general

[sturdy mica]

https://tenor.com/view/half-life-gman-g-man-half-life-2-rise-and-shine-gif-5438670790719714756

[balmy mist]

maybe for kids

[sturdy mica]

i’m leaving this chat

[balmy mist]

it sucks we cant have multiple tools at once

[sturdy mica]

no. no friend, no

[balmy mist]

lol

[sturdy mica]

it does :(

i wish we could, you can with api

[balmy mist]

yeah does not make sense

[sturdy mica]

and with the live streaming thing but that only works with 2.0 flash

[balmy mist]

have you played aroudn with gemini 2.5 and tools?

[sturdy mica]

yes

[balmy mist]

how was it and what did you use?

[sturdy mica]

i use gemini 2.5 with grounding, function calling, code execution

they all work how you think they would

why?

[balmy mist]

i guess this is how you use multiple tools

[balmy mist]

i just was curious because i never played with it

what does code execution od?

do*

just executed what code you give it?

does it pop open view for it?

i need to dive deeper

[sturdy mica]

you can ask it to generate and run code like

“generate and run code that gives a random number “

and it will write python code and show output

ill screenshot rn

ok i cant but

you get yhe idea

just test it urself

[balmy mist]

lmaoo yeah

have you tried fine tuning a model on studio?

[torn mantle]

20 reports/day

quite generous

xd

[drifting thorn]

ok Cline is not that bad

I finally overcome that 2 chapters of my character's throwback

with Cline and a prompt that I've editted multiple times

[balmy mist]

you only get like 20 per day with openai right?

[teal mantle]

Can you still use credits and still qualify for higher request limits?

[balmy mist]

did you hit the 1000 limit?

[keen beacon]

I don't believe so but u should really ask it in OR discord server lol

[teal mantle]

Far from

[torn mantle]

much less

like 2 per day or smth

or 1

xddd

didnt age well

[oblique flint]

meanwhile microsoft still cant produce a frontier model themselves, they're entirely dependent on openai to do so

[alpine coral]

oh i noticed the colour but didn't realise what it meant til now - nice / congrats! 👍
(will be good to have an active mod ha)

[alpine coral]

have you tested it yet? If it's as good as they're saying it is, 20/day is epic value (vs chatgpt's Deep Search offering, which is limited at like 10/month or something it feels like)

[keen fulcrum]

https://storage.googleapis.com/gweb-uniblog-publish-prod/original_videos/ai-mode-animation.mp4

[oblique flint]

https://www.youtube.com/watch?v=_WvtdRtG1aY

wish we had something like this for gemini and other OSes than macos

[alpine coral]

yeah even less (for Plus anyway).. it's 10/month

[keen beacon]

gemini 2.5 flash today apparently

[oblique flint]

sauce?

[sage raptor]

stargazer

[oblique flint]

how do you know its releasing today tho

[north vale]

https://x.com/btibor91/status/1909895821589458989?s=46

Tibor Blaho (@btibor91) on X

gemini-2.5-flash-preview-04-09 + thinking_config/thinking_budget have been added to Google Gen AI Python SDK

X · Apr 9, 2025, 9:07 AM

[hardy pecan]

cool

[alpine coral]

oh interesting so thinking_budget (in this case 10,000) is presumably their language for the parameter that defines token allowance for 'thinking'

sonnet is capped at 32k tokens. oai lets you adjust via low / med / high

[keen beacon]

i think u can do 64k

or with claude u can do even more by taking advantage that its a single model, and continously prefilling it lol

not the exact antml:thinking tag but the behavior is close enough

[alpine coral]

lol indeed

yeah ik what you mean

[keen beacon]

its interestingn to me anthropic didnt make it a special token

so the behavior can 'leak' easily

[alpine coral]

as far as i can tell, there is nothing 'special' going on at all with sonnet3.7 thinking

it just has a 'srcatchpad' thing given to it and some system prompt

there doesn't seem a fundamental difference; it just does CoT reasoning (as part of its regular inference, even if it's rendered in box on the UI) and that informs its 'final' response

when 'thinking' is enabled / usd as the model

[keen beacon]

it doesnt have comprehensive thinking instructions, just an instruction to think in antml:thinking and max thinking length where they probably tuned in several values with differing 'thinking' lengths. it was trained in to think when the response starts with antml:thinking, otherwise its normal

[barren prairie]

I want Gemini 2.5 flash thinking not Gemini 2.5 flash 😭

[keen beacon]

antml:thinking isnt a special token but its sanitized hard so u cant really use it (without degradation) but since it isn't a special token, <thinking> will basically be seen as the same because of pretraining associations instead of adding a new special token to the vocab

[hardy pecan]

a new challenger:

[hazy quest]

We have been saying that every few months for over a year lol. Remember Opus 3.0

[keen beacon]

anthropic doesnt use a regular eot special token even though its present in the tokenizer since claude 1. (this is how this breakage occurs, conditional stop on human:)

they seem to avoid adding special tokens whenever possible to maintain pretraining knowledge throughout the window

(they replace antml:thinking with <thinking>, even though it's actually antml:thinking if the screenshot is confusing)

[sage raptor]

is it good ?

[oblique flint]

please be better than o3 mini at coding

[harsh flume]

https://x.com/BoshiWang2/status/1909772639104540677

Boshi Wang (@BoshiWang2) on X

LLMs exhibit the Reversal Curse, a basic generalization failure where they struggle to learn reversible factual associations (e.g., "A is B" -> "B is A"). But why?

Our new work uncovers that it's a symptom of the long-standing binding problem in AI, and shows that a model design

X · Apr 9, 2025, 12:57 AM

I wonder how much (if any) research is going on in things outside of mainstream transformer architecture on the major companies

[cedar tide]

Its the thinking version today, and we have too thinking budget

[keen beacon]

so 2.5 flash today

but i find it hard to believe that's all we'll be getting

they've literally had like

5 or 6 anonymous models on the text arena

and a few on webdev

[keen beacon]

this is a base model

didn't think with my first test Q (which it got right)

[keen beacon]

they removed all the anon google models on web dev i think

thats interesting

[oblique flint]

what price do yall think 2.5 flash api is gonna be.

Current flash is 0.10$ per 1M input and 0.40$ per 1M output.

o3 mini is 1.10 per 1M input and 4.40 per 1M output

my guess 2.5 flash will be 0.2 per 1M input and 1.0 per 1M output. I think it's prolly still worse than o3 mini at math and coding but it's also a lot cheaper ofc

[keen beacon]

yeah i just checked

looks like those models may be public today, or at least one of them

hopefully nightwhisperer...

[drifting thorn]

Hope nightwhisper is 2.5 Ultra

[keen beacon]

i highly doubt they're working on an ultra model tbh but i would be happy to be proven wrong

[drifting thorn]

Like I won’t hope that nightwhisper is a detuned version of 2.5 Pro

Either increasing parameter or thinking tokens

[barren prairie]

That's great with Meta we were getting 1526366373 Anonymous model to get a stupid one than

[alpine coral]

i think we could see a thinking_budget API param (and like a slider in the aistudio UI)

[balmy mist]

thats pretty interesting

[gentle plinth]

javascript honestly, just tell it to generate everything in one html file, just copy that and open it , no compiling, special software or anything needed and it can run on almost any pc

obviously c++ would be faster but its a tenfold more complicated and highly dependent upon the exact system and hardware configuration

i mean for example gemini 2.5 pro can generate a 3d airplane simulator without a problem in one-shot

all in one html file with html/js/css embeded

it probably uses the three.js lib I think, but dont need to ask it, it will use whatever is appropriate

[brittle tiger]

Sure it's not thinking? Would be first non-thinking model to get one of my test arc-agi problems correct and it went through a bunch of hypotheses and combined them to reach final answer

[keen fulcrum]

Nightwhisper
when , I want it now

[drifting thorn]

I need Deepseek R2

Hope DeepSeek R2 has a million token context window

[keen beacon]

well it started streaming immediately when i tested it

[cloud meadow]

Elon is both naive and simultaneously a narcissist.

You could already do that I think.

Not sure if it's new but Gemini has started asking "Which answer is better?" on some prompts.

[cloud meadow]

The only time an AI will get me to actually le soyjak (mouth wide open) would be if I put in an entire 1 million token project and it efficiently recodes it to another language and another stack. Maybe 3.5 years away?

[drifting thorn]

What if AIs have unlimited thought tokens?

Would AI be able to make decisions on how many tokens they used for thinking? If not, then the above idea is disastrous

[cloud meadow]

Gemini 2.5 is pretty and I'm excited for the future of coding AIs

[hardy pecan]

https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/

TPU dominance

[oblique flint]

https://blog.google/products/google-cloud/next-2025/

Gemini 2.5 Flash, our workhorse model with low latency and cost efficiency, will soon be available in Vertex AI.

wen ai studio lol

[keen beacon]

https://blog.google/products/google-cloud/cloud-next-gen-ai-vertex-ai-updates/

[ivory schooner]

我的24k......但我相信24k的后继者将会是Behemoth ......

但我听中文社群说，恐怕要等很久以后......

（准确一点的话是夏天.......）

期待Behemoth 吧

[torn mantle]

mm

no wonder

[torn mantle]

not yet

[balmy mist]

they changed ui back?

[barren prairie]

Who?

[balmy mist]

nvm my chrome glitching

and studio

[drifting thorn]

I think Anthropic will be able to train a model way better than 3.7, R1, o3-mini or even Gemini 2.5 Pro when they can get a ‘honest’ large multimodal model

The answer of cot models will adhere to what they ‘think’ and we can train it much more efficiently with RL.

[keen beacon]

oh i love the new aistudio ui

[lime coral]

lol

Weird on mobile. Will need some adaptation

[harsh flume]

Anyone can see a feasible path any company surpasses Google this month?

[sturdy mica]

no

also is the new gemini model coming out today

[keen beacon]

yes 2.5 flash

[sturdy mica]

awww

shucks

[oblique flint]

where are the 2.5 flash benchmarks tho

[keen beacon]

when its out in aistudio

woah i just saw the new ai studio

it looks nice

[drifting thorn]

Do you guys think Gemini models are based on MoE architecture?

[keen beacon]

yes

they said so in 1.5 pro's announcement i think.

[torn mantle]

kinda sad we wont have gemini coder

[oblique flint]

moe is the only way their api prices can be so low I think

[torn mantle]

its the only model i was looking forward to tbh

[drifting thorn]

I’m looking for Claude 4.0 after they found out that reasoning models aren’t honest

[keen beacon]

moe is faster/cheaper but i think in high batching/etc the calculus is more complicated and the gains are reduced. i dont know much tho lol

its not just moe thats making the difference

[drifting thorn]

Anthropic is definitely training a new model to be “honest” in showing it’s chain of thought

After their research showing 3.7 thinking hides its actual thoughts

https://www.anthropic.com/research/reasoning-models-dont-say-think

[brittle tiger]

Gemini coder was just a dream, a whisper in the night

[barren prairie]

We will have some stars today

Or maybe luna or some dreams

[drifting thorn]

For every model, does the chain of thought aligns with the answer

[fleet lintel]

close to zero

[balmy mist]

fr im gonna cry tonight

[sturdy mica]

they should release coder instead of flash today…

[balmy mist]

i dont care about flash unless its blazing fast and dirt dirt cheap

[sturdy mica]

yeah

[oblique flint]

o3 mini killer would be great

[barren prairie]

I don t care about flash unless it is flash thinking 🙂 🩷🩵 I love the flash thinking bot

But is it confirmed ? There is no NW?

[keen beacon]

it is thinking dw

[barren prairie]

I mean Gemini coder?

[keen beacon]

2.0 flash was cheaper than 4o mini, i dont expect it to be much more expensive per token

considering each request is generally more expensive because of thinking tokens and they want to be competitive

[sturdy mica]

??

Gemini 2.5 pro is a o3 mini killer

i think

its way better anyway

also when is o4mini and o3 comjng out

[balmy mist]

gg https://x.com/mark_k/status/1909980772338901448

Mark Kretschmann (@mark_k) on X

Google "Ironwood" TPU v7 is a 10x jump in compute over the last generation, and it has vastly more memory. Can anyone even compete with Google's AI hardware at this point?

X · Apr 9, 2025, 2:44 PM

this is y

makes sense now

openai who?

[blazing rune]

well, yeah, but it's also likely a lot bigger, and a good bit more expensive for output

the input is about the same I think

honestly I would say it beats o3 mini considerably, at least for ML stuff

[sage raptor]

what tpu openai uses

[barren prairie]

You mean at coding bcz o3 mini is a trush on other tasks

[torn mantle]

no wonder they are serving all of these models

its just so crazy

[oblique flint]

its way more expensive

[oblique flint]

yea I saw it and I'll definitely try it out, but I doubt it holds up against o3 mini in actual practical coding. This benchmark I believe is more competitive coding

[blazing rune]

that's almost definitely fake

[sage raptor]

it's not fake

[blazing rune]

well, not fake

but it certainly won't be good in the real world

[sage raptor]

idk

[blazing rune]

because of this

competitive coding isn't very useful in the real world, at least for LLMs. it's good for humans because we can apply the concepts in many different ways, but LLMs can't generalize nearly as well

[brittle tiger]

https://www.reddit.com/r/singularity/s/aVgBNOf67x

[lavish orchid]

hey everyone, very happy to share that I got accepted into YC's AI Startup School!! will hopefully see Sam Altman, Elon Musk and others! 🙂

https://x.com/ahmetdedeler101/status/1909988039863959833

Ahmet ☕ (@ahmetdedeler101) on X

Beyond excited to share that I was accepted into @ycombinator's AI Startup School!!

Grateful to be joining at 17, can't wait to learn and build :)

X · Apr 9, 2025, 3:13 PM

[keen beacon]

i'd love to meet sam, dario, demis... leave out elon or i might do something i regret

[sonic tendon]

lmaoo based

[leaden palm]

who watching cloud next

[keen beacon]

watched the opening keynote and will just keep an eye on the blog for everything else

[leaden palm]

lm arena mentioned

ok he just said 2.5 flash

thinking

has reasoning effort

"coming soon"...

lame

should have a giant lever to deploy it to prod

[keen fulcrum]

Shame not the alpha site is shown

[leaden palm]

depends on how soon "soon" is

the work day is still young

[keen beacon]

where's nightwhisperer at 😔

[keen fulcrum]

Did they share nightwhisper yet?

[leaden palm]

theyre not gonna call it that (probably)

[keen beacon]

well yeah

[leaden palm]

so far its just been mentioning 2.5 pro and flash (both thinking, and flash has reasoning effort)

[keen beacon]

either it was a 2.5 pro variant or it was an update

[leaden palm]

ts not serious

[keen beacon]

they done got the mcdonalds ceo on stage 💔

[leaden palm]

oh and dont forget the new tpu announcement

[torn mantle]

when is the google event

how much time left

[leaden palm]

TODAY

NOW

HOP ON

https://www.youtube.com/live/Md4Fs-Zc3tg

[torn mantle]

link?

thanks

[sturdy mica]

minecraft release in 2011

[leaden palm]

3.7 can keep working without breaking things for longer but doesn't think as well

[sturdy mica]

2011

https://tenor.com/view/minecraft-lever-notch-2011-official-release-gif-6894684587692301610

this

[leaden palm]

man i would be so hyped rn if i was a devops engineer

gemini weights leak when /s

~~ is this new~~ no, already made and tested in products, just now in vertex ai

[keen beacon]

depends on how you mean

new on GCP? yes, it was announced earlier today

new as a thing? no

it was already being tested publicly

on musicFX

it's a pretty meh model, we've been spoilt w/ things like suno and udio

[leaden palm]

they just said mindful and demure in the big 25

ehh what should i expect

[balmy mist]

is the event over?

[leaden palm]

no!

one stream

many hours (probably)

[barren prairie]

Still speaking

[balmy mist]

thnx

[lime coral]

Would turn into a generational hater if no update on native audio/img

[balmy mist]

yall notice that gemini works a lot better in api then on studio? is that by design?

might be noob observation

just started using api for gemini lol

[lime coral]

If you want to try nightwhisper this might be the way to go https://x.com/testingcatalog/status/1910010822937698425?s=46

TestingCatalog News 🗞 (@testingcatalog) on X

Google launched Firebase Studio 🔥

There, you can build an app using prompting, and it runs on the web.

"lovable+cursor+replit+bolt+windsurf all in one"

X · Apr 9, 2025, 4:44 PM

[lime coral]

Logan once said ai studio is purposely build to reflect the api behavior. It runs on the api

[torn mantle]

smh

[keen beacon]

for anyone who wants a link - https://studio.firebase.google.com/

[alpine coral]

these events are always so cringe lol

[keen beacon]

third edit is the charm 💀

[balmy mist]

wait what?

what model is powering this?

nw?

or the tools being used with 2.5

makes it like nw

[lime coral]

Obviously I don’t know

[balmy mist]

or nw was just 2.5 with tools all along?

ohh lmaoo

[keen beacon]

using it npw

this thing is fast

it uses next.js

[leaden palm]

ah yes, "war of the checkboxes" needs ai

also buggy

[keen beacon]

yeah i've just got to that stage as well

it ran into an error, said it auto fixed it and it turns out it didn't

[leaden palm]

agentspace is kinda cool though, personalized deep research and tool use (sending mail, analyzing data, generating audio overviews) plus chrome could be useful if i was an employee

[keen beacon]

oh nevermind.. just needed to put in a key for it to let me click fix

[balmy mist]

wow tbh its too much stuff being released lmaoo

like i cant keep up

i gotta test out this firebase studio tho

[oblique flint]

bro release it already darn it

[barren prairie]

This fire studio is for what?? 😅

[lime coral]

Full stack app

[keen beacon]

they're waiting to release 2.5 flash and gemini coder in one go trust me bro trust me 💔

[barren prairie]

When is the "soon" 🥲

[leaden palm]

the voice agent demo is actually really cool

[spare mango]

Is this for real? Do most people not have access to 2.5 Pro (experimental) on the free version of Gemini?

According to Gemini 2.5 Pro itself, only a select few people have access to this experimental model, and I'm one of them?

Is this making stuff up or do any of you guys not have access to this?

[leaden palm]

llms are llms and llms hallucinate

[spare mango]

so the "best" llm in the market is hallucinating about this simple fact, interesting...

I asked it if there's any difference between Gemini Free and Gemini Advanced, since I have access to 2.5 Pro (experimental) on the free version anyway, and it went off on a tangent about how this model does not exist, then went on to say how I'm part of a select cohort that has access to it.

[leaden palm]

if you think that the hallucination problem will never be solved consider going to a prediction market

[spare mango]

Sorry I don't what you mean?

[leaden palm]

gemini advanced still has advantages (first few i can think of are higher usage limits and more deep research)

[keen beacon]

the biggest advantage for gemini advanced rn is 2.5 pro deep research

i am tempted to go for it because of that

otherwise i wouldn't care

[spare mango]

deep research is another model

apparently the non-deep research model is still the best in the market.

[keen beacon]

?

[leaden palm]

is that only for advanced?

[keen beacon]

yeah

[spare mango]

what

[keen beacon]

free deep research is on 2.0 flash thinking

[keen beacon]

it isn't

[leaden palm]

i read this as a general rollout

but yeah no 2.5 pro is a good model for deep research purposes, at least when testing with my own harness

[spare mango]

LMArena did not place 2.5 Pro Experimental as number 1 because of Deep Research is what I'm saying.

[keen beacon]

i never said they did

[spare mango]

you said you wouldn't use it if it weren't for that so I thought you were implying it was.

[keen beacon]

i said i wouldn't use it if it wasn't for 2.5 pro deep research because you can get 2.5 pro with no discernible rate limit for free on ai studio

[leaden palm]

fun fact: if you trust google's deep research evals, their version would be 146 elo above openai's

[keen beacon]

i'd like to see its performance on HLE

at least that way we'd get a more direct comparison

[leaden palm]

(that's like the difference between llama nemotron 49b and gemini 2.5 pro exp)

[balmy mist]

nutss

[keen beacon]

nemotron ultra is now out on nvidia build

it beats R1 in most benchmarks

https://build.nvidia.com/nvidia/llama-3_1-nemotron-ultra-253b-v1

[balmy mist]

is firebase studio free?

if it is gg

[keen beacon]

in my testing it was... meh

[balmy mist]

lol

the thing that is good about it is that it can handle large code

the other apps fail when i give it anything above 32k for some reason

but you are right it is kinda meh, but I would still say its better than gemini 2.5 by itself

im cooking rn, give me a sec

[keen beacon]

I don't even know if it's fully powered by 2.5 pro

some of it may be a flash model

[balmy mist]

bruhh

it better not be, but if it is that is impressive

really shows how tools can boost up a model

nevermind

it keeps failing

[keen beacon]

Isn't this fp8. The other ones don't have native support. there's still a big leap though

slightly off topic but

the more i test chatgpt 4o latest

(the march version)

the higher opinion i have of its creative writing

it feels like R1 quality (great) but it doesn't fall apart after more than a few chapters like R1 does

What about quasar

quasar disappointed me for writing tbh

step back from chatgpt 4o latest

more robotic

for the most part, i agree with this

although c3.7s would def be in my top 5 minimum

[balmy mist]

wow so deepseek taking ws?

[keen beacon]

yeah

i'm very excited for R2

[sturdy mica]

what is quasar

[keen beacon]

we should be getting that quite soon

[sturdy mica]

model

[keen beacon]

anonymous openai model on openrouter

[sturdy mica]

oh

[keen beacon]

its just a gpt 4o update

prob an api dated version

[sturdy mica]

oh 🥀

💔

[keen beacon]

they shouldve released chatgpt 4o the last one as an api dated version

[sturdy mica]

https://tenor.com/view/case-oh-caseoh-waffle-house-waffle-house-gif-10934642274965704175

[keen beacon]

chatgpt 4o and 4o on the api are too different for them to have released it as a 4o dated version

[sturdy mica]

case oh

[keen beacon]

4o is more professional/"serious", chatgpt 4o is optimised for chat

and ofc creative tasks

[sturdy mica]

ok

did u guys see new secret model in battle mode

[keen beacon]

although it did take them too long to release the current latest chatgpt 4o model as a dated version, for a while you could only access it via the -latest endpoint

[sturdy mica]

its called noob pro hacker obby tycoon

[keen beacon]

thats still the case tho?

iirc

[sturdy mica]

it is o6 mini

[keen beacon]

only lmsys has access to older versions

[keen beacon]

oh this is news to me lmao

yeah nevermind

[balmy mist]

wait the event over?

so where is flash?

[keen beacon]

announcement by logan later maybe?

[barren prairie]

"sooner" 🙂

[woeful geyser]

The entire venom system prompt, summarized by Claude:

[keen beacon]

lol this is what i mean

[sturdy mica]

what is venom

[keen beacon]

lmfao

[sturdy mica]

what is this

also what is venom

is that a prompt

[sturdy mica]

ok

[keen beacon]

it was an anonymous meta model on the arena with a long system prompt

[sturdy mica]

oh ok…………

im gonna play roblox now

https://tenor.com/view/minecraft-minecraft-movie-a-minecraft-movie-steve-jack-black-gif-17997916878979753001

actually twm fortress 2

[wintry tinsel]

Deep seek isn’t that good for creative writing it introduces random non sequiters and makes everything overly verbose and dramatic

[keen beacon]

yeah that was the reason i said mostly

sonnet 3.7 is king

[leaden palm]

"optimus alpha"

[ocean vortex]

lol what

[keen beacon]

...for creative writing

in my experience

[balmy mist]

really?

[ocean vortex]

oh. Ok that's less ridiculous lol

[balmy mist]

i cant take anymore torlling

trolling

[wintry tinsel]

It is across the board but in some areas with the right prompt 2.5 pro can be better

[leaden palm]

soon™️

[balmy mist]

i just want my damn nightwhisper

man

[brittle tiger]

Could this be nightwhisper?

https://x.com/_davideast/status/1909984439985229940?t=ddJfOgwZywH0inE1AxyWyg&s=19

David East (@_davideast) on X

Hey Firebase. I'm back.

https://t.co/e8XYdNhGxk

X · Apr 9, 2025, 2:59 PM

[balmy mist]

forget everything else

[wintry tinsel]

2.5 pro is the best at nsfw writing lmao

[balmy mist]

noooo

its bunns

[sage raptor]

nightwhisper there ?

[balmy mist]

jk

its solid

just not nightwhisper

[balmy mist]

2.5 pro maybe with tool calls

[keen beacon]

woah

new model

[balmy mist]

really

give me plz

[ocean vortex]

is it better than opus for that in your experience? @keen beacon

[keen beacon]

let me test this

[keen beacon]

yeah once u get it setup. multi turn, context usage, all their training tricks, etc., are just awesome and result in a really good experience imho

[balmy mist]

it says coming soon

so not out yet

[keen beacon]

oh

hmph

should be up by the end of today then

must be another oai model

[balmy mist]

nooooo

it better be o3

[upper wolf]

2.5 pro still throws so many refusals. It won’t do anything related to web scraping (puppeteer/selenium etc.). on top of that, it can’t even say that it won’t, it says that it’s “not sufficiently trained” on it

[balmy mist]

wait you already test o3 rright and it was mid?

[keen beacon]

i have tested o3 medium

it was pretty good

but there are some things it performs meh on

web development is still a weak point

but significantly better than o1/o3 mini

[barren prairie]

O3 is out? But when ???

[keen beacon]

it's not out

i help labs out sometimes

[wintry tinsel]

Google event goes through 11th right

[balmy mist]

only for chads

[balmy mist]

okay i guess thats okay

ill just stick to 2.5

[wintry tinsel]

Opus has less flavor and prompt adherence but more personality and a “natural feel to it”

There’s no reason to use it though since it’s much more pricy

[keen beacon]

i really like opus even now when price isn't a consideration

just has good vibes

[sturdy mica]

what was that super good coder on lmarena again

it was nightfall or something

[keen beacon]

nightwhisperer

or nightwhisper

or whatever it was

it was webdev arena only

[sturdy mica]

yeah

when is that gonna come out sigh

anyway what are you sll talking about

[wintry tinsel]

The reason google models are so cheap is they are trying to roll them out en masse, if they went the other route and had less instances for higher compute and cost they might be able to have the model give much better outputs relatively, I could not know what I’m talking about tho

[leaden palm]

nah google is now sota

[sturdy mica]

why can i not

react to things

oh

there

i cant react to ktibow

[keen beacon]

hmm sorta weird 2.5 flash isnt released yet, maybe they might do an anon model on openrouter 🤣

[barren prairie]

I can react 🙂

[sturdy mica]

lucky you

https://tenor.com/view/minecraft-minecraft-movie-minecraft-meme-minecraft-villager-villager-gif-5342559956921446299

im going to hit you with my car

that villager is YOU

this is mildly uninteresting and nobody is ralking now

https://tenor.com/view/half-life-gman-g-man-half-life-2-rise-and-shine-gif-5438670790719714756

[keen beacon]

peak shitposter

[torn mantle]

can someone confirm

[torn mantle]

i see

thanks

[wintry tinsel]

They should add auto regressive image capabilities to 2.5 pro

[sturdy mica]

thanks

[sturdy mica]

what’s that

[keen beacon]

native image gen

[keen beacon]

it already has them

they just haven't released it yet

actually no

i may be getting mixed up with 2.0 pro

[keen beacon]

2.5 pro is highly likely to be a cpt of 2.0 pro it should still have them

im curious whether they worked on it at all with 2.5 pro

[sturdy mica]

ok nerd

[brittle tiger]

I don't think it js

[sturdy mica]

whats the best model rn

its 2.5 pro right

ai studio version / api version

[torn mantle]

https://x.com/Kimi_Moonshot/status/1910035354570371082

Kimi.ai (@Kimi_Moonshot) on X

🚀 Meet Kimi-VL and Kimi-VL-Thinking! 🌟 Our latest open source lightweight yet powerful Vision-Language Model with reasoning capability.

✨ Key Highlights:
💡 An MoE VLM and an MoE Reasoning VLM with only ~3B activated parameters
🧠 Strong multimodal reasoning (36.8% on

X · Apr 9, 2025, 6:21 PM

[keen fulcrum]

There is still imagefx which gives you exceptional results

[balmy mist]

as i play with firebase some more i see what truly is

its pretty much a competitor for cursor and every other ide and even claude code tbh

so it basically gemini code

just not in the cli

and this agentspace is nuts

wow

[olive mesa]

are there any models better than 2.5 yet

[balmy mist]

gg open ai

[balmy mist]

there really is no need for any models better than 2.5

[wintry locust]

skibidi toilet rizz

[olive mesa]

there kinda is

[balmy mist]

i would like it

[olive mesa]

usually theres a new best every week to a month

[balmy mist]

but its not necessary for what we need

i would say maybe faster

and larger output and maybe window

[keen beacon]

2.5 flash soon 🤔

[balmy mist]

but thats pretty much it

[keen beacon]

is faster

[balmy mist]

IQ wise no

it would be nice tho

[olive mesa]

it would be nice to have a passive superintelligence

[thorny drum]

i could use some agi rn

[balmy mist]

i mean but do we need that

[olive mesa]

not the allied mastercomputer type

yeah

[balmy mist]

agi is subjective

some people say its already here

some say years away

some say this year

depends on your definition a this point

remember wen it was to pass the turing test lol

[olive mesa]

yeah

[balmy mist]

but this agentspace is very interesting

this is gonna go wild once people get on it both it and fire studio

[olive mesa]

are most ais trained with curriculum learning?

i honestly feel like we would have asi by now if so

like just giving them near-impossible questions and every once in a while they produce an extremely good cot and response

then add that to a dataset

train

and repeat

[keen beacon]

that is sorta being done rn

[barren prairie]

I need deepSeek r2 😁

[balmy mist]

yo firebase is free

wtf

why the hell am i using roo anymore lol

[balmy mist]

yeah i wouldnt mind that

how is it free?

like am i missing something?

gg cline, roocode, augment, cursor, windsurf RIP

bolt

claude code lol, but i still like that its a CLI so claude still got a lil hope

no its using 2.5

why the hell would i use my own api lmaooo

actually nevermind i know what to do

gonna use my own free exp model, buts pretty much like roocode in terms of how you pay

oh i see what you mean

so by defaul the built in is 2.0?

gonna try it with 2.5

[hidden widget]

i can see another clickbait video from IcodeKing here

[oblique flint]

2.5 flash still not released?

[keen beacon]

a little strange right?

[balmy mist]

nahh firebase is expensive

going back to free version lol

[keen beacon]

maybe 2.5 flash doesnt have enough votes in the arena

[oblique flint]

disappointing performance and they want to keep cooking or smth else?

[keen beacon]

if its stargazer i dont think its disappointing

[oblique flint]

Im just confused cause the leaked model string in the python sdk said april 9th right?

[keen beacon]

yes

[balmy mist]

nice

you gonna use that in firebase?

also whats a good prompt for refactoring an app i have to look better visually? should I saw apple design expert and stuff?

[keen fulcrum]

Did they announce nightwhisper?

[brittle tiger]

No

[wet ingot]

Anyone know what the model “dragontail” is? I got it on lmarena and it was good but I can’t find anything about it

[balmy mist]

how good is it?

[vivid oyster]

Yeah

How good

Ima try it rn

[torn mantle]

why would anyone prefer grok 3 over gemini 2.5 pro?>

[vivid oyster]

Cuz elon musk is hot

[torn mantle]

did he do it with you?

[wet ingot]

Idk but it had a better answer to my logic problem than most other models and it was good with an image

[torn mantle]

xd

[vivid oyster]

Yes

[torn mantle]

it felt good?

[vivid oyster]

Yes

[vivid oyster]

What model did it say it was

Did u askit

[wet ingot]

No

Let me see if I can get it again

[vivid oyster]

Im trying rn

Its google @wet ingot

[wet ingot]

I got it

[wet ingot]

Yeah lmao I just got the exact same thing

[vivid oyster]

Only google says 'i am a large language mode trained by google'

Maybe its nightwhisper

Was it better than 2.5 pro

[wet ingot]

Hard to say

[vivid oyster]

Whos

[vivid oyster]

Did u ask wat middle it was

Model

[torn mantle]

Tf

Tf

[brittle tiger]

Shadebrook passes vibe test. dragon tail does not for me

[torn mantle]

2 new models?

[vivid oyster]

Yeah

[brittle tiger]

Shadebrook is Google

[vivid oyster]

Dragontail is google too

[keen ferry]

google

[wet ingot]

Interesting Google is adding so many models

Without saying anything

[vivid oyster]

Maybe they're experimenting

For 2.5 flash

[wet ingot]

Yeah

[vivid oyster]

One of them might be nightwhisper

[torn mantle]

pro? flash?

[vivid oyster]

Idk

Im trying to get it

So I can ask it

Questions

But it said the same prompt all google model says

[torn mantle]

you are trying to debunk it

silly you

[wet ingot]

We don’t know what it is but it’s good with logic and images

[vivid oyster]

Yes

[torn mantle]

shadebrook is not good

probably like flash lite

[upper wolf]

Is it thinking

[torn mantle]

so fast too

no

[keen beacon]

eugh

[keen beacon]

ooh

let me test this stuff

[wet ingot]

Dragontail was good for me

[torn mantle]

its probably flash without thinking

its blazing fast

[brittle tiger]

Actually idk. It just got an arc-agi problem wrong that it got right this morning. It's very fast tho

[keen beacon]

hmm

[torn mantle]

please let dragontail be good

pleaaaaaase

[keen ferry]

can someone give me basic questions

i got him

dragontail

[torn mantle]

how many r's are in strawberry?

[torn mantle]

make a discord clone in one html file.

[vivid oyster]

How tf do u know

[keen ferry]

magic

[vivid oyster]

That u got dragontail

[torn mantle]

shhhhhhh

[upper wolf]

You’re playing Roulette at a casino with a broken wheel that makes it 0.36% more likely to land on Green. What is the new expected value of a $100 bet on the color red?

[keen ferry]

its soo good

There are three 'r's in the word "strawberry".

[torn mantle]

ask him this

[wet ingot]

I mean a lot of newer models know that

[torn mantle]

make a discord clone in one html file.

[keen ferry]

k

[torn mantle]

and give me the html file

[keen ferry]

sec

[torn mantle]

and i will tell if its good or not

[vivid oyster]

This bot is a dumbass

[keen beacon]

oh dear

[olive mesa]

why do models get that specifically wrong

people probably memed it then it got into a dataset

[torn mantle]

dreamtides is meh too

[torn mantle]

because its a tokenization issue

[vivid oyster]

I think I got it

It told me there's three

[keen beacon]

gottem

[zinc ore]

Don't do R questions, they don't see letters like us

[keen beacon]

it's taking a while to start streaming

[torn mantle]

you got his ash

[keen beacon]

okay it just started

[vivid oyster]

I made him give me a discord html

[torn mantle]

gimme gimme

[keen beacon]

im also giving dragontail a web task

[vivid oyster]

This is what it gave me

[torn mantle]

worse than gemini 2.5 pro thinking

[torn mantle]

whats this

are ytou sure its dragontail?

[upper wolf]

Guys

[vivid oyster]

Dragontail

[torn mantle]

its shadebrook?

[upper wolf]

is dragontail thinking

[keen beacon]

yes

[vivid oyster]

Yeah

It's flash thinking

[keen ferry]

mine is still generating

[vivid oyster]

Probably

[keen beacon]

this seems to be flash

[upper wolf]

Thanks

[torn mantle]

its meh

[keen beacon]

yes

[vivid oyster]

It takes a while to start

Sending

[torn mantle]

wtf

[vivid oyster]

Is this dragontail

[torn mantle]

wait a minute

this is so good

[keen ferry]

yeah

[torn mantle]

which model is that

[keen ferry]

dragontail

[torn mantle]

just from that output

[keen beacon]

yeah it's at least on par with 2.5 pro thinking in my very limited testing thus far

which i wouldn't expect from a small model

[torn mantle]

i got it

i will try it

doesnt seem like a thinking model

so fast too

[keen ferry]

it was slow for me

[zinc ore]

Google's models from now on are hybrid models, so if it's something like flash it'll be both

[keen beacon]

it is

it seems dynamic

i have had a request with practically 0 time and another with like 15s

seems pretty good

[zinc ore]

Faster than 2.5 pro?

[keen beacon]

yes

[torn mantle]

its kinda the same as gemini 2.5 pro

a bit worse no?

[keen beacon]

no

they're very similar, i can't really discern much of a performance difference as of yet

which is sorta surprising given this seems to spend half as much time on it and yet matches pro

[zinc ore]

Could also be updated pro if not flash

But my bet is on flash

[vivid oyster]

It's faster

Its probly just flash thinking

[zinc ore]

Flash 2.5

They don't call it thinking anymore

[keen beacon]

i would be kinda surprised

in my (again somewhat limited) testing it doesn't seem worse than pro

[drifting thorn]

All 2.5 models are thinking models now

[keen beacon]

and if this is flash

wtf are the other like 3 anon google models on the arena rn

it doesn't make sense to add flash as an anonymous model today when they're releasing it like tomorrow 😭

[zinc ore]

Yeah depends if we getting flash this week, or more like next week or something

[drifting thorn]

I guess they are Gemma 4?

[torn mantle]

its you

you are giving us all these outputs

[drifting thorn]

Hhhhhhj

[vivid oyster]

.;

[brittle tiger]

This seems plausible bc it nailed the arc-agi problem this morning after running like 4 hypotheses to solve and combining relevant ones and then this afternoon zipped out a totally wrong answer on same problem

[keen beacon]

ffs 🤦‍♂️

the full switch to the new arena couldn't come soon enough

[harsh flume]

Did google remove the option to data train from AI studio or am I just not finding it?

I was gonna test it out today 😭

We should get new LB update today btw

[keen fulcrum]

https://www.bleepingcomputer.com/news/google/google-takes-on-cursor-with-firebase-studio-its-ai-builder-for-vibe-coding/

Microsoft has copilot and vscode (roocode is better), Google has Firebase Studio and then there are third party ones

I wonder if Apple will do something about it

Xcode with AI?

Firebase Studio and Nightwhisper can be the cheapest option honestly.

[alpine coral]

were they the same prompts? perhaps it is dynamic if they were the same (though if they were different, esp in length and / complexity) that might better explain the discrepancy

i just got it. v strong indeed. would've said almost certainly thinking model based on the quality and time(/delay) of the response (+ it was against command-a, which isn't thinking afaik)

[raven void]

grok 3 mini high is looking pretty good

[balmy mist]

thats new?

[raven void]

yes, API

https://docs.x.ai/docs/models#models-and-pricing

[balmy mist]

oh snapp

finally

[keen beacon]

pretty bad pricing lmao

[balmy mist]

lmaoo yeah its like sonnet right?

[raven void]

mini looks like a pareto frontier model though

[brittle tiger]

For being huge and propped up on infra faster than even seemed possible it's not that surprising tho

[ivory schooner]

愿即将到来的Behemoth 是24k......

愿即将到来的Behemoth 是24k......

[sturdy mica]

llama 4 behemoth?

what

[balmy mist]

its coming out?

[vivid oyster]

No

He's saying he wishes it's 24k

24k karat gold

When it comes out

[balmy mist]

oh lol okay

[upper wolf]

It might’ve just been maverick with a modified system prompt it didnt seem that accurate

[vivid oyster]

It is

Behemoth is prob anonymous-test

It's shet

[hardy pecan]

https://x.com/bindureddy/status/1910185483545776630

Bindu Reddy (@bindureddy) on X

Qwen 3 is coming in hours!

Good chance of topping the open source leaderboard

X · Apr 10, 2025, 4:18 AM

[hardy pecan]

Tested Grok 3 Beta in OpenRouter for the 20 public SimpleBench questions, it got 6/20

[torn mantle]

How much did gemini pro get

[oblique flint]

still no 2.5 flash

🥲

[hardy pecan]

9/20

[torn mantle]

does grok 3 use reasoning?

[woeful geyser]

Got Maverick full release! 😌

Quite a surprise, but that's a good thing: I can't pinpoint the exact model because its "vibe" doesn't scream out loud.

Hoping for the actual result.

[torn mantle]

https://x.com/btibor91/status/1910237861674353108

Tibor Blaho (@btibor91) on X

New ChatGPT web app version deployed just now (after midnight San Francisco time) adds mentions of "o4-mini", "o4-mini-high" and "o3"

X · Apr 10, 2025, 7:46 AM

[drifting thorn]

WHY ALL APIs have a context limit of only 1 million

I was just trying to set up a MCP server with Cline

it failed

it says it's out of context window when it's like 80% done

arghhhhhhhhhhh

[cedar tide]

Not yet

[cedar tide]

https://x.com/JustinLin610/status/1910230919346270300?t=y1qXCeqpSD6AIPKMQVNn-w&s=19

Junyang Lin (@JustinLin610) on X

Sorry Bindu, this is not gonna happen that soon. We still need some more time.

X · Apr 10, 2025, 7:18 AM

[drifting thorn]

I need real 10 million context window model

There's been 3 months since the publication of "Titans" architecture by Google

I hope Google can make a reasoning large multimodal reasoning model out of that architecture

[cedar tide]

Wait grok 3 reasoning

[drifting thorn]

Since Grok 3 have no ethics regulations

[tall summit]

hihi

[fleet lintel]

is dragontail as good as nightwhisper?

[torn mantle]

seems like they werent satisfied with the performance

[torn mantle]

coding wise?

no

nightwhisper is finetuned intensively on coding

they probably made a good reward model for web dev

styling wise

[cedar tide]

Dragontail

[noble zinc]

could dragon tail be o4-mini or o3?

[cedar tide]

Its from google

[noble zinc]

hopefully pricing remains similar to 2.0 flash

[torn mantle]

its def not nightwhisper

i got much better results for a discord clone from nw

its on par with gemini 2.5 pro

idk if its better or not

[cedar tide]

sorry for the bad screenshot

screenshot on laptop
dragontail

2.5 pro

i think its gemini 2.5 pro, low thinking

the results are very similar to 2.5 and it thinks but for less time

[oblique flint]

o4 mini 👀 I have no hope for o3 being affordable though lol

[hazy quest]

Do you guys have access to Veo2 in AI Studio? It seems to be rolling out, some have it, some don't. I don't 😦

[hardy pecan]

Yeh I've tried it out

[keen beacon]

it's openai's launch day in the week... we may actually be getting things 👀

[oblique flint]

hmm will o4 mini launch before 2.5 flash, seems like it lol

[keen beacon]

i wonder if o4 mini is based on an updated 4o mini base

or if its just more roids 🤣

seems like google sort of forced their hand

how could it be o4 mini

its literally not thinking

it streams immediately and there is no apparent thinking

yeah

[sage raptor]

maybe its o4 mini low

[keen beacon]

no

and i benchmarked it it must be an insane regression from o3 mini 🤣

there is 0 thinking

nada

zero

people are crazy 🤣

i'm more interested in this

they're adding it to openrouter "tomorrow morning" (EST)

imagine if its 2.5 flash 🤣

same naming scheme as the anon openai model so i don't think so lmao

[keen beacon]

could be named like that to trip people

i find that a little unlikely

[brittle tiger]

Why does open router say o4-mini and have same stats basically?

[keen beacon]

openrouter doesn't say o4 mini

[brittle tiger]

Lmaoo I need to stop trying to use brain 2 min after waking uo

[torn mantle]

im not hyped for any o-serie model

[keen beacon]

suit yourself

i'm still very interested

competition = always good

[alpine coral]

quaser ?

[keen beacon]

ye

[alpine coral]

yeah no thinking there.. but it can't be o4-mini can it?

tf is the point of their naming schema if that is the case ha

[keen beacon]

i measured gpqa diamond and math 500 it must be a severe regression if it was o4 mini (which it can't be because its not thinking)

[alpine coral]

yeah the non-thihking alone

[keen beacon]

#general message

[alpine coral]

but also performance-wise, solid as it is - it's not rubbing up against the frontier in any way

[alpine coral]

ah yup i recall

[keen beacon]

also: o3 mini gpqa diamond: 74.8%, math 500: 97.3%

@keen beacon you should test optimus alpha when it releases

[keen beacon]

i will if rate limits are like quasar

[alpine coral]

some scores from a quiz (~20 questionss in one prompt; same as shared above here somewhere)
just fwiw

wait.. that wasn't sorted right

[tall summit]

which server's that?

[keen beacon]

openrouter

[visual turret]

why is 3.7 sonnet in 23 place? gemini flash lite is in 16

[keen beacon]

use style control..

[visual turret]

well i know but i want to know why tbh.

[visual turret]

still doesn't make sense

[keen beacon]

that's just what something based on human preference will produce 🤷‍♂️

[oblique flint]

man the difference in capability between 2.5 pro in ai studio vs cursor is so huge. I really wish it wasnt so bad in cursor lol

[calm sequoia]

Coyld you explain what does it mean "the open-router server"?

[fleet lintel]

That's disappointing. I am hoping for better models than 2.5 Gemini

[keen fulcrum]

Just use roocode

It deleted my 800k context conversation which is frustrating though

[oblique flint]

I have been avoiding roocode because I hear it can be expensive af

[golden ocean]

I created new chatgpt alt account on new chrome profile and I can use gpt 4o forever it seems unlike my og chatgpt account. Can also generate way more images

Is that a feature for new accounts or something

Will prob return to normal after a while

[balmy mist]

wait who owns dragontail?

[balmy mist]

2.5 is too good

lol

https://x.com/googlecloud/status/1910124837781328117

Google Cloud (@googlecloud) on X

TPU x @ssi news: "OpenAI co-founder and former chief scientist @ilyasut’s new AI startup, Safe Superintelligence, is using Google Cloud’s TPU chips to power its AI research," via @TechCrunch ↓ https://t.co/G0jnB5rEua

X · Apr 10, 2025, 12:17 AM

i wonder what progress he has made

[vague orbit]

Any reports on how Cognito performs?

[sonic tendon]

i think most of it is that claude's terse response style doesn't work well when you're trying to do quick one-shot comparisons between models

which is what lmarena "power users" tend to do en masse, i think

[vague orbit]

What kind of persona is a lmarena power user?

[sonic tendon]

it's nice when you're expecting it though

[sonic tendon]

people that talk a lot on this discord, basically

and/or do more than ~5 chats on lmarena a day on average

(that number is totally arbitrary, i just made it up)

[vague orbit]

I would guess that if you have a problem and wanted to solve it with an LLM, throwing it at LMArena would lead to really mixed results

[sonic tendon]

personally, i do it just to try and explore different models (plus, it's one of the lowest-barrier-of-entry ways to get alpha access to new LLMs as they're being developed)

but yeah, you're right

i don't use it as a general-purpose chatbot, it's more for my own curiosity

[tall summit]

guess im a power user now

[sonic tendon]

welcome to the club :3

[sonic tendon]

what's optimus alpha

[keen beacon]

anonymous model releasing on arena today

[sonic tendon]

ah

openrouter too?

[keen beacon]

sorry i meant openrouter

lmao

[sonic tendon]

all good

[tall summit]

hyped

[sonic tendon]

does openrouter have a discord?

[keen beacon]

yup

look at their footer

[sonic tendon]

joined

oh, grok 3's finally open access now

[tall summit]

i am hearing about so many random models and i've been in the server for 2 minutes

[sonic tendon]

lmaoo

yeah 2025 ai scene in a nutshell

[sonic tendon]

what exactly are you trying to do? most models tend to perform pretty poorly when operating with that much context, even if they do support it officially

especially with coding, in my experience

off the top of my head, i don't know of any models that support much more than 1M context, anyway

[keen beacon]

the new llama models are supposed to but they suck balls

and some of the geminis iirc go to 2M

[sonic tendon]

i forget what that fiction context window benchmark is called

[ocean vortex]

I find it odd this name matching, even the context matches? 🤯

https://x.com/SILXLAB/status/1909475116637208937?t=Jd0rllJnICEeVyKNPuYZ_g&s=19

SILX (@SILXLAB) on X

Quasar this, Quasar that...

It’s finally time to fully announce the new Quasar Series — introducing our 400B parameter models (1m context!!!) and smaller ones, the first of their kind. Built using a new scaling law training pipeline, advanced techniques, and more. By @LambdaAPI

X · Apr 8, 2025, 5:15 AM

[keen beacon]

17 yr old grifter or smthing. just a larper just ignore

[sonic tendon]

lol

[keen beacon]

😭

[ocean vortex]

probably. It's somewhat unusual for openai though as well. It's still not released and only on openrouter

[sonic tendon]

0 weights

[keen beacon]

this guy is actually 17 yrs old

[sonic tendon]

meta in general seems to have been sucking balls lately

/nonsexual

[keen beacon]

behemoth better be good

[sonic tendon]

it's gonna suck

probably

[keen beacon]

it'll probably underperform

we'll see

hopefully we get o3 today

• o4 mini

[sonic tendon]

oh, on that note

man, i need to stop doubting myself

[keen beacon]

wats that for

i kinda doubt it will take top position

idk

[sonic tendon]

https://polymarket.com/event/which-company-has-best-ai-model-end-of-april/will-openai-have-the-top-ai-model-on-april-30?tid=1744290054541

[keen beacon]

maybe joint 1st w/ stylectrl

[north vale]

It’ll matter a lot to how well reasoning scales to superhuman levels how o4-mini performs

It really depends whether they posttrained o3 on their new 4o stuff

imo

[keen beacon]

it also depends on which reasoning effort version they put on the arena

o3 low will be very different from o3 high

its conceivable the new anonymous chatbot could top the leaderboard

[keen beacon]

which one

there are many

[sonic tendon]

i mean, i had a pretty good impression of it, but you may be right

[keen beacon]

anonymous chatbot/quasar by openai. since the leaderboard is mainly human preference

quasar isn't on the arena

and even if it was

[keen beacon]

it is lol

[north vale]

It’s ass compared to 2.5

[keen beacon]

i don't think it would take the top spot tbh

it's not as nice to talk to as chatgpt 4o latest

also seems less creative

[oblique flint]

well if it takes like several minutes to answer a question idk if it will get the top stop lol

[keen beacon]

no it isn't

it's on openrouter

[sonic tendon]

yeah, as we've learned, this matters a lot more than people think it does

[keen beacon]

anonymous chatbot is the same though lol

[sonic tendon]

in the original (gradio) lmarena, it pauses the output of both models until they both start sending tokens

[keen beacon]

anonymous chatbot isn't on the arena

not the same name, but its the same model

[brittle tiger]

making up 30 elo points is harder than it seems too

[sonic tendon]

so, both models get paused until they both finish thinking, basically

[keen beacon]

there's anonymous-test