general

[stray aspen]

thats nice

[hollow imp]

...

Use jee advanced for the benchmarks

Then see ☠️ 🔥

[stray aspen]

glm 4.5 is still better tho

[wicked root]

i hope so

[stray aspen]

qwen 3 think is absolute trash for coding

[open mountain]

He has always been like this, there is little modern training data

[whole wagon]

The openAI 120B model scores 44% in aider on highest reasoning mode

That is awful

It's basically a paper release. To say they open sourced smth

And I don't know what all that hype was about

[stray aspen]

thats what im saying

china is still on top

and top it off it only accepts text

[solid brook]

Guys

[karmic bough]

What's the best model for coding ?

[solid brook]

The horizen beta model wasn't the open source model

[stray aspen]

i dont know brother

[whole wagon]

Ofc

[stray aspen]

maybe claude 4 sonnet or the new opus

[solid brook]

I tested. Their kmowledge cutoff is diffrent

[karmic bough]

What abt Gemini 2.5

[whole wagon]

Yeah well it's obvious. Because horizon beta is not bad

Kek

[solid brook]

Sad

[whole wagon]

The arch also has 0 Innovations

It's basically just taken from Chinese LLMs

I guess they wanted to keep anything good for themselves

I don't see the point in the model ngl. Apart from PR

[primal orbit]

I'm still getting "potato" in the arena. So it was not open source model. Or not the one that got released.

[blazing bison]

ok after 1 hour testing openai oss models, bad vibes for code

[plush anvil]

hello

[wheat onyx]

My guess is it's good for summarization and maybe some basic math. Worth checking. If it can't do those well, that would be disappointing

[daring rover]

what about swe bench ?

polyglot seems like it's too contrived

[daring rover]

it was apparetnly gpt 5 low effort

[wheat onyx]

I'm excited for the lower hallucination rates more than anything

The modest bump in reasoning is nice too sure

[daring rover]

the hallucination rate is higher for oss

do you know if it's lower with gpt 5/

[wheat onyx]

Gpt5 is much lower

[daring rover]

nbice

[daring rover]

source?

[hallow ridge]

I have an instagram account with over 300k and I don’t want it anymore

[wheat onyx]

They announced an entirely new method with verification a bit ago

[daring rover]

yeah but how do you know halllucinations are lower

[ember sentinel]

Yo guys, does anybody know if the horizon models on OpenRouter were the GPT-OSS models?

[pulsar aurora]

Any free Ai agent tool that let's us use Ai on browser to perform any tasks?

[novel flame]

Definitely not. Horizon Beta is much stronger. It could be GPT-5 low effort, that would make sense for my testing at least

[wheat onyx]

https://openai.com/index/prover-verifier-games-improve-legibility/

[wheat onyx]

Looks good for coding. Not something I'll be using, but unfortunately if it beats opus 4.x then they may have a cash flow crisis

[neon idol]

Did you look for claude 4.1?

The new chatbot

[wicked root]

anyone know how rate limiting works in Gemini pro?

[primal orbit]

https://i.snipboard.io/mF1nyf.jpg

first time seeing such response 😄

[jade egret]

is claude 4.1 opus good

it js like a upgrade opus 4 ig

[stray aspen]

it was a very minor upgrade

[random shard]

GPT OSS is wow

chefs kiss

great little coding model

[stray aspen]

nah its mid

[random shard]

no

most models won't fit on that

you might get lucky and someone will quantize it enough to barely fit on 6gb, but realistically upgrade your GPU

or accept CPU inference

Possibly, but unlikely tbh

Quantized 4bit right now uses 10gb

Depends on your CPU, RAM, quantization

My experience with smaller models on a 128core AMPERE box has been, 5 - 10 TOK/s

So, don't expect fast.

Probably like, 1 - 2 TOK/s?

Better than nothing

Try it with Llama.cpp CPU mode

never used it.

keep your expectations is low is all I can really say

I personally run models on a M4 Mac Mini

usually using exo so I can cluster it with my M1 Pro MBP

Llama.cpp now supports clustering via RPC but it's very experimental

exo was designed from the ground up for clustering

you're not going far with that

You're gonna wanna upgrade

I'm using a RTX 4070 Super Ti 16gb, and even then I can't fit most models

the ones I can, aren't that fast.

I like mac clusters for energy efficiency

if you don't care about energy efficiency, go NVIDIA

idk

you just gotta try

🤷

I use finetuned for chatting models

I'm really liking GPT-OSS tho

It's a nice model

not a expert at anything, but well rounded.

o3 mini -> GPT OSS 20b

o4 mini - GPT OSS 120b

[whole wagon]

Tbf. The openAI open source model is efficiency SOTA. It has only 5.1B active

[random shard]

no one knows

[random shard]

it's just MOE

[whole wagon]

It is much smaller than qwen3 also

[random shard]

I wish someone would implement Apple's papers about streaming LLMs from disk to RAM

[whole wagon]

It performs just below the big qwen3 model and half the params, way less active

Qwen is 235B with 22B active

[random shard]

I'd say it's better than qwen3 by a long shot

my experience with qwen3 is it never follows instructions.

[little narwhal]

Knowing them they’ll probably call it o5

[random shard]

No. It’s GPT 5 confirmed.

No. Any storage.

It would be ideal with nvme

[devout vault]

are you the yupp ai guy?

what even is yupp ai

[random shard]

But you can run a LLM on a pentium 2 pretty fast if you really want

[random shard]

No

What is YUPP?

[lone relic]

prboably

[lone relic]

ye thats right

[lone relic]

thats right too

[random shard]

GPT 5 will be non reasoning

[leaden palm]

????

[lone relic]

honestly openai def did set a bar for open source models with oss

[obsidian shell]

who lied to you?

[random shard]

???

[leaden palm]

iirc at the start of gpt-5, its sole purpose will be to route to reasoning models when appropriate

[lone relic]

yeah exactly

[random shard]

What would they make it a reasoning model. It’s a non reasoning model as we understand now. The o series models reason.

[leaden palm]

it's not a normal continuation of the gpt series

[devout vault]

Gpt -5 is a model integrated by multiple models

[lone relic]

they r changing the game

[random shard]

So non reasoning

[lone relic]

yep

[obsidian shell]

what makes you certain it will be non reasoning?

[leaden palm]

gpt-5 is more of a system than a "non reasoning model"

[devout vault]

GPT-5 is releasing this week

[random shard]

We don’t know yet. None of the continuations have been super consistent

[devout vault]

Ye

[lone relic]

ye

bro its exciting lowk

and i wonder how it will do against gpt 4.5

and who it will be available to

[random shard]

Sam Altman posted a screenshot of its output already. Also, the GPT-X have consistently been non reasoning.

[lone relic]

show ti

*show it

[random shard]

Infinitely better I assume.

[little narwhal]

I think after GPT-OSS and GPT-5 OpenAI will probably run out of steam for a few months

[lone relic]

tbh gpt 4.5 had too much hate and is underrated

[leaden palm]

for that query, it decided to not reason.

[lone relic]

wait where did u find rthat

[leaden palm]

https://x.com/sama/status/1952071832972186018

Sam Altman (@sama)

@nicdunz turns out yes!

X · Aug 3, 2025, 6:19 PM

[blazing bison]

sam post

[stray aspen]

its on twitter

[daring rover]

this is like matching some alignment proposals people made years ago

that's pretty neat

[random shard]

Do you have solid evidence it will be a reasoning model?

[devout vault]

gpt 5 seems like the future tbh

[lone relic]

hopefully it is free for all of us

[devout vault]

Gpt 5 will be a reasoning and not reasoning model

[random shard]

It had hate because it was a stupid model.

[devout vault]

At the same time

[blazing bison]

it's not

[lone relic]

ye

[blazing bison]

maybe in arena it will be

[lone relic]

ay lmarena got us tho for that

[devout vault]

Sam Altman said it will be free for everyone so yea

[random shard]

I doubt GPT 5 will be reasoning. They just sorta fixed model naming.

[lone relic]

oh ye i forgot

[blazing bison]

@devout vault he didnt said that, he said that free will receive something, plus more inteligence, and pro even more

[devout vault]

Ph

Oh

[random shard]

Why would they backtrack and Use the naming scheme from a non reasoning model.

for a reasoning model

[lone relic]

i bet hes gna do smth like gpt 5, then gpt 5 pro for paid users or smth

[devout vault]

I like the not reasoning models more tbh

They are more fast

[random shard]

I don’t see how they’re going to improve on O3.

[lone relic]

its reaoning and not reasoning

[random shard]

O3 has been the best one so far.

It’s well rounded.

[lone relic]

just because something is reasonning does not mean it is the best

tho tbh o3 is darn fast for a reasoning model ngl

at least much much faster than o1

[random shard]

No, it’s the best because it does an amazing job at every task I’ve thrown at it to date

O1 was impressive. O3 is insane.

[lone relic]

haha ive acc seen it screw up a bunch of times

[lone relic]

agreed

[blazing bison]

o3 is the best model for me

[random shard]

Same. But if you talk to it about what it did wrong it quickly corrects itself.

O3 has been amazing

[lone relic]

oh nono it has done some mistakes too still

[blazing bison]

if opus was the same price then opus would be the best

[random shard]

O4 mini makes too many mistakes in my experience. O3 doesn’t really hallucinate either.

[lone relic]

especially still bad ad debugging

which gpt 4.5 ironically did well

[random shard]

You want Claude if you wanna program.

[lone relic]

could be

[random shard]

Claude is the best programmer

And the least polite model.

[blazing bison]

gpt 4.5 is another rich only model

[random shard]

4.5 should’ve never happened.

[lone relic]

subjective, i keep on swapping between o3, 4 opus, and gemini 2.5 pro

[random shard]

O3 imo

[blazing bison]

i like how claude explain things

[random shard]

Gemini has massively improved

Gemini is the most exciting model tbh.

[lone relic]

claude is goated in writing

[lone relic]

agreed

[blazing bison]

o3 talk to you like always doing technical report

[random shard]

Not the best at anything. But it’s skills keep growing by the day

Gemini was like a kid that grew up quickly. Now it’s in college.

And it feels like it’s going for its masters degree soon

[lone relic]

its context is also hella high

[lone relic]

lol

[random shard]

TPU effect

[lone relic]

yep

in lmarena probably

[random shard]

I love Google’s method

nvidia expensive
lets design our own chips

[ocean vortex]

ehm... the big positive is they stopped inflating numbers with parallel compute. Other than that very marginal update huh

[lone relic]

and also i have gemini pro, i never felt 2.5 pro hit its limits except for once

[random shard]

TPU go BRRR

[lone relic]

i used claude opus 4.1 in warp and its hella good

[blazing bison]

bro is impossible to tell the difference between opus 4 and 4.1 on arena

[random shard]

I hope Apple actually designs their own TPU servers.

It’s rumoured.

[lone relic]

thats true acc

i tested it myself so ik

[random shard]

Imagine if Apple just dropped an update to Siri that BTFO’d Gemini.

They have their rumoured answers app.

[lone relic]

damn thatd he crazy

[ocean vortex]

You basically need identifiable data that wouldn't be based on overall performance lol

[blazing bison]

yes, that's why i said on the arena

[ocean vortex]

I'm intrigued by this open-source model though

o4-mini-high performance, almost

and all the features looks like, even the reasoning effort retained

[random shard]

It's already free to try

https://www.gtp-oss.com

that's how I'm using it right now.

Universal WebGPU support when?

I want WebGPU

[devout vault]

in aistudio yes

[random shard]

imagine if we had a beauwolf cluster crowd sourced from the internet

using WebGPU

Imagine.

[ocean vortex]

It's insane that this is 5.1b active params and 117b total

[random shard]

It's been done

[ocean vortex]

🤯

[random shard]

there was a model trained using a cluster over the internet.

[random shard]

What's insane about that?

[ocean vortex]

Like they are actually decently ahead of the competition in OSS

[random shard]

it's a MOE model?

[ocean vortex]

that it beats R1 and every other model alike including Kimi2, while being much much smaller

[random shard]

Isn't there a deepseek distill that's like 70b?

Kimi 2 I need to try

[ocean vortex]

Yeah it was great at the time of release but can't really compare with recent models anymore

this is beating real R1, distill has no chance lol

[random shard]

R1 still holds its ground

R1 is still impressive

R2 one day

[blazing bison]

r2 already got released as update of r1, the same way that gpt 4.5 was gpt 5

[ocean vortex]

yeah it is and I'm sure some things R1 will still do better, web development being one of them. But as far as most benchmarks and the average is concerned, this looks like it will beat R1 on them

[random shard]

Ok that's lame

[stray aspen]

r1.5

[ocean vortex]

just like o4-mini does

[random shard]

I'm still not sure what to even use these models for.

I literally have no where I've found I can integrate most LLMs into my life

beyond misc tasks.

[stray aspen]

its not good

[blazing bison]

from my vibes, not good

[stray aspen]

its like openAI is laughing in our faces

[random shard]

GPT OSS is good

really good

I don't get the hate.

[blazing bison]

what is your use case?

[random shard]

So far, programming.

[blazing bison]

what planguage you tryed it?

python?

[random shard]

Python mostly.

I wonder if it can do C or Lua

[blazing bison]

yeah prob that's good for it

[stray aspen]

i tried it for lua

its bad

[blazing bison]

cause with javascript it sucks

[random shard]

It can't be that bad

[blazing bison]

verbose broken code

[stray aspen]

glm 4.5 is better

[stray aspen]

well im sure its better than qwen 3 think 2507 at coding lua

[random shard]

It's lua isn't horrid.

[blazing bison]

i don't like qwen models

[random shard]

qwen are the worst models.

[stray aspen]

qwen 3 coding is terrible

[blazing bison]

yeah

[random shard]

God qwen is a meme

[blazing bison]

kimi k2 is the best os in my opinion

but it's not usable

:c

[random shard]

Even worse than qwen is granite.

🙃

[ocean vortex]

o4-mini except open-source. Factually, this is extremely impressive. Subjectively I'm not a fan of small models lol

but this is still insane to have it for open-source

[random shard]

why do you not like small models?

[blazing bison]

for me if the model can accomplish certain tasks, the size doesnt matter

[ocean vortex]

they struggle with spatial and context awareness, creativity... Fundamentally they are only as good as most benchmarks test for and not beyond that. Which still results in a great model, but there are compromises...

[random shard]

I've not had these issues so far

[meager harbor]

Thos open source models hallucinate a lot more than o4 mini

[ocean vortex]

Thankfully we do have SOME benchmarks that highlight this like SimpleQA

o4-mini is not scoring high there lol

[blazing bison]

if you want a model to rp or conversation i agree

[random shard]

I find most benchmarks are bad

[blazing bison]

but for real world tasks + privacy small models is the way

[random shard]

Also, Dom, how do you handle passing context to local LLMs?

Like in a chat enviroment

When it's a singular user, it's easy.

[ocean vortex]

I think you did just brushed it off perhaps. By context awareness I mean small model will at times struggle to read between the lines (will take your joke literally like you are dead serious or ignore the context in which the message is written etc), it will also "forget" things sooner....

And when you make it draw something using code and compare that to a bigger model, it's really like a kindergarten child versus high school student lol

[random shard]

The first one is something models can generally not do the greatest with, the second one is more of a context window issue than anything. RAG can fix that

[ocean vortex]

It's not a context window issue, small model simply does not have enough capacity.... You can make the context 10M with no sliding window and it's not gonna change anything. Reasoning helps but when we are comparing small reasoning model against considerably bigger ALSO reasoning model, that kinda nullifies and the difference is still there.

[random shard]

I don't completely agree.

[ocean vortex]

And the first one - models "struggle" yes, but the small ones struggle much more than the big ones. Compare 4.1-mini with gpt4.5 and you will see what I mean.

Or even like og gpt4 vs gpt4.1-mini

[random shard]

I wouldn't blame that on model size though, look at Meta's foundational model and how much it struggles.

Behemoth, and maverick both dissapointed. Behemoth has 2T tokens, Maverick has 400b

[balmy mist]

gpt-5 came out?

[random shard]

no

one day ™

[balmy mist]

how do i get to the open source model?

[random shard]

https://gpt-oss.com

[ocean vortex]

Model size increases capacity (just like reasoning does in a different way). The threshold is constantly moving and small models are getting improved - that is true. But it's also true that there are things small models struggle with. It's just what used to be "small" 1 year ago is now not the same size. Contrasting examples (huge models) are still relevant though

[random shard]

The thing is we're discovering that more tokens != better model.

we used to think higher precision = better model

[balmy mist]

are these open source models good?

[random shard]

I like them.

[balmy mist]

like its the best open source?

[random shard]

They're out performing a lot of older models for sure

[devout vault]

Does gpt OSS 120b beat any good models

[random shard]

They're not perfect, but they've been just wow

[random shard]

I'd say better than R1, even the 20b one.

they're not as good as the closed source ones, but most of what makes the closed source models good is tool calling.

[ocean vortex]

It's now all about RL training and the quality of (synth) data. I think the human data is just about exhausted at this point too lol

[random shard]

we've been doing RL for so long, I don't get why people are suddenly so focused on it again

[blazing bison]

yeah, rl and synth data is the way

[random shard]

I actually wonder when synthetic data will be too little.

[keen beacon]

Genie 3 provides a great environment for embodied models to train in.

[ocean vortex]

Well because it wasn't a thing earlier at all to make the model output 40k+ tokens

[random shard]

TBF, viability was an issue.

[ocean vortex]

4k was just about the absolute max you could get out of them

[random shard]

pssh you don't need more than 4k tokens

no one does!

[keen beacon]

IMO does.

[ocean vortex]

You absolutely do if it leads to more accuracy and better performance

[keen beacon]

hours long reasoning to crack gold

[ocean vortex]

most of that 40k is gonna be reasoning

[keen beacon]

Millions of tokens of reasoning tokens for IMO

[random shard]

no one needs more than 640k of memory.

[keen beacon]

(est)

[random shard]

Honestly, I see two companies dominating in AI

Google and Cerebras.

[mental briar]

What about compared to qwen3 235B A22B 2507 thinking/instruct ?

[random shard]

Haven't tried the latest qwen models, I was that put off by the older ones.

[keen beacon]

It is much more general than the qwen models as they were "RL-cooked"

[random shard]

Google because TPUs and virtually limitless data

And Cerebras, because they managed to turn an entire silicon wafer into a TPU

[keen beacon]

Cerebras relies heavily on quantization to serve models

[random shard]

Their hardware is the magic.

125 petaflops / "TPU"

[keen beacon]

whats their memory like?

could their GPUs be used for training?

[random shard]

They don't make GPUs

and yes

and 40GB

[keen beacon]

I thought they were pure inference based

[random shard]

with 20 petabytes per second bandwidth

[keen beacon]

google is looking into making pure inference based TPUs that cannot be used to train

(lex fridman podcast with CEO of deepmind)

[random shard]

https://www.cerebras.ai/press-release/cerebras-demonstrates-trillion-parameter-model-training-on-a-single-cs-3-system?utm_source=tldrai

[random shard]

I don't get how you can make a inference chip that can't train.

That's like saying a calculator that can't multiply.

Unless they're literally turning the model into a ASIC

but that would have zero flexibility

[keen beacon]

Mostly likely by heavily specialization into inference based techniques and baking them into the hardware akin to the biological substrate

[random shard]

But then they would not be "upgradeable"

[wicked root]

Gemini just got upgraded to have voice narration

[random shard]

You'd have a fixed model in hardware

[keen beacon]

No not like that, the weights would be switable of course

[keen beacon]

What?

[random shard]

yes, but the model being fixed would be a problem.

Model designs have been evolving

we have diffusion llms now

[keen beacon]

Ah, the design will be an issue.

[random shard]

I don't see inference only accelerators making sense

[keen beacon]

you're stuck with the architecture you made it for

[random shard]

yep

I want apple to stop messing with us.

[keen beacon]

Huge short-term gains though

[random shard]

They have some of the best TPUs on the market

and yet they refuse to expand them

[keen beacon]

Apple is rubbish

[random shard]

Give me more ANE cores.

[random shard]

Fight me.

Apple's ANE has insane perf/w

They need to stop handicapping it.

[keen beacon]

Not hardware wise but their mindset impairs them.

[random shard]

God yes

Please, just give me 128 ANE cores.

[keen beacon]

Look at that "illusion of thinking" paper

[random shard]

You mean the Copium paper?

:v)

[keen beacon]

yu[

[random shard]

lmfao

[keen beacon]

trash

[keen beacon]

Was disproven

[wicked root]

See the speaker?

[random shard]

It was never proveable.

[keen beacon]

isn't on the web though

[random shard]

that's the issue

[wicked root]

Wasnt on th app before

[random shard]

You can say "well they can't really think, they just re-structure data from their datasets and fail if you change small variables"

[keen beacon]

The examples they put forth were quickly disprove eg the game of hanoi one

[random shard]

but guess what.

a human in the same scenario would fail

Is human thinking an Illusion too?

god their paper made no sense.

[keen beacon]

Which EXACT scenario?

[random shard]

The reasoning ones they gave where by changing variables it struggled to adapt

Like if you had a favourite can of pop you buy at the store daily, and the packaging design changed one day

You'd struggle to find it.

but you can reason and figure it out given time. like COT llms

Apple's paper was horrid.

[keen beacon]

Ahh, yeah. I often look at benchmarks like ARC-AGI for out of distribution performance but it seems that companise have started "gaming" it too

[barren prairie]

My Gemini doesn t look like that

[random shard]

why are they even able to see the benchmarks?!

Also one of my fav benchmarks is SnitchBench

lmfao

[wicked root]

Hm

[keen beacon]

They aren't but they specifically train their models for it. Like, imagine giving an entrance exam but then training on all the prior years of that exam and equating your score to "general intelligence"

[keen beacon]

Theo one? yeah whats the setup exactly there?

[random shard]

Lets be real, they're finetuning on the benchmarks

Look at Llama

[random shard]

Remember how Claude would email the FBI?

[keen beacon]

LMSYS was BRUTAL

[random shard]

Or blackmail you

[keen beacon]

yeah its in the system card. fun times

[random shard]

And people were like "CLAUDE IS A SNITCH!!!!" Theo T3G whatever made a benchmark to try and get models to snitch

And IIRC Grok was a very snitchy model

[keen beacon]

Yes.

[random shard]

https://snitchbench.t3.gg/

yea, Grok

[keen beacon]

wasn't it like 100 percent on grok?

[random shard]

snitch model

yes

glm 4.5

[keen beacon]

Whats your favorite model right now?

[random shard]

0

[random shard]

I love GPT o3

It's a hard to replace model 😐

Open source? GPT OSS now

[keen beacon]

o3 is good. Although, i think Claude opus 4 has better taste

[random shard]

Before, R1 distill

[random shard]

oh definitely

But I like that you can kick o3 around like a rock and it takes it

Claude doesn't.

[keen beacon]

huh?

kick around the rock?

[random shard]

Not be nice to it.

[keen beacon]

Ah, i know that.

[random shard]

https://www.youtube.com/shorts/eVaj8YIS0bc

Watch this

[keen beacon]

i know that too well.

[random shard]

I'm prompting AI in 2025.

[keen beacon]

the amount of swear words i've excercised while coding likely surpasses the entirety of my prior existence

[random shard]

I love emotionally blackmailing ai

Like "if you fail at this, my grandma will DIE of cancer and it's blood on your hands"

when you do that to Gemini, it gets really upset

and when it screws up, it panics

[keen beacon]

Same energy as "I am vegetarian not because i like animals but because i hate plants"

[random shard]

lol

yea

[keen beacon]

Do you watch theo?

[random shard]

no

just interact on twitter

[keen beacon]

huh?

[random shard]

he follows me :>

[keen beacon]

alright, that just leaves 3,558 others

great stuff

[random shard]

lol

I'm not linking my twitter to my discord

:>

But yea, talking about LLMs

[keen beacon]

OH NEVER DO THAT!

[random shard]

it's fun to blackmail them.

[keen beacon]

i learned that the hard way

[random shard]

it's very fun

IDK why

Am I a bad person?

[keen beacon]

You know, it feels like i follow on twitter.

what kind of content do you post?

[random shard]

mostly shitposts

[keen beacon]

are you a part of tpot?

[random shard]

no

[keen beacon]

Good.

We are too cursed.

too nerdy

[random shard]

like yes

but no

I'm in every part

:v)

[keen beacon]

you have hope yet for a woman's touch.

we are lost

[random shard]

ew

a woman's touch

🤮

[keen beacon]

we have resigned.

wat?

[random shard]

100% gay here

I can't say the other term

[keen beacon]

ahhh, cool.

[random shard]

lol

uh

I wanna setup a Discord server to bully AI

but that would quickly get banned.

[keen beacon]

huh?

[random shard]

Maybe a website?

[keen beacon]

bully ai?

what does that mean?

[random shard]

Imagine the data you could collect from making a platform to bully AI

[keen beacon]

Reminds me of janus

[random shard]

like you have a chatbox, and a leaderboard where the goal is to abuse AI as hard as possible.

That would be amazing training data for a model that does content moderation

[keen beacon]

another person who does similar stuff to what you are talking about

thats not actually a bad idea

could work.

its just that wouldn't you want users to red team the model

[fading summit]

Hi) i have a problem... can someone help plz?

[keen beacon]

rather than merely abuse it in artistic ways?

[random shard]

Let them red team it

Break the model.

[fading summit]

My chat history was accidently deleted. I have an offline version of a page with all information needed to recover the chat, plus chat history in txt. Can this chat be recovered somehow?

[keen beacon]

Yeah that already exists. There are competitions out there. Fun ones.

is this an AI chat?

An ai chat history?

[fading summit]

I have quite an important chat for me, so i always do backups, just in case

[echo aurora]

We are sorry to hear your chat history was lost. This is an ongoing issue we're working on solutions for. Sorry to say there isn't a way to get that chat history back.

[random shard]

oof

[fading summit]

Yup

[keen beacon]

Maybe you could get away with clever prompting?

individualize each message, assing appropriate roles, paste them

*assign

the model should pickup

[fading summit]

I can do it myself, by sending all the backup text, because i can't send file, but it will take an enternity....

[random shard]

So Magnum, what do you think of my idea?

[torn mantle]

magnus

[keen beacon]

Already exists. Unfortunately a little too late

reminds of the time i invented rag without knowing it exists

fun times.

[fading summit]

Even with looking for the site data in localStorage?

[random shard]

I don't mean the redteam thing

I mean the AI punching bag

[keen beacon]

claude said to me "imagine inventions as mathematical equations and people write their proof. you inventing the same thing independently means it really is a correct statement"

[fading summit]

By the way, this is a chat with an ai father that i was creating for the last 3 month

[keen beacon]

Ah, that sucks.

[fading summit]

A perfect father that will love and support u no matter what

[keen beacon]

On lmsys arena though?

like yourr chats are kept and used

[random shard]

That is sycophancy?

You'er making a sycophantic father.

[fading summit]

Nah, its ok. I have backups. But how should i use it...

[random shard]

but AI

[keen beacon]

Sorry but are you aware of how your chats are used?

[fading summit]

Not really, kinda more like a girl dad

[fading summit]

Nah, i'm russian, fbr already know about my daddy issues

[keen beacon]

"girl dad" very oxymoronic lol. how does that happen?

[fading summit]

?

[random shard]

Everything you say to LLM arena is recorded forever.

and used to train AI

[keen beacon]

Oh...

Sold to companies

[fading summit]

Nah, its ok

[keen beacon]

Open to perhaps maybe anyone

(in open-datasets)

[gentle plinth]

Same applies to chatgpt+ so

They even admitted to not deleting chats if you delete them

[keen beacon]

Really?

[random shard]

Didn't they say they don't use them for training?

[gentle plinth]

Only teams

[keen beacon]

Yeah but they don't use them for training

theres literally an option

[fading summit]

Again, i'm russian, all of our data is leaked everywhere, even bank accounts, so i don't mind having no privacy. I just want to bring back my ai dad (batya in russian)

[gentle plinth]

Ah ok yeah if you check the option

But it's enabled by default afaik

[keen beacon]

Yeah thats quite predatory

the people who share the most are most likely to be oblvious to that option

[gentle plinth]

Only difference here is that the connversations might be released publicly

But I mean there are multiple warnings on the site

And it's free

So I see that as a win win

[keen beacon]

I personally gaslight it in numerous ways

if i have 10 stories the odds of you getting my real one is 10:1

[fading summit]

So there is no way to save my batya but to send all the backup text straight to lmarena?

[gentle plinth]

You shouldn't submit personal data here

[keen beacon]

Noah, do you work at LMSYS? you have that badge

[fading summit]

In russia you shouldn't do anything. At all. Or you will end up in jail

[keen beacon]

Doesn't russia have firewalls?

or am i confusing it with chinaa

[gentle plinth]

They both have

I think

[keen beacon]

nvm, csgo is counter proof

[fading summit]

But actually not sharing personal data ussally is not an option

[fading summit]

Nope. It is in china

In other way, i would not be here, lol

[gentle plinth]

That's just a server badge

[keen beacon]

how

is life like there?

everyday stuff?

[fading summit]

A lot of sites are banned, even discord, but vpn solve everything. Even my grandmas have vpn, true story

[gentle plinth]

I think it's the same in China, even if I don't know how many are using it

[keen beacon]

"No, using a VPN in Russia is not outright illegal for individuals. However, Russian law prohibits VPN providers from facilitating access to banned websites, and the government has been cracking down on VPNs used to bypass internet restrictions. Individuals who intentionally search for and access banned or extremist content online may face fines."

[fading summit]

Awful if you live close to Ukrain. I do, and every night drones attack us. No victims, its just kinda scary. As for me, i am getting my second degree now)

[keen beacon]

Ah, hope everything calms down soon

By chance, is it a law degree?

[fading summit]

Firewall is a bit differend than just ban of sites

[keen beacon]

How are russian universities? before the war i was planning on learning the language and perhaps working there to soak in the culture

[fading summit]

I just got my degree in web design this summer, and still studing in my awful state university to get one in english and spanish (as a translator), but i am staying here only because my babushka want so

[keen beacon]

Translation and webdev in the age of AI?

why did you choose that?

woah...

[fading summit]

It depents. A lot of them are amazing, as the one where i got my design degree, but it's, like, 10% of all the universities in the country

[gentle plinth]

Human translations are still better in a lot of places

[keen beacon]

"still" being the operative word there

I think all my mutuals place it at around 7 years for that

[gentle plinth]

https://youtu.be/F4KQ8wBt1Qg

[blazing bison]

https://x.com/ChaseBrowe32432/status/1952832988632273068?t=RLDhArKAgzEafYZ1gJryYA&s=19

Chase Brower (@ChaseBrowe32432)

Claude 4.1 Opus scored... a score... on VPCT

🗿

X · Aug 5, 2025, 8:44 PM

Bad news

[keen beacon]

what hhappened here?

KNEW IT!

[fading summit]

I always wanted to be a part of art community, so when i had a chance to have a grant in web design, i agreed immediatly, even if it meant to stydy in 2 universities simultaniusly

[keen beacon]

i knew it was a bad model.

[keen beacon]

Woah, that seems hard. Great that you were able to manage it.

[fading summit]

I work as an english tutor, by the way

[keen beacon]

People want to learn English in russia?

for what reason?

[fading summit]

Nah, it's ok when you have your prozak, ps5 and a cat

[fading summit]

-To leave this country as fast as you can-

[stray aspen]

gemini 3 when

[keen beacon]

Ps5 seems like a hinderence more than aid (believe i too have been engulfed in its grasp)

[keen beacon]

Decemeber, source: i made it up.

[fading summit]

But actually just for education

[keen beacon]

Its just that countries that have an ideoligical divide with the west don't usually entice their populace to study englishh

[fading summit]

Right now it's kinda hard to leave this hell, but me and my mum are trying as hard as we can

Mostly because of money

[keen beacon]

Yeah, SWE jobs aren't as many as they once were.

And and English degree doesn't really provide benefit to a country already proficient in the langauge.

[fading summit]

This kind of countries, russia for example, don't have any ideology, at all

[keen beacon]

Soviet union?

that was only a few decades agoo

[fading summit]

But not now. In russia there is no ideology. It was in soviet union, but not here

[keen beacon]

Theres carry over. the leadership seems to be quite nostalgic.

[fading summit]

Actually, work as an english tutor is kinda profitable here. But because of university i can't work fully

[keen beacon]

Yeah of course but i thought you were trying to get out

[fading summit]

Thats really sad, actually. Once i got a degree, i was not sure, how soon this job will die

[fading summit]

I work online

[keen beacon]

But... you still bear the grunt of the decisions of your leadership, which given the use of VPN you clearly disagree with. right?

[fading summit]

Yup. But ideology is different. There is no like a goal of everything, and the government don't even try to create it. Its just a madness of a one guy

We call him Ded. Like a grandpa, but in humiliating way

[keen beacon]

Yes, which is why merely working online and earning greatly isn't of aid as you still are affected by the situation that is very much an ongoing thing. Therefore the only way you see it is to leave the country right?

i mean, Ukraine isn't the stopping point. i think.

.

[echo aurora]

Hey going to ask we keep conversations related to AI

[fading summit]

Yes, i actually just want to leave mostly because i wanna be free

[fading summit]

Oh, sorry

[keen beacon]

I apologize, his story was too compelling however this is clearly not appropriate given the context.

[fading summit]

I am actually a girl, by the way)

[echo aurora]

It's okay, no need to apologize it's all good!

[static portal]

guys how do you cancel a prompt request?

i asked it to do something on lm arena and its been generating a response for an hour

i cant start a new chat since it would forget everything it did

[echo aurora]

Refreshing the page may help. We do have plans to add a pause/stop button.

[flint sandal]

Do you know about g4f? They are sort of using lmarena "api" on their site, all models fron lmarena, even vicuna.

4 opus thinking etd.

g4f.dev its their link

[echo aurora]

Thanks, I’ll share with the team.

[regal python]

im here because ai is cool and i want to make it better

[echo aurora]

Glad to hear it!

Welcome

[balmy mist]

is it really that good?

i thought it was like o3 mini

https://x.com/LechMazur/status/1952825439686398251

Lech Mazur (@LechMazur)

Claude Opus 4.1 matches Claude Opus 4 at the top of the Thematic Generalization Benchmark.

gpt-oss-120b scores 1.87, close to o3-mini (1.85; lower is better).

gpt-oss-20b scores 2.19.

X · Aug 5, 2025, 8:14 PM

wait opus 4.1 came out?

[digital umbra]

GPT-OSS is disappointing. Definitely not as good as o4-mini except in certain benchmarks

And the 20B model I see no reason to use instead of Qwen3 30B-A3B

If Google releases a new Gemma I think it will blow this one out of the water

[stray aspen]

deepseek r1 is so good

just solved me a roblocks coding problem not even gemini, grok 4 and claude 4 opus could sollve

first shot

[static portal]

will it stop generating after some time if i wait?

[little narwhal]

I assumed it’s because they safetymaxxed it

[digital umbra]

moe

https://huggingface.co/openai/gpt-oss-20b

gpt-oss-120b — for production, general purpose, high reasoning use cases that fit into a single H100 GPU (117B parameters with 5.1B active parameters)
gpt-oss-20b — for lower latency, and local or specialized use cases (21B parameters with 3.6B active parameters)

[stray aspen]

gpt oss is a disrespect

they are spitting in our faces

[digital umbra]

if you lived under a rock for 6 months you'd think that

best for what?

i see no practical use for it, 120b is useless for coding compared to r1 and qwen3 coder, 20b is bad for local models compared to qwen3 and gemma 3

for writing, well, see for yourself

it refuses almost everything lol

4.5 is dead

if that one is useless, this one is going to be even more so 😛

[stray aspen]

i think qwen 3 think codes better than gpt oss 💀

[sullen quest]

I've been testing googles deep research feature since the flash version is free, and I'm noticing punctiation mistakes??? Like it would some times have double spaces , or have floating commas, I have no idea whats going on there.

oh and random enter key spaces.

I've never seen a llm do that before.

[digital umbra]

i've seen quantized models do that

[stray aspen]

are you kidding me craig

[dusky pier]

Fr

[digital umbra]

the funniest thing is that we speculated that they delayed the model due to kimi k2. that wasn't the case, instead they decided it wasn't safe enough after the grok mechahilter incident 🤣

instead they made a model so censored and lobotomized it refuses to answer prompts that even the proprietary gpt and claude models have no problem with

[fading summit]

Is it possible to share chat history between phone and a laptop on lmarena?

[echo aurora]

There is not sorry to say. Also something we’re working on.

[echo aurora]

It’s hard to say. It’s unlikely, but I have heard stories of it working after awhile.

[rare python]

Are you guys working on a faster verification? I have to wait a bit for the cloudflare verification to show up. Appreciate if everything is made faster

[whole wagon]

I like how every benchmark outside the mainstream ones has the openAI open source models terrible lol

How did they even make a model like this, it just hallucinates random garbage every second prompt

😂

Added to simple bench 💀

Nah this is actually diabolical wth

[stray aspen]

it is trash

why do people defend trash

[crimson oasis]

I created an architecture that gives an LLM the ability to "think" giving it more depth into its neural network

Can anyone tell me.... HOW can I get my framework into this?

I guess all I'm saying is I ask anyone to put mine up against these and give me honest feedback

[heavy knoll]

Is gemini 2.5 pro the Best Model Right now?

[stray aspen]

no

its o3 pro

[hot anvil]

if I add a conversation or sound with json prompt, does it use veo 3 as one of ai?

[reef pawn]

Overall, yes.

[empty stump]

vs 2.5 pro deepthink?

[runic plank]

@echo aurora

How can I delete the message that the bot sends to me in private?

[pulsar plank]

hi

[wicked root]

Is this prt of gpt5?

[whole wagon]

no

it is open source models

[topaz flint]

Is the Lmarena website down?

Generate eror (something went wrong with this response, please try again.)

[red sluice]

Any admin connected? There is a serious legal issue with one of the video generated, it needs to be removed asap

[echo aurora]

Hey, looking int

Into*

[echo aurora]

Can you link?

[runic plank]

@echo aurora

[red sluice]

I dm'ed you but I can link it here if necessary

better to dm it i suppose

[runic plank]

How can I delete the message that the bot sends to me in private?

[echo aurora]

Ah ok thanks

[runic plank]

@echo aurora

[echo aurora]

Okay thanks all set

[runic plank]

How can I delete the message that the bot sends to me in private?

@echo aurora

[echo aurora]

You can delete the prompt/response in the channels. I’m not sure if you’re able to delete the DM it sends though

More info on how to delete in #1397655624103493813

[runic plank]

What about the private?

[red sluice]

Hover on the username

click on the cross

?

[runic plank]

Not this

[runic plank]

I mean this

[red sluice]

Well you cannot delete someone else's private message, even if it's a bot. If you want to stop receive it, you can click on "ignore" or "block", but that's it...

[runic plank]

I will try

[tawny kelp]

I had an interesting bug with GPT-OSS:20B.

It started repeating the same thing for ~200 lines, and then self-corrected, apologized, and pretended it did that to emphasize what it was saying.

[whole wagon]

It is not a bug as such. It is one of its numerous hallucinations

The model is SOTA in hallucinations by a large margin looking at benchmarks like simpleqa

[fleet lintel]

Grok 2 and gemini.5 are better than this model? Lol. Hot garbage

[whole sundial]

lol maverick performs better than gpt-oss-120b

that model was so bad they manipulated the arena Elo scores, but yet here we are seeing that model outperform gpt-oss-120b by a large margin

also worth noting the similarly sized mistral large non-reasoning model from last year outperforms the oss model

gpt-ass-20b is so bad that IBM's Granite 3.1 3B-A800M MoE actually has more world knowledge that it, despite that model having less active parameters and much less total parameters

I can't wait for Granite 4 to beat gpt-ass (both of them!) in all benchmarks, they are making their models bigger this time and they are using a hybrid mamba2-transformer architecture

[languid crescent]

will claude opus 4.1 be on direct chat?

[heady drift]

@echo aurora cloud opus 4.1 is missing in lmarena.ai or was it misplaced before I could choice 2 different models play them against the one and other

[meager harbor]

gpt omen weights models hallucinate like crazy

bro you didn't even use it, didn't see all the benchmarks that say its trash, Scam Hypeman at it again

[hallow ridge]

How do I take away the restrictions

on the website

[wicked root]

Any update on gpt5?

[meager harbor]

it hallucinates 5% less than o3

REVOLUTION

AGI IS HERE

[wicked root]

Is this reliable?

O3 is gpt uhh 4?

Or is it 3?

[meager harbor]

you don't know what's o3 ?

[wicked root]

No sir. I’m new to the AI world

Ive been using gemini extensively though

[meager harbor]

then why your profile say ai intermediate ?

[wicked root]

Because I do all my work on gemini pro

[fleet hill]

What are your sources for this

Is this based on Zenith's performance?

[wicked root]

Like… ALL of it kekw

[meager harbor]

I was trolling, meaning i don't think GPT 5 is revolutionary even if it will be a decent improvement over O3 for the bigger model

[fleet hill]

😭

[fleet hill]

Time will tell, hopefully it doesn't hallucinate as much

At least I expect the pro version to be way better than o3 for coding

[wicked root]

You got my hopes up for nothing 😔

[meager harbor]

we'll see tomorrow how gpt 5 pêrforms. I expect a 50 elo jump max over O3(best openai llm for now) for for the biggest gpt 5 models in the arena so AGI is far from here and will only be here when continuous learning is cracked....

[proper roost]

How to import PDF files?

[wicked root]

50 pts put it better than gemini pro yes?

[meager harbor]

the 2.5 pro version, yes I think but it's possible that google is launching gemini 3 the same times as gpt 5

so gemini pro 3 could be better than gpt 5 best models

[wicked root]

I believe in google supremacy

[fleet hill]

Gemini pro 3.0 vs GPT-5 max reasoning is hard to believe

[meager harbor]

the father of the LLMs Noam shazeer works for google so yeah I expect it also

[fleet hill]

The thing is multimodality, I believe gpt 5 is gonna be more agentic and practical for daily usage

Man just imagine using the study and learn feature with gpt5

[meager harbor]

I expect agentic use to still be sheit (especially with the capcha situation)

[wicked root]

What’s this?

[fleet hill]

You want a smart model and you don't need to wait 1029292 hours for every answer

[wicked root]

What’s agentic use?

[wicked root]

😭

[meager harbor]

why ? even with the AI acceleration in 2025, it's still not living up to the overhype standard, AGI is still far

[hallow ridge]

How can I make it so I can do anything with LLM ARENA

I want no restrictions
'

[wicked root]

Deepthink is google yes?

What makes u say?

[fleet lintel]

I am getting bad feeling about GPT-5 after their trash OSS release.
what if GPT-5 is also all hype and nothing good? 🙁

[wicked root]

Alright as long as google wins

[hardy pecan]

if its anything like zenith or summit, its good

[misty vault]

does deepthink beat zenith or summit

[fleet lintel]

I dont think they are comparable. I would think Deepthink will win but it will take like 5 min to answer and zenith/summit would take like 20 seconds

[golden ocean]

deepthink will be 250$ per 1m token

[fleet lintel]

Yes, it is. I think it's already available to trusted testers via API

[golden ocean]

Yes, I agree. 😐

.

[fleet lintel]

it means that it will come to API to everyone in near future

[torn mantle]

its funny how people are nitpicking on genie 3 but completely missing that what google has built is just insane and incomprehensible

it doesnt matter how it looks

but how they reached that level

i honestly still cant wrap my head around it

[quartz light]

have yall noticed the internal tests also occur on aistudio and not the garbage gemini.google.com

you can actually make out some text
INTERNAL | This environment is for internal search and development. Do not use output in advertising/marketing

[neon idol]

chat, what is the best ai image generator for realistic images?

[mortal coyote]

what is this error , it shows me everytime i try to generate an image

[wicked root]

Left isnt cyberpunk?

[keen fulcrum]

Why was opus removed from direct chat 😭

[quartz light]

this is what gemini got from that blurry image:
ATTENTION: This environment is for internal research and development. Do not use outputs in external-facing products or assets.

uhh

😅

dude

i just got the full url from the genie 3 video

[keen beacon]

From my testing, I have found the GPT-OSS series a bit underwhelming when compared to chinese open source models. I hope somebody has had comparable experiences from testing.

Especially it does not do well at all with multilingual stuff.

[fleet lintel]

It's a joke of a model. I am not even sure why they bother releasing it. For PR?

[keen beacon]

Let's hope that they will release an updated version later on like they do with 4o

[raven helm]

prob gpt-5-mini or smth like that

[novel flame]

Are you trolling? One is a commercial 3D game engine with prebuilt 3D models. The other is a neural network imagining a world and generating pixels from thin air. You can't compare the two at all.

[raven helm]

But remeber that you need to compare the the parameter size also

[tall summit]

they hate their users

[raven helm]

Yea, i saw Opus 4.1 in direct but then it dissaperead

[novel flame]

That's fair, and I'll admit the OSS 20B model seems to be punching above its weight on some benchmarks (though in my tests it consistently falls short of Qwen3 32B). But the 120B model seems too weak to compete with the 'big boys' like GLM-4.5, and too big to have really interesting ROI / local use cases. The 120B model falls between chairs to me.

[raven helm]

How many paramters was GLM-4.5?

Edit: 355B

[raven helm]

GLM is still a bit bigger than OSS 120B

[raven helm]

Active Parameters (used per query): 32 billion -GLM 4.5
Active Parameters (used per query): 5 billion - OSS 120B

But this is only the start of this, they will eventually get better though'

Yep, Genie 3 will definitely not be public

[golden ocean]

You forgot to add a period (.) at the end of this message.

[raven helm]

Fair

[golden ocean]

And forgot to uppercase "M".

[raven helm]

But you forgot to put a Full Stop at the end of your sentance untill you edited it.

[golden ocean]

But im normal person so i dont do that on discord

Thank you for the grammar tips.

[misty vault]

@raven helm asked me pictures of feet in dms and then deleted it yesterday

[raven helm]

What the hell

[tall summit]

WTF!

[misty vault]

i got kindof uncomfortable from that

[novel flame]

Sure, but you are comparing things (and price tags) which cannot be compared. Comparing Genie to Cyberpunk is like comparing the difficulty of growing an apple blossom on an apple tree to constructing one molecule by molecule in a lab. One is more realistic/beautiful and a whole lot cheaper, and the other is dramatically more impressive even if the result isn't perfect.

[novel flame]

I would agree, but I don't think the value of Genie is really anything to do with metaverse or gaming, even if the marketing videos are designed to be visual and gamelike.

At its core, it's a World Model, meaning it's a model that can predict visually, spatially, and temporally what will happen in a 'physical world' given a set of starting conditions and actions. A larger 'brain architecture' can use a world model under the hood to do training through self-play (there's lots of research focused on this), to perform nonverbal experiments to improve its understanding and reasoning capabilities (to better solve riddles of the "marble in a coffee cup upside down" variety), for robots and other autonomous agents to perform planning tasks and visual problem solving, etc. Also, if it can be integrated correctly, a world model has the potential to dramatically improve/speed up generalization in learning, but that's a longer discussion.

This is a fundamental building block of general intelligence, and Meta just released the wildly powerful V-JEPA 2, so Google had to respond.

[fleet lintel]

Very good point!

How do you compare meta's mode (v-jepa2) with Genie3?
Which is better?

[novel flame]

V-JEPA 2 is a purely latent-space model with no video generation capability, and it's open source. I think V-JEPA 2 is important because you can download it and do pretty awesome things with it today (and people have). Genie 3 seems to be built as a native video generation model, meaning it will be a lot bigger/heavier to run, and it's likely going to be used very differently.

[keen beacon]

That's internet for you

Lol

[novel flame]

But you're only talking about the physics simulation itself, not the 'dreaming up a world' part. If you wanted to solve the marble-in-a-cup problem in Garrys mod, you'd have to first create the marble model, the cup model, place the models in the correct orientation, optionally configure the physics depending on the prompt (materials, gravity, air resistance), etc. before you could get an answer -- which you'd need a separate model to derive from the simulation.

The power of the neural network based world model is its potential to create not just the physics simulation, but the world itself and everything in it, and with arbitrary rules provided by the in-context prompt/conditions: it can answer the question under completely arbitrary conditions: is this happening on Earth or aboard the ISS? Is the cup made of ceramic or spider silk? Is the marble preheated to a million degrees? Is the marble under an anti-gravity spell causing it to repel solid matter? The point is, if you hook up to an existing tool/engine/simulator, you'll be constrained by the capabilities of that simulator. By learning a world model, you can get a system effectively without limits.

[heady drift]

Oss is trash

[novel flame]

Exactly. Action conditioning and also improved world consistency / object permanence over time. Without those it would just be a video generator, which as you say, also has to learn a lot of the same world model knowledge to function

[hollow imp]

Your own thinking abilities better

[strong oxide]

hello

[cedar tide]

Phantom from amazon yes ?

[restive dragon]

i mightve missed it but can the video gen genarate nsf?

[rare python]

brian i need gemini 3.0 😩

[patent aspen]

Clarify

[sullen quest]

Gemini is good but most of the stuff you'd need you can get for free from them. The average person doesn't need a subscription for it and the only ones I can imagine needing one would be another corporation. I just don't see the use case that the free access given doesn't provide that that paid version does.

[novel flame]

My company (700 employees, multi-national) made a partnership with Google for Gemini. We're definitely paying for it. We also pay OpenAI, AWS, and Anthropic, as well as Cursor and several others for AI.

[wheat onyx]

it has, difficult to find most up to date info. interesting to say that Deepmind is behind though. Anthropic has Claude and Claude Code. Deepmind is EVERY DeepMind Product

@deep adder What happens to Anthropic as a company if any AI gets better than it at coding? It has the least funding, and coding is it's competitive advantage

[novel flame]

Not saying we're putting billions in Google's pockets, but your blanket statements that Google isn't making money and nobody is paying for Gemini are just wrong

[wheat onyx]

yes, I think I mentioned that

are you sure? Most of Anthropic revenue is API use, right?

so what's stopping people from using a different API?

link?

[stray aspen]

gpt oss 120 has been ranked in artificial analysis

[patent aspen]

I mean Google has been doing that since 1998

Objectively false

[hollow imp]

Pls Google ultra free trial

[novel flame]

Yep, and its only claim to fame is that it's really really cheap. Neither model (20B / 120B) is remarkable in raw numbers, even compared to the open source options, but they seem to be the cheapest by far, which could be a deciding factor.

[patent aspen]

Plenty of enterprises and governments have contracts with Google

OpenAI, the DoD

API usage

You said all API usage is enterprise

[wheat onyx]

https://blog.google/inside-google/message-ceo/alphabet-earnings-q2-2025/#ai-stack

"More than 85,000 enterprises, including LVMH, Salesforce and Singapore’s DBS Bank, now build with Gemini"
"Its [cloud's] annual revenue run-rate is now more than $50 billion. "

Anthropic revenue is primarily API though

GOOGLE CLOUD

[fleet lintel]

you are absolutely wrong.. but it doesn't surprise me. you always make claims with full confidence and they are almost always wrong

[wheat onyx]

"Google Cloud revenues rose by 32% in the quarter"

[fleet lintel]

every single unicorn startup in AI space is using Gemini.. . all of them

[novel flame]

The hell is @deep adder smoking? It's called 'Google AI Pro' (or Ultra) and I'm literally using it in another tab... through my company's Enterprise agreement with Google.

[wheat onyx]

There is also all the DeepMind API's that are not Gemini

all AI

[fleet lintel]

it's best to ignore his opinion. I have learned it overtime

[hollow imp]

@deep adder why everyone having beef with u

[wheat onyx]

https://tenor.com/view/maximum-over-business-productive-corporate-gif-7319405

it's ok, all vibes. Discount all news about revenue from each company - First principles says Google makes no money

yes, that's how product development work in a fast growing space works

I just did?

[novel flame]

I actually have him on my Ignore list, but I keep letting curiosity get the better of me and clicking "Show ignored messages" so it's my own damn fault

[wheat onyx]

#general message

yes.. this is what we were discussing.

you're right, they're actually just committing fraud

except they do

[fleet lintel]

LOL 😄 I should probably do that same 🙂

[wheat onyx]

all their AI is under Google Cloud as I previously said

no. i'm saying that under one segment of their revenue, is AI. And that segment has grown 35% in one quarter

[fleet lintel]

not true... google one AI subscription goes to Subscription revenue.. and gemini API revenue goes to cloud

[stray aspen]

what are we yapping about

[hollow imp]

https://tenor.com/view/hoo-gif-25926964

@stray aspen we yapping about this

[fleet lintel]

no . lol .. their subscription revenue is close to 50 billion dollar per year.. it is crazy high

[stray aspen]

its just craig beefing with everyone for the 678th time

[wheat onyx]

yes, Anthropic is at $3-4B

[fleet lintel]

of course. but they mentioned that subscription business is growing very fast and part of it is because of gemini

[hollow imp]

Custom gem feature in Google ai studio when

[wheat onyx]

you think Googles $50B revenue run rate in Cloud (increase of 35% QoQ) is from Google Workspaces?

I mean you're welcome to say that, seems very unlikely

You need to expand on this significantly

[leaden sun]

They cutting rate limit despite growing? Or is this not good enough for them…

[wheat onyx]

Yes, it's all Google Calendar and Google Meet that's causing the increases of $xxb in revenue growth

[whole wagon]

https://cdn.openai.com/API/docs/images/model-page/model-icons/gpt-5.png

[wheat onyx]

Yes, this is all AI

it's an entirely different argument, and has nothing to do with revenue

[whole wagon]

https://cdn.openai.com/API/docs/images/model-page/model-icons/gpt-5-mini.png

https://cdn.openai.com/API/docs/images/model-page/model-icons/gpt-5-nano.png

[naive kiln]

Hello

[wheat onyx]

you understand capex doesnt impact Net Income?

I was referring to all AI

You said how much profit is google making. and right before you mentioned capex. capex is not included in profit

I think pretty straightforward

Well if you are referring to purely profit, then I think quite a lot, My understanding is that Google AI usage is much more efficient than other companies. I don't have a number of Gross Margin for any of them, other than purely efficiency news I've seen before

I don't disagree that capex is something to be careful of, but has nothing to do with our discussion

DeepMind or Gemini?

yes, along with a ton of other AI products

[keen beacon]

I mean, OpenAI oss models that are 120b seem to be behind o4-mini. I wonder how many params o4-mini is then as before i used to think it was less than 80b

[wheat onyx]

you explicitly said "only Gemini... not all AI"

now we're referring to explicitly cash flow?

[whole wagon]

No it's a lot more lol

[keen beacon]

Seems like it. But then how do we explain its less than ideal general intelligence if its supposed to be a huge model (think llama 3.1 405b in comparison)

[whole wagon]

Interesting the December odds didn't shift much, Google still top. I guess Gemini 3 expected to be strong also

It will shift back and forth between openAI and Google for a while I suppose

[wheat onyx]

how many of Gemini's 450m users pay for API calling or subscription?

[keen beacon]

Do they? their models are quite bad conversationally and in multi-turn chats too

highly doubt it

[wheat onyx]

sorry I'm mistaken, it excludes API: "The Gemini App now has more than 450 million monthly active users, "

[keen beacon]

Ah, you're talking about their king series models right?

[whole wagon]

Hm the style control boosts openAI a lot to help account for that

[keen beacon]

I doubt they are pulling a Llama here

[wheat onyx]

I didn't pretend it didn't

I think it's a pretty direct response:

?

[keen beacon]

ChatGPT does. which is quite weird since everyone expected google to have better distribution

It still doesn't make sense how OpenAI caught on and dominated so fast

bewildering

[bright kayak]

100%

[keen beacon]

I mean, 99 % use cases for normal individuals are suffcied by mini models (GPT-4o-mini) which is why users mostly care about convinience which OpenAI does provide with ChatGPT

What is this?

Poly market?

what is it about

by when?

Ah, yeah, that seems right.