general

[ocean vortex]

it's at a clear disadvnatge againt competition in some areas, and this includes all their reasoning models as they are based on it

I wouldn't go as far than that, as the trend lately has been downsizing... So maybe in the long-run the trend/progress will catch up to it and it's not gonna be too small. But for now it does seem suboptimal as people just can't get spatial awareness at a top level without decently big model size

though we also need to keep in mind that we just do not know what could potentially be done with RL training using notably bigger model. Maybe it's diminishing returns, but maybe the gains are actually more substantial

[lavish orchid]

anyone else have the problem of Gemini 2.5 Pro not using terminal in Cursor?

Claude 3.7 runs it 20 times and does everything I ask it, Gemini 2.5 Pro is unsure and asks to install libraries etc

prompted it in user rules too no change

[ocean vortex]

it would unlikely to scale the same way as normal classic chat LLMs

so I would think "perfect size" for that is something different (bigger), even with the current metrics and their limitations

[tall summit]

WHAT

AAAAAAAAA

[small haven]

wen

u just want to see the world burn dont you 😭

[ocean vortex]

how is this real... 😭💀

[small haven]

it was half that last week

o1 pro is unlimited lessss gooo

[wintry locust]

for o1 it's called juice

[ocean vortex]

to my surprise I'm getting the same 8192 score nonsense on playground with completely empty user defined system prompt for o3-high. Now that's 1 way to ruin a model...

[balmy mist]

we still aint got no new stuff?

[ocean vortex]

people gonna have to worry about this stupid yap score now when trying to eval openai models 💀💀

[balmy mist]

whar hppened to r2?

[leaden palm]

there were rumors about it today

probably wont be released super soon tho

polymarket: 9% chance in apr, 90% chance apr/may/jun

manifold: 25% chance in apr, 88% chance in apr/may, 93% chance in apr/may/june

[ocean vortex]

there are only 3 days left in April though lol

[leaden palm]

@keen beacon alt?

[keen beacon]

ive never seen a model do this 🤣

[oblique flint]

It still kinda sucks in cursor unfortunately, just not as good at toolcalling

[small haven]

that discrepancy in apr

how tf is not being arbitraged

[keen beacon]

@olive mesa hi twin

[leaden palm]

ah forgot that exists

[knotty jetty]

Is there a reason why they don’t include Claude’s web search function for the web search leaderboard

[keen beacon]

yea to make sure it hits 1345 words. i didnt request that 🤣

[knotty jetty]

Claude.ai but premium

Ok

But it’s only in premium just so you know

What do you mean

The lmarena api

Ok

So do you know why they haven’t added it

Oh

No it’s really good

I have perplexity pro and I feel like Claude has gotten many more answers right

For real

Oop

Well my dad uses perplexity too

Should I tell him to stop paying for it

[elder rapids]

it's crazy how media is so far off from the reality of these products

and fake news is the reason why a lot of them still live

perplexity should ong be dead rn

dawg

why are you saying it's a scam

because you agree with me

😭

ye

that's what I'm saying

[keen beacon]

u used to get like 600 messages a day on perplexity

[knotty jetty]

My dad uses its API, he runs a music player company and he uses it to find creators numbers and their emails and tells them stuff. Is he cooked or nah

[keen beacon]

with a lot of models, even if u dont use it for search

[elder rapids]

nah

there's always going to be replacements

in the AI industry

maybe even made by openAI

[knotty jetty]

No like do you think it’s getting it wrong

[elder rapids]

oh

depends

that kind of stuff is hard to get wrong

especially for an LLM

[knotty jetty]

Ok thank god bro

[keen beacon]

u cant disable it now?

thats dumb

[knotty jetty]

My dad has been working on ts for 10 years and I’m scared ai is gonna screw him up

[elder rapids]

nah it's alr

if he adapts

he'll be elevated

there's not much downsides

still depends

[knotty jetty]

Dw he’s already a millionaire but I’m just scared he is gonna be sad if this idea fails

[elder rapids]

since it'll get to a point where not taking jobs

becomes unethical

[elder rapids]

he could be impacted yeah, but he might find a whole new interest in AI itself

who knows

ong

[earnest parcel]

gemini is absolutely not impressed by sonnet 3.7:thinking chess moves it seems 😄

[elder rapids]

if you ask 2.5 pro to reason through the spatial task of the board + think like a grandmaster and focus on recollection, itll stomp these other models with the same prompt

[knotty jetty]

He’s in the screen burning industry and he doesn’t use ai for that but now he’s branching to a music player website and app which uses ai for song recommendations but I think he will be fine even if it fails because he is one of the leaders in screen burning services.

[elder rapids]

it seems like the best chess model currently tbh, though as base

the gpt models

might be the best

especially 4.5

[earnest parcel]

i am currently running them against another, with full info and full reasoning, so I'll see. definately a huge difference in behaviour though, gemini is confident and claude is constantly passive/doubting itself

[elder rapids]

but when you tweak them

2.5 pro is the best

nahhh

[knotty jetty]

Are u talking bout studio

[elder rapids]

the industry doesn't allow that

[elder rapids]

AIstudio yeah

but I mean

prompting wise

not temp control

and other dev stuff

[knotty jetty]

Temp is also important tho

[elder rapids]

nice

he'll find his way

[knotty jetty]

Yeah

Disappointing

See this is why screen burning is gonna fail bro

[elder rapids]

ye but when it comes to thinking models this becomes more arbitrary, especially puzzle tasks and spatial tasks like chess

[knotty jetty]

Yeah I only use it for research tasks so I don’t face those issues

[knotty jetty]

https://arenaprints.com/pages/pre-burned-screens-screen-burning-services-and-screen-print-screens

Proudly black owned business if that interests anyone at all💀

[elder rapids]

send that to him not me 💯

[knotty jetty]

Bro come on

[keen beacon]

yea 🤣 qwq 32b preview

[knotty jetty]

Blacks I wild, just call us black people bro😭

[elder rapids]

why proudly smh

[keen beacon]

they added a sh1t ton of rl on top of it probably more than qwq full tbh

[meager sun]

👹 evil

[knotty jetty]

Dw I don’t really care

[elder rapids]

deadass

can't even say type sh**

you can suspect it, but not reasonably believe it imo

they're not in the same position DeepMind is

they don't have the data scientists

they don't have the researchers

ye but I think we can assume they have the researchers, but not insane data scientists

nah not in comparison

just as it is

ye ofc

but it's not necessarily sacrifice

in the way it's suggested

deepmind

anthropic

private institutions

universities

ion think anyone who works at these companies primarily subscribe to the ideals

ye

I would say only the really top guys

that represent those ideals

which is inherent to the ideology themselves

[keen beacon]

twink

[elder rapids]

I mean, if I were a standard worker

I wouldn't care about these things

I'm trying to work hard and get research in lmao

for money

[elder rapids]

since generally, specific researchers aren't valued in the sense they can continually output high level stuff

but can output quality for the direction the company intends

ion know about the specific situation too much with Ilya

but that's prob what happened, and he likely shifted

[ocean vortex]

no "secret sauce". Just a head start when it mattered + userbase and some really smart engineers. Funding helps as well ofc

[elder rapids]

this is the same thing someone sent earlier lmao

"(This content is from public information and is for reference only and does not constitute investment advice) Investment is risky, please be cautious when entering the market!"

just in a different format

or actually prob where they got it from

[keen beacon]

damn u know chinese lol? or just guessed it immediately

[elder rapids]

guessed it immediately

but I know a little Chinese

[keen fulcrum]

I believe this to be the case
lets await next week and hopefulyl get some news
R2 and Qwen 3 are imminent to release soon

[keen beacon]

the qwen 3 release seems to be significant, they did llama cpp prs/transformers prs/vllm prs/mobile apps/etc far before the release of qwen 3

[keen fulcrum]

I believe Google will drop theirs soon after R2

[keen beacon]

im hoping they release a qwen 3 reasoning model off the bat, but im most excited for new pretrained base models for fine-tuning, etc. qwen 2.5 was exceptional

[keen fulcrum]

The coder models

[elder rapids]

it literally says it's not a leak, just an accumulation of already public information

which undermines it's value as a concept stock

since it's not new Information

[keen beacon]

yea theyre releasing smaller models too

maybe a 32b alternative but moe so itll inference faster

[elder rapids]

I truly don't think it's going to get that much better from deepseek

[keen fulcrum]

Oh indeed browser integrated llms soon to be the next thing

[elder rapids]

get rid of 2.5 pro, get rid of o3s and o4 minis release

let r2 release

do you seriously think the gap would've become THAT wide

[keen beacon]

i dont know what to expect with r2 tbh

[elder rapids]

without those 2 crazy releases

nah

[keen fulcrum]

There is the possibility in the room qwen 3 will outperform R2, lets see

[elder rapids]

I can't believe deepseek would've accomplished that

[keen beacon]

i dont think r2 will outperform 2.5 pro at least in simpleqa i think

[elder rapids]

let alone at the level of o3

[keen beacon]

i use 2.5 pro on stuff that requires a lot of world knowledge/niche world knowledge

its exceptional (compared to other reasoning models)

[elder rapids]

especially when r1 wasn't really that good

[keen fulcrum]

R1 forced the industry to release their newest models

[elder rapids]

that's straight up propaganda

😭 🙏

[keen fulcrum]

Why?
AI got significantly better as soon as R1

[elder rapids]

nothing occured when deepseek released that was meaningful

that's literally impossible

the time period is too narrow

that means they weren't planning on releasing o3 mini after the announcement

and it takes a ton of time

for them to prepare it

won't release it on a whim like that

unless it's truly done

especially with how integrated it was

take a look at 2.0 flash thinking

lol

same thing

[keen beacon]

it defo was the reason and the reason why they started working on improving the reasoning summary

[elder rapids]

this is exactly the only thing they did tho

[keen beacon]

i dont recall r1/o3 mini timelines that much tbh ive no idea about timeline

[elder rapids]

since people were whining about it

you cant attribute any of these AI things to the release of deepseek r1

timelines don't add up

• understanding what even goes on behind these ai

and what entails the integration

[keen beacon]

vision model and btw thats a bad way to test lol

[brittle tiger]

grok has good team from what i can tell. i heard they pay way more than other labs, basically a working for elon tax

[elder rapids]

probably ye

but

the thing is, they're not as old

as other labs

[torn mantle]

I cant get enough of o3

[elder rapids]

and that's a serious factor

[elder rapids]

I can, ts cannot comprehend a thing im saying 😭 🙏

jkjkjk

nahhh

I think it's actually starting

you cant get the jump from 2.0 flash thinking to 2.5 pro

without a major breakthrough

oh

wait wym?

era of huge growth

oh ye

if r2 doesn't close the gap tbh

I can reasonably assume

open source is going to be pretty bad

for a little while

until they come up with something

[keen beacon]

nah the qwen team will deliver

its their time this time

hopefully

im not sure about deepseek but the qwen chat website was updated with strings of a qwen plus sub with video gen, image gen, access to qwen 3, etc

[knotty jetty]

Yo what ai should I use for research rn

I’ve tried everything aside from deepseek tbh

No like not deep research

Just like general search

What’s tavily?

How do I use it

[keen beacon]

bruh if u pay those prices lol

paying for o4 mini/o3's api and enabling first party web tools, etc on the api is more reasonable tbh

[ember rapids]

Sam did say r1 made them move several releases

I wonder what impact r2 will have

[earnest parcel]

https://lichess1.org/game/export/gif/white/mLb7m0Zn.gif?theme=brown&piece=cburnett

The game between Gemini 2.5 Pro Preview and Claude 3.7 Sonnet Thinking finally concluded, and with 8 and 7 blunders respectively, ultimately ended in a draw!

[elder rapids]

grok or AI studio 2.5 pro grounding

[elder rapids]

probably to be expected without any prompting

I got 2.5 pro to play at around 1900~ ish

as that's below my elo

[small haven]

bro what is happening to chatgpt, everything is 64k context max

[willow grail]

Cervical Spine Risk: Rotating your head 180 degrees is generally not recommended. It puts significant stress on the cervical vertebrae, discs, ligaments, and potentially the vertebral arteries that run through the neck bones to supply the brain. Doing this forcefully or if you have underlying neck issues (even unknown ones) could risk injury, nerve impingement, or (rarely) vascular problems causing dizziness or pain.

[olive mesa]

hi!

[earnest parcel]

this is with prompting, and also there is no way 2.5 Pro (or any language model) comes even remotely close to 1900 ELO. I have tested and played matches and tournament around 200 times by now (using all types of different methods), and the strongest any LLM ever came was GPT-3.5 Instruct in movetext continuation (aka Chess notation recall from training data). Other language models play more in the 400 Elo range, even SOTA.

[keen beacon]

where were you 💔

[olive mesa]

my discord was muted 💔

[keen beacon]

don't leave me like that smh..

[olive mesa]

sorry lol.. ill try not to in the future

[keen beacon]

(i'm joking)

don't stress 😭

[olive mesa]

lmao ok 😭

[neat apex]

In my calculations gpt 4o have 1000 and o3 mini have 1400 at THE MOST (likely 250 lower)

Rellying a lot in book moves or gimmick moves anyway

[earnest parcel]

tell me where to play against 1000 elo gpt 4o, and I show you its not nearly 1000 elo

[neat apex]

A yes, at bullet

I am saying the most optimist scenario ever

Who said 1900?

[neat apex]

Putted some games i found at the analisis and it said 1000 elo

For some reason at defense it holds very well

[earnest parcel]

which analysis?

[neat apex]

That chess.com one what says the estimated one

Just like the gif shows, at developing it is great but at end game it trash out

It must be the reason it gived an high value

[earnest parcel]

I don't know of any system that can take a game, and determine "ELO" based on it.... that would be super inaccurate. Elo is based on your opponents strenght, and the outcome, not on how good your moves looked in isolation..... Either way, I have tested a ton, and recorded a lot of games, and most SOTA models play around 400 ELO level (when compared to Lichess opponents), and are unable to beat the weakest Stockfish 14 level (sub 800 ELO)

[neat apex]

Oh

Whata twist

I think the 1000 elo number is somehow accurate but limited to hard macths since chess youtubers played against

And for some reason it plays way better at middle game, likely due memorizing moves

[elder rapids]

yeah this isn't true at all lmao, minimum 1400 if you prompt it right + urge it to use opening repertoire, emphasizing move recollection, these models can easily be that good. can you give me your prompts lol

you either don't know what level they really are playing at with lack of experience in chess or you don't know how to prompt

has to be one of those two

[earnest parcel]

talk is cheap, give me your 1400 minimum ELO prompt, and I can directly show you its not 1400 ELO.

(also you just lowered your treshhold by 500 ELO, impressive)

[elder rapids]

yk what just get up a game lmao

[elder rapids]

no?

that's an entirely different claim lmao

I'm not saying 2.5 pro plays at that level

[raven void]

GPT 5 at home

[elder rapids]

I'm saying if you urge recollection, it'll be at least 1400

not restrictive to 2.5 pro

[earnest parcel]

#general message

I got 2.5 pro to play at around 1900~ ish

[elder rapids]

and the longer it goes

it'll definitely deteriorate

[elder rapids]

great now can you quote the claim made in the larger passage

[earnest parcel]

really?`so I can get any model to play 2k elo also. they played E4

[elder rapids]

no, it loses too much context

[earnest parcel]

well unlike you I already provided I have collected data (169 games as of now, between multiple modes). would love to see anything despite baseless claims about your 1900 or 1400 elo LLMs

[elder rapids]

c5

[elder rapids]

🤷

just saying

not a lot of people are that good at understanding models

entire runs can easily be invalidated if you don't adjust prompt techniques respective to the model

[earnest parcel]

i am not interested in troll statements. either provide proof for baseless claims or I got nothing to discuss with you

[elder rapids]

wym?

I thought we were playing it already

that's why I said c5 lmao

[earnest parcel]

saying c5 is not proof of LLM playing 1400 or previously claimed 1900 elo.

[elder rapids]

I'll just send all the thought processes

and outputs

lmao

it's not that deep

dude actually blocked me lmaoooo

😭

anyone want to go against 2.5 pro?

just for fun, and for the sake of testing

[keen beacon]

They need to delete this new 4o personality

[elder rapids]

ye

it's creative

but saying anything to it poisons the well

adjusting to the user is cool, that's what I like about 2.5 pro

but goddamn

I have to prompt it everytime I talk to it

to be the way I like it

[keen beacon]

2.5 pro is great to talk to in comparison. They are trying way too hard with the new 4o etc. I didn't think they'd keep trying to force it this hard

[elder rapids]

I'm seeing people say they really like how warm it is

and how it's better than 3.7 sonnet

etc

but it's kinda of surreal

[keen beacon]

Maybe they like the sycophancy lol

[elder rapids]

knowing how synthetic 4o is

compared to the sonnet models

[leaden palm]

where

[small haven]

what they need to do is release

o3 pro

[elder rapids]

subreddits and stuff

[elder rapids]

ong

back to the pro plan I go

🙏

[small haven]

its gon be worth it

its worth it rn but i want worthier

[leaden palm]

if o3 is deep research lite whats o3 pro

[elder rapids]

if Google releases another model after that tho

ion know what I'm gonna do

imagine anthropic releases 4.0

[leaden palm]

May 20

[small haven]

o4 mini is deep research lite lol

o3 is deep research

[elder rapids]

ye io

[small haven]

o3 pro is deep anus research

[elder rapids]

have you guys tested the deep researches enough

I don't use the Gemini one or the openAI one that much

[small haven]

yes

[elder rapids]

so I'm not sure which one is better

[leaden palm]

somehow my stupid scaffold with exa and gemini flash+pro outperforms all the other free ones

[small haven]

oai deep research is meta rn

nothing beats it

[elder rapids]

damn

[elder rapids]

have you tried the Gemini one?

or nah

[small haven]

yes ..

obviously

[elder rapids]

you have to get the subscription ig

to even use it

[keen beacon]

What do y'all usually use deep research for

[small haven]

its free tiral

*trial first month

[elder rapids]

oh ye

[small haven]

but it just spits out mumbo jumbo, not wurf

[elder rapids]

wym?

its done what I've asked it

very well

[small haven]

oai deep research has high entropy info for every sentence

gemini dr is just stale and a bunch of unneeded detail

[elder rapids]

alr give me a prompt

[small haven]

like any prompt

[elder rapids]

actually nah

I'll just read back

and do one for 2.5 pro with the same prompt

[keen beacon]

Hmm what are you expecting out of "demo results" btw lol

[leaden palm]

it just loads a pregenerated one (im the one who set up that ui)

[elder rapids]

I disagree a ton with the density of info, but the formatting seems more consistent in openAI DR

they seem too similar to compare on that part, or its unnecessary comparison

since there's necessarily limited amount of info for a topic

[small haven]

u are pure coping, its not

gemini dr is like a student trying to fill up the minimum word count

[elder rapids]

wym coping? if I were asking about that I wouldn't be dismissing what you said lmao

[small haven]

oh mb i read it wrong

[elder rapids]

I'm asking which one is better, or which one sufficiently describes the information, not summarize it

[small haven]

sorry

yea i agree

[elder rapids]

in openAI DR it fetches good asf insight

and I've seen the same with Geminis

Gemini seems to verbosify a ton tho

but it doesn't seem like I'm getting less info

just more yap

[leaden palm]

meanwhile perplexity:

[small haven]

"prioritize verbosity" loool

[elder rapids]

prioritize verbosity 😭🙏

nerfing the search

[small haven]

well great o3 solved a unit test where o1 pro couldnt... nice

[raven void]

o3 is just so good

Gemini got the answer to my problem wrong even on the meta synthesis only o3 solved it

[small haven]

yea..

ok so wen full o4 tho 😭

[patent bane]

whats the problem

[raven void]

when they have o5 pro internally 🤣

[raven void]

code for a part of my project

[small haven]

to be working at oai, that must be insane

[woeful geyser]

I feel that the vibe of o3 is too easy to recognize as well.

[fleet lintel]

April was kind of disappointing.. no new top model was released. bunch of hype but nothing materialised.

[earnest parcel]

yea compared to march where we got Gemma 2 27B, Mistral Small 3.1, QwQ, Gemini 2.5 Pro, and Nemotron 49B, it's quite an uneventful month.
Llama 4 and GPT-4.1 are kind of duds, at least for me.

[plain zinc]

Gemini 2.5 Pro, o3, o4-mini-high: 🤨🤨🤨

[drifting thorn]

2.5 Pro is the SOTA...

[solar nebula]

yes

[earnest parcel]

03-25, it doesn't count as April release (even though it got rebranded from exp to preview)

[fleet lintel]

2.5 pro was in march. and nothing really better came after that

o3, o4-mini-high : disappointing. OAI tried to play game by releasing a bit prematurely to take on google but honestly they are just meh compared to 2.5 pro

[plain zinc]

Oh, right. He was in March.

https://x.com/seti_park/status/1915996238979453023

SETI Park (@seti_park) on X

1️⃣ US20250131254A1:

@GoogleDeepMind의 ‘기능 보존형 신경망 확장 기술’ 관련 특허로, 기존에 학습된 지식을 소실하지 않으면서도 AI 모델의 규모를 확장하거나 특정 목적에 맞게 특화시킬 수 있는 혁신적인 방법을 제시합니다. 이는 마치 사람이 새로운 지식을 배울 ...

X · Apr 26, 2025, 5:07 AM

From this post, you can understand where Google gets several models in LMarena.

So much more in such large numbers

[calm sequoia]

Where is this from and what is synthesis

[torn mantle]

@balmy mist i think i may be right again, we may get r2 this monday / next week

https://x.com/ClementDelangue/status/1916345020791001181

clem 🤗 (@ClementDelangue) on X

👀👀👀 https://t.co/mekr0Drodq

X · Apr 27, 2025, 4:13 AM

lot of hints as well

[small haven]

day 11 , where is o3 pro

lemme guess when r2 gets released? smh

[hardy pecan]

he said a few weeks. have more patience.

[fleet lintel]

does anyone know of insider Apple info? Are they planning to compete in AI space at all?

[mossy drum]

New model in Arena: llama-4-scout-17b-16e-instruct

[keen fulcrum]

https://arstechnica.com/tech-policy/2025/04/elon-musks-xai-accused-of-lying-to-black-communities-about-harmful-pollution/

[small haven]

easy to say when ur on a plus sub

[hardy pecan]

Complaining about things you can't change is a wasted effort. Would be good to see it out at the end of this week, provided they are on time

[sage raptor]

Pollution is everywhere

[unborn ocean]

They had this research for a while now apparently: https://arxiv.org/pdf/2308.06103

Maybe the already used it for the early Gemini 2.0 iterations back in 2024 (Like all the ones that where aistudio only)

[ocean vortex]

for what model? I'm absolutely blown away after I found out this silly score applies to playground and API too. If what you are saying is true that means any benchmarks that people did earlier might not even be possible to reproduce now. Changing stuff like that can always have unintended consequences, even if in theory higher value should be better. There is no place in API for crazy sht like that lol

[full kite]

Guys I'm better

[vagrant field]

hi all

hi !

depends , but I use gemini*, o4-mini-high , and claude 3.*

gemini 2.5 pro and claude 3.7

[alpine coral]

the api was what i found most curious about it when we were discussing it yesterday.. you could have been blown away 12hrs earlier if you read my messages lol

[alpine coral]

this is consistent with oai's chain of command i guess

• System / platform (oai)
• Developer (who can add what we call a system message)
• User

if there's a conflict, the higher one has authority

as it explains in its reasoning (and hence why it said 8192;the 2903 in the devloper message was overriden)

[keen beacon]

this seems about right

[brittle tiger]

Is yap score unique for each user preference or are they just changing it for everyone on the fly looking for a sweet spot?

[alpine coral]

i think they might be from the same platform-level prompt (oai's instructions, which [are meant to] override anything given in the Developer Prompt (API) as well as end user messages

the 8196 comes from the platform-level prompt (as do, I think, those instructions about mirroring style etc)

[knotty jetty]

What website is this

[alpine coral]

https://platform.openai.com/playground/prompts?models=o3

[knotty jetty]

Ok thanks bro

[alpine coral]

np

[full kite]

what the fk is a yap score

I NEED SMART MIND HERE Please

I have this test tomorrow

It's like the final test of the year right

But it's a mock one

Idk sht about what the subjects are

I don't go to classes and everything

What would be the best method to know how to solve them

Like I have access to all the past mock tests from the last 4 years

and the correction of them

And also videos about ppl solving them

Like Idk what to do chat

Please 😭

[keen beacon]

Get ready @full kite

[full kite]

ready for what

[full kite]

Are you a smart person

[keen beacon]

To be unemployed

[leaden palm]

read through them then

[full kite]

😭 😭 😭 😭

[full kite]

And then what

[leaden palm]

see what you don't know

what you don't understand

look for patterns in that

[keen beacon]

ur in ai discord bro just cheat

all my classes online

4.0 gpa

[full kite]

Is that good

[keen beacon]

4.0 gpa out of 4

thats the highest u can get

[full kite]

that's like good

[keen beacon]

ye

i have a spare laptop i put in the corner

then use parsec on main desktop

remote into it

gemini free trial gives u unlimited screenshots too

[flint sand]

won't work for people with relative grading

also don't cheat dude

[keen beacon]

whats that

[full kite]

what free trial are you talking about

I'm using google ai studio

[flint sand]

your gpa isn't directly based on your marks, it's based on how much you scored relative to the highest marks obtained in class

[keen beacon]

go to https://gemini.google.com/app

top right corner you should see something like this

click it

1 month free trial no payment

cancel before they charge u

p sure ai studio's 2.5 pro without premium is only like 500k tokens then you get rate limited for a day

[leaden palm]

or use the openrouter + vertex + ai studio + copilot stack

or use the official student free trial if you're in university

[full kite]

bro what tf

Dude 2.5 pro is free and 1 million token per chat

what is pro even about 🙏 😭

I have the flash one too

faster

[leaden palm]

2.5 pro in ai studio is 250k/min and 25/day

[full kite]

gemini.google.com is a scam

ong

[full kite]

what does RPM means

[leaden palm]

requests per minute

[leaden palm]

actually maybe ai studio has higher limits than the ai studio api... idk

[full kite]

so

GUYS

I need to study jessu

please help me

[leaden palm]

what

jesus?

[full kite]

yeahuh

help me studying

[leaden palm]

lmao

[full kite]

😡

[ocean vortex]

yeah sorry didn't see that 👀

[ocean vortex]

I think it's more on the fly thing, especially if this was at 4k earlier. I also very much doubt it was evaled with this in the context. But regardless it's unwanted flood for API as far as I'm concerned which makes "developer message" much less powerful and relevant

[frail thorn]

Guys does anybody have a chatgpt premium account that could be shared with me?

[ocean vortex]

No. But I can run a prompt for you if you want

[full kite]

why do you pay for chatgpt

[frail thorn]

Well a "prompt" wouldn't solve everything

I need full access

😭😭😭😭

[ocean vortex]

well then buy it

[full kite]

what do you need to do lil bro

[frail thorn]

I'M NOT SPENDING 20 DOLLARS ON IT, I ONLY NEED IT FOR LIKE 4 DAYS

Its whatever

[ocean vortex]

well then @full kite could give it to you maybe

[full kite]

WHAT DO YOU NEED CHATGPT FOR

[frail thorn]

REALLY

[full kite]

😡

[ocean vortex]

he does't need his it seems

[frail thorn]

school stuff.....

[full kite]

Ok tell me what

I'll see if I accept

[frail thorn]

well I'm doing a research

I just need chatgpt to help me

the normal one doesn't accpet all of my files

and so on

[full kite]

how many files

[frail thorn]

many...

like 10 at least

[ocean vortex]

have you tried aistudio?

[frail thorn]

yeah

[ocean vortex]

it's free and the king for file uploads

[frail thorn]

but like

[full kite]

No you did not

[frail thorn]

I DID

[full kite]

okay what happen

[frail thorn]

wait

[full kite]

🤨 🫃

[frail thorn]

HEY THATS MY EMOJI

🫃🫃🫃🫃🫃

[full kite]

Dude

listen I'm going to quit

if you don't tell me

😡

[frail thorn]

TELL YOU WHAT

oh my days

you know what

QUIT

[full kite]

WHAT HAPPEN WITH GOOGLE GEMINI

IT HAS 2 000 000 TOKENS FREE

PER CHAT

CHATGPT PRO IS 128 000

[frail thorn]

😇

[full kite]

😡

[full kite]

Ok we should start over

I'll help you

I want to help you

@frail thorn

?

😔

[frail thorn]

😀

[full kite]

Okay so do you need to upload a large ammout of pdfs?

[keen fulcrum]

Are you pregnant?

[frail thorn]

just forget it I already bought a subscription

🙄

[frail thorn]

yes

[timber kiln]

Now you gotta use it overtime so they lose money on your sub

[frail thorn]

hell yeah

[ocean vortex]

I'M NOT SPENDING 20 DOLLARS ON IT, I ONLY NEED IT FOR LIKE 4 DAYS

[frail thorn]

I'll include the thank you

[frail thorn]

yeah well I did

[leaden palm]

wait does the app still lack the 1m context from 4.1 lmao

[ocean vortex]

🫃

[full kite]

And it's free

You fkn looser

[frail thorn]

someones mad

[full kite]

✡️

[frail thorn]

🤗

[full kite]

🥀

[frail thorn]

you greedy juice...

[full kite]

say that to sam the juice

[ocean vortex]

I wouldn't really say chatgpt is better for what you are trying to do even... But ok whatever 😂

[frail thorn]

JUST CHILL OUT GUYS

its just 20 dollars

Its not like I'm going to die from it

[full kite]

yeah 20 dollars or free

[ocean vortex]

kids in Africa could have eaten those dollars

[frail thorn]

LOLLLLLLL

[full kite]

WE GOT ONE

[frail thorn]

now that made me giggle

[full kite]

coins clipper

[frail thorn]

gold detector

[full kite]

the

why have you buy the chatgpt

[frail thorn]

🤗

[full kite]

thing

[frail thorn]

I'm just talking about me

I'm not including any other reference

[full kite]

what does that mean

[frail thorn]

😭

I NEED TO GO BACK TO RESEARCHING

cya

[full kite]

RESEARCHING WHAT

🤨

[frail thorn]

👹

[full kite]

are you doing research about the juice

😔

yes ?

Sez u

aware

[native shore]

diana12493_32 has been warned.

[full kite]

Hello??

Can the other guy be warn too

😡

[keen beacon]

oy vey stop noticing

[native shore]

thekingofnothing_ has been warned.

m0_0d_ai has been warned.

[leaden palm]

which is why if you build anything and want it to be free you have to get like 3 other providers

25 rpd does not serve 2-3 users actively building things

25 rpd is the documented limit

and i've ran into it before

[full kite]

💀

[ocean vortex]

wtf happened here

[frail thorn]

😢 🖕

[full kite]

🫃

[zinc ore]

https://x.com/Teknium1/status/1916398151914950687

Teknium (e/λ) (@Teknium1) on X

Okay come on.. lmao

X · Apr 27, 2025, 7:44 AM

[small haven]

lmao

[pliant cypress]

Logan just hint 2.5 ultra and remove tweet very fast 😆

[keen beacon]

what did he say

[small haven]

screenshot or didnt happen

buddy perma works at google and says this

[pliant cypress]

Something about making custom t-shirts with text "1400+ ELO club", "smell big model", "AGI when". But maybe its just coping

[sage raptor]

maybe he is drunk

[full kite]

1400 elo?

what does that mean

[knotty jetty]

Yo gemini 2.5 pro in perplexity is peak

Also I hate perplexity but its really good with the llm

[full kite]

whats perplexity

[knotty jetty]

New gen

[full kite]

what

dude I know nothing of the sht

I'm using google studio ai

whats perplexity

[knotty jetty]

Just look it up bro

Perplexity.ai

[full kite]

omfg

[leaden palm]

why are you in this server if you arent an arena user lol

[small haven]

deepwiki's deep research is underrated

[wintry tinsel]

Remember how they murdered the old server

🥲

[leaden palm]

the one from devin?

[small haven]

yup

took about 3-4 mins to finish

[full kite]

bro I just use arena to see the scoreboard I'm not jerkng off to it

[leaden palm]

so you do use the leaderboard

which is... elo data

[full kite]

yeah what about it

aok

I thought it was a chess thing, they were talking about that earlier

[leaden palm]

anything with pairwise comparisons can be measured with elo

[full kite]

like llm will never be able to play chess

[full kite]

no but fr what else can we do on arena thing

can we play on it or something

[leaden palm]

you can chat with llms and vote to help build the leaderboard

[wintry tinsel]

LLM’s dead end 😱

LLM’s failed project

[leaden palm]

you can also compare other things:

• llms using github repos
• llms using search
• llms making websites
• llms writing code
• image generation

[wintry tinsel]

Rip LLM’s

Support Yan Le Cun

[torn mantle]

where

[torn mantle]

oh its probably riverhollow/sunstrike

im not impressed with these models tbh

they just seem like gemini 2.5 pro 03

[full kite]

I don't know what a repos is

[leaden palm]

thats ok, a lot of lm arena is for coders and you might not be one

[full kite]

I'm a coder

[leaden palm]

have you used github before?

[full kite]

to download yt dlp

I know about the green board

contributions sht

[small haven]

my iq is dropping

[elder rapids]

where's Nw 🙏😭

[sage raptor]

soon

[keen beacon]

in the void

[elder rapids]

the longer these models are unreleased

the more there's a chance people catch up

and it's unimpressive

[wintry tinsel]

We need human brain computer

Neurons in the circuit

Than true AGI

🔥

[golden ocean]

lets disect urs so we can start the experiments immediately

[wintry tinsel]

That would be cheating I’m already AGI

[misty vault]

more like artificial general stupidity

Ok that sounds more like a direct insult rather than a silly joke

no more funny

[wintry tinsel]

Are LLM’s capable of being stupid?

Perhaps stupidity is a byproduct of intelligence

[leaden palm]

llama 1b:

qwen 500m:

[wintry tinsel]

Well you have a point

[small haven]

ok but wen o3 pro

baited, just wanted to spawn my man craig

[elder rapids]

nah ion think so

a major part of what made o1 pro so good was it's ability for pure longer context reasoning

[small haven]

ye im using o3 more than o1 pro..

what we know is its 10x compute (so thinking for 10x longer than o1)

[keen beacon]

lol

[small haven]

and the fact it is outputting as a one shot answer rather than streaming tokens, means it is using an internal canvas in the backend

so it is constantly iterating its final answer, thats my hypothesis

[elder rapids]

we can infer

longer reasoning

whereas pro is probably fine tuned

[small haven]

and internal canvas

[elder rapids]

specifically to force a longer reasoning chain + better initial instruction retention

[small haven]

internal canvas, checking its own answer on a pass@1, then reiterates it, until it is satisfied with a hard limit of 10x compute

[elder rapids]

there won't be an o3 pro if it's not better than o1 pro

just saying lmao

[willow grail]

great for cleaning my ass. yes. thats it.

[small haven]

can sam stop riding his husband and release o3 pro

[willow grail]

ceos dont have a healthy relationship

its a ceo lol?

watch movies

then u know

rich people are very unhappy

thats why i am happy

nonetheless you are not rich

cause its adverse to not say it

youre stuffed with adverse lies

/s

ceo's dont live on prairies, therefore they are unhappy creatures.

[terse shuttle]

Should we expect a new image generation model from openAI in the arena?

[willow grail]

you have cute attributes, federighi

ceos dont have small things.

see it as necessary.

we said the same word... lol

._. we are models?

i will weigh 81kg after water flunctuations in exactly 30 days

now i am 82.6kg

having diseases which makes loosing weight slower is bad

i think u just wanna beleive they are as happy as a nanny in a village

cause u wanan be rich too

so u have a goal to go for right now

[ocean vortex]

No it’s more like multiple attempts and consensus system. You can still set low-med-high for pro.

[tall summit]

oh hell yeah #announcements cute

[small haven]

omg sam tweeted, plz be o3 pro

[golden ocean]

its him riding his husband

[golden ocean]

hi

[leaden palm]

probably a refreshed pretrain

[keen beacon]

its likely to be a cpt of 2.0 pro

it isnt. even openai cant increase simpleqa that much yet thru reasoning

[small haven]

i love o3

but i would love o3 pro even more

[dapper aspen]

hi

[balmy mist]

we got deepseek yet?

imma go to china and fight Winnie if we dont get r2 this week

[alpine coral]

just got folsom-exp-v1, which i haven't seen or heard of before - new anon model?

presumably related to cobalt, apricot

so amazon ig

[torn mantle]

https://x.com/ZhaoTing1024/status/1916678514180497416

Zhao Tianyu (@ZhaoTing1024) on X

Official announcement: Qwen 3 this week. Reasoning and non-reasoning in one.

X · Apr 28, 2025, 2:19 AM

[small haven]

when o3 shows these traces in the cot, i kinda leak a bit

[cedar tide]

Amazon reasoning model no?

[keen beacon]

you do what now

[calm sequoia]

Any news on possible specs?

[alpine coral]

don't think so

it seemed quite fast and dumb

[cedar tide]

https://x.com/btibor91/status/1916756247422124353?t=nmHdYoncI3U1Sdle6XHOlg&s=19

Tibor Blaho (@btibor91) on X

Amazon Bedrock already lists "Amazon Nova Premier" with a release date of 2025-04-30, along with a new model "Writer Palmyra X5" (2025-04-28) and "Llama 4 Maverick 17B Instruct" and "Llama 4 Scout 17B Instruct" (both 2025-04-28)

X · Apr 28, 2025, 7:27 AM

[small haven]

why is o3 limited at 64k context, absolute dogwater

[keen beacon]

omg qwen 3

apparently pre trained on 36 trillion tokens 😮 (2x qwen 2.5). multiple moe models?

[stuck orchid]

Hi.
Is there a chat limit in https://beta.lmarena.ai? Or is it possible to test ai even on long requests (10K+ tokens)?

[keen beacon]

curious about the super low amnt of active params on the moe models. qwen 15b in the huggingface pr has 2b active, qwen 30b has 3b active

[torn mantle]

https://x.com/OpulentByte/status/1916794351025352859

Opulent Byte (@OpulentByte) on X

@ai_for_success Looks like smaller Qwen 3 models today only.

X · Apr 28, 2025, 9:59 AM

[keen fulcrum]

https://hiddenlayer.com/innovation-hub/novel-universal-bypass-for-all-major-llms/

[torn mantle]

So llama 4 reasoning models werent added on lmarena giving the recent controversies

And its kinda weird too qwen3 isnt added yet in an anonymous battle mode

[ocean vortex]

em?

gpt-4.1-2025-04-14 --> 41.6%
o3-low --> 49.4%

in both cases just shy of 20% of the earlier score increase

[keen beacon]

2.5 flash got a worse simpleqa tho compared to gemini 2 flash, i would expect openai to lead there in terms of factual reasoning. but yeah i guess its out of date, since we have o3. i was comparing it with o1 when i formed that opinion initially

[keen fulcrum]

They were

[torn mantle]

https://x.com/JustinLin610/status/1916805525171494965

Junyang Lin ✈️ ICLR (@JustinLin610) on X

see if we can finish the job tonight #Qwen3 👻

X · Apr 28, 2025, 10:43 AM

[humble sonnet]

https://code-arena.fly.dev
what is this link ?

[balmy mist]

Omggg

[neon anchor]

Sunstrike is not generating code at all

[balmy mist]

Anybody tested qwen3?

[keen beacon]

its not out yet

the 235b moe (apparently) with reasoning mode might be very impressive

[sage raptor]

No eletricity in europe 😭

[keen beacon]

no lmao

it's just that those were the ones that accidentally appeared on modelscope briefly

anyway, to summarise, this week we are getting:

• qwen 3 (likely today)
• amazon nova premier (wed)
• deepseek R2 (somewhat likely, depends on qwen 3)

[brittle tiger]

Would be cool if r2 went on arena before debuting

[torn mantle]

Yea the big one is 235b

[keen beacon]

i expect that to be a very good model

[balmy mist]

ah okay, its like we get a fire week, then break, then back to fire lol

i kinda like that, give us time to collect ourselves

[torn mantle]

Qwen 3 is what people expected llama 4 to be like

[calm sequoia]

The GPT 4o doesn't allow altering real photos due to their stupid policies. Anyone have the jailbreak code or the alternative tool?

[sonic tendon]

oh damn
was wondering if qwen3 was going to include a big model
seems sort of odd that they didn't include it in the huggingface pr from a week or two ago, but maybe they're trying to keep it closed?

[keen beacon]

a lot of models werent in the pr

[sonic tendon]

ah

[keen beacon]

they just briefly put out a fp8 quantized qwen 3 0.6b on hf and removed it

[torn mantle]

Yea

Will 235b be the biggest model released by them so far?

Or they released smth bigger before

[keen beacon]

yea. 110b dense was their largest before

im so hyped rn

[calm sequoia]

I think nothing will be better for at least a month

[calm sequoia]

If anyone will need the same: https://sider.ai

[ocean vortex]

it should at least be mostly as good as 2.5 I think.

probably. But it's not gonna be worth using it given the price that's for sure lol

technically sota now is o3 anyway

[keen beacon]

i dont think they will put o3 pro in the arena

[barren prairie]

Maybe to beat Gemini 🙂

[keen beacon]

imho the most likely thing to dethrone gemini 2.5 pro will be gemini 2.5 pro (at least in the arena) 🤣

[torn mantle]

In all seriousness o3 model is a lot better than gemini 2.5 pro at general tasks

You can still feel the robotic vibes from gemini

[full kite]

Just got r2 it's good

[ocean vortex]

old model I'm already using r3

[full kite]

bro what

Guys can IA do my homework pls

it's a math test

[calm sequoia]

Why did you change your username from Mango to Diana? 🙂

[full kite]

cause diana rules

https://tenor.com/view/shoebill-diana-hello-gif-25289056

[calm sequoia]

if R2 > Behemot, Zuck is cooked

[keen beacon]

thats obviously gonna be the case tbh

[full kite]

what is behemot

[calm sequoia]

Obvious for you, but not for Zuck or his investors 😄

[full kite]

what is behemot

[keen beacon]

behemot is agi

[alpine coral]

yeah zero chance

[calm sequoia]

Behemot is minotaur shrek

[severe tinsel]

Hi, I’m not a pro, will Qwen3 compete with Qwen 2.5 Max at the top of the leaderboard or its not the same category?

[full kite]

agi doesn't exist

[full kite]

nuhuh

[keen beacon]

qwen 3 is a line of models its not even released

or in the arena

[severe tinsel]

Yea i mean one of them

[keen beacon]

yes at least one of them will compete i think

probably the big one will be competitive in the leaderboard

[severe tinsel]

Okay thanks!

[stuck orchid]

Qwen 3 may be better than Gemini-2.5-pro 👍

[full kite]

qwen is slow asf

[ocean vortex]

this

[full kite]

this is a bear

[ocean vortex]

it's behemoth

[full kite]

like a bear with teef

4 legged bear

[ocean vortex]

https://sciifii.fandom.com/wiki/Behemoth

[alpine coral]

bro... it's trolling....

(i assume / hope lol)

[ocean vortex]

I think we both are. I hope he does too lol

[alpine coral]

lol

[keen beacon]

try it yourself lol

direct chat / side by side?

i think side by side was configured differently a while back it might have better limits

yea

[alpine coral]

i dunno about coding (let alone for a specifc programming language), but i feel 2.5 pro is just solid as af all round

[keen beacon]

its incredible i still main it over anything else

[alpine coral]

it's more usable than o3 (esp o3 high)

yes

[keen beacon]

its definitely seen less c++ than the others and its more things to manage

yea

if look at it historically probably. it depends on how much of it was curated, but i think its likely to be even less

im not very sure about the others. but 1 and 2 is highly likely to be python/javascript (not sure which one is which though)

it depends really but with a gc and such its generally slower/much slower i think

did u try?

2.5 pro has the best context retention/usage in a model i think anyway, it helps a lot

you should probably upload the code into the repo instead of it being in a zip

do u have git installed?

i think u can also upload folders directly on the website

maybe ask 2.5 pro to teach u git

its more convenient and allows u to have version control

[alpine coral]

i'm getting so many errors in the arena atm

i feel like the yap score has been there all along (and at 8192).. recently discovered / noticed, rather than added..

also the other models' responses.. 3.7-sonn and v3 do well; sunstrike also (though verbose af)

folsom-exp-v1 assumes it's a bitcoin or something - pretty terrible response imo

[tall summit]

what.

[alpine coral]

oai reasoning models have a top-level prompt that gives guidance about how to intereact and includes this part about a 'yap score' (which has always been at 8192, so far as i can tell)

[tall summit]

is that real.

[alpine coral]

yeah i mean, to the extent there's some oai-imposed instrucitons that include this thing called a yap score, it's real

[tall summit]

wow okay.

[alpine coral]

whether it's concerning though im not really sure tbh

like esp if it's been there all along

it may just be for stylistic purposes

but yeah that said.. my intial reaction was to think it was a way for oai to dynamically throttle outputs on o models in chagtgpt, to like manage costs / compute

[keen beacon]

iirc it was a thing since launch

o3/o4 mini launch

[alpine coral]

yeah i only learnt about it here a couple of days ago

but now i think it's likely been there all along (and am not really bothered by it.. like there's no indication of it being used to nerf the models or whatever... yet anyway aha)

i kinda wonder if, in cases where lots of reasoning tokens are used especially, the final outputs could be super lengthy, and this was just their solution (prompting), rather than it actually meant to be dynamic

[brittle tiger]

2.5 Pro tops another benchmark

https://geobench.org/

[cedar tide]

Qwen 3 253b will be better than deepseek 3.1 671b ? (And Maverick 400b)

[torn mantle]

Probably

The question is : will the smaller models be even better than Maverick?

Imagine 30b > 400b

[cedar tide]

below 253b there is only a dense of 14b and a moe of 30b
I don't think it can be better than maverick

apart from their version with reasoning

[keen beacon]

i think all/most of them are hybrid reasoning models

[cedar tide]

👀
https://x.com/alexalbert__/status/1916874027756666981?t=eC1sQXsZWHold8XGlIlxbw&s=19
https://x.com/alexalbert__/status/1916874039769120904?t=UGLhieKvS3ariM5WPQxCwQ&s=19

Alex Albert (@alexalbert__) on X

"let's train our model to get higher chat slop ELO scores"

*model starts exclusively outputting pure chat slop*

X · Apr 28, 2025, 3:15 PM

Alex Albert (@alexalbert__) on X

not meant to be a jab at any one lab in particular, just highlighting a particularly bad incentive structure I see rn. there's a reason you don't find Claude at #1 on chat slop leaderboards. it's the LLM equivalent of optimizing for video watch time in a social media algo.

X · Apr 28, 2025, 3:15 PM

[cedar tide]

Yes, it has already been leaked, we know that.

[keen beacon]

there is 15b moe according to the hf pr

might be a placeholder but the pr also mentioned the 8b model which was confirmed

[keen beacon]

LOOL he deleted them

[cedar tide]

Yes, I was thinking the same thing and I said it to several people, but in the end, whether it's the leaks from ModelScope or Hugging Face, there’s no trace of this model anywhere.

[keen beacon]

ya i guess its a placeholder

wont be surprised if its real tho

[terse shuttle]

.

[keen fulcrum]

Qwen 3 dropping

[keen beacon]

yup they're trickling in

[torn mantle]

Gemini 2.5 pro is a solid model

[keen beacon]

will probably be all done in the next hour or two

I'm just waiting for the big one

[torn mantle]

Its either o3 or gemini 2.5 pro that deserves #1 spot tbh

[keen fulcrum]

https://fixupx.com/kimmonismus/status/1916818352485413038

Chubby♨️ (@kimmonismus)

Qwen 3 released
︀︀
︀︀Qwen3-8B
︀︀
︀︀Qwen3 Highlights
︀︀
︀︀Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Building upon extensive advancements in training data, model architecture, and optimization techniques, Qwen3 delivers the following key improvements over the previously released Qwen2.5:
︀︀
︀︀•⁠Expanded Higher-Quality Pre-training Corpus: Qwen3 is pre-trained on 36 trillion tokens across 119 languages — tripling the language coverage of Qwen2.5 — with a much richer mix of high-quality data, including coding, STEM, reasoning, book, multilingual, and synthetic data.
︀︀•⁠Training Techniques and Model Architecture: Qwen3 incorporates a series of training techiques and architectural refinements, including global-batch load balancing loss for MoE models and qk layernorm for all models, leading to improved stability and ov…

FixupX · Apr 28, 2025, 11:34 AM

[keen beacon]

lmao someone leaked qwen 3 32b it seems

seemingly one of the randos (non qwen team) in the hf qwen org

[cedar tide]

32b where ?

[keen beacon]

eh they've started dropping them now anyway

doesn't matter much

apparently: https://huggingface.co/second-state/Qwen3-32B-GGUF

not many details there tho

why are there random people in the qwen hf org lol

it'll be officially released literally in the next hour probably

I wonder if we also see something from deepseek this week

it would make sense but who knows

the 235b possibly

depends on how strong their reasoner is

[keen beacon]

yeah

I do expect it to beat R1 minimum really

if they can't even do that it's a flop

llama 4 reasoning releasing at llamacon tomorrow by the looks of it, frontend is ready

i would also expect behemoth

[tall summit]

i think o3 is better at translation than g2.5pro

[tall summit]

sounds fun

[keen beacon]

its gonna be a huge flop

if it's anything like the rest

there will be a lot of memes about qwen and llama i suspect

idk about qwen

i have a lot more faith in them than i do meta

yann lecooked

[keen beacon]

yea i meant clowning on meta lol

how qwen was the llama 4 people expected

oh

yeah nevermind

[keen fulcrum]

https://modelscope.cn/collections/Qwen3-9743180bdc6b48
still no entries

[keen beacon]

its a shame that guy is uploading the ggufs publicly before qwen officially announces it

[keen fulcrum]

Must be some employee with access to the system

[keen beacon]

no its a random guy in the hf org i believe

[keen fulcrum]

Oh lol

[keen beacon]

I also hope that 15b moe is real, it's an awesome size

It's insane how many models they are gonna release. With 36 trillion tokens in pretraining not to mention the reasoning training etc

[keen fulcrum]

Is R2 coming out this week too?

[keen beacon]

Idk lol

Its qwens week probably

[barren prairie]

deepSeek team always work silently no one knows what they are planning 🙃

[keen fulcrum]

There was a leak for deepseek too
if its not this week, must be next one

[sonic tendon]

qwen 2.5 is already higher than maverick lmao

[keen fulcrum]

Behemoth

[sonic tendon]

yeah but

eh

[sonic tendon]

one of the chief researchers is tweeting aggressively about it

i would personally expect q3 within the next 24 hours

[torn mantle]

Yea

Weird they didnt add any new models on lmarena