when you take into account the existence of models like AlphaFold

[leaden sun]

in short, it's a multifaceted problem, beginning with what "understanding" even truly means for a machine, to the problem of translation between formal logic and natural language, to the fact that most if not all traditionally trained mathematicians work more with intuition rather than pure information retrieval, connecting the dots works often subconsciously that happens to surface into conscious understanding, leading to the Eureka moment. As far as i know, the current ai architecture is still too limiting?

in case you're interested, one of the current frontier ai research is about the connection between consciousness and high intelligence, it's still an open problem, but a very fascinating one compared to those hopeless millennium prize problems...

[novel crater]

what is the fastest model on lmarena?

[harsh flume]

I understand when it comes to tier 4, but AlphaGO in 2017 kinda solved the dilemma of navigating a giant state space (10^170), I am kinda dumb but it feels like problems in tier 1-3 of FrontierMath would be a lot easier and lower search space than that since they are all solvable.

It seems like they are only testing LLMs tho which makes sense to have a low score, altough i'd assume that LLMs could implement math-driven tools like alphaproof where the LLM layer would translate a problem into pure math and call in the solver

[leaden sun]

i think proof assistants are already being integrated into the architecture to make it more deterministic, the thing is, those theorem provers are not complete and still an area of active research

they only testing LLMs? so they have figured another way already? dont tell me it's an artificially grown organic hybrid brain hahah

[wicked root]

Is there a new model that's being tested right now in LMArena?

Word on the street is GPT5 is being tested rn

[drifting thorn]

nah it's great in creative writing(writing lyrics)

[whole sundial]

time to make your pfp a picture of cliff richard lol

[drifting thorn]

the never gonna give you up hallucination is the LLM joke of the year

[quiet moss]

If GPT-5 releases by July 31, is it likely it will be on LMArena on the same day?

[whole sundial]

(the correct answer to the prompt is "Nothing's Gonna Stop Us Now" by Starship. "Never Gonna Give You Up" was the number one song of 1987 in the UK, but it's not by Cliff Richard.)
large reasoning llms (o3, 2.5 pro, claude opus, grok 4) get this right.

GLM 4.5 gets this right as well

[drifting thorn]

actually the whole answer hallucinates...

GLM 4.5 has bad lyric writing

It doesn't even rhyme with the line I gave it

he gave me 4 answers, but none of them rhymes

[gusty night]

You are quick at model integration 👏

[drifting thorn]

What's the provider of kraken-072125-1\

[whole sundial]

amazon

[harsh flume]

well yea

I read through what I could find of information on their website and apparently the bench is done with the models using tools, so it'd be possible to integrate a native math AI that an LLM could call on

[digital umbra]

this came up in openrouter discord

[leaden sun]

yeah define math ai first 😅 i've never looked at those frontiermath questions so i assume it's a broard selection across the entire mathematical discipline, good luck building a math ai who can afford all those vast math tools

i know it's difficult for people outside math to imagine how...fundamentally different the areas in maths actually are

three examples i'd personally love the llms to be able to use:
https://dealii.org/
https://www.sagemath.org/tour.html
https://rocq-prover.org/
and those are just one of the many out there

[leaden sun]

obviously, llms need to understand the problem first, recall knowledge needed (theorems, lemmatas, corollaries etc), connect the dots and use the tools correctly to get the final answer

[harsh flume]

alphaproof would be an example

alpha geometry2 another

[leaden sun]

those are not general math ai, they are specialized if am not mistaken, but yeah, you can always build a swarm of specialized ones and call it a general ai

[harsh flume]

yea that's why I was positing in the usage of them as tool calls

AGI will prob be a form of that anyways as I dont think general intelligence will come from a pure next-token-predictor model with infinite scaling

[leaden sun]

the coordination between those agents within a swarm will be a challenge, it's studied also in dynamical systems

[harsh flume]

the interesting thing is that these are a whole other transformers achitecture so integrating them within the answer scope of a LLM would be really dope

lol it seems like they are on it already

here I was proposing the invention of fire whilst they are already on blowtorch schematics lol

man, I wish LMArena would organize a sorts of AMA with top AI researchers from these labs, they must be in direct contact with the industry's forefront and that'd make some great content given how invested this server's users are

People here would formulate more interesting questions than 90% of podcast hosts

[leaden sun]

"there is nothing new under the sun", we're simply rediscovering them all...😊

[agile bloom]

based on the response glm 4.5 gave me, it needs to be worked on

like damn, glm 4.5 told me it's mental state

[slim mesa]

hii

the grok 4, on the part of direct chat, is really grok 4?

mine say him is the grok 1 xd

sorry bad english

[calm sequoia]

What does this even mean

[nimble trail]

Yep it is. The model itself tends to hallucinate about it's model.

[whole sundial]

it's interesting that the current model here is 4o. must be filler for gpt-5 (which, considering they have already added this, should be coming very soon)

[calm sequoia]

THis selection also exist on o3

And I wouldn't say this is much longer.

[ashen mauve]

What is GLM anyways?

[cedar tide]

Nemotron v1.5 on Artificial analysis
Its best score for an open source model that can be deployed on a single h100

Go Upvote this model
https://discord.com/channels/1340554757349179412/1398515764448989304

[golden ocean]

bing chat sydney

[cedar tide]

First of all, I want to clarify that I don't trust this score at all to predict their overall performance.

Kimi k2 is 2nd best Model without Reasoning so no problem with his score, and you can't compare him with reasoning models

For glm They themselves shared the score of their model on the same benchmarks as artificial analysis and these are the right places

It's certain that if he had infinite money he would have set many other benchmarks

[humble sonnet]

What is GLM 4.5 ?

[keen beacon]

Chinese's company's new model

[reef pawn]

GLM is proprietary model, right?

[teal mantle]

I am mostly API only but should I renew GPT Plus or Supergrok
One for agent, one for grok 4

[torn mantle]

is it really good or they are just benchmaxing again

[teal mantle]

MIT

[reef pawn]

Oh then the scores are good

[cedar tide]

What does proprietary model mean?

[torn mantle]

david

is it good or nah

[reef pawn]

Not Open Weight

[cedar tide]

Open, mit licence

[torn mantle]

proprietary means ownership @reef pawn

[reef pawn]

Oh okay

[torn mantle]

not open source

[teal mantle]

Which one I should get again

[cedar tide]

So very good

[torn mantle]

good?

[reef pawn]

What is opposite of Not open weight?

[cedar tide]

Il speaking about that is open mit

[torn mantle]

closed source

[torn mantle]

i see

[humble sonnet]

Are there any special features?

[reef pawn]

Are proprietary model closed source or open?

[torn mantle]

they can be both

[humble sonnet]

Are there any special features?

[reef pawn]

💀

[torn mantle]

proprietary is owned by the ones who made it

[cedar tide]

@humble sonnet salut

[torn mantle]

what are you talking about?

[reef pawn]

How you gonna make money from open source model

[keen beacon]

I don't know that much. Haven't tried GLM models before

[torn mantle]

i think you are confusing it with another word or something

[keen beacon]

Funding

[cedar tide]

From your api and chatbot if he have a subcription

[reef pawn]

But that is not allowed as business

[teal mantle]

GPT plus or supergrok btw

[reef pawn]

Both sucks

Gemini better

[teal mantle]

How much usage I could milk

[cedar tide]

but the problem is that if the open source model there will often be APIs much cheaper than yours

[reef pawn]

GPT-1 IMAGE is good tho

[teal mantle]

I already milk CLI and AIstudio like anyone decent

[torn mantle]

you are right from a business perspective

[reef pawn]

1 million context window

[teal mantle]

Message limits I mean😂

[reef pawn]

Thanks but I get your point, you was right too!

[reef pawn]

Then it's horrible, I got 1 year free Gemini AI pro student membership here in India!

https://tenor.com/view/dead-chat-dead-chat-discord-revive-gif-21258645

[torn mantle]

no you are right

if i say you are right then you are right

@cedar tide you are wrong

[reef pawn]

Aight 🙏

[cedar tide]

What ?

[keen beacon]

Did not know

lol

[calm sequoia]

Wtf guys, what are you using gemini for

Just run out of o3 request

Tried gemini 2.5 Pro max thinking budget

Failed at all of my requests miserably (o3 successful 90%)

Is Gemini always like this? 💀

[blazing bison]

GPT-5 has been spotted https://x.com/ryolu_/status/1950163428389040431

Ryo Lu (@ryolu_)

here’s my prompt
one shot + details

X · Jul 29, 2025, 11:56 AM

Cursor staff is already using it

Now I'm pretty sure Zenith was GPT 5

[mortal lynx]

To me o3 and 2.5 pro are both pretty hit or miss

Gemini 2.5 Pro was vastly superior at some tasks and garbage at others, same for o3

for coding tasks atleast

[blazing bison]

gpt 5 is good

it's agi, believe

[mortal lynx]

If zenith was GPT-5 it's still not quite AGI, but much closer than o3 and o4-mini were

[blazing bison]

i'm just kidding bro

[mortal lynx]

I know, i'm just commenting on it

I do think Zenith and o3-alpha were a considerably improvement over what we have today, atleast for what I've tested

much more than the "20%+ points in HLE and ARC-AGI!" models we got these past few months

[blazing bison]

o3-alpha was the best one, idk why people said zenith was better

maybe zenith is o3 alpha but the feeling that i got trying o3 alpha, the results, it was better than anything i ever tryed for coding

[keen fulcrum]

they obviously posted it

[blazing bison]

yes, for the hype

[keen fulcrum]

the blur is horrible

[blazing bison]

it's on purpose

but it means that gpt-5 is ready, idk openai is waiting for

[keen fulcrum]

this is a correct blur

[blazing bison]

and after this week rate limits from anthropic uugh

I really want OpenAI to dethrone them

[keen fulcrum]

in late august introduced

they ran out of gpus

[mortal lynx]

They're probably just preparing the blog posts, demos, videos and research papers

hopefully the demos are better than their usual "Look how our model can order a new shoe! AGI is here!"

Google and xAI do a much better job at that

[blazing bison]

Its gonna be a travel plan

[floral comet]

I heard there's (GPT-5) model also known as Zenith, is it still in the LLM Arena?

[blazing bison]

No

Removed

[floral comet]

Ohh, damn.. How do people can even use it.. I guess I'm not lucky enough

[odd shard]

hello

[stray aspen]

craig will gpt 5 be AGI

[quiet moss]

no

[blazing bison]

it's removed, it was avaliable 2 days ago

[keen beacon]

it's a bitcoin wallet

[quiet moss]

No it says sk

[keen fulcrum]

Its bait anyway

[stray aspen]

is it just me or is gemini 2.5 pro getting worse each day

[torn mantle]

who voted no

lets talk

[keen beacon]

I strongly hope to see that. I'd try all sorts of things

[sour spindle]

Completely unusable for me outside of summarizing long documents

It’s tool use is completely broken

You would think they would be dominating in this regard

[cedar tide]

You voted no

Batard 🤣

[keen beacon]

@echo aurora

[echo aurora]

This request is for sure on our radar! Was chatting with a few coworkers yesterday about it

But don't forget to use #1372230675914031105 ! Best place to make these kinds of requests.

[cedar tide]

https://fxtwitter.com/Alibaba_Qwen/status/1950227114793586867?t=PAPpkXrdI3340UJiuu30oA&s=19

Qwen (@Alibaba_Qwen)

🚀 Qwen3-30B-A3B Small Update: Smarter, faster, and local deployment-friendly.
︀︀
︀︀✨ Key Enhancements:
︀︀✅ Enhanced reasoning, coding, and math skills
︀︀✅ Broader multilingual knowledge
︀︀✅ Improved long-context understanding (up to 256K tokens)
︀︀✅ Better alignment with user intent and open-ended tasks
︀︀✅ No more blocks — now operating exclusively in non-thinking mode
︀︀
︀︀🔧 With 3B activated parameters, it's approaching the performance of GPT-4o and Qwen3-235B-A22B Non-Thinking
︀︀
︀︀Qwen Chat: chat.qwen.ai/?model=Qwen3-30B-A3B-2507
︀︀
︀︀HF:huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507 or huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8
︀︀
︀︀ModelScope: modelscope.cn/models/Qwen/Qwen3-30B-A3B-Instruct-2507 or modelscope.cn/models/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8

**💬 12 🔁 21 ❤️ 194 👁️ 4.8K **

FxTwitter · Jul 29, 2025, 4:09 PM

Anyone can make a request ?

[drifting crow]

I feel like they made that sk just to help us with our regex

[cedar tide]

qwen 3 coder arrived in the leaderboard and its the 3th overall open source model
(Kimi 2 and old qwen 3 no think better)

[keen fulcrum]

David do you have news about grok 4 coder

when will it release

[cedar tide]

Officialy august

[cedar tide]

average oh the 25 benchmark of qwen 30b a3b

by category

[fleet lintel]

Based on all the hype, I think gpt-5 is much much better than o3 / o4... isn't that the case?

[keen beacon]

I may sound dumb but what does spatial awareness mean in LLM models? Vision capabilities?

I see, thanks

[blazing bison]

openai is genius with this study together release lol

[civic flame]

ZENITH IS BACK

😍

[blazing bison]

really?

[civic flame]

YESS

okay perchance false alarm

[keen beacon]

:((

[blazing bison]

100% false alarm

[stray aspen]

guys whats the most trustworthy ai Benchmark

[blazing bison]

openai will collect so much reasoning data with this study together

this mode actually asks a lot about your reasoning

funny

[stray aspen]

craig wil gpt 5 be agi

[civic flame]

it's half back

can't share much but

[blazing bison]

?

[civic flame]

all it needs now is for them to flick a switch to enable it in battle

it's been re-added as if ready

[blazing bison]

claude 5 will be agi

with weekly rate limits

after 2 prompts

[jade egret]

gpt 5 when ):

[blazing bison]

thursday

[jade egret]

fr?

this?

[blazing bison]

no

[civic flame]

Craig

Neptune isn't a new model dawg

🥀

[keen beacon]

Apple guy was wrong after all

[jade egret]

bruh

[civic flame]

well yeah opus is a good model

😭

[cedar tide]

Go Upvote this very good model
https://discord.com/channels/1340554757349179412/1399812659800445131

How do you know ?

Go upvote this request
https://discord.com/channels/1340554757349179412/1394703782255788122

Very good artificial analysis

Its 32b sota

[digital umbra]

Doesn't require internet connection

[cedar tide]

I didn't say it was a sota and that it was better than o3 I don't know what you're talking about

in the arena you even have 1b models, the arena is not only for sota models

[digital umbra]

Still I think EXAONE (which btw isn't a chinese model) is problematic because its license basically forbids you from doing anything at all useful with it

Sure you can benchmark but that's about it lol

[cedar tide]

Yes exaone its non commercial permissive
We have on just one api with 1/1$ input output

[digital umbra]

Which is stupidly expensive for a 32B

[cedar tide]

@deep adder I don't understand anything you're saying

[cedar tide]

Yes average price for qwen 32b its 0.15 0.45

[stray aspen]

damn craig is educating everyone

[digital umbra]

yes, why would anyone use an open source model they can run locally when they could give their personal data to openai, be forced to use a web interface and rate limits

[stray aspen]

didnt expect qwen would get this far on the artificial analysis leaderboard

[digital umbra]

i can feed how much sensitive data i want into my gpu with no regrets. i mean it already sees everything i have on my screen anyway 😛

[keen beacon]

they are doing it because of the nyt thing right?

sam is trying to bring attention to it to win that lawsuit i guess

[digital umbra]

you're typing this on the discord of a site that provides user prompts to AI companies to improve their models...

and even if the data is useless for training it would still be useful to sell to data brokers

[devout vault]

is glm 4.5 even good

is it better than other smart models like gemini 2.5 pro?

[stray aspen]

hell no

are you serious

[devout vault]

people say it is

weird

[blazing bison]

This craig is just a rage baiter yapper

[stray aspen]

welcome to the internet bro

[torn mantle]

Ong

Onnng

[blazing bison]

It is, just ignore him

Bait again

[keen beacon]

That's just capitalism

[digital umbra]

your original point was that open source models were useless because the chatgpt free tier existed

[torn mantle]

Thanks

[keen beacon]

That sounds dumb

[digital umbra]

it is

[keen beacon]

There's more to life than ChatGPT

I use deepseek and Kimi

for example

why use chatgpt instead of aistudio atp btw

[keen beacon]

More data collection

Even more than in gemini.app

ok but youre basically already accepting theyre collecting your data

use a frontier reasoning model and make it worth it 🤣

[balmy mist]

when do yall think gpt5 is coming out?

[keen beacon]

I heard some news though that Sam Altman revealed that people say all kinds of personal info on Chatgpt

[digital umbra]

fun thing that's a requirement for using o3 then

[keen beacon]

?

That's some braindead thinking

to do

on Twitter

[primal orbit]

@echo aurora thank you for bringing rate limit notification in the direct chat! very much appreciated.

[echo aurora]

Glad to hear it!

[primal orbit]

If we could edit the message in chat and reroll, would be a great next update. Like in the Google AI Studio. Sometimes you make mistake and the chat goes off rails.

[digital umbra]

yes

A "tournament" mode where you can keep using the winning model from the previous turn would also be nice

[blazing rune]

https://tenor.com/view/michael-jackson-comendo-picoca-gif-9669437860846841235

[digital umbra]

ah yes, let me just set up a shell company in panama so i can use chatgpt without letting them know my identity

[keen beacon]

Does the EU's GDPR help in how AI companies can collect data? Just curious if people here would know more

[unborn ocean]

well idk about the specifics, but there are a lot of data collection things that are turned off for eu consumers

[keen beacon]

EU is based

[unborn ocean]

e.g. training on data with aistudio free tier

(in the api only)

no i meant the api

has a free tier

aistudio as a webapp is a different quota that is completely free, separate of the api free tier

[keen beacon]

Well, it's good I heard about that usement of data too

👍

[novel crater]

yoooooooooooooo

your parents gave you a great name

haha

[stray aspen]

wassup billy

[blazing bison]

So let's bring zenith back?

[stray aspen]

yes

[torn mantle]

you're still thinking of zenith

quite the obsession

[blazing bison]

it has the potential to be 1500 elo

[stray aspen]

I'm gonna play hugging face

[grizzled bobcat]

Guys

Help me

It limit massage

It not unlimited massage

[echo aurora]

There is a rate limit to how often you use models

[keen beacon]

Hey, I have a question.

I saw zenith got re-added, yesterday.

Can't find it in battle mode.

I'm new to all this, when can it be found again?

Anyone?

[verbal nimbus]

Talking to Gemini 2.5 Pro is a bit frustrating sometimes. It doesn't notify me that I provided the same attachment twice.

On AIStudio it likes to use flowery language for open-ended questions like it's inventing marketing terms, but it's great on STEM questions.

[zinc ore]

I'd love to believe it

[sturdy mica]

how come you can't add attachments to searching models!?!?

does anyone know a workaround or something

[keen beacon]

Anyone spotted zenith yet?

[stray aspen]

craig do you think gpt 5 will smoke all the other models

and will remain for a long time

[keen beacon]

https://x.com/hunoematic/status/1944090606625792085

invincibleHunter (@hunoematic)

To clarify, these are my predictions for GPT-5, and insider Satoshi confirms most are accurate, or somewhat accurate. Those are mostly based on rumors.

X · Jul 12, 2025, 5:44 PM

[sturdy mica]

yo how do i have attachments and internet access at the same time

cuz this low-key annoying

[stray aspen]

you dont

[sturdy mica]

cool............. :-[

[stray aspen]

make a feedback

and maybe theyll add it

[sturdy mica]

is there some free service where i could use some models like grok 4 with internet and also attachments

@stray aspen do you know of one

sigh

bzzzzzzzz

bzbzbzbzbz

aaaaaaaaahhhh

bbbbbbbbbb

[keen beacon]

Lmarena, no?

[sturdy mica]

with search AND attachments

lmarena supports only one or the other

i need both at same time

[lime coral]

Since GPT5 uses tools by default they should be compared with Deep Research version

[fleet lintel]

i need GPT5 now. when are they launching?

[keen beacon]

Early August as was rumoured on some sources

[cedar tide]

a good cleaning is nice

[torn mantle]

happy david

[cedar tide]

We need more cleaning
There are still 2 kraken, cuttlefish, clownfish, octopus, stephen

[cedar tide]

Go upvote for the best 32B model (and also the best one with fewer than 235B total parameters !)

https://discord.com/channels/1340554757349179412/1396370899342725253

[frosty lark]

elo are relative, one cannot compare between playerpools.

[cedar tide]

Ernie 4.5 is underrated 😵‍💫

Average of the 20 benchmark that baidu shared (Chinese benchmark excluded)
Upvote here https://discord.com/channels/1340554757349179412/1392140140662489108

[cedar tide]

Now that we have cleaned these 11 models, add these 10 models 😶

Qwen 30b A3b 25 07
Gemini 2.5 no think
Open reasoning nemotron 32b
Ernie 4.5 300b
Glm 4.5 no think and on webdev
Solar pro 2
Exaone 4.0 32b
Hunyuan 80b a13b
Intern S1 (241b vision)
Reka flash 3.1

[torn mantle]

look i appreciate benchmarks, but they dont reflect how the model is practically

[cedar tide]

have you tried it?

I don't think so

[torn mantle]

where can i try it

lol

https://aistudio.baidu.com/overview

let me see

[cedar tide]

https://openrouter.ai/baidu/ernie-4.5-300b-a47b

And on novita ai

[torn mantle]

the first thing i try them on is : french -> eng and eng -> french

for multilingual vibe check

if its not fluent and feels native then its a big -1

they all sound robotic and ai gen

@cedar tide whats your first benchmark

or what do you try it on

[cedar tide]

@torn mantle The truth is I haven't tried it like everyone else, but just if it has good benchmarks we should give it a chance in the arena so we can try it.

[languid crescent]

@echo aurora am so sorry 🙁 I received a warning about advertising didn't know that I can't share it :((

[torn mantle]

ive tried it

its meh

[cedar tide]

Where ?

[torn mantle]

here

[ornate agate]

I think benchmarks are still a lot better than random vibes or assuming it’s bad

[torn mantle]

no they are not

vibes check is superior

[cedar tide]

@torn mantle for you deepseek v3 is much better?

[languid crescent]

i am such a disappointment 🙁

[torn mantle]

yes

[torn mantle]

its fine

you have the same pfp picture as david

why

[ornate agate]

Exaone and nemotron AA benchmark at 32b size makes them very compelling for further analysis

[cedar tide]

Yes

but glm is mostly good at webdev and he's not on it yet

and there are only think versions of glm, but sometimes people prefer no think versions, for example qwen 3 no think is much higher than think version in the leaderboard

[tall summit]

which lmarena direct chat models have ratelimits?

[blazing bison]

The update of the chatgpt Mac app with preparations for gpt 5 basically confirmed that it's gonna be a router

🤓

[stray aspen]

Bro the Baidu ernie playground is so trash

[cedar tide]

https://fixupx.com/ArtificialAnlys/status/1950555928014844415?t=fD2hm_gS7PlfvD8_R9kljQ&s=19

Artificial Analysis (@ArtificialAnlys)

Announcing the Artificial Analysis Music Arena Leaderboard: with >5k votes, Suno v4.5 is the leading Music Generation model followed by Riffusion’s FUZZ-1.1 Pro.
︀︀
︀︀Google’s Lyria 2 places third in our Instrumental leaderboard, and Udio’s v1.5 Allegro places third in our Vocals leaderboard.
︀︀
︀︀The Instrumental Leaderboard is as follows:
︀︀🥇 @SunoMusic V4.5
︀︀🥈 @riffusionai FUZZ-1.1 Pro
︀︀🥉 @GoogleDeepMind Lyria 2
︀︀@udiomusic v1.5 Allegro
︀︀@StabilityAI Stable Audio 2.0
︀︀@metaai MusicGen
︀︀
︀︀Rankings are based on community votes across a diverse range of genres and prompts. Want to see your prompt featured? You can submit prompts in the arena today.
︀︀
︀︀👇 See below for the Vocals Leaderboard and link to participate!

**💬 4 🔁 6 ❤️ 38 👁️ 1.2K **

FixupX · Jul 30, 2025, 1:55 PM

[whole wagon]

Did I hallucinate, I swear on chatGPT the switch model option had gpt5 for a second 😂

[finite pollen]

hey in our battles, models that are removed get relabed back to Assistant A so we dont know what they were.. can this be fixed?

[jade egret]

when gpt 5

[echo aurora]

Interesting. I’ll flag to the team and see if there is a fix that’ll keep those names even if removed.

[finite pollen]

e.g

[tall summit]

which lmarena direct chat models have ratelimits?

[torn mantle]

@cedar tide https://x.com/Alibaba_Qwen/status/1950570969036361799

Qwen (@Alibaba_Qwen)

🚀 Qwen3-30B-A3B-Thinking-2507, a medium-size model that can think!

• Nice performance on reasoning tasks, including math, science, code & beyond
• Good at tool use, competitive with larger models
• Native support of 256K-token context, extendable to 1M

Qwen Chat: Go to

X · Jul 30, 2025, 2:55 PM

[cedar tide]

Already see

[torn mantle]

but you didnt share it

xd

[cedar tide]

Yes im busy now

Soon the average of the benchmark

@torn mantle officialy coder 30b a13b tomorow

[cedar tide]

Just i need to go to my pc

[cedar tide]

21 vote for create agent arena

[torn mantle]

😮

[digital umbra]

interesting chain of thought

[cedar tide]

average of the 24 benchmark

by category

@torn mantle @ancient reef

[torn mantle]

Nice

Not bad

[cedar tide]

In my opinion, Gemini 2.5 Flash is about this size.

size of deepseek r1 ?

[digital umbra]

https://youtu.be/0obMRztklqU?t=25 speculating on model size

[cedar tide]

@ornate agatethe only hint we have is that there was a gemini 1.5 flash 8b version

go upvote the new qwen thinking https://discord.com/channels/1340554757349179412/1399812659800445131

[keen beacon]

Zenith in lmarena, anyone?

Or openrouter horizon alpha?

[digital umbra]

zenith was removed, horizon is not out yet

[wicked root]

any update on GPT5?

[lapis light]

Can I just ask though, why are there three Video Arena channels?

[echo aurora]

To spread generations out a bit. If it was all in one channel it'd be a bit much,

[lapis light]

Okay, fair enough. I guess that makes sense.

[tawdry sapphire]

yo

[echo aurora]

hello

[tawdry sapphire]

wassup bro

[mint cape]

wasn't video arena already here?

[echo aurora]

It's been here for a little bit, but we wanted to soft launch it first before dropping the @ everyone

[mint cape]

I forgot to disable @everyone pings on the server and was midly annoyed 🙁

[tribal glacier]

hello .)

[brittle tiger]

Pretty crazy anthropic API revenue is higher than openai now

https://x.com/aj_kourabi/status/1950577614772662325

AJ (@aj_kourabi)

anthro’s API rev has over taken OpenAI’s

and once I internalised that, and its implications, I joined their ranks in believing code is the Only Thing That Matters

X · Jul 30, 2025, 3:21 PM

[humble sonnet]

So there's still a limit? Like the video limit per day.

[digital umbra]

you have to go through hoops to use o3 through API so it's not too surprising

[echo aurora]

Yup

[humble sonnet]

Okay

Do you have a limit with image with bot?

[void tusk]

oh boy, i bet there are going to be a bunch of new people here xd

even newer than me XD

[echo aurora]

It should all count the same regardless if you use image, video, or image to video

[humble sonnet]

Oh , but image is unlimited on website

[wicked root]

how do people know this? I can't find official statements anywhere

[reef pawn]

It's not official just rumours

[wicked root]

and is the new model going to be better than gemini pro?

I might switch over to gpt5 if that's the case

[nimble trail]

they should be.

[ionic idol]

Hi

[wicked root]

What about specifically for text?

[wintry tinsel]

I eagerly await August for gpt 5

[dusky ore]

#share-prompts create video, on a subway platform, the chubby raccoon is running away, clutching the three ducklings in its arms. Behind him, the mother duck is chasing after the raccoon, wings spread and beak open in panic. The scene is dynamic and full of action. vertical, 9:16

[echo aurora]

You need to use the /video command, this can only be done in #video-arena-1 #video-arena-2 so on. More info in #1397655624103493813

[golden ocean]

reallllll

[reef pawn]

How can I enable my Gemini AI pro membership in Google AI Studio? I already have this membership and want to use Veo 3 and Imagen Ultra but I'm unable to do so

[merry pond]

@amber warren Hi !

[amber warren]

helloo

[dusky ore]

/video create video, on a subway platform, the chubby raccoon is running away, clutching the three ducklings in its arms. Behind him, the mother duck is chasing after the raccoon, wings spread and beak open in panic. The scene is dynamic and full of action. vertical, 9:16

[merry pond]

Congratulations for intern stage at LMArena

[echo aurora]

Need to be in one of the video-arena channels, try in #video-arena-4

[merry pond]

Is there a role for Lmarena staff to recognize employees who work there and prevent people from being misled by imposters?

[echo aurora]

There isn't

[wheat hamlet]

hello guys

[merry pond]

is it planned or not?

[echo aurora]

welcome!

[echo aurora]

TBD, not having one is intentional

[wheat hamlet]

would love to know whats ur personal fav model for text, music and video

i feel like people have different preferred models nowadays

[merry pond]

Okay

[trail ember]

🤗

[stray aspen]

@echo aurora Mr why so many video arena chats

[echo aurora]

Spread out the generations, else it'd be a bit crazy

[digital umbra]

i wonder who is paying for it

[stray aspen]

We are

Ok but fr how does lm arena profit

@echo aurora

[echo aurora]

I'd encourage you to read this blog post - https://news.lmarena.ai/new-lmarena/

[shut quiver]

Um... hi there. Just joined today. Nice to meet you guys

[golden ocean]

REAL

[keen beacon]

create the videos on #video-arena-1 to 4

main chat is not the place.

[echo aurora]

welcome welcome!

[shut quiver]

Thanks! I've been using images for a while. It's honestly pretty neat

[old ginkgo]

Do you guys think meta will make a pay2win comeback?

[digital umbra]

considering how much they invested on stealing talent from other companies, i think they have to

[old ginkgo]

Yeah i think so too. I think mark will just let them do whatever. Literally 0 safeguards, just make sure you win. Also here is 100 billion dollars in salary and datacenters lol

[woven harness]

video arena will be on the web?

[echo aurora]

it's possible, be sure to share any feedback related to this in #bot-feedback

[wintry tinsel]

The question is will it matter? This is a massive gamble really, and there is a strong possibility of it not being good enough to justify that huge expenses other companies already have their teams assembled the dynamics established and the momentum rolling, it’s a little late to come in and do something new by 2025..

To me it feels do or die, they think their future value as a company is banking on this promise so they are willing to burn all their money on the chance they succeed regardless of how high that chance is, even if there’s a real possibility it doesn’t pan out

[keen beacon]

I love this

But how much does this cost to run?

I wonder if companies gave credits to these guys to advertise their services

[verbal sorrel]

You guys need to turn off sound so its not a dead giveaway that the model is Veo 3.... 🤣

Also doesnt seem like an apples to apples comparison at the point anymore too

[echo aurora]

Fair feedback! I'm going to move this to #bot-feedback so we can keep it all in one place.

[torn mantle]

Do it

Ah you already did

Free money

Hax

Smh

[patent aspen]

It is apples to apples. The other video models just lack sound and should be punished for it accordingly

[torn mantle]

Yea

You need to look for arbitrage

Between diff platforms

[verbal sorrel]

Poor take

[torn mantle]

Hax

You have a script for it?

[verbal sorrel]

You arent measuring video models anymore at that point

[torn mantle]

Or Is it manual

Liar

Smh

[patent aspen]

Sure you are. All video models except one have sound though

[verbal sorrel]

The entire point of LMArena is blind testing, if you know the model is Veo 3 right away then it defeats the purpose.

[verbal sorrel]

Thats the point...

[torn mantle]

Craig just let it out

Don't worry

You are safe here

Im skydiving rn

Wish me lick

Luck i mean

[verbal sorrel]

For LMArena to be effective you need to remove bias, which is why its blind. But if you know the model is Veo 3 out the gates that obviously doesnt work anymore.

[patent aspen]

It's too bad the others don't but it's not like this is the first time it's been possible to deduce what a model is based on its responses

[torn mantle]

Orabazes is mad

[verbal sorrel]

True but this is just too obvious and detrimental to testing

[torn mantle]

Angry

[echo aurora]

same same

[verbal sorrel]

Im not mad just trying to make it a better leaderboard for everybody

[torn mantle]

I have never used any gambling website

Like ever

[echo aurora]

It's valid feedback for sure

[patent aspen]

It would be silly to punish the best video model because the other models don't have feature parity

[real whale]

/image-to-video /image-to-video

[echo aurora]

Need to use the video arena channels, like #video-arena-4

[patent aspen]

Worse video models should also be incentivised to compete for user preference holistically

[real whale]

Thanks you @echo aurora

@echo aurora I'm little nob on this site

[visual panther]

yes

[merry reef]

hola

[blazing bison]

Bring zenith back

😢

[deep valve]

hello

[fading kraken]

hello

how can i check out the code named models

clownfish, nettle, etc

[tired herald]

Can't directly, you need to go to the battle and have luck ig

[fading kraken]

ah got it

[warm fulcrum]

when gpt-5 releases, do u think they will bring it in lmarena?

[fading kraken]

no

[tired herald]

Maybe to battle for a day or two, but not to direct chat

[warm fulcrum]

i dont want to pay openai 20$ 😭

they got enough money out of me already

[tired herald]

I can't help you with that

[warm fulcrum]

im just saying

[tired herald]

And I'm just saying too 😭

[warm fulcrum]

👍

[tired herald]

You'll probably still have some free messages tho

On OpenAI

[warm fulcrum]

they would water it down hella tho so

[tired herald]

Very likely yes

[warm fulcrum]

im so jealous i never got to try zenith ngl

[tired herald]

So am I

[warm fulcrum]

🤷

[tired herald]

Just have to wait then ig

[tall summit]

they might

gpt-5 will be huge no matter what

even just by reputation

it'd be very idiotic of them not to

[digital umbra]

gpt-5 will be huge because it's the first openai model released in a long time with a name that actually makes some sense

[blissful sluice]

Its. Called gpt-five

[ocean vortex]

AGI confirmed

[digital umbra]

the first AGI model will be called gpt5.1o-max-pro-alpha

[civic flame]

i'm going for no

🍿 this next 7 days is going to be HOT

[stray aspen]

No

[civic flame]

in demis we trust

[zinc ore]

InB4 drop at the same hr

[ocean vortex]

lmfao. Ultra 1.0 when

[blazing bison]

English only

[silent flume]

new model in arena

potato and dino

[digital umbra]

dino claims to be by anthropic

and potato by openai

[silent flume]

they are good?

[coarse flame]

Potato keep trying to use imgur link

[digital umbra]

they might also be distilled from those companies, apparently

no idea if they're good or not

[ebon jacinth]

bros gpt 5 tomorrow?

[long depot]

Hi, new here! I'm curious to know if anyone has measured whether people are inherently more likely to pick option A or B in the arena, because of recency bias. I know that when I get long responses I read through one and establish an opinion, then read through the other but can't help comparing as I go, which might bias my vote.

[bronze veldt]

hello

[echo aurora]

Hello! Welcome! Our blog here has articles you may find interesting. https://news.lmarena.ai/. Iirc there was a section related to recency bias. I'll double check with the team and let you know.

[whole wagon]

https://openrouter.ai/openrouter/horizon-alpha

This is OpenAI

256k context, smells like open source model

[digital umbra]

i had a suspicion it would be that

if it's the open source model, 256k context would be nice

[whole wagon]

They aren't even hiding it tbh

Like didn't even bother with the system prompt lol

[whole wagon]

That's a tokenizer issue I'm pretty sure. There's some explanation to why it's a crap test

Same issue caused the r in strawberry thing

[digital umbra]

it doesn't seem like it's a reasoning model though

[whole wagon]

Could be they just turned off the reasoning

[digital umbra]

maybe

[whole wagon]

For the demo

[digital umbra]

hm

it doesn't seem to be terrible at the one trivia question i asked it

which even sonnet 4, deepseek and glm-4.5 fails at (but kimi k2 gets right)

[whole wagon]

The training cutoff date is strange

It's long ago for some reason

Hm well the open source model was delayed a lot. So maybe it does add up

The openAI models on LM arena know the current president

[digital umbra]

qwen3 says biden is president too

[whole wagon]

Yeah. It doesn't instantly learn the cutoff date has to be like may onwards

For it to say trump is president

[digital umbra]

i think it's a large model, too big for 1 gpu

[torn mantle]

@cedar tide is probably sleeping and missed horizon alpha

[cedar tide]

Jamais

[torn mantle]

lmao

xdddd

[cedar tide]

He say october cutoff but he know deepseek r1

[torn mantle]

its probably the open sourced model from openai

deepseek r1 update could be potato

new model added to lmarena

[digital umbra]

it is actually really good at trivia.

close to gpt 4.1 for sure

[torn mantle]

people are not liking this horizon alpha model at all

makes me wonder if deepseek really hit a wall or nah

potato was ok-ish but nothing crazy

[digital umbra]

potato isn't horizon alpha

[torn mantle]

yea ik ik

potato could be a chinese model

horizon alpha is def from oai

[runic axle]

From initial testing Horizon Alpha has the same writing style as Zenith/Summit

[torn mantle]

yea coding wise, it has similarities to summit/zenith but not that good tho

[digital umbra]

it could be a gpt5 variant for sure, i'd assume it would be for the free tier of chatgpt in that case (replacing 4o)

[runic axle]

I meant for creative writing. IIRC the Zenith/Sumit models in LMArena had a thinking/reasoning budget, but Horizon Alpha doesn't.

[torn mantle]

surely not a gpt5 variant

[digital umbra]

it also makes sense if it's the open source model if kimi was indeed the reason why they delayed it, because based on how it answers i think it's probably in the same size range, around 1T parameters (and yes i know kimi is kinda bad for its size)

[torn mantle]

nah

the one who got access said its much much smaller

can run in a single H100(?) gpu

[digital umbra]

if it's dense it would be much smaller

[torn mantle]

i dont know if he meant H100 or B-serie

you mean moe?

dense it will just activate the whole params

[digital umbra]

kimi is moe

[torn mantle]

yea it is

[digital umbra]

so i think if it's a dense model it would surely be smaller for the same performance

[torn mantle]

yea could be

[runic axle]

If this is the open-weights model maybe it was distilled from GPT-5?

[digital umbra]

nah

distilled from gpt-4 variants

[runic axle]

Yeah probably more likely

[torn mantle]

this is also possible

you know what

everything is possible

lets just wait and see

[cedar tide]

Horizon alpha

[torn mantle]

the video needs more work

but gl

i would just brainstorm ideas -> run it on notebooklm video overviews

and make a similar presentation

you are a cutie paws

😖

its a good one

👍

[cedar tide]

https://fixupx.com/Angaisb_/status/1950722641314066607?t=nCMQDnM-riw8eZcmAScfew&s=19

Angel Bogado 🌻 (@Angaisb_)

Doodle Jump test
︀︀
︀︀Horizon Alpha

**💬 1 🔁 1 ❤️ 7 👁️ 274 **

▶ Play video

FixupX · Jul 31, 2025, 12:58 AM

[hardy pecan]

quick pass of simplebench for Horizon-Alpha: 3/20 lmao

[finite pollen]

its probably that OSS one they keep hyping :p

[ebon jacinth]

it could be GPT5-nano

[lapis light]

Reminds me of the Google Lamda moment

[misty vault]

stop

[paper nimbus]

where can i track new models

[mild vapor]

Does "LMArena" not support setting the Aspect Ratio when creating images? I've given the commands as detailed as possible, but the result is still a 1:1 image.

[echo aurora]

We don't currently have this functionality; however, this is very much on our radar.

[gritty flare]

What is it/

[mild vapor]

In "video-arena", is there a limit to the number of videos can make?

[echo aurora]

Yeah

[mild vapor]

how much? and daily reset?

[echo aurora]

Yup daily, it's currently @ 8 but we may change it.

[junior shuttle]

hallo everyone

[languid crescent]

Will video arena be ever in lmarena? Or it just stays in discord for good?

[echo aurora]

That's TBD, that's why we're considering this experimental. Be sure to use #bot-feedback to let us know what you'd like to see happen!

[languid crescent]

GPT just released a new mode, will this be possible incorporated in LMarena?

https://openai.com/index/chatgpt-study-mode/

[languid crescent]

probably not 😭

[nova frost]

is there a option to use veo 3 in generating a video because i like it when sound effect is available.

[echo aurora]

No it's battle mode only atm, there isn't a way to select a specific model.

[nova frost]

how many credits we generate a video here?

[echo aurora]

It isn't going to be consistent, but it's currently set to 8 generations a day. Note you can only do so in the video-arena channels like #video-arena-3

[nova frost]

ohh thank you so much

[echo aurora]

no problem!

[whole sundial]

it's just a system prompt, it's easy to find on the internet and you can just make it the first message

https://www.reddit.com/r/ChatGPTJailbreak/comments/1mcqe7l/chatgpt_study_mode_system_prompt/

[languid crescent]

you can do it by prompt engineering itright?

[whole sundial]

well, that's how they made the "mode" lol

[languid crescent]

oh damn 😭

@echo aurora it possible for the video generation arena to send the result directly to the person who prompted it? (Assuming it's not a chatbot you can interact with.) The idea is: you type your prompt in the #video-arena channel, but only you can see the generated video result? idk if discord can even do this 😭

[echo aurora]

Both in DMs to the person who generated & the server, or just DM? Be sure to share this in #bot-feedback

[languid crescent]

gotcha imma share this to #bot-feedback thanks!!

[limber kiln]

Hello ! Best wishes for all.

[tight sedge]

camping value video

[echo aurora]

Note the #1397655624103493813 channel will give you info on how to use the bot

[primal fern]

How to generta evideo in this channel

[echo aurora]

Info in #1397655624103493813

[nocturne bear]

@icy forge great prompt buddy

[tight sedge]

A serene lakeside campsite at dawn, golden sunlight filtering through pine trees. A tent is pitched near the water, with a small campfire smoldering. A coffee pot steams on a rustic wooden table. Slow drone shot moving from the lake to the campsite."

[echo aurora]

#1397655624103493813 has info on how to use the bot.

[pseudo summit]

could it b a grok model?

haven't gotten to mess w/ it too much, but some of the responses i got were similar to past grok responses

[pseudo summit]

this is not consistent for potato btw, not sure if that points towards chinese model

[hallow ridge]

How can I use AI to take ove r

[halcyon vortex]

yo all of my chats just got wiped...

[still bramble]

Yeah, same

And this error now appears

[halcyon vortex]

DUDEEE IM COOKED MY GAME SYSTEM RELIED ON ITTTT

[pseudo summit]

interesting

observed some very weird behavior by models right before the error

maybe they're doing maintenance or smth?

[gloomy knot]

hello everyone, joining here to try out the video arena 🙂

[halcyon vortex]

hopefully

[still bramble]

rip(

[halcyon vortex]

I think the devs are trying to fix lmarena tho

they better 🙏

[nova edge]

hello

[echo aurora]

I'm not able to repro this. Do you know if this is only happening on mobile?

[echo aurora]

welcome! be sure to check out #1397655624103493813 for more info!

[pseudo summit]

no, i have same error on pc

[echo aurora]

just with gpt-4.1 or are all models giving you this error?

[pseudo summit]

all models

nvm, seems to be back!

[echo aurora]

glad to hear it! but keep me updated if things seem broken again

[velvet iris]

hi

[void perch]

can i invite the bot to my private channel, easy lost track in public chanel

[flint mortar]

they lowkey did me a favor had too many i didn’t use lol

[calm sequoia]

The o3 suddenly started showing code changes without any actual differences. Anyone noticed this?

[still bramble]

On PC everything is ok (chats are there), I just checked. On smartphone the error went away, but chats did not return.

[pseudo summit]

chats did not return for me on PC ...

[torn mantle]

https://x.com/xai/status/1950828488405254268

xAI (@xai)

xAI supports AI safety and will be signing the EU AI Act’s Code of Practice Chapter on Safety and Security. While the AI Act and the Code have a portion that promotes AI safety, its other parts contain requirements that are profoundly detrimental to innovation and its copyright

X · Jul 31, 2025, 7:58 AM

[still bramble]

I didn't check the chats on PC right away, only now. Maybe the bug that caused this is gone now, idk.

[wind briar]

hello, do you know any way to jailbreak gpt? like give you information it shouldn't give you ( tax fraud, fake ids, etc)

askin for a friend😅

[halcyon vortex]

@echo aurora Will I ever get my chats back? They just dissapeared randomly

[still bramble]

This article looks pretty convincing (https://habr.com/ru/articles/923084/), I once added it to my bookmarks, but never tried the advice from there. It is in Russian, but the screenshots with examples of all jailbreaks are in English, so I think everything will be clear.

[golden onyx]

GM beautiful people

[paper nimbus]

gippity 5 wen chat

[dusky aurora]

I remember when it was still habrahabr

[civic flame]

Tuesday

[weak sluice]

Hi!

[paper nimbus]

a lotta ppl sayin July 31st tho

https://tenor.com/view/miau-adobe-after-effects-glass-breaking-default-preset-gif-5777596260912161139

[primal sun]

hi

[main gulch]

he just uses pliny jbs

[blazing narwhal]

hi

[keen beacon]

Hello

[teal crypt]

Hi

[quiet pollen]

Been hearing that Gemini got nerfed, is it true? But when I see the leaderboard, they are still first in most of the fields

[deft vigil]

is there any unreleased model on lmarena right now ?

[cedar tide]

Well, I was busy with Horizon. What are Dino and Potato worth? Is it worth going to the arena?

[deft vigil]

im on dino now but it really slow though. horizon is openai model ?

[torn mantle]

im really starting to think deepseek is ded

[civic flame]

they're both mid

chinese distils by the looks of it

[cedar tide]

stephen
kraken
kraken 2
folsom
nightride
nightride v2
dino
potato
octopus
clownfish
cuttlefish

[civic flame]

cuttlefish and clownfish were removed

[cedar tide]

reasoning ? deepseek ?

[civic flame]

no lol

[deft vigil]

wow the only way to access it through lmarena battle right

[cedar tide]

not reasoning ?

[torn mantle]

people on chinese forums/servers are talking about anothe 2 months delay for deepseek r2

[civic flame]

no

[civic flame]

deepseek are kinda cooked

[cedar tide]

@torn mantleWell, I slept, what did I miss on Horizon?

[torn mantle]

they hit a wall and they have some technicall issues

[civic flame]

lol they probably realised they're too far below gpt-5 if they release r2 soon

[torn mantle]

i slept as well

i tried it on 1 prompt only

[civic flame]

tldr is that it's a very small model good at code and svgs but not much else

terrible at math

[torn mantle]

idk its only good at svg

[deft vigil]

Dino is keep on generating crazy

[cedar tide]

yes i know all this

[civic flame]

then why ask what you missed 😭

[torn mantle]

its the same thing, when something became trendy they will finetune like crazy on it

and thats what happened with svg

they are focusing on the wrong thing

[cedar tide]

@civic flamehave people run benchmarks?

apart gpqa and math 500

and arc agi

and eq bench, creative writing

[civic flame]

it got ~67% on aider

[torn mantle]

the thing that we should focus on is aristotle x1

[civic flame]

which puts it just below claude 4 opus

[cedar tide]

yes i saw this too

[torn mantle]

https://x.com/ai_for_success/status/1950826316712013843

AshutoshShrivastava (@ai_for_success)

92.4% on GPQA Diamond and 96.1% on SimpleQA, what the heck?

Aristotle X1 Verify is the new AI co-scientist from Autopoiesis that topped one of the toughest scientific reasoning benchmarks GPQA and scored 96.1 percent on SimpleQA.

No moat for big AI labs, lol.

Will GPT 5 really

X · Jul 31, 2025, 7:50 AM

[civic flame]

what the

never heard of that

oh wait

[torn mantle]

isnt it the math model

[civic flame]

that's aristotle

[torn mantle]

that acer was using?

[civic flame]

yeah

he said it's definitely not "mathematical superintelligence" as they were selling it

[torn mantle]

https://autopoiesis.science/blog/92-4-gpqa-diamond

[rare python]

do you have link?

[civic flame]

apparently it's only good at very specific kinds of math

[torn mantle]

they blog can have some insights

their*

[rare python]

Can I use it?

[torn mantle]

Our system achieves state-of-the-art performance on both benchmarks because we've built systematic verification directly into the reasoning process. Rather than emotional doubt, our models apply procedural self-skepticism to their outputs, making the skepticism reliable and scalable rather than unpredictable.

Other models embody the opposite of scientific thinking. They're trained to sound confident about everything, when science requires knowing when you don't know.

[torn mantle]

not yet

seems to me they are looking for more fundings

typical small business -> big business

[civic flame]

https://aristotle.harmonic.fun/

acer and another guy got a testflight invite pretty quickly

[torn mantle]

im not sure if this model is powering their app or nah

[civic flame]

it is

[torn mantle]

yea that one

kinda curious if you can just ask it something unrelated to math

like inject some random math question inside an actual question

[rare python]

Oh I thought this one is more general

[torn mantle]

well they said they saturated both benchmarks